diff --git a/CLAUDE.md b/CLAUDE.md
index 6c39425..5ed4ecf 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -27,9 +27,15 @@ majordomo        Registry, Parse, env-DSN loading, chain executor, re-exports
   llm/           canonical contract: Message/Part/Request/Response/Option,
                  Tool/Toolbox, Capabilities, Stream, Model, Provider, errors
   health/        clock-injected health tracker (bench/backoff)
-  media/         image normalization to target capabilities   (Phase 3)
+  media/         image normalization to target capabilities (sniff real
+                 format, downscale, transcode, byte ladder; ErrUnsupported
+                 for what can't fit) — chains normalize PER TARGET
   provider/fake/ scriptable in-memory provider for hermetic tests
-  provider/{openai,anthropic,ollama,google}/                  (Phases 3-4)
+  provider/openai/    Chat Completions client (+ all OpenAI-compat targets)
+  provider/anthropic/ Messages API client (+ Anthropic-compat targets)
+  provider/ollama/    one native /api/chat client serving the ollama,
+                      ollama-cloud, and foreman built-ins via presets
+  provider/google/    Gemini on the official genai SDK         (Phase 4)
   agent/         Agent run loop                               (Phase 5)
   skill/         Skill interface + composition                (Phase 6)
   examples/      one runnable example per hard requirement    (Phase 7-8)
diff --git a/README.md b/README.md
index 8356ed8..cccdbbf 100644
--- a/README.md
+++ b/README.md
@@ -100,12 +100,25 @@ Chains are health-tracked per target:
 
 | Provider | Spec name | Key env var | Default endpoint |
 |----------|-----------|-------------|------------------|
-| OpenAI (+compatible) | `openai` | `OPENAI_API_KEY` | api.openai.com *(pending)* |
-| Anthropic (+compatible) | `anthropic` | `ANTHROPIC_API_KEY` | api.anthropic.com *(pending)* |
+| OpenAI (+compatible) | `openai` | `OPENAI_API_KEY` | https://api.openai.com/v1 |
+| Anthropic (+compatible) | `anthropic` | `ANTHROPIC_API_KEY` | https://api.anthropic.com |
 | Google (Gemini) | `google` | `GOOGLE_API_KEY` / `GEMINI_API_KEY` | Gen AI API *(pending)* |
-| Ollama Cloud | `ollama-cloud` | `OLLAMA_API_KEY` | https://ollama.com *(pending)* |
-| Ollama (local) | `ollama` | — | `OLLAMA_HOST` or http://localhost:11434 *(pending)* |
-| foreman | `foreman` | — (token via DSN) | requires DSN/base URL *(pending)* |
+| Ollama Cloud | `ollama-cloud` | `OLLAMA_API_KEY` | https://ollama.com |
+| Ollama (local) | `ollama` | — | `OLLAMA_HOST` or http://localhost:11434 |
+| foreman | `foreman` | — (token via DSN) | requires an LLM_* DSN or `ollama.Foreman(url, token)` |
+
+OpenAI-compatible / Anthropic-compatible endpoints: construct the provider
+with a name and base URL and register it —
+
+```go
+reg.RegisterProvider(openai.New(
+    openai.WithName("groq"),
+    openai.WithBaseURL("https://api.groq.com/openai/v1"),
+    openai.WithAPIKey(key),
+    // openai.WithLegacyMaxTokens(), // for servers that only honor max_tokens
+))
+// now "groq/llama-3.3-70b" works in Parse, chains, and aliases
+```
 
 ### `LLM_*` env-DSN provider definitions
 
@@ -139,11 +152,15 @@ Implement the two-method `Provider` interface and register it:
 reg.RegisterProvider(myProvider) // now "myprovider/model-x" parses, chains, aliases
 ```
 
-## Multimodality *(pending — Phase 3)*
+## Multimodality
 
-Attach images without knowing the target's limits; majordomo normalizes
-(downscale, re-encode, count/size limits) against the resolved target's
-declared capabilities and rejects clearly what cannot fit.
+Attach images without knowing the target's limits. Before each attempt the
+request is normalized against the **actual serving target's** declared
+capabilities: the real format is sniffed from the bytes, oversize images
+are downscaled (aspect preserved), disallowed formats are re-encoded, and
+byte budgets are enforced by a quality ladder. What cannot be made to fit
+is rejected with a clear `ErrUnsupported` error — and in a chain, the
+request simply advances to the next (e.g. vision-capable) element.
 
 ```go
 resp, err := m.Generate(ctx, majordomo.Request{
@@ -154,7 +171,7 @@ resp, err := m.Generate(ctx, majordomo.Request{
 })
 ```
 
-## Tool calls *(canonical API ready; provider wiring pending — Phase 3)*
+## Tool calls
 
 ```go
 weather := majordomo.Tool{
@@ -171,14 +188,20 @@ resp, _ := m.Generate(ctx, req, majordomo.WithTools(weather))
 // resp.ToolCalls → execute → append ToolResultsMessage → continue
 ```
 
-## Structured output *(canonical API ready; provider wiring pending — Phase 3)*
+Each provider maps this one shape to its native function-calling format
+(OpenAI tools/tool_calls, Anthropic tool_use/tool_result, Ollama tools with
+object arguments). Tool-call ids are synthesized when a backend omits them;
+streaming buffers tool-call arguments until they parse.
+
+## Structured output
 
 ```go
 resp, _ := m.Generate(ctx, req, majordomo.WithSchema(schemaJSON, "answer"))
 ```
 
-A generic `Generate[T]` helper (schema from your struct, unmarshal into it)
-lands with the agent phase.
+Maps to OpenAI `response_format: json_schema`, Anthropic
+`output_config.format`, and Ollama `format`. A generic `Generate[T]` helper
+(schema from your struct, unmarshal into it) lands with the agent phase.
 
 ## Agents & skills *(pending — Phases 5–6)*
 
@@ -189,17 +212,25 @@ skills = reusable instruction+tool bundles attachable to any agent.
 
 | Provider | Resolve/Parse | Chat | Streaming | Tools | Structured | Images | Env DSN |
 |----------------------|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
-| OpenAI (+compatible) | ✅ | pending | pending | pending | pending | pending | ✅ |
-| Anthropic (+compat) | ✅ | pending | pending | pending | pending | pending | ✅ |
+| OpenAI (+compatible) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| Anthropic (+compat) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
 | Google (Gemini) | ✅ | pending | pending | pending | pending | pending | ✅ |
-| Ollama Cloud | ✅ | pending | pending | pending | pending | pending | ✅ |
-| Ollama (local) | ✅ | pending | pending | pending | pending | pending | ✅ |
-| foreman | ✅ | pending | pending | pending | pending | pending | ✅ |
+| Ollama Cloud | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| Ollama (local) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| foreman | ✅ | ✅ | ✅¹ | ✅ | ✅ | ✅ | ✅ |
 | fake (testing) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | — |
 
+¹ foreman's daemon currently buffers sync chat responses (no token-by-token
+streaming); majordomo's stream API works against it and delivers the
+response as a single delta plus final event.
+
+Notes: Ollama has no native tool_choice — `"none"` drops the tools;
+`"required"`/named choices are best-effort ignored there.
+
 Cross-cutting: Parse grammar ✅ · aliases/tiers ✅ · failover chains ✅ ·
-health tracking/backoff ✅ · LLM_* env DSNs ✅ · media pipeline pending ·
-agent loop pending · skills pending · `Generate[T]` pending.
+health tracking/backoff ✅ · LLM_* env DSNs ✅ · media pipeline ✅
+(per-target normalization in chains) · agent loop pending · skills pending
+· `Generate[T]` pending.
 
 ## Development
 
diff --git a/builtin.go b/builtin.go
index 0b83908..0109097 100644
--- a/builtin.go
+++ b/builtin.go
@@ -3,8 +3,12 @@ package majordomo
 import (
 	"context"
 	"fmt"
+	"net/http"
 
 	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/provider/anthropic"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/provider/ollama"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/provider/openai"
 )
 
 // Built-in provider names. Real client implementations land per-phase
@@ -20,23 +24,78 @@ const (
 )
 
 // registerBuiltins installs the built-in providers and env-DSN scheme
-// factories into a fresh registry.
-func registerBuiltins(r *Registry) {
-	stub := func(kind string) SchemeFactory {
-		return func(name string, dsn DSN) (llm.Provider, error) {
-			return &stubProvider{name: name, kind: kind, baseURL: dsn.BaseURL(), token: dsn.Token}, nil
+// factories into a fresh registry. httpClient, when non-nil, is used by
+// every provider and factory the registry itself constructs.
+func registerBuiltins(r *Registry, httpClient *http.Client) {
+	ollamaOpts := func(extra ...ollama.Option) []ollama.Option {
+		if httpClient != nil {
+			extra = append(extra, ollama.WithHTTPClient(httpClient))
 		}
+		return extra
 	}
 
-	for _, kind := range []string{
-		ProviderOpenAI, ProviderAnthropic, ProviderGoogle,
-		ProviderOllama, ProviderOllamaCloud, ProviderForeman,
-	} {
-		r.providers[kind] = &stubProvider{name: kind, kind: kind}
-		r.schemes[kind] = stub(kind)
+	// Native-Ollama family: three names over one client with presets.
+	r.providers[ProviderOllama] = ollama.Local(ollamaOpts()...)
+	r.providers[ProviderOllamaCloud] = ollama.Cloud(ollamaOpts()...)
+	// foreman has no default URL; the no-DSN registration resolves but
+	// errors on use with a clear message (use an LLM_* DSN or
+	// ollama.Foreman(...) + RegisterProvider).
+	r.providers[ProviderForeman] = ollama.New(ollamaOpts(ollama.WithName(ProviderForeman))...)
+
+	ollamaScheme := func(name string, dsn DSN) (llm.Provider, error) {
+		return ollama.New(ollamaOpts(
+			ollama.WithName(name),
+			ollama.WithBaseURL(dsn.BaseURL()),
+			ollama.WithToken(dsn.Token),
+		)...), nil
 	}
+	r.schemes[ProviderOllama] = ollamaScheme
+	r.schemes[ProviderOllamaCloud] = ollamaScheme
+	r.schemes[ProviderForeman] = ollamaScheme
+
+	// OpenAI and OpenAI-compatible endpoints.
+	openaiOpts := func(extra ...openai.Option) []openai.Option {
+		if httpClient != nil {
+			extra = append(extra, openai.WithHTTPClient(httpClient))
+		}
+		return extra
+	}
+	r.providers[ProviderOpenAI] = openai.New(openaiOpts()...)
+	r.schemes[ProviderOpenAI] = func(name string, dsn DSN) (llm.Provider, error) {
+		return openai.New(openaiOpts(
+			openai.WithName(name),
+			openai.WithBaseURL(dsn.BaseURL()),
+			openai.WithAPIKey(dsn.Token),
+		)...), nil
+	}
+
+	// Anthropic and Anthropic-compatible endpoints.
+	anthropicOpts := func(extra ...anthropic.Option) []anthropic.Option {
+		if httpClient != nil {
+			extra = append(extra, anthropic.WithHTTPClient(httpClient))
+		}
+		return extra
+	}
+	r.providers[ProviderAnthropic] = anthropic.New(anthropicOpts()...)
+	r.schemes[ProviderAnthropic] = func(name string, dsn DSN) (llm.Provider, error) {
+		return anthropic.New(anthropicOpts(
+			anthropic.WithName(name),
+			anthropic.WithBaseURL(dsn.BaseURL()),
+			anthropic.WithAPIKey(dsn.Token),
+		)...), nil
+	}
+
+	// Google lands in its own phase; stub until then.
+	r.providers[ProviderGoogle] = &stubProvider{name: ProviderGoogle, kind: ProviderGoogle}
+	r.schemes[ProviderGoogle] = stubScheme(ProviderGoogle)
 	// "gemini" is an alternate scheme for the Google provider.
-	r.schemes["gemini"] = stub(ProviderGoogle)
+	r.schemes["gemini"] = stubScheme(ProviderGoogle)
+}
+
+func stubScheme(kind string) SchemeFactory {
+	return func(name string, dsn DSN) (llm.Provider, error) {
+		return &stubProvider{name: name, kind: kind, baseURL: dsn.BaseURL(), token: dsn.Token}, nil
+	}
 }
 
 // stubProvider stands in for a provider implementation that lands in a
diff --git a/chain.go b/chain.go
index 42eb2e0..26928de 100644
--- a/chain.go
+++ b/chain.go
@@ -7,6 +7,7 @@ import (
 
 	"gitea.stevedudenhoeffer.com/steve/majordomo/health"
 	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/media"
 )
 
 // ErrChainExhausted reports that every element of a failover chain failed
@@ -65,8 +66,8 @@ func (c *chain) Capabilities() llm.Capabilities {
 // Generate tries each target per the chain semantics above.
 func (c *chain) Generate(ctx context.Context, req llm.Request, opts ...llm.Option) (*llm.Response, error) {
 	req = req.Apply(opts...)
-	return chainDo(ctx, c, func(ctx context.Context, t chainTarget) (*llm.Response, error) {
-		return t.model.Generate(ctx, req)
+	return chainDo(ctx, c, req, func(ctx context.Context, t chainTarget, nreq llm.Request) (*llm.Response, error) {
+		return t.model.Generate(ctx, nreq)
 	})
 }
 
@@ -76,14 +77,17 @@ func (c *chain) Generate(ctx context.Context, req llm.Request, opts ...llm.Optio
 // (replaying half-delivered output would duplicate content).
 func (c *chain) Stream(ctx context.Context, req llm.Request, opts ...llm.Option) (llm.Stream, error) {
 	req = req.Apply(opts...)
-	return chainDo(ctx, c, func(ctx context.Context, t chainTarget) (llm.Stream, error) {
-		return t.model.Stream(ctx, req)
+	return chainDo(ctx, c, req, func(ctx context.Context, t chainTarget, nreq llm.Request) (llm.Stream, error) {
+		return t.model.Stream(ctx, nreq)
 	})
 }
 
 // chainDo runs the head-to-tail failover algorithm around an attempt
-// function, generic over the result type (response vs stream).
-func chainDo[T any](ctx context.Context, c *chain, attempt func(context.Context, chainTarget) (T, error)) (T, error) {
+// function, generic over the result type (response vs stream). Before each
+// target is tried, the request's media is normalized against THAT target's
+// capabilities (ADR-0008/0009) — a request that cannot be made to fit one
+// target advances to the next without a health penalty.
+func chainDo[T any](ctx context.Context, c *chain, req llm.Request, attempt func(context.Context, chainTarget, llm.Request) (T, error)) (T, error) {
 	var zero T
 	var failures []error
 
@@ -94,12 +98,20 @@ func chainDo[T any](ctx context.Context, c *chain, attempt func(context.Context,
 			continue
 		}
 
+		nreq, err := media.Normalize(req, t.model.Capabilities())
+		if err != nil {
+			// Always ErrUnsupported-wrapped: this target cannot take the
+			// request by declaration. Advance, no health penalty.
+			failures = append(failures, fmt.Errorf("%s: %w", t.key, err))
+			continue
+		}
+
 		retries := c.cfg.retries()
 		for attemptN := 0; ; attemptN++ {
 			if err := ctx.Err(); err != nil {
 				return zero, err
 			}
-			result, err := attempt(ctx, t)
+			result, err := attempt(ctx, t, nreq)
 			if err == nil {
 				c.tracker.ReportSuccess(t.key)
 				return result, nil
diff --git a/docs/adr/0009-multimodal-strategy.md b/docs/adr/0009-multimodal-strategy.md
new file mode 100644
index 0000000..ca32db4
--- /dev/null
+++ b/docs/adr/0009-multimodal-strategy.md
@@ -0,0 +1,57 @@
+# ADR-0009: Multimodal strategy — normalize per target, enforce at the provider
+
+**Status:** Accepted — 2026-06-10
+
+## Context
+
+Every provider (and some models) imposes different image rules: max
+dimensions/bytes, allowed MIME types, max images per request. A caller must
+be able to attach an image without knowing the eventual target — especially
+with failover chains, where the serving target isn't known until runtime.
+
+## Decision
+
+Two cooperating layers:
+
+1. **`media.Normalize(req, caps)`** — the transformation point. The chain
+   executor calls it **per target, per attempt**, against the actual
+   target's capabilities, before the provider sees the request:
+   - The real format is **sniffed from magic bytes** and wins over the
+     declared MIME (callers lie; jpeg/png/gif/webp recognized).
+   - Already-fitting images pass through untouched (fast path: zero copies).
+   - Oversize dimensions downscale (aspect-preserving) with a hand-rolled
+     box-filter — stdlib has no scaler and `x/image` stays out per
+     ADR-0007; box-average quality is ample for vision input.
+   - Disallowed MIME re-encodes: original format if allowed, else JPEG
+     (q85), else PNG, else the first allowed encodable type.
+   - Byte budgets enforce via a quality ladder (jpeg 85→65→45→30) then
+     dimension halving; ~6 attempts before giving up.
+   - WebP cannot be decoded by stdlib: it passes through when it fits and
+     is allowed; any needed transform is a clear error.
+   - Everything that cannot be made to fit errors **wrapping
+     `llm.ErrUnsupported`** — never silently dropped.
+2. **Provider backstop** — each provider cheaply enforces its effective
+   capabilities at request time (image count/MIME/bytes, plus
+   tools/structured/streaming support flags) and rejects with
+   `ErrUnsupported`. This keeps providers honest for expert callers who
+   build models directly without the registry.
+
+Chain semantics: a normalization failure for one target **advances** to the
+next element with no health penalty (the target isn't sick, it's just
+incapable) — so `fp/text-only,fp/vision` serves an image request from the
+vision element automatically.
+
+Canonical image content stays **bytes + MIME** (ADR-0002); no URL fetching.
+
+## Consequences
+
+- A 100×50 PNG sent at a 32px-cap target arrives as a 32×16 PNG; the same
+  request served by an 8000px target arrives untouched.
+- Conditional provider rules (e.g. Anthropic's 2000px cap above 20 images)
+  are approximated by the flat declared caps — conservative and simple.
+
+## Alternatives considered
+
+- Normalize once against chain-intersection caps: over-restricts every
+  request for the sake of rarely-used fallbacks. Rejected (ADR-0008).
+- `x/image/draw` scalers: a dependency for one function. Rejected.
diff --git a/docs/adr/0010-tools-structured-output-mapping.md b/docs/adr/0010-tools-structured-output-mapping.md
new file mode 100644
index 0000000..94d4751
--- /dev/null
+++ b/docs/adr/0010-tools-structured-output-mapping.md
@@ -0,0 +1,49 @@
+# ADR-0010: Tools and structured output — one canonical shape, native mappings
+
+**Status:** Accepted — 2026-06-10
+
+## Context
+
+Tool calling and schema-constrained output exist on every target but with
+different wire shapes (verified against current docs, June 2026; shapes
+recorded in each provider's package doc). The canonical API must hide all
+of it.
+
+## Decision
+
+Canonical: `Tool{Name, Description, Parameters (JSON Schema), Handler}`;
+`Response.ToolCalls[]{ID, Name, Arguments json.RawMessage}`; results return
+as `ToolResultsMessage(ToolResult{ID, Name, Content, IsError})`. Structured
+output via `WithSchema(schema, name)`. Per-provider mapping:
+
+| Concern | OpenAI(+compat) | Anthropic(+compat) | Ollama/foreman | Google (Phase 4) |
+|---|---|---|---|---|
+| Tool def | `tools[].function{name,description,parameters}` | `tools[]{name,description,input_schema}` | `tools[].function` | `FunctionDeclaration.ParametersJsonSchema` |
+| Call args | JSON **string** → RawMessage | `tool_use.input` object | `arguments` **object** | `FunctionCall.Args` map |
+| Results | one `role:tool` msg per result (`tool_call_id`) | one **user** msg, `tool_result` blocks (`is_error` native) | `role:tool` + `tool_name` | `FunctionResponse` parts |
+| IsError | `"ERROR: "` content prefix | `is_error: true` | `"ERROR: "` prefix | response payload field |
+| Forced choice | `tool_choice` string / named object | `{"type":"any"/"tool"/"none"}` | none → drop tools; others best-effort ignored | `FunctionCallingConfig` modes |
+| Structured | `response_format json_schema` (no strict flag) | `output_config.format json_schema` (GA mechanism) | `format: <schema>` | `ResponseJsonSchema` + JSON MIME |
+
+Cross-cutting decisions:
+
+- **Missing call ids are synthesized** (`call_<n>`) — Ollama and some
+  compat servers omit them; the agent loop needs ids to match results.
+- **Streaming buffers tool-call arguments to completion** (ADR-0002):
+  OpenAI fragments accumulate by index, Anthropic `input_json_delta`
+  fragments accumulate per block; consumers only ever see parseable calls.
+- **No strict-mode flag is sent** to OpenAI: strict mode imposes schema
+  constraints (every property required, additionalProperties:false) that
+  caller-supplied schemas may not satisfy. The `Generate[T]` reflector
+  (Phase 5) emits strict-compatible schemas anyway.
+- `SchemaName` feeds providers that need a name (OpenAI; default
+  "response"); others ignore it.
+- Tool handlers never panic the loop: `Toolbox.Execute`/`ExecuteTool`
+  recover panics and JSON-encode results (ADR to agent loop, Phase 5).
+
+## Consequences
+
+- One test matrix per provider asserts the exact wire JSON both directions;
+  drift is caught by httptest fixtures, not in production.
+- Ollama's missing tool_choice means "required" cannot be enforced there —
+  documented in the README matrix rather than emulated.
diff --git a/docs/adr/README.md b/docs/adr/README.md
index 6346769..d7bf294 100644
--- a/docs/adr/README.md
+++ b/docs/adr/README.md
@@ -12,3 +12,5 @@ One decision per file, append-only; supersede rather than rewrite.
 | [0006](0006-health-and-backoff.md) | Model health tracking and backoff | Accepted |
 | [0007](0007-dependency-policy.md) | Dependency policy — stdlib-first, hand-rolled REST clients | Accepted |
 | [0008](0008-chain-semantics.md) | Failover-chain execution semantics | Accepted |
+| [0009](0009-multimodal-strategy.md) | Multimodal strategy — normalize per target, enforce at provider | Accepted |
+| [0010](0010-tools-structured-output-mapping.md) | Tools and structured output — canonical shape, native mappings | Accepted |
diff --git a/env_test.go b/env_test.go
index dda67bd..fcbede9 100644
--- a/env_test.go
+++ b/env_test.go
@@ -1,10 +1,17 @@
 package majordomo
 
 import (
+	"context"
+	"encoding/json"
 	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
 	"slices"
 	"strings"
 	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/provider/ollama"
 )
 
 func TestParseDSN(t *testing.T) {
@@ -71,19 +78,16 @@ func TestLoadEnvForeman(t *testing.T) {
 		if !ok {
 			t.Fatalf("provider %q not registered", name)
 		}
-		sp, ok := p.(*stubProvider)
+		op, ok := p.(*ollama.Provider)
 		if !ok {
-			t.Fatalf("provider %q is %T, want *stubProvider (phase 1)", name, p)
+			t.Fatalf("provider %q is %T, want *ollama.Provider (foreman scheme)", name, p)
 		}
-		if sp.kind != ProviderForeman {
-			t.Errorf("provider %q kind = %q, want foreman", name, sp.kind)
+		if op.Name() != name {
+			t.Errorf("provider name = %q, want %q", op.Name(), name)
 		}
 		wantURL := "https://foreman-" + name + ".orgrimmar.dudenhoeffer.casa"
-		if sp.baseURL != wantURL {
-			t.Errorf("provider %q baseURL = %q, want %q", name, sp.baseURL, wantURL)
-		}
-		if sp.token != "test-token-change-me" {
-			t.Errorf("provider %q token = %q, want the DSN userinfo", name, sp.token)
+		if op.BaseURL() != wantURL {
+			t.Errorf("provider %q baseURL = %q, want %q", name, op.BaseURL(), wantURL)
 		}
 	}
 
@@ -193,3 +197,54 @@ func TestNewLoadsProcessEnv(t *testing.T) {
 		t.Error("New() should eagerly load LLM_ENVTEST from the process environment")
 	}
 }
+
+// TestEnvForemanChatRoundTrip is the required end-to-end case: an LLM_*
+// foreman DSN resolves through Parse and serves a real chat over the wire
+// (TLS test server, since env DSNs always dial https), with the DSN token
+// arriving as the bearer credential.
+func TestEnvForemanChatRoundTrip(t *testing.T) {
+	var gotAuth, gotPath string
+	var gotModel string
+	ts := httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotAuth = r.Header.Get("Authorization")
+		gotPath = r.URL.Path
+		var body struct {
+			Model string `json:"model"`
+		}
+		_ = json.NewDecoder(r.Body).Decode(&body)
+		gotModel = body.Model
+		_, _ = io.WriteString(w, `{"message":{"role":"assistant","content":"hi from foreman"},"done":true,"done_reason":"stop","prompt_eval_count":2,"eval_count":3}`)
+	}))
+	defer ts.Close()
+
+	host := strings.TrimPrefix(ts.URL, "https://")
+	r := New(
+		WithoutEnvProviders(),
+		WithEnvLookup(func(string) string { return "" }),
+		WithHTTPClient(ts.Client()),
+	)
+	if err := r.LoadEnv(map[string]string{"LLM_FM": "foreman://round-trip-token@" + host}); err != nil {
+		t.Fatalf("LoadEnv: %v", err)
+	}
+
+	m, err := r.Parse("fm/qwen3:30b")
+	if err != nil {
+		t.Fatalf("Parse: %v", err)
+	}
+	resp, err := m.Generate(context.Background(), Request{Messages: []Message{UserText("hello")}})
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if resp.Text() != "hi from foreman" {
+		t.Errorf("text = %q", resp.Text())
+	}
+	if resp.Model != "fm/qwen3:30b" {
+		t.Errorf("resp.Model = %q, want fm/qwen3:30b", resp.Model)
+	}
+	if gotAuth != "Bearer round-trip-token" {
+		t.Errorf("auth = %q, want the DSN token as bearer", gotAuth)
+	}
+	if gotPath != "/api/chat" || gotModel != "qwen3:30b" {
+		t.Errorf("path=%q model=%q", gotPath, gotModel)
+	}
+}
diff --git a/failover_test.go b/failover_test.go
index dc0a976..625c21c 100644
--- a/failover_test.go
+++ b/failover_test.go
@@ -1,8 +1,12 @@
 package majordomo
 
 import (
+	"bytes"
 	"context"
 	"errors"
+	"image"
+	"image/color"
+	"image/png"
 	"slices"
 	"strings"
 	"sync"
@@ -325,6 +329,95 @@ func TestSingleTargetGetsChainSemantics(t *testing.T) {
 	}
 }
 
+// pngImage encodes a width×height PNG for media tests.
+func pngImage(t *testing.T, width, height int) []byte {
+	t.Helper()
+	img := image.NewRGBA(image.Rect(0, 0, width, height))
+	for y := range height {
+		for x := range width {
+			img.Set(x, y, color.RGBA{R: uint8(x), G: uint8(y), B: 128, A: 255})
+		}
+	}
+	var buf bytes.Buffer
+	if err := png.Encode(&buf, img); err != nil {
+		t.Fatalf("encode png: %v", err)
+	}
+	return buf.Bytes()
+}
+
+// TestChainNormalizesMediaPerTarget: the request's image is downscaled to
+// the capabilities of the target that actually serves it.
+func TestChainNormalizesMediaPerTarget(t *testing.T) {
+	r := newTestRegistry(t)
+	fp := fake.New("fp",
+		fake.WithModelCapabilities("small-vision", llm.Capabilities{
+			MaxImagesPerReq:   2,
+			MaxImageDimension: 32,
+			AllowedImageMIME:  []string{"image/png"},
+		}),
+	)
+	r.RegisterProvider(fp)
+
+	m, _ := r.Parse("fp/small-vision")
+	_, err := m.Generate(context.Background(), Request{Messages: []Message{
+		UserParts(Text("describe"), Image("image/png", pngImage(t, 100, 50))),
+	}})
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+
+	calls := fp.Calls()
+	if len(calls) != 1 {
+		t.Fatalf("calls = %d", len(calls))
+	}
+	var img llm.ImagePart
+	for _, part := range calls[0].Request.Messages[0].Parts {
+		if ip, ok := part.(llm.ImagePart); ok {
+			img = ip
+		}
+	}
+	if img.Data == nil {
+		t.Fatal("no image reached the provider")
+	}
+	cfg, err := png.DecodeConfig(bytes.NewReader(img.Data))
+	if err != nil {
+		t.Fatalf("decode delivered image: %v", err)
+	}
+	if cfg.Width != 32 || cfg.Height != 16 {
+		t.Errorf("delivered image = %dx%d, want 32x16 (downscaled to target cap)", cfg.Width, cfg.Height)
+	}
+}
+
+// TestChainAdvancesPastImagelessTarget: a text-only head can't take an
+// image request; the chain advances to a vision-capable element with no
+// health penalty.
+func TestChainAdvancesPastImagelessTarget(t *testing.T) {
+	r := newTestRegistry(t)
+	fp := fake.New("fp",
+		fake.WithModelCapabilities("text-only", llm.Capabilities{SupportsTools: true}),
+		fake.WithModelCapabilities("vision", llm.Capabilities{MaxImagesPerReq: 4}),
+	)
+	r.RegisterProvider(fp)
+	fp.Enqueue("vision", fake.Reply("a tasteful png"))
+
+	m, _ := r.Parse("fp/text-only,fp/vision")
+	resp, err := m.Generate(context.Background(), Request{Messages: []Message{
+		UserParts(Text("what is this?"), Image("image/png", pngImage(t, 8, 8))),
+	}})
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if resp.Text() != "a tasteful png" {
+		t.Errorf("text = %q", resp.Text())
+	}
+	if got := fp.CallCount("text-only"); got != 0 {
+		t.Errorf("text-only target saw %d calls, want 0 (normalization rejects pre-send)", got)
+	}
+	if !r.Health().Available("fp/text-only") {
+		t.Error("media rejection must not penalize health")
+	}
+}
+
 // TestHTTP529ClassifiedTransient: Anthropic's "overloaded" status fails
 // over like any other transient error.
 func TestHTTP529FailsOver(t *testing.T) {
diff --git a/majordomo.go b/majordomo.go
index 9a77942..d8c4f4d 100644
--- a/majordomo.go
+++ b/majordomo.go
@@ -80,26 +80,26 @@ var ErrModelNotFound = llm.ErrModelNotFound
 var ErrUnsupported = llm.ErrUnsupported
 
 // Re-exported content and message constructors.
-func Text(s string) Part                              { return llm.Text(s) }
-func Image(mime string, data []byte) Part             { return llm.Image(mime, data) }
-func SystemText(s string) Message                     { return llm.SystemText(s) }
-func UserText(s string) Message                       { return llm.UserText(s) }
-func UserParts(parts ...Part) Message                 { return llm.UserParts(parts...) }
-func AssistantText(s string) Message                  { return llm.AssistantText(s) }
+func Text(s string) Part                               { return llm.Text(s) }
+func Image(mime string, data []byte) Part              { return llm.Image(mime, data) }
+func SystemText(s string) Message                      { return llm.SystemText(s) }
+func UserText(s string) Message                        { return llm.UserText(s) }
+func UserParts(parts ...Part) Message                  { return llm.UserParts(parts...) }
+func AssistantText(s string) Message                   { return llm.AssistantText(s) }
 func ToolResultsMessage(results ...ToolResult) Message { return llm.ToolResultsMessage(results...) }
-func NewToolbox(name string, tools ...Tool) *Toolbox  { return llm.NewToolbox(name, tools...) }
+func NewToolbox(name string, tools ...Tool) *Toolbox   { return llm.NewToolbox(name, tools...) }
 
 // Re-exported request options.
-func WithSystem(s string) Option                        { return llm.WithSystem(s) }
-func WithTools(tools ...Tool) Option                    { return llm.WithTools(tools...) }
-func WithToolbox(b *Toolbox) Option                     { return llm.WithToolbox(b) }
-func WithToolChoice(choice string) Option               { return llm.WithToolChoice(choice) }
+func WithSystem(s string) Option                            { return llm.WithSystem(s) }
+func WithTools(tools ...Tool) Option                        { return llm.WithTools(tools...) }
+func WithToolbox(b *Toolbox) Option                         { return llm.WithToolbox(b) }
+func WithToolChoice(choice string) Option                   { return llm.WithToolChoice(choice) }
 func WithSchema(schema json.RawMessage, name string) Option { return llm.WithSchema(schema, name) }
-func WithTemperature(t float64) Option                  { return llm.WithTemperature(t) }
-func WithTopP(p float64) Option                         { return llm.WithTopP(p) }
-func WithMaxTokens(n int) Option                        { return llm.WithMaxTokens(n) }
-func WithStopSequences(stops ...string) Option          { return llm.WithStopSequences(stops...) }
-func WithReasoningEffort(level string) Option           { return llm.WithReasoningEffort(level) }
+func WithTemperature(t float64) Option                      { return llm.WithTemperature(t) }
+func WithTopP(p float64) Option                             { return llm.WithTopP(p) }
+func WithMaxTokens(n int) Option                            { return llm.WithMaxTokens(n) }
+func WithStopSequences(stops ...string) Option              { return llm.WithStopSequences(stops...) }
+func WithReasoningEffort(level string) Option               { return llm.WithReasoningEffort(level) }
 
 // WithModelCapabilities re-exports llm.WithCapabilities for Provider.Model
 // calls made through this package.
diff --git a/media/media.go b/media/media.go
new file mode 100644
index 0000000..e6e0bc8
--- /dev/null
+++ b/media/media.go
@@ -0,0 +1,293 @@
+// Package media fits request images to a target's declared capabilities.
+//
+// Normalize sniffs each image's real format from magic bytes (declared MIME
+// types lie), corrects the part's MIME, and passes through anything that
+// already satisfies the target's llm.Capabilities. Images that do not fit
+// are decoded, downscaled (never upscaled), and re-encoded into an allowed
+// format and byte budget. Anything that cannot honestly be made to fit —
+// undecodable formats, impossible byte budgets, too many images, images for
+// a text-only target — fails with an error wrapping llm.ErrUnsupported so a
+// failover chain can advance to a more capable target without a health
+// penalty.
+//
+// Why a separate package: every provider would otherwise duplicate the same
+// decode/scale/encode pipeline. Providers keep only a cheap capability
+// enforcement backstop; this package performs the actual transformation,
+// once, against whichever target a chain is currently attempting.
+package media
+
+import (
+	"bytes"
+	"fmt"
+	"image"
+	"image/gif"
+	"image/jpeg"
+	"image/png"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// Normalize returns a copy of req whose images fit caps, transforming
+// (downscale, re-encode) where needed. The input request is never mutated.
+//
+// Fast paths: a request with no image parts, or whose images already satisfy
+// caps, is returned unchanged with all underlying slices shared. When any
+// image transforms, the Messages slice and the Parts slices of affected
+// messages are copied (copy-on-write); untouched parts stay shared.
+//
+// Images that cannot be made to fit return an error wrapping
+// llm.ErrUnsupported.
+func Normalize(req llm.Request, caps llm.Capabilities) (llm.Request, error) {
+	total := 0
+	for i := range req.Messages {
+		for _, p := range req.Messages[i].Parts {
+			if _, ok := p.(llm.ImagePart); ok {
+				total++
+			}
+		}
+	}
+	if total == 0 {
+		return req, nil
+	}
+	if !caps.SupportsImages() {
+		return llm.Request{}, fmt.Errorf("media: %w: target does not accept image input (request carries %d image(s))", llm.ErrUnsupported, total)
+	}
+	// Why error instead of dropping the overflow: silently removing an image
+	// changes the question the caller asked; the honest move is to refuse and
+	// let a chain try a roomier target.
+	if total > caps.MaxImagesPerReq {
+		return llm.Request{}, fmt.Errorf("media: %w: request carries %d images, target allows at most %d per request", llm.ErrUnsupported, total, caps.MaxImagesPerReq)
+	}
+
+	out := req
+	copiedMessages := false
+	for mi := range req.Messages {
+		copiedParts := false
+		for pi, part := range req.Messages[mi].Parts {
+			ip, ok := part.(llm.ImagePart)
+			if !ok {
+				continue
+			}
+			norm, changed, err := normalizeImage(ip, caps)
+			if err != nil {
+				return llm.Request{}, fmt.Errorf("media: message %d, part %d: %w", mi, pi, err)
+			}
+			if !changed {
+				continue
+			}
+			if !copiedMessages {
+				out.Messages = make([]llm.Message, len(req.Messages))
+				copy(out.Messages, req.Messages)
+				copiedMessages = true
+			}
+			if !copiedParts {
+				parts := make([]llm.Part, len(req.Messages[mi].Parts))
+				copy(parts, req.Messages[mi].Parts)
+				out.Messages[mi].Parts = parts
+				copiedParts = true
+			}
+			out.Messages[mi].Parts[pi] = norm
+		}
+	}
+	return out, nil
+}
+
+// Info reports an image part's sniffed format ("jpeg", "png", "gif", or
+// "webp") and pixel dimensions. It is a cheap metadata read — the pixels are
+// never decoded. webp is recognized by signature but not decodable with the
+// standard library, so it reports format "webp" with zero dimensions and a
+// nil error.
+func Info(p llm.ImagePart) (format string, width, height int, err error) {
+	format = sniff(p.Data)
+	switch format {
+	case "":
+		return "", 0, 0, fmt.Errorf("media: image bytes match no known format (jpeg, png, gif, webp)")
+	case "webp":
+		return "webp", 0, 0, nil
+	}
+	cfg, _, err := image.DecodeConfig(bytes.NewReader(p.Data))
+	if err != nil {
+		return format, 0, 0, fmt.Errorf("media: decode %s config: %w", format, err)
+	}
+	return format, cfg.Width, cfg.Height, nil
+}
+
+// normalizeImage fits one image part to caps. It returns the (possibly
+// transformed) part and whether it differs from the input. A corrected MIME
+// with untouched bytes still counts as changed so Normalize copy-on-writes
+// the containing slices.
+func normalizeImage(p llm.ImagePart, caps llm.Capabilities) (llm.ImagePart, bool, error) {
+	// Why sniff instead of trusting p.MIME: callers routinely mislabel image
+	// bytes, and providers reject mismatches; the bytes are the truth.
+	format := sniff(p.Data)
+	if format == "" {
+		return p, false, fmt.Errorf("%w: image bytes (declared %q) match no known format (jpeg, png, gif, webp)", llm.ErrUnsupported, p.MIME)
+	}
+	realMIME := "image/" + format
+	changed := false
+	if p.MIME != realMIME {
+		p.MIME = realMIME
+		changed = true
+	}
+
+	mimeOK := caps.MIMEAllowed(realMIME)
+	fitsBytes := caps.MaxImageBytes == 0 || len(p.Data) <= caps.MaxImageBytes
+	fitsDims := true
+	if caps.MaxImageDimension > 0 && format != "webp" {
+		// Cheap header-only dimension read; a failure forces the transform
+		// path, which surfaces the real decode error.
+		cfg, _, err := image.DecodeConfig(bytes.NewReader(p.Data))
+		if err != nil {
+			fitsDims = false
+		} else {
+			fitsDims = cfg.Width <= caps.MaxImageDimension && cfg.Height <= caps.MaxImageDimension
+		}
+	}
+	// Why webp skips the dimension check: the stdlib cannot read webp
+	// headers, so dimensions are unverifiable; if MIME and bytes fit we pass
+	// it through rather than reject a possibly-fine image.
+	if mimeOK && fitsBytes && fitsDims {
+		return p, changed, nil
+	}
+
+	// Transformation required from here on, which needs a real decode.
+	if format == "webp" {
+		return p, false, fmt.Errorf("%w: image is webp (%d bytes), which the Go standard library cannot decode; provide jpeg, png, or gif instead", llm.ErrUnsupported, len(p.Data))
+	}
+	img, _, err := image.Decode(bytes.NewReader(p.Data))
+	if err != nil {
+		return p, false, fmt.Errorf("%w: cannot decode %s image for transformation: %v", llm.ErrUnsupported, format, err)
+	}
+
+	if caps.MaxImageDimension > 0 {
+		b := img.Bounds()
+		if b.Dx() > caps.MaxImageDimension || b.Dy() > caps.MaxImageDimension {
+			nw, nh := fitDims(b.Dx(), b.Dy(), caps.MaxImageDimension)
+			img = downscale(img, nw, nh)
+		}
+	}
+
+	target, err := targetMIME(realMIME, caps)
+	if err != nil {
+		return p, false, err
+	}
+	data, err := encodeFit(img, target, caps.MaxImageBytes)
+	if err != nil {
+		return p, false, err
+	}
+	return llm.ImagePart{MIME: target, Data: data}, true, nil
+}
+
+// sniff identifies an image format from its magic bytes, returning "jpeg",
+// "png", "gif", "webp", or "" when nothing matches.
+func sniff(data []byte) string {
+	switch {
+	case len(data) >= 3 && data[0] == 0xFF && data[1] == 0xD8 && data[2] == 0xFF:
+		return "jpeg"
+	case len(data) >= 4 && data[0] == 0x89 && data[1] == 'P' && data[2] == 'N' && data[3] == 'G':
+		return "png"
+	case len(data) >= 4 && string(data[:4]) == "GIF8":
+		return "gif"
+	case len(data) >= 12 && string(data[:4]) == "RIFF" && string(data[8:12]) == "WEBP":
+		return "webp"
+	default:
+		return ""
+	}
+}
+
+// encodableMIME reports whether the stdlib can encode the given image type.
+func encodableMIME(mime string) bool {
+	switch mime {
+	case "image/jpeg", "image/png", "image/gif":
+		return true
+	}
+	return false
+}
+
+// targetMIME picks the re-encode format: the original when allowed, else
+// jpeg, else png, else the first allowed encodable type (gif). When nothing
+// allowed is encodable (e.g. only webp), it errors with llm.ErrUnsupported.
+func targetMIME(original string, caps llm.Capabilities) (string, error) {
+	if encodableMIME(original) && caps.MIMEAllowed(original) {
+		return original, nil
+	}
+	// Why jpeg before png: vision inputs are photographs more often than
+	// screenshots, and jpeg's quality knob is the only size lever we have
+	// for the byte-budget loop.
+	for _, m := range []string{"image/jpeg", "image/png"} {
+		if caps.MIMEAllowed(m) {
+			return m, nil
+		}
+	}
+	// An empty allow-list permits everything and was caught above, so the
+	// list is non-empty here: take its first encodable entry.
+	for _, m := range caps.AllowedImageMIME {
+		if encodableMIME(m) {
+			return m, nil
+		}
+	}
+	return "", fmt.Errorf("%w: none of the allowed image types %v can be encoded with the Go standard library", llm.ErrUnsupported, caps.AllowedImageMIME)
+}
+
+// encodeFit encodes img as mime within maxBytes (0 = no limit), trading
+// jpeg quality first and then resolution for size. The ladder is fixed
+// (jpeg: q85, q65, q45, q30, then half and quarter dimensions at q65;
+// png/gif: full, half, quarter dimensions) — at most six attempts, since an
+// image that survives a 16x pixel reduction over budget will not be saved
+// by further fiddling.
+func encodeFit(img image.Image, mime string, maxBytes int) ([]byte, error) {
+	type attempt struct {
+		div     int // divide both dimensions by this
+		quality int // jpeg quality; ignored for png/gif
+	}
+	var ladder []attempt
+	if mime == "image/jpeg" {
+		ladder = []attempt{{1, 85}, {1, 65}, {1, 45}, {1, 30}, {2, 65}, {4, 65}}
+	} else {
+		ladder = []attempt{{1, 0}, {2, 0}, {4, 0}}
+	}
+
+	scaled := map[int]image.Image{1: img}
+	smallest := -1
+	for _, a := range ladder {
+		cur, ok := scaled[a.div]
+		if !ok {
+			b := img.Bounds()
+			nw, nh := max(b.Dx()/a.div, 1), max(b.Dy()/a.div, 1)
+			cur = downscale(img, nw, nh)
+			scaled[a.div] = cur
+		}
+		data, err := encodeImage(cur, mime, a.quality)
+		if err != nil {
+			return nil, fmt.Errorf("encode %s: %w", mime, err)
+		}
+		if maxBytes == 0 || len(data) <= maxBytes {
+			return data, nil
+		}
+		if smallest == -1 || len(data) < smallest {
+			smallest = len(data)
+		}
+	}
+	return nil, fmt.Errorf("%w: image cannot be reduced to the %d-byte limit; smallest achievable %s encoding is %d bytes", llm.ErrUnsupported, maxBytes, mime, smallest)
+}
+
+// encodeImage encodes img into the given MIME type. quality applies to jpeg
+// only.
+func encodeImage(img image.Image, mime string, quality int) ([]byte, error) {
+	var buf bytes.Buffer
+	var err error
+	switch mime {
+	case "image/jpeg":
+		err = jpeg.Encode(&buf, img, &jpeg.Options{Quality: quality})
+	case "image/png":
+		err = png.Encode(&buf, img)
+	case "image/gif":
+		err = gif.Encode(&buf, img, nil)
+	default:
+		return nil, fmt.Errorf("no stdlib encoder for %q", mime)
+	}
+	if err != nil {
+		return nil, err
+	}
+	return buf.Bytes(), nil
+}
diff --git a/media/media_test.go b/media/media_test.go
new file mode 100644
index 0000000..2e3e92c
--- /dev/null
+++ b/media/media_test.go
@@ -0,0 +1,513 @@
+package media
+
+import (
+	"bytes"
+	"errors"
+	"image"
+	"image/color"
+	"image/gif"
+	"image/jpeg"
+	"image/png"
+	"math/rand/v2"
+	"strings"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// --- test image builders -------------------------------------------------
+
+// gradient builds a smooth w x h RGBA image (compresses well).
+func gradient(w, h int) *image.RGBA {
+	img := image.NewRGBA(image.Rect(0, 0, w, h))
+	for y := 0; y < h; y++ {
+		for x := 0; x < w; x++ {
+			img.SetRGBA(x, y, color.RGBA{
+				R: uint8(x * 255 / max(w-1, 1)),
+				G: uint8(y * 255 / max(h-1, 1)),
+				B: 128,
+				A: 255,
+			})
+		}
+	}
+	return img
+}
+
+// noisy builds a w x h image of deterministic random pixels (compresses
+// terribly — ideal for exercising the byte-budget ladder).
+func noisy(w, h int) *image.RGBA {
+	rng := rand.New(rand.NewPCG(1, 2))
+	img := image.NewRGBA(image.Rect(0, 0, w, h))
+	for i := range img.Pix {
+		img.Pix[i] = uint8(rng.UintN(256))
+	}
+	return img
+}
+
+func encPNG(t *testing.T, img image.Image) []byte {
+	t.Helper()
+	var buf bytes.Buffer
+	if err := png.Encode(&buf, img); err != nil {
+		t.Fatalf("png encode: %v", err)
+	}
+	return buf.Bytes()
+}
+
+func encJPEG(t *testing.T, img image.Image) []byte {
+	t.Helper()
+	var buf bytes.Buffer
+	if err := jpeg.Encode(&buf, img, &jpeg.Options{Quality: 90}); err != nil {
+		t.Fatalf("jpeg encode: %v", err)
+	}
+	return buf.Bytes()
+}
+
+func encGIF(t *testing.T, img image.Image) []byte {
+	t.Helper()
+	var buf bytes.Buffer
+	if err := gif.Encode(&buf, img, nil); err != nil {
+		t.Fatalf("gif encode: %v", err)
+	}
+	return buf.Bytes()
+}
+
+// webpBlob is a minimal byte sequence carrying the RIFF/WEBP signature.
+// The stdlib cannot decode webp, so sniffing is all that ever reads it.
+func webpBlob() []byte {
+	b := []byte("RIFF")
+	b = append(b, 0x1a, 0x00, 0x00, 0x00)
+	b = append(b, "WEBPVP8 "...)
+	b = append(b, make([]byte, 18)...)
+	return b
+}
+
+func imgReq(parts ...llm.Part) llm.Request {
+	return llm.Request{Messages: []llm.Message{llm.UserParts(parts...)}}
+}
+
+// firstImage returns the first image part in the request.
+func firstImage(t *testing.T, req llm.Request) llm.ImagePart {
+	t.Helper()
+	for _, m := range req.Messages {
+		for _, p := range m.Parts {
+			if ip, ok := p.(llm.ImagePart); ok {
+				return ip
+			}
+		}
+	}
+	t.Fatal("no image part in request")
+	return llm.ImagePart{}
+}
+
+// --- fast paths -----------------------------------------------------------
+
+func TestNormalizeFastPathNoImages(t *testing.T) {
+	req := llm.Request{Messages: []llm.Message{llm.UserText("hello")}}
+	got, err := Normalize(req, llm.Capabilities{}) // even a no-image target
+	if err != nil {
+		t.Fatalf("Normalize: %v", err)
+	}
+	if &got.Messages[0] != &req.Messages[0] {
+		t.Error("messages slice was copied on the no-image fast path")
+	}
+	if &got.Messages[0].Parts[0] != &req.Messages[0].Parts[0] {
+		t.Error("parts slice was copied on the no-image fast path")
+	}
+}
+
+func TestNormalizeFastPathFittingImages(t *testing.T) {
+	data := encPNG(t, gradient(20, 10))
+	req := imgReq(llm.Text("look:"), llm.Image("image/png", data))
+	caps := llm.Capabilities{
+		MaxImagesPerReq:   4,
+		MaxImageBytes:     len(data) + 100,
+		MaxImageDimension: 64,
+		AllowedImageMIME:  []string{"image/png"},
+	}
+	got, err := Normalize(req, caps)
+	if err != nil {
+		t.Fatalf("Normalize: %v", err)
+	}
+	if &got.Messages[0] != &req.Messages[0] {
+		t.Error("messages slice was copied although every image already fits")
+	}
+	if &got.Messages[0].Parts[1] != &req.Messages[0].Parts[1] {
+		t.Error("parts slice was copied although every image already fits")
+	}
+}
+
+// --- rejection paths ------------------------------------------------------
+
+func TestNormalizeImagesUnsupported(t *testing.T) {
+	req := imgReq(llm.Image("image/png", encPNG(t, gradient(4, 4))))
+	_, err := Normalize(req, llm.Capabilities{MaxImagesPerReq: 0})
+	if !errors.Is(err, llm.ErrUnsupported) {
+		t.Fatalf("err = %v, want ErrUnsupported", err)
+	}
+	if !strings.Contains(err.Error(), "does not accept image input") {
+		t.Errorf("err message %q lacks explanation", err)
+	}
+}
+
+func TestNormalizeTooManyImages(t *testing.T) {
+	img := llm.Image("image/png", encPNG(t, gradient(4, 4)))
+	req := llm.Request{Messages: []llm.Message{
+		llm.UserParts(img, img),
+		llm.UserParts(img),
+	}}
+	_, err := Normalize(req, llm.Capabilities{MaxImagesPerReq: 2})
+	if !errors.Is(err, llm.ErrUnsupported) {
+		t.Fatalf("err = %v, want ErrUnsupported", err)
+	}
+	if !strings.Contains(err.Error(), "3 images") || !strings.Contains(err.Error(), "at most 2") {
+		t.Errorf("err message %q lacks the counts", err)
+	}
+}
+
+func TestNormalizeGarbageBytes(t *testing.T) {
+	req := imgReq(llm.Image("image/png", []byte("certainly not an image")))
+	_, err := Normalize(req, llm.Capabilities{MaxImagesPerReq: 1})
+	if !errors.Is(err, llm.ErrUnsupported) {
+		t.Fatalf("err = %v, want ErrUnsupported", err)
+	}
+	if !strings.Contains(err.Error(), "no known format") {
+		t.Errorf("err message %q lacks a clear explanation", err)
+	}
+}
+
+// --- MIME sniffing & correction --------------------------------------------
+
+func TestNormalizeMIMECorrection(t *testing.T) {
+	data := encPNG(t, gradient(8, 8))
+	req := imgReq(llm.Image("image/jpeg", data)) // caller lies: bytes are png
+	got, err := Normalize(req, llm.Capabilities{MaxImagesPerReq: 1})
+	if err != nil {
+		t.Fatalf("Normalize: %v", err)
+	}
+	ip := firstImage(t, got)
+	if ip.MIME != "image/png" {
+		t.Errorf("MIME = %q, want sniff-corrected %q", ip.MIME, "image/png")
+	}
+	if !bytes.Equal(ip.Data, data) {
+		t.Error("image bytes changed although only the MIME needed correcting")
+	}
+	if orig := firstImage(t, req); orig.MIME != "image/jpeg" {
+		t.Errorf("input request mutated: MIME now %q", orig.MIME)
+	}
+}
+
+func TestNormalizeCopyOnWrite(t *testing.T) {
+	data := encPNG(t, gradient(8, 8))
+	req := llm.Request{Messages: []llm.Message{
+		llm.UserText("untouched message"),
+		llm.UserParts(llm.Text("untouched part"), llm.Image("image/jpeg", data)),
+	}}
+	got, err := Normalize(req, llm.Capabilities{MaxImagesPerReq: 1})
+	if err != nil {
+		t.Fatalf("Normalize: %v", err)
+	}
+	if &got.Messages[0] == &req.Messages[0] {
+		t.Error("messages slice shared although a part changed (mutation hazard)")
+	}
+	if &got.Messages[0].Parts[0] != &req.Messages[0].Parts[0] {
+		t.Error("parts slice of the untouched message was copied")
+	}
+	if &got.Messages[1].Parts[0] == &req.Messages[1].Parts[0] {
+		t.Error("parts slice of the changed message is still shared (mutation hazard)")
+	}
+}
+
+// --- dimension capping ------------------------------------------------------
+
+func TestNormalizeDownscale(t *testing.T) {
+	req := imgReq(llm.Image("image/png", encPNG(t, gradient(200, 100))))
+	caps := llm.Capabilities{MaxImagesPerReq: 1, MaxImageDimension: 50}
+	got, err := Normalize(req, caps)
+	if err != nil {
+		t.Fatalf("Normalize: %v", err)
+	}
+	format, w, h, err := Info(firstImage(t, got))
+	if err != nil {
+		t.Fatalf("Info: %v", err)
+	}
+	if format != "png" {
+		t.Errorf("format = %q, want original format %q preserved", format, "png")
+	}
+	if w != 50 || h != 25 {
+		t.Errorf("dimensions = %dx%d, want 50x25 (aspect preserved)", w, h)
+	}
+}
+
+func TestNormalizeDownscalePortrait(t *testing.T) {
+	req := imgReq(llm.Image("image/png", encPNG(t, gradient(100, 200))))
+	caps := llm.Capabilities{MaxImagesPerReq: 1, MaxImageDimension: 50}
+	got, err := Normalize(req, caps)
+	if err != nil {
+		t.Fatalf("Normalize: %v", err)
+	}
+	_, w, h, err := Info(firstImage(t, got))
+	if err != nil {
+		t.Fatalf("Info: %v", err)
+	}
+	if w != 25 || h != 50 {
+		t.Errorf("dimensions = %dx%d, want 25x50 (aspect preserved)", w, h)
+	}
+}
+
+// --- transcoding -------------------------------------------------------------
+
+func TestNormalizeTranscode(t *testing.T) {
+	tests := []struct {
+		name    string
+		data    []byte
+		mime    string
+		allowed []string
+		want    string
+	}{
+		{"png to jpeg", encPNG(t, gradient(16, 16)), "image/png", []string{"image/jpeg"}, "jpeg"},
+		{"jpeg to png", encJPEG(t, gradient(16, 16)), "image/jpeg", []string{"image/png"}, "png"},
+		{"gif to png", encGIF(t, gradient(16, 16)), "image/gif", []string{"image/png"}, "png"},
+		{"png to gif fallback", encPNG(t, gradient(16, 16)), "image/png", []string{"image/gif"}, "gif"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			req := imgReq(llm.Image(tt.mime, tt.data))
+			caps := llm.Capabilities{MaxImagesPerReq: 1, AllowedImageMIME: tt.allowed}
+			got, err := Normalize(req, caps)
+			if err != nil {
+				t.Fatalf("Normalize: %v", err)
+			}
+			ip := firstImage(t, got)
+			if ip.MIME != "image/"+tt.want {
+				t.Errorf("MIME = %q, want %q", ip.MIME, "image/"+tt.want)
+			}
+			format, w, h, err := Info(ip)
+			if err != nil {
+				t.Fatalf("Info: %v", err)
+			}
+			if format != tt.want {
+				t.Errorf("sniffed format = %q, want %q", format, tt.want)
+			}
+			if w != 16 || h != 16 {
+				t.Errorf("dimensions = %dx%d, want 16x16 (no resize needed)", w, h)
+			}
+		})
+	}
+}
+
+func TestNormalizeNoEncodableAllowedType(t *testing.T) {
+	// png needs transcoding but the only allowed type is webp, which the
+	// stdlib cannot encode.
+	req := imgReq(llm.Image("image/png", encPNG(t, gradient(8, 8))))
+	caps := llm.Capabilities{MaxImagesPerReq: 1, AllowedImageMIME: []string{"image/webp"}}
+	_, err := Normalize(req, caps)
+	if !errors.Is(err, llm.ErrUnsupported) {
+		t.Fatalf("err = %v, want ErrUnsupported", err)
+	}
+	if !strings.Contains(err.Error(), "image/webp") {
+		t.Errorf("err message %q does not name the unencodable allowed types", err)
+	}
+}
+
+// --- byte budget ---------------------------------------------------------------
+
+func TestNormalizeByteBudgetFits(t *testing.T) {
+	// Random noise defeats q85 jpeg at full size; the ladder must walk down
+	// quality and then resolution until the encoding fits.
+	req := imgReq(llm.Image("image/png", encPNG(t, noisy(256, 256))))
+	caps := llm.Capabilities{
+		MaxImagesPerReq:  1,
+		AllowedImageMIME: []string{"image/jpeg"},
+		MaxImageBytes:    8 * 1024,
+	}
+	got, err := Normalize(req, caps)
+	if err != nil {
+		t.Fatalf("Normalize: %v", err)
+	}
+	ip := firstImage(t, got)
+	if len(ip.Data) > caps.MaxImageBytes {
+		t.Errorf("len(Data) = %d, exceeds budget %d", len(ip.Data), caps.MaxImageBytes)
+	}
+	if format, _, _, err := Info(ip); err != nil || format != "jpeg" {
+		t.Errorf("Info = %q, %v; want jpeg, nil", format, err)
+	}
+}
+
+func TestNormalizeByteBudgetImpossible(t *testing.T) {
+	req := imgReq(llm.Image("image/png", encPNG(t, noisy(256, 256))))
+	caps := llm.Capabilities{
+		MaxImagesPerReq:  1,
+		AllowedImageMIME: []string{"image/jpeg"},
+		MaxImageBytes:    10, // no image fits in 10 bytes
+	}
+	_, err := Normalize(req, caps)
+	if !errors.Is(err, llm.ErrUnsupported) {
+		t.Fatalf("err = %v, want ErrUnsupported", err)
+	}
+	if !strings.Contains(err.Error(), "10-byte limit") {
+		t.Errorf("err message %q lacks the budget", err)
+	}
+	if !strings.Contains(err.Error(), "smallest achievable") {
+		t.Errorf("err message %q lacks the achieved size", err)
+	}
+}
+
+// --- webp ---------------------------------------------------------------------
+
+func TestNormalizeWebPPassThrough(t *testing.T) {
+	data := webpBlob()
+	req := imgReq(llm.Image("image/webp", data))
+	caps := llm.Capabilities{
+		MaxImagesPerReq:   1,
+		MaxImageBytes:     1024,
+		MaxImageDimension: 50, // unverifiable for webp; must not force a transform
+		AllowedImageMIME:  []string{"image/webp"},
+	}
+	got, err := Normalize(req, caps)
+	if err != nil {
+		t.Fatalf("Normalize: %v", err)
+	}
+	if &got.Messages[0] != &req.Messages[0] {
+		t.Error("request copied although the webp image passes through")
+	}
+	if ip := firstImage(t, got); !bytes.Equal(ip.Data, data) {
+		t.Error("webp bytes changed on pass-through")
+	}
+}
+
+func TestNormalizeWebPNeedsTransform(t *testing.T) {
+	req := imgReq(llm.Image("image/webp", webpBlob()))
+	caps := llm.Capabilities{MaxImagesPerReq: 1, AllowedImageMIME: []string{"image/jpeg"}}
+	_, err := Normalize(req, caps)
+	if !errors.Is(err, llm.ErrUnsupported) {
+		t.Fatalf("err = %v, want ErrUnsupported", err)
+	}
+	if !strings.Contains(err.Error(), "webp") {
+		t.Errorf("err message %q does not name the format", err)
+	}
+	if !strings.Contains(err.Error(), "jpeg, png, or gif") {
+		t.Errorf("err message %q does not say what to provide instead", err)
+	}
+}
+
+// --- input immutability ----------------------------------------------------------
+
+func TestNormalizeInputNotMutated(t *testing.T) {
+	data := encPNG(t, gradient(200, 100))
+	snapshot := bytes.Clone(data)
+	req := llm.Request{
+		System: "sys",
+		Messages: []llm.Message{
+			llm.UserParts(llm.Text("scale me"), llm.Image("image/jpeg", data)),
+		},
+	}
+	caps := llm.Capabilities{
+		MaxImagesPerReq:   1,
+		MaxImageDimension: 50,
+		AllowedImageMIME:  []string{"image/jpeg"},
+	}
+	got, err := Normalize(req, caps)
+	if err != nil {
+		t.Fatalf("Normalize: %v", err)
+	}
+	orig := firstImage(t, req)
+	if orig.MIME != "image/jpeg" {
+		t.Errorf("input MIME mutated to %q", orig.MIME)
+	}
+	if !bytes.Equal(orig.Data, snapshot) {
+		t.Error("input image bytes mutated")
+	}
+	if txt := req.Messages[0].Parts[0].(llm.TextPart); txt.Text != "scale me" {
+		t.Errorf("input text part mutated: %q", txt.Text)
+	}
+	if ip := firstImage(t, got); bytes.Equal(ip.Data, snapshot) {
+		t.Error("output image was expected to transform but is byte-identical")
+	}
+}
+
+// --- alpha handling ----------------------------------------------------------------
+
+func TestNormalizeAlphaPNGToJPEG(t *testing.T) {
+	img := image.NewRGBA(image.Rect(0, 0, 32, 32))
+	for y := 0; y < 32; y++ {
+		for x := 0; x < 32; x++ {
+			img.SetRGBA(x, y, color.RGBA{R: 200, G: 60, B: 30, A: uint8(x * 8)})
+		}
+	}
+	req := imgReq(llm.Image("image/png", encPNG(t, img)))
+	caps := llm.Capabilities{MaxImagesPerReq: 1, AllowedImageMIME: []string{"image/jpeg"}}
+	got, err := Normalize(req, caps)
+	if err != nil {
+		t.Fatalf("Normalize: %v", err)
+	}
+	ip := firstImage(t, got)
+	decoded, err := jpeg.Decode(bytes.NewReader(ip.Data))
+	if err != nil {
+		t.Fatalf("decoding transcoded jpeg: %v", err)
+	}
+	if b := decoded.Bounds(); b.Dx() != 32 || b.Dy() != 32 {
+		t.Errorf("decoded dimensions = %dx%d, want 32x32", b.Dx(), b.Dy())
+	}
+}
+
+// --- Info ----------------------------------------------------------------------------
+
+func TestInfo(t *testing.T) {
+	pngData := encPNG(t, gradient(10, 7))
+	jpegData := encJPEG(t, gradient(5, 9))
+	gifData := encGIF(t, gradient(6, 4))
+
+	tests := []struct {
+		name    string
+		part    llm.ImagePart
+		format  string
+		w, h    int
+		wantErr bool
+	}{
+		{"png", llm.ImagePart{MIME: "image/png", Data: pngData}, "png", 10, 7, false},
+		{"jpeg", llm.ImagePart{MIME: "image/jpeg", Data: jpegData}, "jpeg", 5, 9, false},
+		{"gif", llm.ImagePart{MIME: "image/gif", Data: gifData}, "gif", 6, 4, false},
+		{"mislabeled png", llm.ImagePart{MIME: "image/jpeg", Data: pngData}, "png", 10, 7, false},
+		{"webp", llm.ImagePart{MIME: "image/webp", Data: webpBlob()}, "webp", 0, 0, false},
+		{"garbage", llm.ImagePart{MIME: "image/png", Data: []byte("nope")}, "", 0, 0, true},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			format, w, h, err := Info(tt.part)
+			if tt.wantErr {
+				if err == nil {
+					t.Fatal("Info: expected error, got nil")
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("Info: %v", err)
+			}
+			if format != tt.format || w != tt.w || h != tt.h {
+				t.Errorf("Info = %q, %d, %d; want %q, %d, %d", format, w, h, tt.format, tt.w, tt.h)
+			}
+		})
+	}
+}
+
+// --- byte-cap pass-through interplay ----------------------------------------------------
+
+func TestNormalizeOversizeBytesTriggersTransform(t *testing.T) {
+	// A fitting MIME and dimension but an over-budget payload must re-encode,
+	// not pass through.
+	data := encPNG(t, noisy(64, 64))
+	req := imgReq(llm.Image("image/png", data))
+	caps := llm.Capabilities{
+		MaxImagesPerReq:  1,
+		MaxImageBytes:    len(data) / 2,
+		AllowedImageMIME: []string{"image/png", "image/jpeg"},
+	}
+	got, err := Normalize(req, caps)
+	if err != nil {
+		t.Fatalf("Normalize: %v", err)
+	}
+	ip := firstImage(t, got)
+	if len(ip.Data) > caps.MaxImageBytes {
+		t.Errorf("len(Data) = %d, exceeds budget %d", len(ip.Data), caps.MaxImageBytes)
+	}
+}
diff --git a/media/scale.go b/media/scale.go
new file mode 100644
index 0000000..f21a8ab
--- /dev/null
+++ b/media/scale.go
@@ -0,0 +1,54 @@
+package media
+
+import "image"
+
+// fitDims scales (w, h) so the longer side equals limit, preserving aspect
+// ratio with round-half-up on the shorter side, floored at 1 pixel.
+func fitDims(w, h, limit int) (int, int) {
+	if w >= h {
+		return limit, max((h*limit+w/2)/w, 1)
+	}
+	return max((w*limit+h/2)/h, 1), limit
+}
+
+// downscale resizes src to dw x dh using area averaging (a box filter): each
+// destination pixel is the mean of its corresponding source region.
+//
+// Why hand-rolled: the stdlib has no scaler and ADR-0007 bars
+// golang.org/x/image without a new ADR. Area averaging is dependency-free,
+// alias-resistant when shrinking (every source pixel contributes exactly
+// once), and entirely adequate quality for vision-model input. It is only
+// ever called to shrink — Normalize never upscales.
+func downscale(src image.Image, dw, dh int) *image.RGBA {
+	b := src.Bounds()
+	sw, sh := b.Dx(), b.Dy()
+	dst := image.NewRGBA(image.Rect(0, 0, dw, dh))
+	for dy := 0; dy < dh; dy++ {
+		// Integer box edges: destination pixel dy covers source rows
+		// [dy*sh/dh, (dy+1)*sh/dh), widened to at least one row.
+		sy0 := dy * sh / dh
+		sy1 := max((dy+1)*sh/dh, sy0+1)
+		for dx := 0; dx < dw; dx++ {
+			sx0 := dx * sw / dw
+			sx1 := max((dx+1)*sw/dw, sx0+1)
+			var r, g, bl, a uint64
+			for sy := sy0; sy < sy1; sy++ {
+				for sx := sx0; sx < sx1; sx++ {
+					pr, pg, pb, pa := src.At(b.Min.X+sx, b.Min.Y+sy).RGBA()
+					r += uint64(pr)
+					g += uint64(pg)
+					bl += uint64(pb)
+					a += uint64(pa)
+				}
+			}
+			n := uint64((sy1 - sy0) * (sx1 - sx0))
+			i := dst.PixOffset(dx, dy)
+			// RGBA() returns 16-bit channels; average, then drop to 8 bits.
+			dst.Pix[i+0] = uint8(r / n >> 8)
+			dst.Pix[i+1] = uint8(g / n >> 8)
+			dst.Pix[i+2] = uint8(bl / n >> 8)
+			dst.Pix[i+3] = uint8(a / n >> 8)
+		}
+	}
+	return dst
+}
diff --git a/parse_test.go b/parse_test.go
index 6b52847..53d2bc3 100644
--- a/parse_test.go
+++ b/parse_test.go
@@ -63,8 +63,8 @@ func TestParseModelIDIsVerbatim(t *testing.T) {
 	// Everything after the first slash, up to the next comma, is the model
 	// id: colons and additional slashes pass through untouched.
 	for spec, want := range map[string]string{
-		"ollama-cloud/minimax-m3:cloud":  "ollama-cloud/minimax-m3:cloud",
-		"google/models/gemini-3.0-pro":   "google/models/gemini-3.0-pro",
+		"ollama-cloud/minimax-m3:cloud":       "ollama-cloud/minimax-m3:cloud",
+		"google/models/gemini-3.0-pro":        "google/models/gemini-3.0-pro",
 		"ollama-cloud/qwen3-coder:480b-cloud": "ollama-cloud/qwen3-coder:480b-cloud",
 	} {
 		m, err := r.Parse(spec)
diff --git a/progress.md b/progress.md
index f332bb1..982a020 100644
--- a/progress.md
+++ b/progress.md
@@ -1,5 +1,45 @@
 # progress
 
+## 2026-06-10 — Phase 3: REST providers (OpenAI, Anthropic, Ollama×3) + media
+
+**Landed:**
+- `provider/openai`: Chat Completions client for OpenAI and every
+  OpenAI-compatible endpoint (tools with string-arguments mapping, strict
+  SSE streaming incl. by-index tool-call assembly and the empty-choices
+  usage chunk, response_format json_schema, max_completion_tokens with a
+  WithLegacyMaxTokens compat option, reasoning_effort).
+- `provider/anthropic`: Messages API client (anthropic-version 2023-06-01,
+  required-max_tokens defaulting, tool_use/tool_result blocks with native
+  is_error, GA structured output via output_config.format, full SSE event
+  parser with input_json_delta buffering, 529-overloaded classified
+  transient, usage sums cache tokens).
+- `provider/ollama`: ONE native /api/chat client serving ollama (local,
+  OLLAMA_HOST normalization), ollama-cloud (https://ollama.com + bearer
+  OLLAMA_API_KEY), and foreman (base URL + bearer; tolerates its
+  buffered-single-object "streaming"). Object tool arguments, tool_name
+  results, format-schema structured output, think-level mapping, NDJSON
+  streaming with 16MB lines.
+- `media/`: normalization pipeline per ADR-0009 (magic-byte sniffing,
+  box-filter downscale, transcode preference ladder, byte-budget quality
+  ladder, webp passthrough-or-reject, copy-on-write, everything-unfittable
+  wraps ErrUnsupported).
+- Chain executor now normalizes media PER TARGET before each attempt and
+  advances penalty-free past targets that can't take the request (proven:
+  text-only head + vision fallback; per-target downscale assertions).
+- Registry: real providers + scheme factories wired for openai, anthropic,
+  ollama, ollama-cloud, foreman (google still stubbed, Phase 4);
+  WithHTTPClient registry option; required env-foreman TLS chat round-trip
+  test (LLM_FM=foreman://token@host → Parse("fm/qwen3:30b") → bearer
+  arrives, chat answers).
+- ADR-0009 (multimodal), ADR-0010 (tools/structured mapping); README
+  matrix flipped to ✅ for the four landed provider families; ~70 new
+  hermetic tests across the three provider packages + media.
+- Run note: openai/anthropic/media were built by three parallel
+  subagents against the frozen llm contract; ollama/foreman, chain wiring,
+  and registry integration done in the main line. All gates green.
+
+**Next:** Phase 4 — Google provider on google.golang.org/genai.
+
 ## 2026-06-10 — Phase 2: health + failover chain, proven
 
 **Landed:** the full deterministic failover test matrix over the fake
diff --git a/provider/anthropic/anthropic.go b/provider/anthropic/anthropic.go
new file mode 100644
index 0000000..b9b0309
--- /dev/null
+++ b/provider/anthropic/anthropic.go
@@ -0,0 +1,319 @@
+// Package anthropic implements llm.Provider for the Anthropic Messages API
+// and Anthropic-compatible endpoints.
+//
+// API surface targeted: POST {base}/v1/messages with headers x-api-key,
+// anthropic-version: 2023-06-01, and content-type: application/json, per the
+// platform.claude.com Messages API reference as of June 2026. Streaming uses
+// the documented SSE event sequence (message_start, content_block_start,
+// content_block_delta, content_block_stop, message_delta, message_stop).
+// Structured output uses the GA output_config.format mechanism with
+// {"type":"json_schema"}; the result arrives as JSON text in the first text
+// content block.
+//
+// Why a hand-rolled client (no SDK): ADR-0007 — majordomo is stdlib-first,
+// and the canonical llm contract needs only a narrow slice of the API.
+package anthropic
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+const (
+	defaultName    = "anthropic"
+	defaultBaseURL = "https://api.anthropic.com"
+
+	// apiVersion is the anthropic-version header value. 2023-06-01 remains
+	// the current (and only) stable version string as of June 2026.
+	apiVersion = "2023-06-01"
+
+	// defaultMaxTokens is used when Request.MaxTokens is 0, because the
+	// Messages API requires max_tokens on every request.
+	defaultMaxTokens = 4096
+)
+
+// defaultCapabilities reflects the documented first-party API image limits:
+// 100 images per request (200K-context models), 10 MB per image, 8000 px per
+// side, and the four supported media types.
+func defaultCapabilities() llm.Capabilities {
+	return llm.Capabilities{
+		SupportsTools:      true,
+		SupportsStructured: true,
+		SupportsStreaming:  true,
+		MaxImagesPerReq:    100,
+		MaxImageBytes:      10 << 20,
+		MaxImageDimension:  8000,
+		AllowedImageMIME: []string{
+			"image/jpeg", "image/png", "image/gif", "image/webp",
+		},
+	}
+}
+
+// Provider is an llm.Provider backed by the Anthropic Messages API.
+type Provider struct {
+	name      string
+	apiKey    string
+	baseURL   string
+	client    *http.Client
+	caps      llm.Capabilities
+	maxTokens int
+}
+
+// Option configures the provider at construction.
+type Option func(*Provider)
+
+// WithAPIKey sets the API key explicitly, bypassing the ANTHROPIC_API_KEY
+// environment default.
+func WithAPIKey(key string) Option {
+	return func(p *Provider) { p.apiKey = key }
+}
+
+// WithBaseURL points the provider at an Anthropic-compatible endpoint. A
+// trailing slash is trimmed; "/v1/messages" is appended per request.
+func WithBaseURL(u string) Option {
+	return func(p *Provider) { p.baseURL = strings.TrimRight(u, "/") }
+}
+
+// WithHTTPClient replaces the HTTP client (timeouts, proxies, test doubles).
+func WithHTTPClient(c *http.Client) Option {
+	return func(p *Provider) { p.client = c }
+}
+
+// WithName overrides the registry name. Why: an Anthropic-compatible
+// endpoint registered under its own name must surface that name in
+// Response.Model and errors, not "anthropic".
+func WithName(name string) Option {
+	return func(p *Provider) { p.name = name }
+}
+
+// WithDefaultCapabilities replaces the provider-default capabilities.
+func WithDefaultCapabilities(caps llm.Capabilities) Option {
+	return func(p *Provider) { p.caps = caps }
+}
+
+// WithDefaultMaxTokens overrides the max_tokens value used when
+// Request.MaxTokens is 0. Why: the Messages API rejects requests without
+// max_tokens, so the provider must always send something.
+func WithDefaultMaxTokens(n int) Option {
+	return func(p *Provider) { p.maxTokens = n }
+}
+
+// New creates an Anthropic provider. It never fails: a missing API key
+// (no WithAPIKey and no ANTHROPIC_API_KEY in the environment) surfaces as a
+// 401-style *llm.APIError at request time, not at construction.
+func New(opts ...Option) *Provider {
+	p := &Provider{
+		name:      defaultName,
+		baseURL:   defaultBaseURL,
+		client:    http.DefaultClient,
+		caps:      defaultCapabilities(),
+		maxTokens: defaultMaxTokens,
+	}
+	for _, opt := range opts {
+		opt(p)
+	}
+	if p.apiKey == "" {
+		p.apiKey = os.Getenv("ANTHROPIC_API_KEY")
+	}
+	return p
+}
+
+// Name implements llm.Provider.
+func (p *Provider) Name() string { return p.name }
+
+// Model implements llm.Provider. The id is passed through verbatim — it is
+// never validated against a catalog.
+func (p *Provider) Model(id string, opts ...llm.ModelOption) (llm.Model, error) {
+	cfg := llm.ApplyModelOptions(opts)
+	caps := p.caps
+	if cfg.Capabilities != nil {
+		caps = *cfg.Capabilities
+	}
+	return &model{provider: p, id: id, caps: caps}, nil
+}
+
+type model struct {
+	provider *Provider
+	id       string
+	caps     llm.Capabilities
+}
+
+// Capabilities implements llm.Model.
+func (m *model) Capabilities() llm.Capabilities { return m.caps }
+
+// fullName is the "provider/model" identifier used in Response.Model.
+func (m *model) fullName() string { return m.provider.name + "/" + m.id }
+
+// Generate implements llm.Model.
+func (m *model) Generate(ctx context.Context, req llm.Request, opts ...llm.Option) (*llm.Response, error) {
+	req = req.Apply(opts...)
+	if err := m.enforceCapabilities(req); err != nil {
+		return nil, err
+	}
+	httpResp, err := m.do(ctx, req, false)
+	if err != nil {
+		return nil, err
+	}
+	defer httpResp.Body.Close()
+	if httpResp.StatusCode/100 != 2 {
+		return nil, m.apiError(httpResp)
+	}
+	var wr wireResponse
+	if err := json.NewDecoder(httpResp.Body).Decode(&wr); err != nil {
+		return nil, fmt.Errorf("%s: decode response: %w", m.provider.name, err)
+	}
+	return m.toResponse(&wr), nil
+}
+
+// Stream implements llm.Model. A non-2xx status is returned as an error from
+// Stream itself, before any events are delivered.
+func (m *model) Stream(ctx context.Context, req llm.Request, opts ...llm.Option) (llm.Stream, error) {
+	req = req.Apply(opts...)
+	if err := m.enforceCapabilities(req); err != nil {
+		return nil, err
+	}
+	httpResp, err := m.do(ctx, req, true)
+	if err != nil {
+		return nil, err
+	}
+	if httpResp.StatusCode/100 != 2 {
+		defer httpResp.Body.Close()
+		return nil, m.apiError(httpResp)
+	}
+	return newStream(m, httpResp.Body), nil
+}
+
+// enforceCapabilities is the honest backstop behind the media layer: it
+// rejects (rather than silently mutates) requests the target cannot serve.
+// Why: a separate media layer resizes/transcodes images BEFORE requests
+// reach the provider, so anything still out of bounds here is a real error.
+func (m *model) enforceCapabilities(req llm.Request) error {
+	images := 0
+	for _, msg := range req.Messages {
+		for _, part := range msg.Parts {
+			img, ok := part.(llm.ImagePart)
+			if !ok {
+				continue
+			}
+			images++
+			if !m.caps.SupportsImages() {
+				return fmt.Errorf("%w: %s does not accept image input", llm.ErrUnsupported, m.fullName())
+			}
+			if !m.caps.MIMEAllowed(img.MIME) {
+				return fmt.Errorf("%w: %s does not accept image MIME %q", llm.ErrUnsupported, m.fullName(), img.MIME)
+			}
+			if m.caps.MaxImageBytes > 0 && len(img.Data) > m.caps.MaxImageBytes {
+				return fmt.Errorf("%w: image of %d bytes exceeds %s limit of %d bytes",
+					llm.ErrUnsupported, len(img.Data), m.fullName(), m.caps.MaxImageBytes)
+			}
+		}
+	}
+	if m.caps.MaxImagesPerReq > 0 && images > m.caps.MaxImagesPerReq {
+		return fmt.Errorf("%w: request carries %d images, %s allows at most %d",
+			llm.ErrUnsupported, images, m.fullName(), m.caps.MaxImagesPerReq)
+	}
+	return nil
+}
+
+// do builds and executes one Messages API call. Transport errors are wrapped
+// with context but NOT converted to *llm.APIError, so llm.Classify still
+// sees the underlying net.Error / syscall errno.
+func (m *model) do(ctx context.Context, req llm.Request, streaming bool) (*http.Response, error) {
+	p := m.provider
+	if p.apiKey == "" {
+		// Why request-time, not construction-time: New never fails by
+		// convention, and a 401-shaped APIError classifies permanent so
+		// chains fail fast past a misconfigured target.
+		return nil, &llm.APIError{
+			Provider: p.name,
+			Model:    m.id,
+			Status:   http.StatusUnauthorized,
+			Code:     "authentication_error",
+			Message:  "no API key configured: set ANTHROPIC_API_KEY or use WithAPIKey",
+		}
+	}
+
+	body, err := json.Marshal(buildWireRequest(m.id, req, p.maxTokens, streaming))
+	if err != nil {
+		return nil, fmt.Errorf("%s: encode request: %w", p.name, err)
+	}
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, p.baseURL+"/v1/messages", bytes.NewReader(body))
+	if err != nil {
+		return nil, fmt.Errorf("%s: build request: %w", p.name, err)
+	}
+	httpReq.Header.Set("x-api-key", p.apiKey)
+	httpReq.Header.Set("anthropic-version", apiVersion)
+	httpReq.Header.Set("content-type", "application/json")
+	if streaming {
+		httpReq.Header.Set("accept", "text/event-stream")
+	}
+
+	resp, err := p.client.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("%s: do request: %w", p.name, err)
+	}
+	return resp, nil
+}
+
+// apiError converts a non-2xx response into *llm.APIError, filling Code and
+// Message from the documented {"type":"error","error":{...}} body when it
+// parses, and falling back to the raw body text when it does not.
+func (m *model) apiError(resp *http.Response) error {
+	apiErr := &llm.APIError{
+		Provider: m.provider.name,
+		Model:    m.id,
+		Status:   resp.StatusCode,
+	}
+	body, err := io.ReadAll(io.LimitReader(resp.Body, 1<<20))
+	if err != nil {
+		return apiErr
+	}
+	var we wireErrorEnvelope
+	if json.Unmarshal(body, &we) == nil && we.Error.Type != "" {
+		apiErr.Code = we.Error.Type
+		apiErr.Message = we.Error.Message
+	} else {
+		apiErr.Message = strings.TrimSpace(string(body))
+	}
+	return apiErr
+}
+
+// toResponse maps a wire response onto the canonical llm.Response. Thinking
+// and other unrecognized block types are tolerated and skipped — they are
+// not part of the canonical content vocabulary.
+func (m *model) toResponse(wr *wireResponse) *llm.Response {
+	resp := &llm.Response{
+		FinishReason: mapStopReason(wr.StopReason),
+		Usage:        wr.Usage.toUsage(),
+		Model:        m.fullName(),
+		Raw:          wr,
+	}
+	for _, block := range wr.Content {
+		switch block.Type {
+		case "text":
+			resp.Parts = append(resp.Parts, llm.TextPart{Text: block.Text})
+		case "tool_use":
+			args := block.Input
+			if len(args) == 0 {
+				args = json.RawMessage("{}")
+			}
+			resp.ToolCalls = append(resp.ToolCalls, llm.ToolCall{
+				ID:        block.ID,
+				Name:      block.Name,
+				Arguments: args,
+			})
+		default:
+			// thinking, redacted_thinking, server-tool blocks, and any
+			// future types are skipped, not surfaced as parts.
+		}
+	}
+	return resp
+}
diff --git a/provider/anthropic/anthropic_test.go b/provider/anthropic/anthropic_test.go
new file mode 100644
index 0000000..73dde5d
--- /dev/null
+++ b/provider/anthropic/anthropic_test.go
@@ -0,0 +1,774 @@
+package anthropic
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"sync"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// okBody is a minimal successful Messages API response.
+const okBody = `{
+	"id": "msg_01",
+	"type": "message",
+	"role": "assistant",
+	"model": "claude-test",
+	"content": [{"type": "text", "text": "ok"}],
+	"stop_reason": "end_turn",
+	"usage": {"input_tokens": 3, "output_tokens": 5}
+}`
+
+// capture records the last request the test server received.
+type capture struct {
+	mu     sync.Mutex
+	hits   int
+	method string
+	path   string
+	header http.Header
+	body   []byte
+}
+
+func (c *capture) handler(status int, respBody string) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		body, _ := io.ReadAll(r.Body)
+		c.mu.Lock()
+		c.hits++
+		c.method = r.Method
+		c.path = r.URL.Path
+		c.header = r.Header.Clone()
+		c.body = body
+		c.mu.Unlock()
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(status)
+		_, _ = w.Write([]byte(respBody))
+	}
+}
+
+// bodyMap decodes the captured request body for key-presence assertions.
+func (c *capture) bodyMap(t *testing.T) map[string]any {
+	t.Helper()
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	var m map[string]any
+	if err := json.Unmarshal(c.body, &m); err != nil {
+		t.Fatalf("decode captured body: %v\nbody: %s", err, c.body)
+	}
+	return m
+}
+
+// newTestProvider spins up an httptest server and a provider pointed at it.
+func newTestProvider(t *testing.T, h http.Handler, opts ...Option) *Provider {
+	t.Helper()
+	srv := httptest.NewServer(h)
+	t.Cleanup(srv.Close)
+	return New(append([]Option{WithAPIKey("test-key"), WithBaseURL(srv.URL)}, opts...)...)
+}
+
+func mustModel(t *testing.T, p *Provider, id string, opts ...llm.ModelOption) llm.Model {
+	t.Helper()
+	m, err := p.Model(id, opts...)
+	if err != nil {
+		t.Fatalf("Model(%q): %v", id, err)
+	}
+	return m
+}
+
+func generate(t *testing.T, m llm.Model, req llm.Request, opts ...llm.Option) *llm.Response {
+	t.Helper()
+	resp, err := m.Generate(context.Background(), req, opts...)
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	return resp
+}
+
+func TestRequestHeadersAndPath(t *testing.T) {
+	var c capture
+	p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+	m := mustModel(t, p, "claude-test")
+
+	generate(t, m, llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+
+	if c.method != http.MethodPost {
+		t.Errorf("method = %q, want POST", c.method)
+	}
+	if c.path != "/v1/messages" {
+		t.Errorf("path = %q, want /v1/messages", c.path)
+	}
+	for header, want := range map[string]string{
+		"x-api-key":         "test-key",
+		"anthropic-version": "2023-06-01",
+		"content-type":      "application/json",
+	} {
+		if got := c.header.Get(header); got != want {
+			t.Errorf("header %s = %q, want %q", header, got, want)
+		}
+	}
+}
+
+func TestSystemFold(t *testing.T) {
+	var c capture
+	p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+	m := mustModel(t, p, "claude-test")
+
+	generate(t, m, llm.Request{
+		System: "base prompt",
+		Messages: []llm.Message{
+			llm.SystemText("first extra"),
+			llm.UserText("hi"),
+			llm.SystemText("second extra"),
+		},
+	})
+
+	body := c.bodyMap(t)
+	if got, want := body["system"], "base prompt\n\nfirst extra\n\nsecond extra"; got != want {
+		t.Errorf("system = %q, want %q", got, want)
+	}
+	msgs := body["messages"].([]any)
+	if len(msgs) != 1 {
+		t.Fatalf("messages length = %d, want 1 (system messages must be excluded)", len(msgs))
+	}
+	if role := msgs[0].(map[string]any)["role"]; role != "user" {
+		t.Errorf("remaining message role = %q, want user", role)
+	}
+}
+
+func TestNoSystemOmitsField(t *testing.T) {
+	var c capture
+	p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+	m := mustModel(t, p, "claude-test")
+
+	generate(t, m, llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+
+	if _, ok := c.bodyMap(t)["system"]; ok {
+		t.Error("system key present, want omitted when empty")
+	}
+}
+
+func TestMaxTokens(t *testing.T) {
+	t.Run("default 4096", func(t *testing.T) {
+		var c capture
+		p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+		generate(t, mustModel(t, p, "m"), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+		if got := c.bodyMap(t)["max_tokens"].(float64); got != 4096 {
+			t.Errorf("max_tokens = %v, want 4096", got)
+		}
+	})
+	t.Run("explicit wins", func(t *testing.T) {
+		var c capture
+		p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+		generate(t, mustModel(t, p, "m"), llm.Request{
+			Messages:  []llm.Message{llm.UserText("hi")},
+			MaxTokens: 123,
+		})
+		if got := c.bodyMap(t)["max_tokens"].(float64); got != 123 {
+			t.Errorf("max_tokens = %v, want 123", got)
+		}
+	})
+	t.Run("WithDefaultMaxTokens overrides default", func(t *testing.T) {
+		var c capture
+		p := newTestProvider(t, c.handler(http.StatusOK, okBody), WithDefaultMaxTokens(99))
+		generate(t, mustModel(t, p, "m"), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+		if got := c.bodyMap(t)["max_tokens"].(float64); got != 99 {
+			t.Errorf("max_tokens = %v, want 99", got)
+		}
+	})
+}
+
+func TestImageBlock(t *testing.T) {
+	var c capture
+	p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+	m := mustModel(t, p, "claude-test")
+
+	raw := []byte{0x01, 0x02, 0x03}
+	generate(t, m, llm.Request{Messages: []llm.Message{
+		llm.UserParts(llm.Text("look at this"), llm.Image("image/png", raw)),
+	}})
+
+	msgs := c.bodyMap(t)["messages"].([]any)
+	content := msgs[0].(map[string]any)["content"].([]any)
+	if len(content) != 2 {
+		t.Fatalf("content blocks = %d, want 2", len(content))
+	}
+	img := content[1].(map[string]any)
+	if img["type"] != "image" {
+		t.Fatalf("block type = %v, want image", img["type"])
+	}
+	src := img["source"].(map[string]any)
+	if src["type"] != "base64" {
+		t.Errorf("source type = %v, want base64", src["type"])
+	}
+	if src["media_type"] != "image/png" {
+		t.Errorf("media_type = %v, want image/png", src["media_type"])
+	}
+	if want := base64.StdEncoding.EncodeToString(raw); src["data"] != want {
+		t.Errorf("data = %v, want %q", src["data"], want)
+	}
+}
+
+func TestToolUseToolResultRoundTrip(t *testing.T) {
+	var c capture
+	p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+	m := mustModel(t, p, "claude-test")
+
+	generate(t, m, llm.Request{Messages: []llm.Message{
+		llm.UserText("weather?"),
+		{
+			Role:  llm.RoleAssistant,
+			Parts: []llm.Part{llm.Text("checking")},
+			ToolCalls: []llm.ToolCall{
+				{ID: "toolu_1", Name: "get_weather", Arguments: json.RawMessage(`{"location":"Paris"}`)},
+				{ID: "toolu_2", Name: "noop"}, // empty args must become {}
+			},
+		},
+		llm.ToolResultsMessage(
+			llm.ToolResult{ID: "toolu_1", Name: "get_weather", Content: "72F and sunny"},
+			llm.ToolResult{ID: "toolu_2", Name: "noop", Content: "boom", IsError: true},
+		),
+	}})
+
+	msgs := c.bodyMap(t)["messages"].([]any)
+	if len(msgs) != 3 {
+		t.Fatalf("messages = %d, want 3", len(msgs))
+	}
+
+	asst := msgs[1].(map[string]any)
+	if asst["role"] != "assistant" {
+		t.Errorf("messages[1].role = %v, want assistant", asst["role"])
+	}
+	asstContent := asst["content"].([]any)
+	if len(asstContent) != 3 {
+		t.Fatalf("assistant blocks = %d, want 3 (text + 2 tool_use)", len(asstContent))
+	}
+	tu := asstContent[1].(map[string]any)
+	if tu["type"] != "tool_use" || tu["id"] != "toolu_1" || tu["name"] != "get_weather" {
+		t.Errorf("tool_use block = %v", tu)
+	}
+	if loc := tu["input"].(map[string]any)["location"]; loc != "Paris" {
+		t.Errorf("tool_use input.location = %v, want Paris", loc)
+	}
+	if input := asstContent[2].(map[string]any)["input"].(map[string]any); len(input) != 0 {
+		t.Errorf("empty-args tool_use input = %v, want {}", input)
+	}
+
+	// RoleTool → ONE user message with one tool_result block per result.
+	toolMsg := msgs[2].(map[string]any)
+	if toolMsg["role"] != "user" {
+		t.Errorf("messages[2].role = %v, want user", toolMsg["role"])
+	}
+	results := toolMsg["content"].([]any)
+	if len(results) != 2 {
+		t.Fatalf("tool_result blocks = %d, want 2", len(results))
+	}
+	first := results[0].(map[string]any)
+	if first["type"] != "tool_result" || first["tool_use_id"] != "toolu_1" || first["content"] != "72F and sunny" {
+		t.Errorf("first tool_result = %v", first)
+	}
+	if _, ok := first["is_error"]; ok {
+		t.Error("first tool_result has is_error, want omitted when false")
+	}
+	second := results[1].(map[string]any)
+	if second["tool_use_id"] != "toolu_2" || second["is_error"] != true {
+		t.Errorf("second tool_result = %v, want is_error true", second)
+	}
+}
+
+func TestToolDefinitions(t *testing.T) {
+	var c capture
+	p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+	m := mustModel(t, p, "claude-test")
+
+	schema := json.RawMessage(`{"type":"object","properties":{"q":{"type":"string"}},"required":["q"]}`)
+	generate(t, m, llm.Request{
+		Messages: []llm.Message{llm.UserText("hi")},
+		Tools: []llm.Tool{
+			{Name: "search", Description: "Search the web.", Parameters: schema},
+			{Name: "ping"}, // nil Parameters → default empty object schema
+		},
+	})
+
+	tools := c.bodyMap(t)["tools"].([]any)
+	if len(tools) != 2 {
+		t.Fatalf("tools = %d, want 2", len(tools))
+	}
+	search := tools[0].(map[string]any)
+	if search["name"] != "search" || search["description"] != "Search the web." {
+		t.Errorf("tool[0] = %v", search)
+	}
+	if typ := search["input_schema"].(map[string]any)["type"]; typ != "object" {
+		t.Errorf("input_schema.type = %v, want object", typ)
+	}
+	ping := tools[1].(map[string]any)
+	if typ := ping["input_schema"].(map[string]any)["type"]; typ != "object" {
+		t.Errorf("nil-Parameters input_schema.type = %v, want object", typ)
+	}
+}
+
+func TestToolChoiceForms(t *testing.T) {
+	cases := []struct {
+		choice   string
+		wantType string // "" means the field must be absent
+		wantName string
+	}{
+		{choice: "", wantType: ""},
+		{choice: "auto", wantType: "auto"},
+		{choice: "required", wantType: "any"},
+		{choice: "none", wantType: "none"},
+		{choice: "get_weather", wantType: "tool", wantName: "get_weather"},
+	}
+	for _, tc := range cases {
+		t.Run("choice="+tc.choice, func(t *testing.T) {
+			var c capture
+			p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+			generate(t, mustModel(t, p, "m"), llm.Request{
+				Messages:   []llm.Message{llm.UserText("hi")},
+				ToolChoice: tc.choice,
+			})
+			body := c.bodyMap(t)
+			raw, present := body["tool_choice"]
+			if tc.wantType == "" {
+				if present {
+					t.Fatalf("tool_choice present (%v), want omitted", raw)
+				}
+				return
+			}
+			choice := raw.(map[string]any)
+			if choice["type"] != tc.wantType {
+				t.Errorf("tool_choice.type = %v, want %q", choice["type"], tc.wantType)
+			}
+			if tc.wantName != "" && choice["name"] != tc.wantName {
+				t.Errorf("tool_choice.name = %v, want %q", choice["name"], tc.wantName)
+			}
+		})
+	}
+}
+
+func TestOutputConfigFormat(t *testing.T) {
+	var c capture
+	p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+	m := mustModel(t, p, "claude-test")
+
+	schema := json.RawMessage(`{"type":"object","properties":{"name":{"type":"string"}},"required":["name"],"additionalProperties":false}`)
+	generate(t, m, llm.Request{Messages: []llm.Message{llm.UserText("hi")}},
+		llm.WithSchema(schema, "person"))
+
+	body := c.bodyMap(t)
+	format := body["output_config"].(map[string]any)["format"].(map[string]any)
+	if format["type"] != "json_schema" {
+		t.Errorf("output_config.format.type = %v, want json_schema", format["type"])
+	}
+	// Normalize both sides through any → Marshal (sorted keys) to compare.
+	got, _ := json.Marshal(format["schema"])
+	var want any
+	_ = json.Unmarshal(schema, &want)
+	wantJSON, _ := json.Marshal(want)
+	if string(got) != string(wantJSON) {
+		t.Errorf("schema = %s, want %s", got, wantJSON)
+	}
+}
+
+func TestOutputConfigOmittedWithoutSchema(t *testing.T) {
+	var c capture
+	p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+	generate(t, mustModel(t, p, "m"), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if _, ok := c.bodyMap(t)["output_config"]; ok {
+		t.Error("output_config present, want omitted when Schema is nil")
+	}
+}
+
+func TestSamplingKnobs(t *testing.T) {
+	t.Run("omitted when unset", func(t *testing.T) {
+		var c capture
+		p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+		generate(t, mustModel(t, p, "m"), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+		body := c.bodyMap(t)
+		if _, ok := body["temperature"]; ok {
+			t.Error("temperature present, want omitted when unset")
+		}
+		if _, ok := body["top_p"]; ok {
+			t.Error("top_p present, want omitted when unset")
+		}
+		if _, ok := body["stop_sequences"]; ok {
+			t.Error("stop_sequences present, want omitted when unset")
+		}
+	})
+	t.Run("present when set", func(t *testing.T) {
+		var c capture
+		p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+		generate(t, mustModel(t, p, "m"),
+			llm.Request{Messages: []llm.Message{llm.UserText("hi")}},
+			llm.WithTemperature(0), // explicit zero must still be sent
+			llm.WithTopP(0.9),
+			llm.WithStopSequences("END"))
+		body := c.bodyMap(t)
+		if got, ok := body["temperature"]; !ok || got.(float64) != 0 {
+			t.Errorf("temperature = %v (present=%v), want explicit 0", got, ok)
+		}
+		if got := body["top_p"].(float64); got != 0.9 {
+			t.Errorf("top_p = %v, want 0.9", got)
+		}
+		stops := body["stop_sequences"].([]any)
+		if len(stops) != 1 || stops[0] != "END" {
+			t.Errorf("stop_sequences = %v, want [END]", stops)
+		}
+	})
+}
+
+func TestStreamFieldOmittedOnGenerate(t *testing.T) {
+	var c capture
+	p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+	generate(t, mustModel(t, p, "m"), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if _, ok := c.bodyMap(t)["stream"]; ok {
+		t.Error("stream key present on Generate, want omitted")
+	}
+}
+
+func TestResponseParse(t *testing.T) {
+	const body = `{
+		"id": "msg_02",
+		"type": "message",
+		"role": "assistant",
+		"model": "claude-test",
+		"content": [
+			{"type": "thinking", "thinking": "pondering...", "signature": "sig"},
+			{"type": "text", "text": "I'll check the weather."},
+			{"type": "tool_use", "id": "toolu_9", "name": "get_weather", "input": {"location": "Paris"}}
+		],
+		"stop_reason": "tool_use",
+		"usage": {
+			"input_tokens": 3,
+			"output_tokens": 7,
+			"cache_creation_input_tokens": 10,
+			"cache_read_input_tokens": 20
+		}
+	}`
+	var c capture
+	p := newTestProvider(t, c.handler(http.StatusOK, body))
+	resp := generate(t, mustModel(t, p, "claude-test"), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+
+	if len(resp.Parts) != 1 {
+		t.Fatalf("parts = %d, want 1 (thinking blocks must be skipped)", len(resp.Parts))
+	}
+	if got := resp.Text(); got != "I'll check the weather." {
+		t.Errorf("text = %q", got)
+	}
+	if len(resp.ToolCalls) != 1 {
+		t.Fatalf("tool calls = %d, want 1", len(resp.ToolCalls))
+	}
+	call := resp.ToolCalls[0]
+	if call.ID != "toolu_9" || call.Name != "get_weather" {
+		t.Errorf("tool call = %+v", call)
+	}
+	var args map[string]any
+	if err := json.Unmarshal(call.Arguments, &args); err != nil || args["location"] != "Paris" {
+		t.Errorf("arguments = %s (err %v), want location Paris", call.Arguments, err)
+	}
+	if resp.FinishReason != llm.FinishToolCalls {
+		t.Errorf("finish = %q, want %q", resp.FinishReason, llm.FinishToolCalls)
+	}
+	// Total real input = input + cache_creation + cache_read.
+	if resp.Usage.InputTokens != 33 || resp.Usage.OutputTokens != 7 {
+		t.Errorf("usage = %+v, want {33 7}", resp.Usage)
+	}
+	if resp.Model != "anthropic/claude-test" {
+		t.Errorf("model = %q, want anthropic/claude-test", resp.Model)
+	}
+	if resp.Raw == nil {
+		t.Error("Raw = nil, want wire response")
+	}
+}
+
+func TestStopReasonMapping(t *testing.T) {
+	cases := map[string]llm.FinishReason{
+		"end_turn":                      llm.FinishStop,
+		"stop_sequence":                 llm.FinishStop,
+		"max_tokens":                    llm.FinishLength,
+		"model_context_window_exceeded": llm.FinishLength,
+		"tool_use":                      llm.FinishToolCalls,
+		"refusal":                       llm.FinishContentFilter,
+		"pause_turn":                    llm.FinishOther,
+		"some_future_reason":            llm.FinishOther,
+	}
+	for stop, want := range cases {
+		if got := mapStopReason(stop); got != want {
+			t.Errorf("mapStopReason(%q) = %q, want %q", stop, got, want)
+		}
+	}
+}
+
+func TestHTTPErrorMapping(t *testing.T) {
+	cases := []struct {
+		name      string
+		status    int
+		body      string
+		wantCode  string
+		wantClass llm.ErrorClass
+	}{
+		{
+			name:     "429 rate limit is transient",
+			status:   http.StatusTooManyRequests,
+			body:     `{"type":"error","error":{"type":"rate_limit_error","message":"slow down"}}`,
+			wantCode: "rate_limit_error", wantClass: llm.ClassTransient,
+		},
+		{
+			name:     "529 overloaded is transient",
+			status:   529,
+			body:     `{"type":"error","error":{"type":"overloaded_error","message":"Overloaded"}}`,
+			wantCode: "overloaded_error", wantClass: llm.ClassTransient,
+		},
+		{
+			name:     "401 auth is permanent",
+			status:   http.StatusUnauthorized,
+			body:     `{"type":"error","error":{"type":"authentication_error","message":"invalid x-api-key"}}`,
+			wantCode: "authentication_error", wantClass: llm.ClassPermanent,
+		},
+		{
+			name:     "404 is permanent",
+			status:   http.StatusNotFound,
+			body:     `{"type":"error","error":{"type":"not_found_error","message":"model: nope"}}`,
+			wantCode: "not_found_error", wantClass: llm.ClassPermanent,
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			var c capture
+			p := newTestProvider(t, c.handler(tc.status, tc.body))
+			_, err := mustModel(t, p, "claude-test").Generate(context.Background(),
+				llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+			if err == nil {
+				t.Fatal("Generate succeeded, want error")
+			}
+			apiErr, ok := errors.AsType[*llm.APIError](err)
+			if !ok {
+				t.Fatalf("error %T (%v), want *llm.APIError", err, err)
+			}
+			if apiErr.Provider != "anthropic" || apiErr.Model != "claude-test" {
+				t.Errorf("provider/model = %s/%s", apiErr.Provider, apiErr.Model)
+			}
+			if apiErr.Status != tc.status {
+				t.Errorf("status = %d, want %d", apiErr.Status, tc.status)
+			}
+			if apiErr.Code != tc.wantCode {
+				t.Errorf("code = %q, want %q", apiErr.Code, tc.wantCode)
+			}
+			if apiErr.Message == "" {
+				t.Error("message empty, want provider message")
+			}
+			if got := llm.Classify(err); got != tc.wantClass {
+				t.Errorf("Classify = %v, want %v", got, tc.wantClass)
+			}
+		})
+	}
+
+	t.Run("404 unwraps to ErrModelNotFound", func(t *testing.T) {
+		var c capture
+		p := newTestProvider(t, c.handler(http.StatusNotFound,
+			`{"type":"error","error":{"type":"not_found_error","message":"model: nope"}}`))
+		_, err := mustModel(t, p, "missing").Generate(context.Background(),
+			llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+		if !errors.Is(err, llm.ErrModelNotFound) {
+			t.Errorf("errors.Is(err, ErrModelNotFound) = false for %v", err)
+		}
+	})
+
+	t.Run("non-JSON error body falls back to raw text", func(t *testing.T) {
+		var c capture
+		p := newTestProvider(t, c.handler(http.StatusBadGateway, "upstream exploded"))
+		_, err := mustModel(t, p, "m").Generate(context.Background(),
+			llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+		apiErr, ok := errors.AsType[*llm.APIError](err)
+		if !ok {
+			t.Fatalf("error %T, want *llm.APIError", err)
+		}
+		if apiErr.Status != http.StatusBadGateway || apiErr.Message != "upstream exploded" {
+			t.Errorf("apiErr = %+v", apiErr)
+		}
+	})
+}
+
+func TestMissingAPIKey(t *testing.T) {
+	t.Setenv("ANTHROPIC_API_KEY", "") // isolate from any real environment
+
+	var c capture
+	srv := httptest.NewServer(c.handler(http.StatusOK, okBody))
+	t.Cleanup(srv.Close)
+
+	p := New(WithBaseURL(srv.URL)) // construction must not fail
+	_, err := mustModel(t, p, "claude-test").Generate(context.Background(),
+		llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	apiErr, ok := errors.AsType[*llm.APIError](err)
+	if !ok {
+		t.Fatalf("error %T (%v), want *llm.APIError", err, err)
+	}
+	if apiErr.Status != http.StatusUnauthorized || apiErr.Code != "authentication_error" {
+		t.Errorf("apiErr = %+v, want 401 authentication_error", apiErr)
+	}
+	if llm.Classify(err) != llm.ClassPermanent {
+		t.Error("missing key must classify permanent")
+	}
+	if c.hits != 0 {
+		t.Errorf("server hits = %d, want 0 (no request without a key)", c.hits)
+	}
+}
+
+func TestAPIKeyFromEnv(t *testing.T) {
+	t.Setenv("ANTHROPIC_API_KEY", "env-key")
+
+	var c capture
+	srv := httptest.NewServer(c.handler(http.StatusOK, okBody))
+	t.Cleanup(srv.Close)
+
+	p := New(WithBaseURL(srv.URL))
+	generate(t, mustModel(t, p, "m"), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if got := c.header.Get("x-api-key"); got != "env-key" {
+		t.Errorf("x-api-key = %q, want env-key", got)
+	}
+}
+
+func TestCapabilityEnforcement(t *testing.T) {
+	img := func(mime string, n int) llm.Part { return llm.Image(mime, make([]byte, n)) }
+
+	cases := []struct {
+		name string
+		caps *llm.Capabilities // nil = provider defaults
+		req  llm.Request
+	}{
+		{
+			name: "images unsupported",
+			caps: &llm.Capabilities{}, // MaxImagesPerReq 0 = no images
+			req:  llm.Request{Messages: []llm.Message{llm.UserParts(img("image/png", 4))}},
+		},
+		{
+			name: "too many images",
+			caps: &llm.Capabilities{MaxImagesPerReq: 1},
+			req: llm.Request{Messages: []llm.Message{
+				llm.UserParts(img("image/png", 4), img("image/png", 4)),
+			}},
+		},
+		{
+			name: "disallowed MIME",
+			req:  llm.Request{Messages: []llm.Message{llm.UserParts(img("image/bmp", 4))}},
+		},
+		{
+			name: "image too large",
+			caps: &llm.Capabilities{MaxImagesPerReq: 1, MaxImageBytes: 2},
+			req:  llm.Request{Messages: []llm.Message{llm.UserParts(img("image/png", 3))}},
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			var c capture
+			p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+			var opts []llm.ModelOption
+			if tc.caps != nil {
+				opts = append(opts, llm.WithCapabilities(*tc.caps))
+			}
+			m := mustModel(t, p, "claude-test", opts...)
+
+			_, err := m.Generate(context.Background(), tc.req)
+			if !errors.Is(err, llm.ErrUnsupported) {
+				t.Errorf("Generate err = %v, want ErrUnsupported", err)
+			}
+			_, err = m.Stream(context.Background(), tc.req)
+			if !errors.Is(err, llm.ErrUnsupported) {
+				t.Errorf("Stream err = %v, want ErrUnsupported", err)
+			}
+			if c.hits != 0 {
+				t.Errorf("server hits = %d, want 0 (rejected before sending)", c.hits)
+			}
+		})
+	}
+
+	t.Run("within limits passes", func(t *testing.T) {
+		var c capture
+		p := newTestProvider(t, c.handler(http.StatusOK, okBody))
+		generate(t, mustModel(t, p, "m"), llm.Request{
+			Messages: []llm.Message{llm.UserParts(llm.Text("ok"), img("image/jpeg", 16))},
+		})
+		if c.hits != 1 {
+			t.Errorf("server hits = %d, want 1", c.hits)
+		}
+	})
+}
+
+func TestCompatEndpointWithNameAndBaseURL(t *testing.T) {
+	var c capture
+	p := newTestProvider(t, c.handler(http.StatusOK, okBody), WithName("compat"))
+	if p.Name() != "compat" {
+		t.Errorf("Name() = %q, want compat", p.Name())
+	}
+	resp := generate(t, mustModel(t, p, "claude-test"), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if resp.Model != "compat/claude-test" {
+		t.Errorf("resp.Model = %q, want compat/claude-test", resp.Model)
+	}
+
+	var ec capture
+	pe := newTestProvider(t, ec.handler(http.StatusTooManyRequests,
+		`{"type":"error","error":{"type":"rate_limit_error","message":"x"}}`), WithName("compat"))
+	_, err := mustModel(t, pe, "m").Generate(context.Background(),
+		llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	apiErr, ok := errors.AsType[*llm.APIError](err)
+	if !ok || apiErr.Provider != "compat" {
+		t.Errorf("error provider = %v, want compat (err %v)", apiErr, err)
+	}
+}
+
+func TestCapabilitiesDefaultsAndOverrides(t *testing.T) {
+	p := New(WithAPIKey("k"))
+	m := mustModel(t, p, "m")
+	caps := m.Capabilities()
+	if !caps.SupportsTools || !caps.SupportsStructured || !caps.SupportsStreaming {
+		t.Errorf("default feature flags = %+v, want all true", caps)
+	}
+	if caps.MaxImagesPerReq != 100 || caps.MaxImageBytes != 10<<20 || caps.MaxImageDimension != 8000 {
+		t.Errorf("default image limits = %+v", caps)
+	}
+	wantMIME := []string{"image/jpeg", "image/png", "image/gif", "image/webp"}
+	if len(caps.AllowedImageMIME) != len(wantMIME) {
+		t.Fatalf("AllowedImageMIME = %v, want %v", caps.AllowedImageMIME, wantMIME)
+	}
+	for i, mime := range wantMIME {
+		if caps.AllowedImageMIME[i] != mime {
+			t.Errorf("AllowedImageMIME[%d] = %q, want %q", i, caps.AllowedImageMIME[i], mime)
+		}
+	}
+
+	custom := llm.Capabilities{SupportsStreaming: true, MaxImagesPerReq: 1}
+	p2 := New(WithAPIKey("k"), WithDefaultCapabilities(custom))
+	if got := mustModel(t, p2, "m").Capabilities(); got.MaxImagesPerReq != 1 || got.SupportsTools {
+		t.Errorf("WithDefaultCapabilities not applied: %+v", got)
+	}
+
+	perModel := llm.Capabilities{SupportsTools: true}
+	if got := mustModel(t, p2, "m", llm.WithCapabilities(perModel)).Capabilities(); !got.SupportsTools || got.MaxImagesPerReq != 0 {
+		t.Errorf("per-model capabilities not applied: %+v", got)
+	}
+}
+
+func TestTransportErrorNotAPIError(t *testing.T) {
+	// Point at a server that is immediately closed: the connection failure
+	// must surface as a wrapped transport error, not *llm.APIError.
+	srv := httptest.NewServer(http.HandlerFunc(func(http.ResponseWriter, *http.Request) {}))
+	url := srv.URL
+	srv.Close()
+
+	p := New(WithAPIKey("k"), WithBaseURL(url))
+	_, err := mustModel(t, p, "m").Generate(context.Background(),
+		llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err == nil {
+		t.Fatal("Generate succeeded, want transport error")
+	}
+	if _, ok := errors.AsType[*llm.APIError](err); ok {
+		t.Errorf("transport error wrapped in APIError: %v", err)
+	}
+	if llm.Classify(err) != llm.ClassTransient {
+		t.Errorf("connection failure must classify transient: %v", err)
+	}
+}
diff --git a/provider/anthropic/stream.go b/provider/anthropic/stream.go
new file mode 100644
index 0000000..cd6cc2a
--- /dev/null
+++ b/provider/anthropic/stream.go
@@ -0,0 +1,247 @@
+package anthropic
+
+import (
+	"bufio"
+	"encoding/json"
+	"fmt"
+	"io"
+	"strings"
+	"sync"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// wireStreamEvent is the union of all SSE data payloads the Messages API
+// emits. Dispatch is on Type (the data always carries one), so the SSE
+// "event:" line is informational only.
+type wireStreamEvent struct {
+	Type  string `json:"type"`
+	Index int    `json:"index"`
+
+	// message_start
+	Message *struct {
+		Usage wireUsage `json:"usage"`
+	} `json:"message"`
+
+	// content_block_start
+	ContentBlock *struct {
+		Type string `json:"type"`
+		ID   string `json:"id"`
+		Name string `json:"name"`
+	} `json:"content_block"`
+
+	// content_block_delta / message_delta
+	Delta struct {
+		Type        string `json:"type"`
+		Text        string `json:"text"`
+		PartialJSON string `json:"partial_json"`
+		StopReason  string `json:"stop_reason"`
+	} `json:"delta"`
+
+	// message_delta
+	Usage *wireUsage `json:"usage"`
+
+	// error
+	Error *struct {
+		Type    string `json:"type"`
+		Message string `json:"message"`
+	} `json:"error"`
+}
+
+// stream adapts the Messages API SSE stream to llm.Stream.
+//
+// Why single-threaded pull (no reader goroutine): Next is already the
+// consumer's pull point, so parsing lazily inside Next keeps cancellation,
+// buffering, and error propagation trivial — Close just closes the body and
+// the next read fails.
+type stream struct {
+	provider string
+	model    string
+	full     string // provider/model
+	body     io.ReadCloser
+	scanner  *bufio.Scanner
+
+	// accumulated response
+	parts     []llm.Part
+	toolCalls []llm.ToolCall
+	usage     llm.Usage
+	finish    llm.FinishReason
+
+	// current content block state
+	blockType string
+	textBuf   strings.Builder
+	toolID    string
+	toolName  string
+	argsBuf   strings.Builder
+
+	done      bool // final Response event emitted
+	closeOnce sync.Once
+	closeErr  error
+}
+
+func newStream(m *model, body io.ReadCloser) *stream {
+	sc := bufio.NewScanner(body)
+	// Why a large limit: one SSE line carries one whole delta; default 64K
+	// can be exceeded by large structured-output or tool-argument deltas.
+	sc.Buffer(make([]byte, 0, 64*1024), 10*1024*1024)
+	return &stream{
+		provider: m.provider.name,
+		model:    m.id,
+		full:     m.fullName(),
+		body:     body,
+		scanner:  sc,
+		finish:   llm.FinishOther,
+	}
+}
+
+// Close implements llm.Stream. Safe to call at any time and more than once.
+func (s *stream) Close() error {
+	s.closeOnce.Do(func() { s.closeErr = s.body.Close() })
+	return s.closeErr
+}
+
+// Next implements llm.Stream. It emits TextDelta fragments as they arrive,
+// fully-assembled ToolCalls at content_block_stop, exactly one final
+// Response event at message_stop, then io.EOF.
+func (s *stream) Next() (llm.StreamEvent, error) {
+	if s.done {
+		return llm.StreamEvent{}, io.EOF
+	}
+	for {
+		data, err := s.nextData()
+		if err != nil {
+			return llm.StreamEvent{}, err
+		}
+		var ev wireStreamEvent
+		if err := json.Unmarshal([]byte(data), &ev); err != nil {
+			return llm.StreamEvent{}, fmt.Errorf("%s: decode stream event: %w", s.provider, err)
+		}
+
+		switch ev.Type {
+		case "message_start":
+			if ev.Message != nil {
+				s.usage = ev.Message.Usage.toUsage()
+			}
+
+		case "content_block_start":
+			s.blockType = ""
+			s.textBuf.Reset()
+			s.argsBuf.Reset()
+			if ev.ContentBlock != nil {
+				s.blockType = ev.ContentBlock.Type
+				if s.blockType == "tool_use" {
+					s.toolID = ev.ContentBlock.ID
+					s.toolName = ev.ContentBlock.Name
+				}
+			}
+
+		case "content_block_delta":
+			switch ev.Delta.Type {
+			case "text_delta":
+				s.textBuf.WriteString(ev.Delta.Text)
+				return llm.StreamEvent{TextDelta: ev.Delta.Text}, nil
+			case "input_json_delta":
+				// Buffer partial JSON internally; consumers never see it.
+				s.argsBuf.WriteString(ev.Delta.PartialJSON)
+			default:
+				// thinking_delta / signature_delta: tolerated, skipped.
+			}
+
+		case "content_block_stop":
+			if event, ok := s.finishBlock(); ok {
+				return event, nil
+			}
+
+		case "message_delta":
+			if ev.Delta.StopReason != "" {
+				s.finish = mapStopReason(ev.Delta.StopReason)
+			}
+			if ev.Usage != nil {
+				// Output tokens arrive cumulatively in the final delta;
+				// input tokens were reported in message_start.
+				s.usage.OutputTokens = ev.Usage.OutputTokens
+			}
+
+		case "message_stop":
+			s.done = true
+			return llm.StreamEvent{Response: &llm.Response{
+				Parts:        s.parts,
+				ToolCalls:    s.toolCalls,
+				FinishReason: s.finish,
+				Usage:        s.usage,
+				Model:        s.full,
+			}}, nil
+
+		case "error":
+			// Mid-stream failure after the 200 (e.g. overloaded_error).
+			// Status stays 0: there is no HTTP status for it, and the
+			// default Classify treats it as transient, which fits overload.
+			apiErr := &llm.APIError{Provider: s.provider, Model: s.model}
+			if ev.Error != nil {
+				apiErr.Code = ev.Error.Type
+				apiErr.Message = ev.Error.Message
+			}
+			return llm.StreamEvent{}, apiErr
+
+		default:
+			// ping and unknown event types: ignored.
+		}
+	}
+}
+
+// finishBlock closes out the current content block, appending its result to
+// the accumulated response. Tool-use blocks produce a stream event.
+func (s *stream) finishBlock() (llm.StreamEvent, bool) {
+	defer func() {
+		s.blockType = ""
+		s.textBuf.Reset()
+		s.argsBuf.Reset()
+	}()
+	switch s.blockType {
+	case "text":
+		if s.textBuf.Len() > 0 {
+			s.parts = append(s.parts, llm.TextPart{Text: s.textBuf.String()})
+		}
+	case "tool_use":
+		args := s.argsBuf.String()
+		if args == "" {
+			// A tool called with no arguments streams zero (or empty)
+			// input_json_delta fragments; the canonical form is "{}".
+			args = "{}"
+		}
+		call := llm.ToolCall{ID: s.toolID, Name: s.toolName, Arguments: json.RawMessage(args)}
+		s.toolCalls = append(s.toolCalls, call)
+		return llm.StreamEvent{ToolCall: &call}, true
+	}
+	return llm.StreamEvent{}, false
+}
+
+// nextData reads SSE lines until one complete event's data is assembled
+// (multi-line data fields are joined with "\n" per the SSE spec). "event:"
+// lines and comments are ignored; dispatch keys off the JSON "type" field.
+func (s *stream) nextData() (string, error) {
+	var data strings.Builder
+	for s.scanner.Scan() {
+		line := s.scanner.Text()
+		if line == "" {
+			if data.Len() > 0 {
+				return data.String(), nil
+			}
+			continue
+		}
+		if rest, ok := strings.CutPrefix(line, "data:"); ok {
+			if data.Len() > 0 {
+				data.WriteByte('\n')
+			}
+			data.WriteString(strings.TrimPrefix(rest, " "))
+		}
+	}
+	if err := s.scanner.Err(); err != nil {
+		return "", fmt.Errorf("%s: read stream: %w", s.provider, err)
+	}
+	if data.Len() > 0 {
+		return data.String(), nil
+	}
+	// EOF before message_stop: the connection dropped mid-response.
+	return "", fmt.Errorf("%s: stream ended before message_stop: %w", s.provider, io.ErrUnexpectedEOF)
+}
diff --git a/provider/anthropic/stream_test.go b/provider/anthropic/stream_test.go
new file mode 100644
index 0000000..1445471
--- /dev/null
+++ b/provider/anthropic/stream_test.go
@@ -0,0 +1,324 @@
+package anthropic
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// sse joins data payloads into an SSE body. Each payload becomes one event
+// ("event:" name derived from the JSON type field is what the real API
+// sends, but the client dispatches on the data, so a generic name is fine).
+func sse(payloads ...string) string {
+	var b strings.Builder
+	for _, p := range payloads {
+		b.WriteString("event: event\n")
+		b.WriteString("data: ")
+		b.WriteString(p)
+		b.WriteString("\n\n")
+	}
+	return b.String()
+}
+
+func sseServer(t *testing.T, c *capture, body string) *Provider {
+	t.Helper()
+	return newTestProvider(t, http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		raw, _ := io.ReadAll(r.Body)
+		c.mu.Lock()
+		c.hits++
+		c.header = r.Header.Clone()
+		c.body = raw
+		c.mu.Unlock()
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = io.WriteString(w, body)
+	}))
+}
+
+// drain collects all events until io.EOF, failing the test on any error.
+func drain(t *testing.T, s llm.Stream) []llm.StreamEvent {
+	t.Helper()
+	var events []llm.StreamEvent
+	for {
+		ev, err := s.Next()
+		if err == io.EOF {
+			return events
+		}
+		if err != nil {
+			t.Fatalf("Next: %v", err)
+		}
+		events = append(events, ev)
+	}
+}
+
+func openStream(t *testing.T, p *Provider, modelID string) llm.Stream {
+	t.Helper()
+	s, err := mustModel(t, p, modelID).Stream(context.Background(),
+		llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err != nil {
+		t.Fatalf("Stream: %v", err)
+	}
+	t.Cleanup(func() { _ = s.Close() })
+	return s
+}
+
+func TestStreamTextDeltas(t *testing.T) {
+	body := sse(
+		`{"type":"message_start","message":{"id":"msg_1","type":"message","role":"assistant","content":[],"model":"m","usage":{"input_tokens":10,"cache_creation_input_tokens":2,"cache_read_input_tokens":3,"output_tokens":1}}}`,
+		`{"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}`,
+		`{"type":"ping"}`,
+		`{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Hel"}}`,
+		`{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"lo"}}`,
+		`{"type":"content_block_stop","index":0}`,
+		`{"type":"content_block_start","index":1,"content_block":{"type":"text","text":""}}`,
+		`{"type":"content_block_delta","index":1,"delta":{"type":"text_delta","text":" world"}}`,
+		`{"type":"content_block_stop","index":1}`,
+		`{"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"output_tokens":12}}`,
+		`{"type":"message_stop"}`,
+	)
+	var c capture
+	p := sseServer(t, &c, body)
+	s := openStream(t, p, "claude-test")
+	events := drain(t, s)
+
+	if len(events) != 4 {
+		t.Fatalf("events = %d, want 4 (3 deltas + final response)", len(events))
+	}
+	for i, want := range []string{"Hel", "lo", " world"} {
+		if events[i].TextDelta != want {
+			t.Errorf("event[%d].TextDelta = %q, want %q", i, events[i].TextDelta, want)
+		}
+	}
+
+	final := events[3].Response
+	if final == nil {
+		t.Fatal("last event has no Response")
+	}
+	if len(final.Parts) != 2 {
+		t.Fatalf("final parts = %d, want 2 (one per text block)", len(final.Parts))
+	}
+	if final.Text() != "Hello world" {
+		t.Errorf("final text = %q, want %q", final.Text(), "Hello world")
+	}
+	if final.FinishReason != llm.FinishStop {
+		t.Errorf("finish = %q, want stop", final.FinishReason)
+	}
+	// Input = 10+2+3 from message_start; output = 12 from message_delta.
+	if final.Usage.InputTokens != 15 || final.Usage.OutputTokens != 12 {
+		t.Errorf("usage = %+v, want {15 12}", final.Usage)
+	}
+	if final.Model != "anthropic/claude-test" {
+		t.Errorf("model = %q, want anthropic/claude-test", final.Model)
+	}
+
+	// Past EOF, Next keeps returning io.EOF.
+	if _, err := s.Next(); err != io.EOF {
+		t.Errorf("Next after EOF = %v, want io.EOF", err)
+	}
+
+	// The request must carry "stream": true.
+	if streamFlag := c.bodyMap(t)["stream"]; streamFlag != true {
+		t.Errorf("request stream = %v, want true", streamFlag)
+	}
+}
+
+func TestStreamToolCallAssembly(t *testing.T) {
+	body := sse(
+		`{"type":"message_start","message":{"id":"msg_1","usage":{"input_tokens":8,"output_tokens":1}}}`,
+		`{"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}`,
+		`{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Checking."}}`,
+		`{"type":"content_block_stop","index":0}`,
+		`{"type":"content_block_start","index":1,"content_block":{"type":"tool_use","id":"toolu_9","name":"get_weather","input":{}}}`,
+		`{"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":""}}`,
+		`{"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"{\"location\":"}}`,
+		`{"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":" \"San Francisco, CA\"}"}}`,
+		`{"type":"content_block_stop","index":1}`,
+		`{"type":"content_block_start","index":2,"content_block":{"type":"tool_use","id":"toolu_10","name":"noop","input":{}}}`,
+		`{"type":"content_block_stop","index":2}`,
+		`{"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"output_tokens":21}}`,
+		`{"type":"message_stop"}`,
+	)
+	var c capture
+	p := sseServer(t, &c, body)
+	events := drain(t, openStream(t, p, "claude-test"))
+
+	if len(events) != 4 {
+		t.Fatalf("events = %d, want 4 (text, 2 tool calls, final)", len(events))
+	}
+	if events[0].TextDelta != "Checking." {
+		t.Errorf("event[0] = %+v, want text delta", events[0])
+	}
+
+	call := events[1].ToolCall
+	if call == nil {
+		t.Fatal("event[1] has no ToolCall")
+	}
+	if call.ID != "toolu_9" || call.Name != "get_weather" {
+		t.Errorf("tool call = %+v", call)
+	}
+	var args map[string]any
+	if err := json.Unmarshal(call.Arguments, &args); err != nil {
+		t.Fatalf("assembled arguments invalid JSON: %v (%s)", err, call.Arguments)
+	}
+	if args["location"] != "San Francisco, CA" {
+		t.Errorf("arguments = %v", args)
+	}
+
+	empty := events[2].ToolCall
+	if empty == nil || empty.ID != "toolu_10" {
+		t.Fatalf("event[2] = %+v, want second tool call", events[2])
+	}
+	if string(empty.Arguments) != "{}" {
+		t.Errorf("empty tool call arguments = %s, want {}", empty.Arguments)
+	}
+
+	final := events[3].Response
+	if final == nil {
+		t.Fatal("last event has no Response")
+	}
+	if len(final.ToolCalls) != 2 {
+		t.Errorf("final tool calls = %d, want 2", len(final.ToolCalls))
+	}
+	if final.FinishReason != llm.FinishToolCalls {
+		t.Errorf("finish = %q, want tool_calls", final.FinishReason)
+	}
+	if final.Text() != "Checking." {
+		t.Errorf("final text = %q", final.Text())
+	}
+	if final.Usage.InputTokens != 8 || final.Usage.OutputTokens != 21 {
+		t.Errorf("usage = %+v, want {8 21}", final.Usage)
+	}
+}
+
+func TestStreamThinkingSkipped(t *testing.T) {
+	body := sse(
+		`{"type":"message_start","message":{"id":"msg_1","usage":{"input_tokens":5,"output_tokens":1}}}`,
+		`{"type":"content_block_start","index":0,"content_block":{"type":"thinking","thinking":""}}`,
+		`{"type":"content_block_delta","index":0,"delta":{"type":"thinking_delta","thinking":"hmm"}}`,
+		`{"type":"content_block_delta","index":0,"delta":{"type":"signature_delta","signature":"sig"}}`,
+		`{"type":"content_block_stop","index":0}`,
+		`{"type":"content_block_start","index":1,"content_block":{"type":"text","text":""}}`,
+		`{"type":"content_block_delta","index":1,"delta":{"type":"text_delta","text":"hi"}}`,
+		`{"type":"content_block_stop","index":1}`,
+		`{"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"output_tokens":2}}`,
+		`{"type":"message_stop"}`,
+	)
+	var c capture
+	p := sseServer(t, &c, body)
+	events := drain(t, openStream(t, p, "claude-test"))
+
+	if len(events) != 2 {
+		t.Fatalf("events = %d, want 2 (thinking produces none)", len(events))
+	}
+	if events[0].TextDelta != "hi" {
+		t.Errorf("event[0] = %+v, want TextDelta hi", events[0])
+	}
+	final := events[1].Response
+	if final == nil || len(final.Parts) != 1 || final.Text() != "hi" {
+		t.Errorf("final = %+v, want single text part %q", final, "hi")
+	}
+}
+
+func TestStreamMidStreamError(t *testing.T) {
+	body := sse(
+		`{"type":"message_start","message":{"id":"msg_1","usage":{"input_tokens":5,"output_tokens":1}}}`,
+		`{"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}`,
+		`{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"par"}}`,
+		`{"type":"error","error":{"type":"overloaded_error","message":"Overloaded"}}`,
+	)
+	var c capture
+	p := sseServer(t, &c, body)
+	s := openStream(t, p, "claude-test")
+
+	ev, err := s.Next()
+	if err != nil || ev.TextDelta != "par" {
+		t.Fatalf("first Next = (%+v, %v), want text delta", ev, err)
+	}
+	_, err = s.Next()
+	if err == nil {
+		t.Fatal("second Next succeeded, want mid-stream error")
+	}
+	apiErr, ok := errors.AsType[*llm.APIError](err)
+	if !ok {
+		t.Fatalf("error %T (%v), want *llm.APIError", err, err)
+	}
+	if apiErr.Code != "overloaded_error" || apiErr.Message != "Overloaded" || apiErr.Status != 0 {
+		t.Errorf("apiErr = %+v", apiErr)
+	}
+	if llm.Classify(err) != llm.ClassTransient {
+		t.Error("overloaded_error must classify transient")
+	}
+}
+
+func TestStreamHTTPErrorBeforeEvents(t *testing.T) {
+	var c capture
+	p := newTestProvider(t, c.handler(529,
+		`{"type":"error","error":{"type":"overloaded_error","message":"Overloaded"}}`))
+	_, err := mustModel(t, p, "claude-test").Stream(context.Background(),
+		llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err == nil {
+		t.Fatal("Stream succeeded, want APIError before any events")
+	}
+	apiErr, ok := errors.AsType[*llm.APIError](err)
+	if !ok {
+		t.Fatalf("error %T (%v), want *llm.APIError", err, err)
+	}
+	if apiErr.Status != 529 || apiErr.Code != "overloaded_error" {
+		t.Errorf("apiErr = %+v, want 529 overloaded_error", apiErr)
+	}
+	if llm.Classify(err) != llm.ClassTransient {
+		t.Error("529 must classify transient")
+	}
+}
+
+func TestStreamTruncatedBody(t *testing.T) {
+	// Stream ends without message_stop: Next must surface unexpected EOF.
+	body := sse(
+		`{"type":"message_start","message":{"id":"msg_1","usage":{"input_tokens":5,"output_tokens":1}}}`,
+		`{"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}`,
+		`{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"hi"}}`,
+	)
+	var c capture
+	p := sseServer(t, &c, body)
+	s := openStream(t, p, "claude-test")
+
+	if ev, err := s.Next(); err != nil || ev.TextDelta != "hi" {
+		t.Fatalf("first Next = (%+v, %v)", ev, err)
+	}
+	if _, err := s.Next(); !errors.Is(err, io.ErrUnexpectedEOF) {
+		t.Errorf("Next on truncated stream = %v, want io.ErrUnexpectedEOF", err)
+	}
+}
+
+func TestStreamCloseIsSafe(t *testing.T) {
+	body := sse(
+		`{"type":"message_start","message":{"id":"msg_1","usage":{"input_tokens":5,"output_tokens":1}}}`,
+		`{"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}`,
+		`{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"hi"}}`,
+		`{"type":"content_block_stop","index":0}`,
+		`{"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"output_tokens":2}}`,
+		`{"type":"message_stop"}`,
+	)
+	var c capture
+	p := sseServer(t, &c, body)
+	s := openStream(t, p, "claude-test")
+
+	if err := s.Close(); err != nil {
+		t.Errorf("first Close: %v", err)
+	}
+	if err := s.Close(); err != nil {
+		t.Errorf("second Close: %v", err)
+	}
+
+	// After EOF, Close is still fine.
+	s2 := openStream(t, p, "claude-test")
+	drain(t, s2)
+	if err := s2.Close(); err != nil {
+		t.Errorf("Close after EOF: %v", err)
+	}
+}
diff --git a/provider/anthropic/wire.go b/provider/anthropic/wire.go
new file mode 100644
index 0000000..d80b5da
--- /dev/null
+++ b/provider/anthropic/wire.go
@@ -0,0 +1,299 @@
+package anthropic
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// Wire types mirror the Messages API JSON shapes (June 2026 docs). Only the
+// fields majordomo uses are modeled; unknown response fields are ignored by
+// encoding/json.
+
+type wireRequest struct {
+	Model         string            `json:"model"`
+	MaxTokens     int               `json:"max_tokens"`
+	System        string            `json:"system,omitempty"`
+	Messages      []wireMessage     `json:"messages"`
+	Stream        bool              `json:"stream,omitempty"`
+	Tools         []wireTool        `json:"tools,omitempty"`
+	ToolChoice    *wireToolChoice   `json:"tool_choice,omitempty"`
+	Temperature   *float64          `json:"temperature,omitempty"`
+	TopP          *float64          `json:"top_p,omitempty"`
+	StopSequences []string          `json:"stop_sequences,omitempty"`
+	OutputConfig  *wireOutputConfig `json:"output_config,omitempty"`
+}
+
+type wireMessage struct {
+	Role    string      `json:"role"`
+	Content []wireBlock `json:"content"`
+}
+
+// wireBlock is a request-side content block. Exactly one shape is populated
+// per block, keyed by Type: text, image, tool_use, or tool_result.
+type wireBlock struct {
+	Type string `json:"type"`
+
+	// text
+	Text string `json:"text,omitempty"`
+
+	// image
+	Source *wireImageSource `json:"source,omitempty"`
+
+	// tool_use
+	ID    string          `json:"id,omitempty"`
+	Name  string          `json:"name,omitempty"`
+	Input json.RawMessage `json:"input,omitempty"`
+
+	// tool_result
+	ToolUseID string `json:"tool_use_id,omitempty"`
+	Content   string `json:"content,omitempty"`
+	IsError   bool   `json:"is_error,omitempty"`
+}
+
+type wireImageSource struct {
+	Type      string `json:"type"`
+	MediaType string `json:"media_type"`
+	Data      string `json:"data"`
+}
+
+type wireTool struct {
+	Name        string          `json:"name"`
+	Description string          `json:"description,omitempty"`
+	InputSchema json.RawMessage `json:"input_schema"`
+}
+
+type wireToolChoice struct {
+	Type string `json:"type"`
+	Name string `json:"name,omitempty"`
+}
+
+type wireOutputConfig struct {
+	Format *wireOutputFormat `json:"format,omitempty"`
+}
+
+type wireOutputFormat struct {
+	Type   string          `json:"type"`
+	Schema json.RawMessage `json:"schema"`
+}
+
+type wireResponse struct {
+	ID         string          `json:"id"`
+	Type       string          `json:"type"`
+	Role       string          `json:"role"`
+	Model      string          `json:"model"`
+	Content    []wireRespBlock `json:"content"`
+	StopReason string          `json:"stop_reason"`
+	Usage      wireUsage       `json:"usage"`
+}
+
+type wireRespBlock struct {
+	Type  string          `json:"type"`
+	Text  string          `json:"text"`
+	ID    string          `json:"id"`
+	Name  string          `json:"name"`
+	Input json.RawMessage `json:"input"`
+}
+
+type wireUsage struct {
+	InputTokens              int `json:"input_tokens"`
+	OutputTokens             int `json:"output_tokens"`
+	CacheCreationInputTokens int `json:"cache_creation_input_tokens"`
+	CacheReadInputTokens     int `json:"cache_read_input_tokens"`
+}
+
+// toUsage maps API token accounting onto the canonical Usage. Why the sum:
+// the API's input_tokens counts only tokens after the last cache breakpoint;
+// real total input is input + cache_creation + cache_read.
+func (u wireUsage) toUsage() llm.Usage {
+	return llm.Usage{
+		InputTokens:  u.InputTokens + u.CacheCreationInputTokens + u.CacheReadInputTokens,
+		OutputTokens: u.OutputTokens,
+	}
+}
+
+type wireErrorEnvelope struct {
+	Type  string `json:"type"`
+	Error struct {
+		Type    string `json:"type"`
+		Message string `json:"message"`
+	} `json:"error"`
+}
+
+// buildWireRequest translates the canonical request into the Messages API
+// shape.
+//
+// Request.ReasoningEffort is intentionally ignored: the current Messages API
+// has no low/medium/high reasoning knob — thinking is adaptive on current
+// models, and the legacy budget/disable parameters 400 on them. The llm
+// contract says providers ignore ReasoningEffort where no mapping exists.
+//
+// Request.SchemaName is likewise ignored: output_config.format takes a bare
+// schema with no name field.
+func buildWireRequest(modelID string, req llm.Request, defaultMax int, stream bool) wireRequest {
+	maxTokens := req.MaxTokens
+	if maxTokens == 0 {
+		// max_tokens is required by the API; 0 means "provider default".
+		maxTokens = defaultMax
+	}
+
+	wr := wireRequest{
+		Model:         modelID,
+		MaxTokens:     maxTokens,
+		System:        foldSystem(req),
+		Messages:      toWireMessages(req.Messages),
+		Stream:        stream,
+		Tools:         toWireTools(req.Tools),
+		ToolChoice:    toWireToolChoice(req.ToolChoice),
+		Temperature:   req.Temperature,
+		TopP:          req.TopP,
+		StopSequences: req.StopSequences,
+	}
+	if req.Schema != nil {
+		wr.OutputConfig = &wireOutputConfig{Format: &wireOutputFormat{
+			Type:   "json_schema",
+			Schema: req.Schema,
+		}}
+	}
+	return wr
+}
+
+// foldSystem joins Request.System with the text of every RoleSystem message
+// (System field first, original order, "\n\n" separators). Why: the API
+// takes the system prompt as a top-level field and rejects system roles
+// inside messages, so canonical RoleSystem messages must fold in here.
+func foldSystem(req llm.Request) string {
+	parts := make([]string, 0, 2)
+	if req.System != "" {
+		parts = append(parts, req.System)
+	}
+	for _, msg := range req.Messages {
+		if msg.Role != llm.RoleSystem {
+			continue
+		}
+		if text := msg.Text(); text != "" {
+			parts = append(parts, text)
+		}
+	}
+	return strings.Join(parts, "\n\n")
+}
+
+func toWireMessages(msgs []llm.Message) []wireMessage {
+	out := make([]wireMessage, 0, len(msgs))
+	for _, msg := range msgs {
+		switch msg.Role {
+		case llm.RoleSystem:
+			// Folded into the top-level system field by foldSystem.
+			continue
+
+		case llm.RoleTool:
+			// One user message carrying one tool_result block per result.
+			blocks := make([]wireBlock, 0, len(msg.ToolResults))
+			for _, res := range msg.ToolResults {
+				blocks = append(blocks, wireBlock{
+					Type:      "tool_result",
+					ToolUseID: res.ID,
+					Content:   res.Content,
+					IsError:   res.IsError,
+				})
+			}
+			out = append(out, wireMessage{Role: "user", Content: blocks})
+
+		case llm.RoleAssistant:
+			blocks := toWireBlocks(msg.Parts)
+			for _, call := range msg.ToolCalls {
+				args := call.Arguments
+				if len(args) == 0 {
+					// The API requires input to be a JSON object.
+					args = json.RawMessage("{}")
+				}
+				blocks = append(blocks, wireBlock{
+					Type:  "tool_use",
+					ID:    call.ID,
+					Name:  call.Name,
+					Input: args,
+				})
+			}
+			out = append(out, wireMessage{Role: "assistant", Content: blocks})
+
+		default: // llm.RoleUser and anything unrecognized
+			out = append(out, wireMessage{Role: "user", Content: toWireBlocks(msg.Parts)})
+		}
+	}
+	return out
+}
+
+func toWireBlocks(parts []llm.Part) []wireBlock {
+	blocks := make([]wireBlock, 0, len(parts))
+	for _, part := range parts {
+		switch p := part.(type) {
+		case llm.TextPart:
+			blocks = append(blocks, wireBlock{Type: "text", Text: p.Text})
+		case llm.ImagePart:
+			blocks = append(blocks, wireBlock{Type: "image", Source: &wireImageSource{
+				Type:      "base64",
+				MediaType: p.MIME,
+				Data:      base64.StdEncoding.EncodeToString(p.Data),
+			}})
+		}
+	}
+	return blocks
+}
+
+func toWireTools(tools []llm.Tool) []wireTool {
+	if len(tools) == 0 {
+		return nil
+	}
+	out := make([]wireTool, 0, len(tools))
+	for _, t := range tools {
+		schema := t.Parameters
+		if len(schema) == 0 {
+			// Why: input_schema is required by the API; a tool with no
+			// arguments still needs an (empty) object schema.
+			schema = json.RawMessage(`{"type":"object","properties":{}}`)
+		}
+		out = append(out, wireTool{
+			Name:        t.Name,
+			Description: t.Description,
+			InputSchema: schema,
+		})
+	}
+	return out
+}
+
+// toWireToolChoice maps the canonical tool-choice policy. "" omits the field
+// (API default is auto); any value other than the three keywords names the
+// one tool the model must call.
+func toWireToolChoice(choice string) *wireToolChoice {
+	switch choice {
+	case "":
+		return nil
+	case "auto":
+		return &wireToolChoice{Type: "auto"}
+	case "required":
+		return &wireToolChoice{Type: "any"}
+	case "none":
+		return &wireToolChoice{Type: "none"}
+	default:
+		return &wireToolChoice{Type: "tool", Name: choice}
+	}
+}
+
+// mapStopReason maps the API stop_reason onto the canonical FinishReason.
+func mapStopReason(stop string) llm.FinishReason {
+	switch stop {
+	case "end_turn", "stop_sequence":
+		return llm.FinishStop
+	case "max_tokens", "model_context_window_exceeded":
+		return llm.FinishLength
+	case "tool_use":
+		return llm.FinishToolCalls
+	case "refusal":
+		return llm.FinishContentFilter
+	default:
+		// pause_turn and any future provider-specific reasons.
+		return llm.FinishOther
+	}
+}
diff --git a/provider/ollama/ollama.go b/provider/ollama/ollama.go
new file mode 100644
index 0000000..c78e29a
--- /dev/null
+++ b/provider/ollama/ollama.go
@@ -0,0 +1,168 @@
+// Package ollama implements majordomo's provider contract over Ollama's
+// native chat API (POST {base}/api/chat), targeted at three backends that
+// share one wire protocol:
+//
+//   - a local Ollama instance (preset Local: OLLAMA_HOST or
+//     http://localhost:11434, no auth),
+//   - Ollama Cloud (preset Cloud: https://ollama.com, bearer key from
+//     OLLAMA_API_KEY), and
+//   - foreman, Steve's native-Ollama queue daemon (preset Foreman: explicit
+//     base URL + bearer token).
+//
+// Wire surface verified against docs.ollama.com and ollama/ollama
+// docs/api.md + api/types.go (June 2026): NDJSON streaming (stream defaults
+// true server-side — Generate always sends stream:false explicitly);
+// tool_calls carry arguments as a JSON OBJECT (not a string); tool results
+// return as {"role":"tool","content",...,"tool_name"}; structured output
+// via "format" (a full JSON-schema object); thinking via the bool-or-string
+// "think" field; errors as {"error":"message"} with a non-2xx status.
+//
+// foreman deviation (verified in its source): sync /api/chat does not
+// stream — a stream:true request yields ONE buffered application/json
+// object. The NDJSON reader here handles that transparently (a single JSON
+// line parses as the final chunk).
+package ollama
+
+import (
+	"fmt"
+	"net/http"
+	"os"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// DefaultLocalBaseURL is the default base URL for a locally-running Ollama.
+const DefaultLocalBaseURL = "http://localhost:11434"
+
+// DefaultCloudBaseURL is the base URL for Ollama Cloud.
+const DefaultCloudBaseURL = "https://ollama.com"
+
+// defaultCapabilities is the conservative provider-wide default; individual
+// models (e.g. high-resolution vision tags) override via llm.WithCapabilities.
+var defaultCapabilities = llm.Capabilities{
+	SupportsTools:      true,
+	SupportsStructured: true,
+	SupportsStreaming:  true,
+	MaxImagesPerReq:    8,
+	MaxImageBytes:      20 << 20,
+	MaxImageDimension:  2048,
+	AllowedImageMIME:   []string{"image/jpeg", "image/png"},
+}
+
+// Provider is a native-Ollama chat client bound to one base URL.
+type Provider struct {
+	name    string
+	baseURL string
+	token   string
+	client  *http.Client
+	caps    llm.Capabilities
+}
+
+// Option configures the provider.
+type Option func(*Provider)
+
+// WithName overrides the registry name (default "ollama").
+func WithName(name string) Option { return func(p *Provider) { p.name = name } }
+
+// WithBaseURL sets the backend base URL (scheme://host[:port][/path]).
+func WithBaseURL(u string) Option {
+	return func(p *Provider) { p.baseURL = strings.TrimRight(u, "/") }
+}
+
+// WithToken sets the bearer token (Ollama Cloud key / foreman token).
+// Empty means no Authorization header (local mode).
+func WithToken(token string) Option { return func(p *Provider) { p.token = token } }
+
+// WithHTTPClient overrides the HTTP client (proxies, test TLS, timeouts —
+// note foreman sync chat long-polls; prefer context deadlines over client
+// timeouts).
+func WithHTTPClient(c *http.Client) Option { return func(p *Provider) { p.client = c } }
+
+// WithDefaultCapabilities overrides the provider-wide default capabilities.
+func WithDefaultCapabilities(caps llm.Capabilities) Option {
+	return func(p *Provider) { p.caps = caps }
+}
+
+// New creates a generic native-Ollama provider. Most callers want one of
+// the presets (Local, Cloud, Foreman) or an LLM_* env DSN instead.
+// Construction never fails; a missing base URL surfaces at request time.
+func New(opts ...Option) *Provider {
+	p := &Provider{
+		name:   "ollama",
+		client: &http.Client{},
+		caps:   defaultCapabilities,
+	}
+	for _, opt := range opts {
+		opt(p)
+	}
+	return p
+}
+
+// Local returns the local-Ollama preset: name "ollama", base URL from
+// OLLAMA_HOST (normalized per Ollama conventions) or localhost:11434.
+func Local(opts ...Option) *Provider {
+	base := DefaultLocalBaseURL
+	if h := os.Getenv("OLLAMA_HOST"); h != "" {
+		base = NormalizeHost(h)
+	}
+	return New(append([]Option{WithBaseURL(base)}, opts...)...)
+}
+
+// Cloud returns the Ollama Cloud preset: name "ollama-cloud",
+// https://ollama.com, bearer key from OLLAMA_API_KEY.
+func Cloud(opts ...Option) *Provider {
+	return New(append([]Option{
+		WithName("ollama-cloud"),
+		WithBaseURL(DefaultCloudBaseURL),
+		WithToken(os.Getenv("OLLAMA_API_KEY")),
+	}, opts...)...)
+}
+
+// Foreman returns a foreman preset bound to the given daemon.
+func Foreman(baseURL, token string, opts ...Option) *Provider {
+	return New(append([]Option{
+		WithName("foreman"),
+		WithBaseURL(baseURL),
+		WithToken(token),
+	}, opts...)...)
+}
+
+// NormalizeHost turns an OLLAMA_HOST-style value into a base URL:
+// "host" → http://host:11434, "host:port" → http://host:port, full URLs
+// pass through (trailing slash trimmed).
+func NormalizeHost(h string) string {
+	h = strings.TrimRight(strings.TrimSpace(h), "/")
+	if strings.Contains(h, "://") {
+		return h
+	}
+	if !strings.Contains(h, ":") {
+		h += ":11434"
+	}
+	return "http://" + h
+}
+
+// Name implements llm.Provider.
+func (p *Provider) Name() string { return p.name }
+
+// BaseURL reports the configured backend base URL (diagnostics).
+func (p *Provider) BaseURL() string { return p.baseURL }
+
+// Model implements llm.Provider; the id passes through verbatim.
+func (p *Provider) Model(id string, opts ...llm.ModelOption) (llm.Model, error) {
+	cfg := llm.ApplyModelOptions(opts)
+	caps := p.caps
+	if cfg.Capabilities != nil {
+		caps = *cfg.Capabilities
+	}
+	return &model{provider: p, id: id, caps: caps}, nil
+}
+
+// checkReady reports a usable configuration (a base URL is the only hard
+// requirement; auth problems surface as 401s from the backend).
+func (p *Provider) checkReady() error {
+	if p.baseURL == "" {
+		return fmt.Errorf("ollama provider %q: no base URL configured (set one via the preset, WithBaseURL, or an LLM_* env DSN)", p.name)
+	}
+	return nil
+}
diff --git a/provider/ollama/ollama_test.go b/provider/ollama/ollama_test.go
new file mode 100644
index 0000000..197adeb
--- /dev/null
+++ b/provider/ollama/ollama_test.go
@@ -0,0 +1,492 @@
+package ollama
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// capture spins up an httptest server that records the request and replies
+// with the given handler.
+type captured struct {
+	auth        string
+	contentType string
+	path        string
+	body        map[string]any
+	raw         []byte
+}
+
+func serve(t *testing.T, status int, respond func(w http.ResponseWriter)) (*Provider, *captured) {
+	t.Helper()
+	cap := &captured{}
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		cap.auth = r.Header.Get("Authorization")
+		cap.contentType = r.Header.Get("Content-Type")
+		cap.path = r.URL.Path
+		cap.raw, _ = io.ReadAll(r.Body)
+		_ = json.Unmarshal(cap.raw, &cap.body)
+		w.WriteHeader(status)
+		respond(w)
+	}))
+	t.Cleanup(ts.Close)
+	return New(WithBaseURL(ts.URL), WithToken("test-token")), cap
+}
+
+func jsonReply(obj string) func(w http.ResponseWriter) {
+	return func(w http.ResponseWriter) { _, _ = io.WriteString(w, obj) }
+}
+
+func basicRequest() llm.Request {
+	return llm.Request{Messages: []llm.Message{llm.UserText("hi")}}
+}
+
+func TestGenerateRoundTrip(t *testing.T) {
+	p, cap := serve(t, 200, jsonReply(`{
+		"model":"qwen3:30b",
+		"message":{"role":"assistant","content":"hello there"},
+		"done":true,"done_reason":"stop",
+		"prompt_eval_count":12,"eval_count":7
+	}`))
+
+	m, _ := p.Model("qwen3:30b")
+	temp := 0.2
+	resp, err := m.Generate(context.Background(), llm.Request{
+		System:        "be terse",
+		Messages:      []llm.Message{llm.SystemText("extra sys"), llm.UserText("hi")},
+		Temperature:   &temp,
+		MaxTokens:     64,
+		StopSequences: []string{"END"},
+	})
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+
+	// Wire assertions.
+	if cap.path != "/api/chat" {
+		t.Errorf("path = %q", cap.path)
+	}
+	if cap.auth != "Bearer test-token" {
+		t.Errorf("auth = %q", cap.auth)
+	}
+	if cap.body["model"] != "qwen3:30b" {
+		t.Errorf("model = %v", cap.body["model"])
+	}
+	if stream, ok := cap.body["stream"].(bool); !ok || stream {
+		t.Errorf("stream must be explicit false, got %v", cap.body["stream"])
+	}
+	msgs := cap.body["messages"].([]any)
+	first := msgs[0].(map[string]any)
+	if first["role"] != "system" || first["content"] != "be terse\n\nextra sys" {
+		t.Errorf("system fold = %v", first)
+	}
+	second := msgs[1].(map[string]any)
+	if second["role"] != "user" || second["content"] != "hi" {
+		t.Errorf("user msg = %v", second)
+	}
+	opts := cap.body["options"].(map[string]any)
+	if opts["temperature"] != 0.2 || opts["num_predict"] != float64(64) {
+		t.Errorf("options = %v", opts)
+	}
+
+	// Response assertions.
+	if resp.Text() != "hello there" {
+		t.Errorf("text = %q", resp.Text())
+	}
+	if resp.FinishReason != llm.FinishStop {
+		t.Errorf("finish = %v", resp.FinishReason)
+	}
+	if resp.Usage.InputTokens != 12 || resp.Usage.OutputTokens != 7 {
+		t.Errorf("usage = %+v", resp.Usage)
+	}
+	if resp.Model != "ollama/qwen3:30b" {
+		t.Errorf("resp.Model = %q", resp.Model)
+	}
+}
+
+func TestImagesEncodeAsBase64(t *testing.T) {
+	p, cap := serve(t, 200, jsonReply(`{"message":{"role":"assistant","content":"a cat"},"done":true,"done_reason":"stop"}`))
+	imgBytes := []byte{0xFF, 0xD8, 0xFF, 0xE0, 1, 2, 3}
+
+	m, _ := p.Model("llava")
+	_, err := m.Generate(context.Background(), llm.Request{
+		Messages: []llm.Message{llm.UserParts(llm.Text("describe"), llm.Image("image/jpeg", imgBytes))},
+	})
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	msgs := cap.body["messages"].([]any)
+	user := msgs[0].(map[string]any)
+	images := user["images"].([]any)
+	if len(images) != 1 || images[0] != base64.StdEncoding.EncodeToString(imgBytes) {
+		t.Errorf("images = %v", images)
+	}
+	if strings.Contains(images[0].(string), "data:") {
+		t.Error("images must be raw base64 without data: prefix")
+	}
+}
+
+func TestToolsAndToolCallRoundTrip(t *testing.T) {
+	p, cap := serve(t, 200, jsonReply(`{
+		"message":{"role":"assistant","content":"","tool_calls":[
+			{"function":{"index":0,"name":"get_weather","arguments":{"city":"Tokyo"}}}
+		]},
+		"done":true,"done_reason":"stop"
+	}`))
+
+	tool := llm.Tool{
+		Name: "get_weather", Description: "weather",
+		Parameters: json.RawMessage(`{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}`),
+	}
+	m, _ := p.Model("qwen3")
+	resp, err := m.Generate(context.Background(), basicRequest(), llm.WithTools(tool))
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+
+	// Tools serialize with parameters as an object.
+	tools := cap.body["tools"].([]any)
+	fn := tools[0].(map[string]any)["function"].(map[string]any)
+	if fn["name"] != "get_weather" {
+		t.Errorf("tool fn = %v", fn)
+	}
+	if _, ok := fn["parameters"].(map[string]any); !ok {
+		t.Errorf("parameters must be an object, got %T", fn["parameters"])
+	}
+
+	// Tool call comes back with arguments as a JSON object → RawMessage.
+	if len(resp.ToolCalls) != 1 {
+		t.Fatalf("tool calls = %v", resp.ToolCalls)
+	}
+	tc := resp.ToolCalls[0]
+	if tc.Name != "get_weather" || tc.ID == "" {
+		t.Errorf("call = %+v (id must be synthesized)", tc)
+	}
+	var args struct {
+		City string `json:"city"`
+	}
+	if err := json.Unmarshal(tc.Arguments, &args); err != nil || args.City != "Tokyo" {
+		t.Errorf("arguments = %s (%v)", tc.Arguments, err)
+	}
+	if resp.FinishReason != llm.FinishToolCalls {
+		t.Errorf("finish = %v, want tool_calls", resp.FinishReason)
+	}
+}
+
+func TestToolResultsAndHistoryToolCalls(t *testing.T) {
+	p, cap := serve(t, 200, jsonReply(`{"message":{"role":"assistant","content":"21C"},"done":true,"done_reason":"stop"}`))
+
+	m, _ := p.Model("qwen3")
+	_, err := m.Generate(context.Background(), llm.Request{
+		Messages: []llm.Message{
+			llm.UserText("weather?"),
+			{Role: llm.RoleAssistant, ToolCalls: []llm.ToolCall{
+				{ID: "call_0", Name: "get_weather", Arguments: json.RawMessage(`{"city":"Tokyo"}`)},
+			}},
+			llm.ToolResultsMessage(
+				llm.ToolResult{ID: "call_0", Name: "get_weather", Content: `{"temp":21}`},
+				llm.ToolResult{ID: "call_1", Name: "broken_tool", Content: "boom", IsError: true},
+			),
+		},
+	})
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	msgs := cap.body["messages"].([]any)
+	if len(msgs) != 4 {
+		t.Fatalf("messages = %d, want 4 (user, assistant, 2 tool results)", len(msgs))
+	}
+	asst := msgs[1].(map[string]any)
+	calls := asst["tool_calls"].([]any)
+	args := calls[0].(map[string]any)["function"].(map[string]any)["arguments"]
+	if _, ok := args.(map[string]any); !ok {
+		t.Errorf("history tool-call arguments must be an object, got %T", args)
+	}
+	tr1 := msgs[2].(map[string]any)
+	if tr1["role"] != "tool" || tr1["tool_name"] != "get_weather" || tr1["content"] != `{"temp":21}` {
+		t.Errorf("tool result 1 = %v", tr1)
+	}
+	tr2 := msgs[3].(map[string]any)
+	if tr2["content"] != "ERROR: boom" {
+		t.Errorf("error result content = %v", tr2["content"])
+	}
+}
+
+func TestStructuredOutputFormat(t *testing.T) {
+	p, cap := serve(t, 200, jsonReply(`{"message":{"role":"assistant","content":"{\"name\":\"Ada\"}"},"done":true,"done_reason":"stop"}`))
+	schema := json.RawMessage(`{"type":"object","properties":{"name":{"type":"string"}},"required":["name"]}`)
+
+	m, _ := p.Model("qwen3")
+	resp, err := m.Generate(context.Background(), basicRequest(), llm.WithSchema(schema, "person"))
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	format, ok := cap.body["format"].(map[string]any)
+	if !ok || format["type"] != "object" {
+		t.Errorf("format = %v, want the schema object", cap.body["format"])
+	}
+	if resp.Text() != `{"name":"Ada"}` {
+		t.Errorf("text = %q", resp.Text())
+	}
+}
+
+func TestThinkMapping(t *testing.T) {
+	p, cap := serve(t, 200, jsonReply(`{"message":{"role":"assistant","content":"ok"},"done":true,"done_reason":"stop"}`))
+	m, _ := p.Model("gpt-oss:120b")
+	_, err := m.Generate(context.Background(), basicRequest(), llm.WithReasoningEffort("high"))
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if cap.body["think"] != "high" {
+		t.Errorf("think = %v", cap.body["think"])
+	}
+
+	if _, err := m.Generate(context.Background(), basicRequest(), llm.WithReasoningEffort("max")); err == nil {
+		t.Error("invalid reasoning effort should error")
+	}
+}
+
+func TestToolChoiceNoneDropsTools(t *testing.T) {
+	p, cap := serve(t, 200, jsonReply(`{"message":{"role":"assistant","content":"ok"},"done":true,"done_reason":"stop"}`))
+	m, _ := p.Model("qwen3")
+	_, err := m.Generate(context.Background(), basicRequest(),
+		llm.WithTools(llm.Tool{Name: "t"}), llm.WithToolChoice("none"))
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if _, present := cap.body["tools"]; present {
+		t.Error("tool_choice none must omit tools")
+	}
+}
+
+func TestStreamingNDJSON(t *testing.T) {
+	p, _ := serve(t, 200, func(w http.ResponseWriter) {
+		w.Header().Set("Content-Type", "application/x-ndjson")
+		_, _ = io.WriteString(w, `{"message":{"role":"assistant","content":"Hel"},"done":false}
+{"message":{"role":"assistant","content":"lo"},"done":false}
+{"message":{"role":"assistant","content":"","tool_calls":[{"function":{"name":"ping","arguments":{}}}]},"done":false}
+{"message":{"role":"assistant","content":""},"done":true,"done_reason":"stop","prompt_eval_count":5,"eval_count":9}
+`)
+	})
+
+	m, _ := p.Model("qwen3")
+	s, err := m.Stream(context.Background(), basicRequest())
+	if err != nil {
+		t.Fatalf("Stream: %v", err)
+	}
+	defer s.Close()
+
+	var text strings.Builder
+	var toolCalls []llm.ToolCall
+	var final *llm.Response
+	for {
+		ev, err := s.Next()
+		if errors.Is(err, io.EOF) {
+			break
+		}
+		if err != nil {
+			t.Fatalf("Next: %v", err)
+		}
+		text.WriteString(ev.TextDelta)
+		if ev.ToolCall != nil {
+			toolCalls = append(toolCalls, *ev.ToolCall)
+		}
+		if ev.Response != nil {
+			final = ev.Response
+		}
+	}
+	if text.String() != "Hello" {
+		t.Errorf("text = %q", text.String())
+	}
+	if len(toolCalls) != 1 || toolCalls[0].Name != "ping" {
+		t.Errorf("tool calls = %+v", toolCalls)
+	}
+	if final == nil {
+		t.Fatal("no final response event")
+	}
+	if final.Usage.InputTokens != 5 || final.Usage.OutputTokens != 9 {
+		t.Errorf("final usage = %+v", final.Usage)
+	}
+	if final.FinishReason != llm.FinishToolCalls {
+		t.Errorf("final finish = %v", final.FinishReason)
+	}
+	if final.Text() != "Hello" {
+		t.Errorf("final text = %q", final.Text())
+	}
+}
+
+// TestStreamingForemanSingleObject: foreman returns one buffered JSON
+// object to a stream:true request; the stream must still deliver the text
+// and a final response.
+func TestStreamingForemanSingleObject(t *testing.T) {
+	p, cap := serve(t, 200, func(w http.ResponseWriter) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = io.WriteString(w, `{"message":{"role":"assistant","content":"queued answer"},"done":true,"done_reason":"stop","prompt_eval_count":3,"eval_count":4}`)
+	})
+
+	m, _ := p.Model("qwen3:30b")
+	s, err := m.Stream(context.Background(), basicRequest())
+	if err != nil {
+		t.Fatalf("Stream: %v", err)
+	}
+	defer s.Close()
+
+	if stream, ok := cap.body["stream"].(bool); !ok || !stream {
+		t.Errorf("stream flag = %v, want true", cap.body["stream"])
+	}
+
+	var text strings.Builder
+	var final *llm.Response
+	for {
+		ev, err := s.Next()
+		if errors.Is(err, io.EOF) {
+			break
+		}
+		if err != nil {
+			t.Fatalf("Next: %v", err)
+		}
+		text.WriteString(ev.TextDelta)
+		if ev.Response != nil {
+			final = ev.Response
+		}
+	}
+	if text.String() != "queued answer" || final == nil || final.Usage.OutputTokens != 4 {
+		t.Errorf("text=%q final=%+v", text.String(), final)
+	}
+}
+
+func TestErrorMapping(t *testing.T) {
+	t.Run("404 is model-not-found", func(t *testing.T) {
+		p, _ := serve(t, 404, jsonReply(`{"error":"model not found"}`))
+		m, _ := p.Model("nope")
+		_, err := m.Generate(context.Background(), basicRequest())
+		if !errors.Is(err, llm.ErrModelNotFound) {
+			t.Errorf("error = %v, want ErrModelNotFound", err)
+		}
+		if llm.Classify(err) != llm.ClassPermanent {
+			t.Error("404 must classify permanent")
+		}
+	})
+
+	t.Run("503 transient with message", func(t *testing.T) {
+		p, _ := serve(t, 503, jsonReply(`{"error":"request cancelled while waiting"}`))
+		m, _ := p.Model("qwen3")
+		_, err := m.Generate(context.Background(), basicRequest())
+		var apiErr *llm.APIError
+		if !errors.As(err, &apiErr) || apiErr.Status != 503 || !strings.Contains(apiErr.Message, "cancelled") {
+			t.Errorf("error = %v", err)
+		}
+		if llm.Classify(err) != llm.ClassTransient {
+			t.Error("503 must classify transient")
+		}
+	})
+
+	t.Run("non-JSON error body", func(t *testing.T) {
+		p, _ := serve(t, 500, jsonReply(`upstream exploded`))
+		m, _ := p.Model("qwen3")
+		_, err := m.Generate(context.Background(), basicRequest())
+		var apiErr *llm.APIError
+		if !errors.As(err, &apiErr) || !strings.Contains(apiErr.Message, "upstream exploded") {
+			t.Errorf("error = %v", err)
+		}
+	})
+}
+
+func TestCapabilityEnforcement(t *testing.T) {
+	p, _ := serve(t, 200, jsonReply(`{"message":{"content":"x"},"done":true}`))
+
+	t.Run("too many images", func(t *testing.T) {
+		m, _ := p.Model("llava", llm.WithCapabilities(llm.Capabilities{MaxImagesPerReq: 1, AllowedImageMIME: []string{"image/png"}}))
+		_, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{
+			llm.UserParts(llm.Image("image/png", []byte{1}), llm.Image("image/png", []byte{2})),
+		}})
+		if !errors.Is(err, llm.ErrUnsupported) {
+			t.Errorf("error = %v, want ErrUnsupported", err)
+		}
+	})
+
+	t.Run("images on text-only model", func(t *testing.T) {
+		m, _ := p.Model("qwen3", llm.WithCapabilities(llm.Capabilities{}))
+		_, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{
+			llm.UserParts(llm.Image("image/png", []byte{1})),
+		}})
+		if !errors.Is(err, llm.ErrUnsupported) {
+			t.Errorf("error = %v, want ErrUnsupported", err)
+		}
+	})
+
+	t.Run("disallowed mime", func(t *testing.T) {
+		m, _ := p.Model("llava") // default caps: jpeg/png only
+		_, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{
+			llm.UserParts(llm.Image("image/tiff", []byte{1})),
+		}})
+		if !errors.Is(err, llm.ErrUnsupported) {
+			t.Errorf("error = %v, want ErrUnsupported", err)
+		}
+	})
+}
+
+func TestNoBaseURL(t *testing.T) {
+	p := New(WithBaseURL(""))
+	m, _ := p.Model("x")
+	if _, err := m.Generate(context.Background(), basicRequest()); err == nil ||
+		!strings.Contains(err.Error(), "no base URL") {
+		t.Errorf("error = %v, want a clear no-base-URL message", err)
+	}
+}
+
+func TestNormalizeHost(t *testing.T) {
+	for in, want := range map[string]string{
+		"myhost":              "http://myhost:11434",
+		"myhost:8080":         "http://myhost:8080",
+		"http://myhost:8080/": "http://myhost:8080",
+		"https://ollama.com":  "https://ollama.com",
+		" 127.0.0.1:11434 ":   "http://127.0.0.1:11434",
+	} {
+		if got := NormalizeHost(in); got != want {
+			t.Errorf("NormalizeHost(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+func TestPresets(t *testing.T) {
+	t.Run("cloud", func(t *testing.T) {
+		t.Setenv("OLLAMA_API_KEY", "cloud-key")
+		p := Cloud()
+		if p.Name() != "ollama-cloud" || p.baseURL != DefaultCloudBaseURL || p.token != "cloud-key" {
+			t.Errorf("cloud preset = %+v", p)
+		}
+	})
+	t.Run("local respects OLLAMA_HOST", func(t *testing.T) {
+		t.Setenv("OLLAMA_HOST", "box.lan:9999")
+		p := Local()
+		if p.Name() != "ollama" || p.baseURL != "http://box.lan:9999" || p.token != "" {
+			t.Errorf("local preset = %+v", p)
+		}
+	})
+	t.Run("foreman", func(t *testing.T) {
+		p := Foreman("http://foreman-m1:8080", "tok")
+		if p.Name() != "foreman" || p.baseURL != "http://foreman-m1:8080" || p.token != "tok" {
+			t.Errorf("foreman preset = %+v", p)
+		}
+	})
+}
+
+func TestLocalNoAuthHeader(t *testing.T) {
+	p, cap := serve(t, 200, jsonReply(`{"message":{"content":"x"},"done":true}`))
+	p.token = "" // simulate local mode on the test server
+	m, _ := p.Model("llama3")
+	if _, err := m.Generate(context.Background(), basicRequest()); err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if cap.auth != "" {
+		t.Errorf("auth header = %q, want none in local mode", cap.auth)
+	}
+}
diff --git a/provider/ollama/stream.go b/provider/ollama/stream.go
new file mode 100644
index 0000000..944c8db
--- /dev/null
+++ b/provider/ollama/stream.go
@@ -0,0 +1,140 @@
+package ollama
+
+import (
+	"bufio"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"strconv"
+	"sync"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// Stream implements llm.Model over Ollama's NDJSON streaming. It also
+// transparently handles foreman's non-streaming degradation (a single
+// buffered JSON object): one JSON line parses as the final chunk.
+func (m *model) Stream(ctx context.Context, req llm.Request, opts ...llm.Option) (llm.Stream, error) {
+	req = req.Apply(opts...)
+	if err := m.enforceCapabilities(req); err != nil {
+		return nil, err
+	}
+	wireReq, err := m.buildRequest(req, true)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := m.do(ctx, wireReq)
+	if err != nil {
+		return nil, err
+	}
+
+	sc := bufio.NewScanner(resp.Body)
+	// Single NDJSON lines can far exceed the 64KB default (thinking dumps,
+	// tool payloads, foreman's whole-response-as-one-line degradation).
+	sc.Buffer(make([]byte, 64<<10), 16<<20)
+
+	return &stream{model: m, body: resp.Body, scanner: sc}, nil
+}
+
+type stream struct {
+	model   *model
+	body    io.Closer
+	scanner *bufio.Scanner
+
+	mu         sync.Mutex
+	closed     bool
+	finished   bool
+	toolCalls  []llm.ToolCall
+	text       []byte
+	pending    []llm.StreamEvent
+	usage      llm.Usage
+	doneReason string
+}
+
+func (s *stream) Next() (llm.StreamEvent, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	for {
+		if len(s.pending) > 0 {
+			ev := s.pending[0]
+			s.pending = s.pending[1:]
+			return ev, nil
+		}
+		if s.finished {
+			return llm.StreamEvent{}, io.EOF
+		}
+		if !s.scanner.Scan() {
+			if err := s.scanner.Err(); err != nil {
+				return llm.StreamEvent{}, fmt.Errorf("ollama %s: read stream: %w", s.model.qualified(), err)
+			}
+			// EOF without a done chunk: synthesize the final response from
+			// what we accumulated rather than losing it.
+			s.queueFinal()
+			continue
+		}
+		line := s.scanner.Bytes()
+		if len(line) == 0 {
+			continue
+		}
+
+		var chunk chatResponse
+		if err := json.Unmarshal(line, &chunk); err != nil {
+			return llm.StreamEvent{}, fmt.Errorf("ollama %s: decode stream chunk: %w", s.model.qualified(), err)
+		}
+
+		if chunk.Message.Content != "" {
+			s.text = append(s.text, chunk.Message.Content...)
+			s.pending = append(s.pending, llm.StreamEvent{TextDelta: chunk.Message.Content})
+		}
+		// Tool calls arrive complete per chunk (no partial-argument deltas
+		// in the native protocol).
+		base := len(s.toolCalls)
+		for i, tc := range chunk.Message.ToolCalls {
+			id := tc.ID
+			if id == "" {
+				id = "call_" + strconv.Itoa(base+i)
+			}
+			args := tc.Function.Arguments
+			if len(args) == 0 {
+				args = json.RawMessage("{}")
+			}
+			call := llm.ToolCall{ID: id, Name: tc.Function.Name, Arguments: args}
+			s.toolCalls = append(s.toolCalls, call)
+			s.pending = append(s.pending, llm.StreamEvent{ToolCall: &s.toolCalls[len(s.toolCalls)-1]})
+		}
+		if chunk.Done {
+			s.usage = llm.Usage{InputTokens: chunk.PromptEvalCount, OutputTokens: chunk.EvalCount}
+			s.doneReason = chunk.DoneReason
+			s.queueFinal()
+		}
+	}
+}
+
+// queueFinal appends the final Response event and marks the stream done.
+func (s *stream) queueFinal() {
+	resp := &llm.Response{
+		Model:        s.model.qualified(),
+		Usage:        s.usage,
+		FinishReason: finishReason(s.doneReason, len(s.toolCalls) > 0),
+	}
+	if len(s.text) > 0 {
+		resp.Parts = append(resp.Parts, llm.Text(string(s.text)))
+	}
+	if len(s.toolCalls) > 0 {
+		resp.ToolCalls = append([]llm.ToolCall(nil), s.toolCalls...)
+	}
+	s.pending = append(s.pending, llm.StreamEvent{Response: resp})
+	s.finished = true
+}
+
+func (s *stream) Close() error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.closed {
+		return nil
+	}
+	s.closed = true
+	return s.body.Close()
+}
diff --git a/provider/ollama/wire.go b/provider/ollama/wire.go
new file mode 100644
index 0000000..72ad8f0
--- /dev/null
+++ b/provider/ollama/wire.go
@@ -0,0 +1,343 @@
+package ollama
+
+import (
+	"bytes"
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strconv"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// ---- wire types (field names per ollama api/types.go) ----
+
+type chatRequest struct {
+	Model    string          `json:"model"`
+	Messages []chatMessage   `json:"messages"`
+	Tools    []toolDef       `json:"tools,omitempty"`
+	Format   json.RawMessage `json:"format,omitempty"`
+	Options  map[string]any  `json:"options,omitempty"`
+	// Stream has no omitempty on purpose: the server default is true, so
+	// Generate must send an explicit false.
+	Stream bool `json:"stream"`
+	// Think is bool-or-string on the wire ("low"/"medium"/"high" or a bool).
+	Think json.RawMessage `json:"think,omitempty"`
+}
+
+type chatMessage struct {
+	Role      string     `json:"role"`
+	Content   string     `json:"content"`
+	Images    []string   `json:"images,omitempty"` // raw base64, no data: prefix
+	ToolCalls []toolCall `json:"tool_calls,omitempty"`
+	ToolName  string     `json:"tool_name,omitempty"` // on role:"tool" results
+}
+
+type toolDef struct {
+	Type     string      `json:"type"`
+	Function toolDefFunc `json:"function"`
+}
+
+type toolDefFunc struct {
+	Name        string          `json:"name"`
+	Description string          `json:"description,omitempty"`
+	Parameters  json.RawMessage `json:"parameters,omitempty"`
+}
+
+type toolCall struct {
+	ID       string       `json:"id,omitempty"`
+	Function toolCallFunc `json:"function"`
+}
+
+type toolCallFunc struct {
+	Index int    `json:"index,omitempty"`
+	Name  string `json:"name"`
+	// Arguments is a JSON OBJECT on the wire (unlike OpenAI's string).
+	Arguments json.RawMessage `json:"arguments"`
+}
+
+type chatResponse struct {
+	Model           string      `json:"model"`
+	Message         respMessage `json:"message"`
+	Done            bool        `json:"done"`
+	DoneReason      string      `json:"done_reason"`
+	PromptEvalCount int         `json:"prompt_eval_count"`
+	EvalCount       int         `json:"eval_count"`
+}
+
+type respMessage struct {
+	Role      string     `json:"role"`
+	Content   string     `json:"content"`
+	Thinking  string     `json:"thinking"`
+	ToolCalls []toolCall `json:"tool_calls"`
+}
+
+type errorBody struct {
+	Error string `json:"error"`
+}
+
+// ---- model ----
+
+type model struct {
+	provider *Provider
+	id       string
+	caps     llm.Capabilities
+}
+
+func (m *model) Capabilities() llm.Capabilities { return m.caps }
+
+func (m *model) qualified() string { return m.provider.name + "/" + m.id }
+
+// enforceCapabilities is the backstop check (the media layer normalizes
+// before requests get here; see ADR-0009).
+func (m *model) enforceCapabilities(req llm.Request) error {
+	count := 0
+	for _, msg := range req.Messages {
+		for _, part := range msg.Parts {
+			img, ok := part.(llm.ImagePart)
+			if !ok {
+				continue
+			}
+			count++
+			if !m.caps.SupportsImages() {
+				return fmt.Errorf("%w: %s does not accept image input", llm.ErrUnsupported, m.qualified())
+			}
+			if !m.caps.MIMEAllowed(img.MIME) {
+				return fmt.Errorf("%w: %s does not accept %s images", llm.ErrUnsupported, m.qualified(), img.MIME)
+			}
+			if m.caps.MaxImageBytes > 0 && len(img.Data) > m.caps.MaxImageBytes {
+				return fmt.Errorf("%w: image of %d bytes exceeds %s limit of %d",
+					llm.ErrUnsupported, len(img.Data), m.qualified(), m.caps.MaxImageBytes)
+			}
+		}
+	}
+	if count > 0 && m.caps.MaxImagesPerReq > 0 && count > m.caps.MaxImagesPerReq {
+		return fmt.Errorf("%w: %d images exceed %s limit of %d",
+			llm.ErrUnsupported, count, m.qualified(), m.caps.MaxImagesPerReq)
+	}
+	return nil
+}
+
+// buildRequest maps the canonical request onto the wire shape.
+func (m *model) buildRequest(req llm.Request, stream bool) (*chatRequest, error) {
+	out := &chatRequest{Model: m.id, Stream: stream}
+
+	// System prompt: dedicated field first, then folded RoleSystem messages.
+	var sys []string
+	if req.System != "" {
+		sys = append(sys, req.System)
+	}
+	for _, msg := range req.Messages {
+		if msg.Role == llm.RoleSystem {
+			if t := msg.Text(); t != "" {
+				sys = append(sys, t)
+			}
+		}
+	}
+	if len(sys) > 0 {
+		out.Messages = append(out.Messages, chatMessage{
+			Role: "system", Content: strings.Join(sys, "\n\n"),
+		})
+	}
+
+	for _, msg := range req.Messages {
+		switch msg.Role {
+		case llm.RoleSystem:
+			// Already folded above.
+		case llm.RoleTool:
+			for _, res := range msg.ToolResults {
+				content := res.Content
+				if res.IsError {
+					content = "ERROR: " + content
+				}
+				out.Messages = append(out.Messages, chatMessage{
+					Role: "tool", Content: content, ToolName: res.Name,
+				})
+			}
+		default:
+			cm := chatMessage{Role: string(msg.Role), Content: msg.Text()}
+			for _, part := range msg.Parts {
+				if img, ok := part.(llm.ImagePart); ok {
+					cm.Images = append(cm.Images, base64.StdEncoding.EncodeToString(img.Data))
+				}
+			}
+			for _, tc := range msg.ToolCalls {
+				args := tc.Arguments
+				if len(args) == 0 {
+					args = json.RawMessage("{}")
+				}
+				cm.ToolCalls = append(cm.ToolCalls, toolCall{
+					ID:       tc.ID,
+					Function: toolCallFunc{Name: tc.Name, Arguments: args},
+				})
+			}
+			out.Messages = append(out.Messages, cm)
+		}
+	}
+
+	// Tools. Ollama has no tool_choice: "none" maps to omitting the tools;
+	// "required"/named choices have no wire equivalent and are best-effort
+	// ignored (documented in the README support matrix).
+	if req.ToolChoice != "none" {
+		for _, t := range req.Tools {
+			params := t.Parameters
+			if len(params) == 0 {
+				params = json.RawMessage(`{"type":"object","properties":{}}`)
+			}
+			out.Tools = append(out.Tools, toolDef{
+				Type:     "function",
+				Function: toolDefFunc{Name: t.Name, Description: t.Description, Parameters: params},
+			})
+		}
+	}
+
+	if len(req.Schema) > 0 {
+		out.Format = req.Schema
+	}
+
+	opts := make(map[string]any)
+	if req.Temperature != nil {
+		opts["temperature"] = *req.Temperature
+	}
+	if req.TopP != nil {
+		opts["top_p"] = *req.TopP
+	}
+	if req.MaxTokens > 0 {
+		opts["num_predict"] = req.MaxTokens
+	}
+	if len(req.StopSequences) > 0 {
+		opts["stop"] = req.StopSequences
+	}
+	if len(opts) > 0 {
+		out.Options = opts
+	}
+
+	switch req.ReasoningEffort {
+	case "":
+	case "low", "medium", "high":
+		out.Think = json.RawMessage(strconv.Quote(req.ReasoningEffort))
+	default:
+		return nil, fmt.Errorf("ollama: invalid reasoning effort %q (want low/medium/high)", req.ReasoningEffort)
+	}
+
+	return out, nil
+}
+
+// do POSTs /api/chat and returns the response body on 2xx, or a classified
+// error.
+func (m *model) do(ctx context.Context, wireReq *chatRequest) (*http.Response, error) {
+	p := m.provider
+	if err := p.checkReady(); err != nil {
+		return nil, err
+	}
+	body, err := json.Marshal(wireReq)
+	if err != nil {
+		return nil, fmt.Errorf("ollama: encode request: %w", err)
+	}
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, p.baseURL+"/api/chat", bytes.NewReader(body))
+	if err != nil {
+		return nil, fmt.Errorf("ollama: build request: %w", err)
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+	if p.token != "" {
+		httpReq.Header.Set("Authorization", "Bearer "+p.token)
+	}
+
+	resp, err := p.client.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("ollama %s: do request: %w", m.qualified(), err)
+	}
+	if resp.StatusCode/100 != 2 {
+		defer resp.Body.Close()
+		raw, _ := io.ReadAll(io.LimitReader(resp.Body, 8<<10))
+		var eb errorBody
+		_ = json.Unmarshal(raw, &eb)
+		msg := eb.Error
+		if msg == "" {
+			msg = strings.TrimSpace(string(raw))
+		}
+		return nil, &llm.APIError{
+			Provider: p.name, Model: m.id,
+			Status: resp.StatusCode, Message: msg,
+		}
+	}
+	return resp, nil
+}
+
+// Generate implements llm.Model.
+func (m *model) Generate(ctx context.Context, req llm.Request, opts ...llm.Option) (*llm.Response, error) {
+	req = req.Apply(opts...)
+	if err := m.enforceCapabilities(req); err != nil {
+		return nil, err
+	}
+	wireReq, err := m.buildRequest(req, false)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := m.do(ctx, wireReq)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+
+	var cr chatResponse
+	if err := json.NewDecoder(resp.Body).Decode(&cr); err != nil {
+		return nil, fmt.Errorf("ollama %s: decode response: %w", m.qualified(), err)
+	}
+	return m.toResponse(&cr), nil
+}
+
+// toResponse converts a final wire chunk into the canonical response.
+func (m *model) toResponse(cr *chatResponse) *llm.Response {
+	out := &llm.Response{
+		Model: m.qualified(),
+		Usage: llm.Usage{InputTokens: cr.PromptEvalCount, OutputTokens: cr.EvalCount},
+		Raw:   cr,
+	}
+	if cr.Message.Content != "" {
+		out.Parts = append(out.Parts, llm.Text(cr.Message.Content))
+	}
+	out.ToolCalls = convertToolCalls(cr.Message.ToolCalls)
+	out.FinishReason = finishReason(cr.DoneReason, len(out.ToolCalls) > 0)
+	return out
+}
+
+// convertToolCalls maps wire tool calls, synthesizing ids where the model
+// omitted them (ids are optional in Ollama's shape but required by our
+// agent loop to match results to calls).
+func convertToolCalls(calls []toolCall) []llm.ToolCall {
+	out := make([]llm.ToolCall, 0, len(calls))
+	for i, tc := range calls {
+		id := tc.ID
+		if id == "" {
+			id = "call_" + strconv.Itoa(i)
+		}
+		args := tc.Function.Arguments
+		if len(args) == 0 {
+			args = json.RawMessage("{}")
+		}
+		out = append(out, llm.ToolCall{ID: id, Name: tc.Function.Name, Arguments: args})
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}
+
+func finishReason(doneReason string, hasToolCalls bool) llm.FinishReason {
+	if hasToolCalls {
+		return llm.FinishToolCalls
+	}
+	switch doneReason {
+	case "stop", "":
+		return llm.FinishStop
+	case "length":
+		return llm.FinishLength
+	default:
+		return llm.FinishOther
+	}
+}
diff --git a/provider/openai/model.go b/provider/openai/model.go
new file mode 100644
index 0000000..2744a25
--- /dev/null
+++ b/provider/openai/model.go
@@ -0,0 +1,222 @@
+package openai
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// model is one provider-bound target.
+type model struct {
+	p    *Provider
+	id   string
+	caps llm.Capabilities
+}
+
+// Capabilities implements llm.Model.
+func (m *model) Capabilities() llm.Capabilities { return m.caps }
+
+// Generate implements llm.Model.
+func (m *model) Generate(ctx context.Context, req llm.Request, opts ...llm.Option) (*llm.Response, error) {
+	req = req.Apply(opts...)
+	if err := checkRequest(m.caps, req); err != nil {
+		return nil, err
+	}
+	httpResp, err := m.do(ctx, req, false)
+	if err != nil {
+		return nil, err
+	}
+	defer httpResp.Body.Close()
+	if httpResp.StatusCode/100 != 2 {
+		return nil, m.apiError(httpResp)
+	}
+	var wire chatResponse
+	if err := json.NewDecoder(httpResp.Body).Decode(&wire); err != nil {
+		return nil, fmt.Errorf("openai: decode response: %w", err)
+	}
+	return m.toResponse(&wire), nil
+}
+
+// Stream implements llm.Model.
+func (m *model) Stream(ctx context.Context, req llm.Request, opts ...llm.Option) (llm.Stream, error) {
+	req = req.Apply(opts...)
+	if !m.caps.SupportsStreaming {
+		return nil, fmt.Errorf("%w: streaming not supported by %s/%s", llm.ErrUnsupported, m.p.name, m.id)
+	}
+	if err := checkRequest(m.caps, req); err != nil {
+		return nil, err
+	}
+	httpResp, err := m.do(ctx, req, true)
+	if err != nil {
+		return nil, err
+	}
+	if httpResp.StatusCode/100 != 2 {
+		defer httpResp.Body.Close()
+		return nil, m.apiError(httpResp)
+	}
+	sc := bufio.NewScanner(httpResp.Body)
+	// Why: a single SSE data line carries a whole JSON chunk; tool-call
+	// argument fragments can make lines far larger than Scanner's 64 KiB
+	// default cap.
+	sc.Buffer(make([]byte, 0, 64*1024), 16<<20)
+	return &stream{m: m, body: httpResp.Body, sc: sc}, nil
+}
+
+// do builds and performs the HTTP request. Transport failures are wrapped
+// raw (never as *llm.APIError) so llm.Classify still sees net.Error,
+// syscall errnos, and context errors underneath.
+func (m *model) do(ctx context.Context, req llm.Request, stream bool) (*http.Response, error) {
+	if m.p.apiKey == "" {
+		// Why a synthetic 401: the constructor never fails, so a missing
+		// key must surface at request time as the auth failure it is —
+		// permanent under llm.Classify, like a real 401.
+		return nil, &llm.APIError{
+			Provider: m.p.name,
+			Model:    m.id,
+			Status:   http.StatusUnauthorized,
+			Code:     "missing_api_key",
+			Message:  "no API key configured: set OPENAI_API_KEY or use WithAPIKey",
+		}
+	}
+	body, err := json.Marshal(m.buildRequest(req, stream))
+	if err != nil {
+		return nil, fmt.Errorf("openai: encode request: %w", err)
+	}
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, m.p.baseURL+"/chat/completions", bytes.NewReader(body))
+	if err != nil {
+		return nil, fmt.Errorf("openai: build request: %w", err)
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+	httpReq.Header.Set("Authorization", "Bearer "+m.p.apiKey)
+	if stream {
+		httpReq.Header.Set("Accept", "text/event-stream")
+	}
+	httpResp, err := m.p.client.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("openai: do request: %w", err)
+	}
+	return httpResp, nil
+}
+
+// apiError converts a non-2xx response into *llm.APIError, pulling code and
+// message from the {"error":{...}} body when it parses.
+func (m *model) apiError(httpResp *http.Response) error {
+	apiErr := &llm.APIError{Provider: m.p.name, Model: m.id, Status: httpResp.StatusCode}
+	body, _ := io.ReadAll(io.LimitReader(httpResp.Body, 1<<20))
+	var env errorEnvelope
+	if err := json.Unmarshal(body, &env); err == nil &&
+		(env.Error.Message != "" || env.Error.Type != "" || env.Error.Code != "") {
+		apiErr.Message = env.Error.Message
+		apiErr.Code = env.Error.Code
+		if apiErr.Code == "" {
+			apiErr.Code = env.Error.Type
+		}
+	} else {
+		// Why: compat servers emit all sorts of error bodies; a raw snippet
+		// beats silence when the canonical envelope is absent.
+		apiErr.Message = strings.TrimSpace(string(body))
+	}
+	return apiErr
+}
+
+// toResponse maps the wire response onto the canonical llm.Response.
+func (m *model) toResponse(wire *chatResponse) *llm.Response {
+	resp := &llm.Response{Model: m.p.name + "/" + m.id, Raw: wire}
+	if wire.Usage != nil {
+		resp.Usage = llm.Usage{
+			InputTokens:  wire.Usage.PromptTokens,
+			OutputTokens: wire.Usage.CompletionTokens,
+		}
+	}
+	if len(wire.Choices) == 0 {
+		resp.FinishReason = llm.FinishOther
+		return resp
+	}
+	choice := wire.Choices[0]
+	if choice.Message.Content != "" {
+		resp.Parts = append(resp.Parts, llm.TextPart{Text: choice.Message.Content})
+	}
+	for i, tc := range choice.Message.ToolCalls {
+		id := tc.ID
+		if id == "" {
+			// Why: ToolResult.ID must echo ToolCall.ID, so calls from compat
+			// servers that omit ids get synthesized ones.
+			id = fmt.Sprintf("call_%d", i)
+		}
+		resp.ToolCalls = append(resp.ToolCalls, llm.ToolCall{
+			ID:        id,
+			Name:      tc.Function.Name,
+			Arguments: json.RawMessage(tc.Function.Arguments),
+		})
+	}
+	resp.FinishReason = mapFinish(choice.FinishReason, len(resp.ToolCalls) > 0)
+	return resp
+}
+
+// mapFinish maps a wire finish_reason to the canonical enum. Tool-call
+// presence wins over the reported reason: a forced (named tool_choice) call
+// can finish with "stop" while still carrying tool_calls.
+func mapFinish(reason string, hasToolCalls bool) llm.FinishReason {
+	if hasToolCalls {
+		return llm.FinishToolCalls
+	}
+	switch reason {
+	case "stop":
+		return llm.FinishStop
+	case "length":
+		return llm.FinishLength
+	case "tool_calls":
+		return llm.FinishToolCalls
+	case "content_filter":
+		return llm.FinishContentFilter
+	default:
+		return llm.FinishOther
+	}
+}
+
+// checkRequest enforces the model's effective capabilities. Why enforcement
+// rather than normalization: a separate media layer resizes/transcodes
+// images BEFORE requests reach the provider; this check is the honest
+// backstop that refuses, with llm.ErrUnsupported, what the target
+// declaredly cannot serve (chains advance past it penalty-free).
+func checkRequest(caps llm.Capabilities, req llm.Request) error {
+	if len(req.Tools) > 0 && !caps.SupportsTools {
+		return fmt.Errorf("%w: tools not supported", llm.ErrUnsupported)
+	}
+	if len(req.Schema) > 0 && !caps.SupportsStructured {
+		return fmt.Errorf("%w: structured output not supported", llm.ErrUnsupported)
+	}
+	images := 0
+	for _, msg := range req.Messages {
+		for _, part := range msg.Parts {
+			img, ok := part.(llm.ImagePart)
+			if !ok {
+				continue
+			}
+			images++
+			if !caps.SupportsImages() {
+				return fmt.Errorf("%w: image input not supported", llm.ErrUnsupported)
+			}
+			if !caps.MIMEAllowed(img.MIME) {
+				return fmt.Errorf("%w: image MIME type %q not allowed (allowed: %s)",
+					llm.ErrUnsupported, img.MIME, strings.Join(caps.AllowedImageMIME, ", "))
+			}
+			if caps.MaxImageBytes > 0 && len(img.Data) > caps.MaxImageBytes {
+				return fmt.Errorf("%w: image is %d bytes, limit is %d",
+					llm.ErrUnsupported, len(img.Data), caps.MaxImageBytes)
+			}
+		}
+	}
+	if images > caps.MaxImagesPerReq {
+		return fmt.Errorf("%w: request carries %d images, limit is %d",
+			llm.ErrUnsupported, images, caps.MaxImagesPerReq)
+	}
+	return nil
+}
diff --git a/provider/openai/openai.go b/provider/openai/openai.go
new file mode 100644
index 0000000..f8956dd
--- /dev/null
+++ b/provider/openai/openai.go
@@ -0,0 +1,133 @@
+// Package openai implements llm.Provider for the OpenAI Chat Completions
+// API and, via WithBaseURL/WithName, any OpenAI-compatible endpoint
+// (vLLM, Groq, Together, LM Studio, Ollama's /v1 shim, ...).
+//
+// Targeted API surface (verified against developers.openai.com, June 2026):
+// POST {base}/chat/completions with
+//   - messages: plain-string content for text-only turns, part arrays with
+//     base64 data-URL image_url entries for multimodal turns, assistant
+//     tool_calls history, and {"role":"tool","tool_call_id",...} results;
+//   - tools as {"type":"function","function":{...}} with tool_choice
+//     "auto"/"none"/"required" or a named-function object;
+//   - response_format {"type":"json_schema",...} structured output;
+//   - max_completion_tokens (or legacy max_tokens via WithLegacyMaxTokens
+//     for compat servers), temperature, top_p, stop, reasoning_effort;
+//   - data-only SSE streaming with stream_options.include_usage, the
+//     "data: [DONE]" sentinel, and tool-call deltas accumulated by index.
+//
+// Newer response fields (refusal, annotations, usage *_details, delta
+// obfuscation) are tolerated and ignored so both api.openai.com and older
+// compat servers decode cleanly.
+package openai
+
+import (
+	"net/http"
+	"os"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+const defaultBaseURL = "https://api.openai.com/v1"
+
+// Provider is an llm.Provider backed by an OpenAI Chat Completions endpoint.
+type Provider struct {
+	name            string
+	apiKey          string
+	baseURL         string
+	client          *http.Client
+	caps            llm.Capabilities
+	legacyMaxTokens bool
+}
+
+// Option configures the provider at construction.
+type Option func(*Provider)
+
+// WithAPIKey sets the API key. When absent, New reads OPENAI_API_KEY from
+// the environment at construction time.
+func WithAPIKey(key string) Option {
+	return func(p *Provider) { p.apiKey = key }
+}
+
+// WithBaseURL points the client at a different endpoint (compat servers).
+// The path "/chat/completions" is appended; a trailing slash is trimmed.
+func WithBaseURL(u string) Option {
+	return func(p *Provider) { p.baseURL = u }
+}
+
+// WithHTTPClient substitutes the HTTP client (timeouts, proxies, tests).
+func WithHTTPClient(c *http.Client) Option {
+	return func(p *Provider) {
+		if c != nil {
+			p.client = c
+		}
+	}
+}
+
+// WithName overrides the registry name ("openai" by default). Why: the same
+// client serves many OpenAI-compatible endpoints, and each needs a distinct
+// name in "provider/model" specs and error reporting.
+func WithName(name string) Option {
+	return func(p *Provider) { p.name = name }
+}
+
+// WithDefaultCapabilities replaces the provider-default capabilities.
+// Per-model overrides via llm.WithCapabilities still take precedence.
+func WithDefaultCapabilities(caps llm.Capabilities) Option {
+	return func(p *Provider) { p.caps = caps }
+}
+
+// WithLegacyMaxTokens sends Request.MaxTokens as "max_tokens" instead of
+// "max_completion_tokens". Why: OpenAI deprecated max_tokens, but many
+// third-party compat servers still only honor the legacy field.
+func WithLegacyMaxTokens() Option {
+	return func(p *Provider) { p.legacyMaxTokens = true }
+}
+
+// defaultCapabilities reflects OpenAI's current vision-capable chat models.
+// Why these limits: the published per-request caps (1500 images, 512 MB)
+// are far beyond what compat servers accept; 100 images / 20 MB each is a
+// conservative envelope, and the MIME list is the documented set (PNG,
+// JPEG, WEBP, non-animated GIF).
+func defaultCapabilities() llm.Capabilities {
+	return llm.Capabilities{
+		SupportsTools:      true,
+		SupportsStructured: true,
+		SupportsStreaming:  true,
+		MaxImagesPerReq:    100,
+		MaxImageBytes:      20 << 20,
+		AllowedImageMIME:   []string{"image/jpeg", "image/png", "image/webp", "image/gif"},
+	}
+}
+
+// New creates a Provider. It never fails: a missing API key surfaces as a
+// 401-style *llm.APIError at request time, not at construction.
+func New(opts ...Option) *Provider {
+	p := &Provider{
+		name:    "openai",
+		apiKey:  os.Getenv("OPENAI_API_KEY"),
+		baseURL: defaultBaseURL,
+		client:  http.DefaultClient,
+		caps:    defaultCapabilities(),
+	}
+	for _, opt := range opts {
+		opt(p)
+	}
+	p.baseURL = strings.TrimRight(p.baseURL, "/")
+	return p
+}
+
+// Name implements llm.Provider.
+func (p *Provider) Name() string { return p.name }
+
+// Model implements llm.Provider. The id is passed through verbatim — no
+// catalog validation; unknown models fail at request time with the
+// backend's own error.
+func (p *Provider) Model(id string, opts ...llm.ModelOption) (llm.Model, error) {
+	cfg := llm.ApplyModelOptions(opts)
+	caps := p.caps
+	if cfg.Capabilities != nil {
+		caps = *cfg.Capabilities
+	}
+	return &model{p: p, id: id, caps: caps}, nil
+}
diff --git a/provider/openai/openai_test.go b/provider/openai/openai_test.go
new file mode 100644
index 0000000..09843a0
--- /dev/null
+++ b/provider/openai/openai_test.go
@@ -0,0 +1,614 @@
+package openai
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"reflect"
+	"strings"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+var (
+	_ llm.Provider = (*Provider)(nil)
+	_ llm.Model    = (*model)(nil)
+	_ llm.Stream   = (*stream)(nil)
+)
+
+const textResponse = `{
+  "id": "chatcmpl-1", "object": "chat.completion", "created": 1741570283, "model": "gpt-test",
+  "choices": [{
+    "index": 0,
+    "message": {"role": "assistant", "content": "hello", "refusal": null, "annotations": []},
+    "logprobs": null,
+    "finish_reason": "stop"
+  }],
+  "usage": {
+    "prompt_tokens": 19, "completion_tokens": 10, "total_tokens": 29,
+    "prompt_tokens_details": {"cached_tokens": 0, "audio_tokens": 0},
+    "completion_tokens_details": {"reasoning_tokens": 0}
+  },
+  "service_tier": "default", "system_fingerprint": "fp_x"
+}`
+
+// recorded captures the last request a test server received.
+type recorded struct {
+	body   map[string]any
+	header http.Header
+	path   string
+	hits   int
+}
+
+// newServer starts a test server that records the request and replies with
+// a fixed status and body.
+func newServer(t *testing.T, status int, respBody string) (*httptest.Server, *recorded) {
+	t.Helper()
+	rec := &recorded{}
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		rec.hits++
+		rec.header = r.Header.Clone()
+		rec.path = r.URL.Path
+		raw, err := io.ReadAll(r.Body)
+		if err != nil {
+			t.Errorf("read request body: %v", err)
+		}
+		if len(raw) > 0 {
+			if err := json.Unmarshal(raw, &rec.body); err != nil {
+				t.Errorf("request body is not JSON: %v\n%s", err, raw)
+			}
+		}
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(status)
+		io.WriteString(w, respBody)
+	}))
+	t.Cleanup(srv.Close)
+	return srv, rec
+}
+
+func testModel(t *testing.T, srv *httptest.Server, popts []Option, mopts ...llm.ModelOption) llm.Model {
+	t.Helper()
+	opts := append([]Option{WithAPIKey("test-key"), WithBaseURL(srv.URL)}, popts...)
+	m, err := New(opts...).Model("gpt-test", mopts...)
+	if err != nil {
+		t.Fatalf("Model: %v", err)
+	}
+	return m
+}
+
+func fptr(f float64) *float64 { return &f }
+
+func TestGenerateRequestShape(t *testing.T) {
+	srv, rec := newServer(t, http.StatusOK, textResponse)
+	m := testModel(t, srv, nil)
+
+	req := llm.Request{
+		System: "base system",
+		Messages: []llm.Message{
+			llm.SystemText("folded system"),
+			llm.UserParts(llm.Text("look:"), llm.Image("image/png", []byte{1, 2, 3})),
+			{
+				Role:  llm.RoleAssistant,
+				Parts: []llm.Part{llm.Text("checking")},
+				ToolCalls: []llm.ToolCall{
+					{ID: "call_1", Name: "get_weather", Arguments: json.RawMessage(`{"city":"Boston"}`)},
+				},
+			},
+			llm.ToolResultsMessage(
+				llm.ToolResult{ID: "call_1", Name: "get_weather", Content: "72F"},
+				llm.ToolResult{ID: "call_2", Name: "get_weather", Content: "boom", IsError: true},
+			),
+			llm.UserText("thanks"),
+		},
+		Tools: []llm.Tool{{
+			Name:        "get_weather",
+			Description: "Get current weather",
+			Parameters:  json.RawMessage(`{"type":"object","properties":{"city":{"type":"string"}}}`),
+		}},
+		ToolChoice:      "auto",
+		Temperature:     fptr(0.5),
+		TopP:            fptr(0.9),
+		MaxTokens:       256,
+		StopSequences:   []string{"END"},
+		ReasoningEffort: "high",
+		Schema:          json.RawMessage(`{"type":"object","properties":{"ok":{"type":"boolean"}}}`),
+		SchemaName:      "verdict",
+	}
+	if _, err := m.Generate(context.Background(), req); err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+
+	want := map[string]any{
+		"model": "gpt-test",
+		"messages": []any{
+			map[string]any{"role": "system", "content": "base system\n\nfolded system"},
+			map[string]any{"role": "user", "content": []any{
+				map[string]any{"type": "text", "text": "look:"},
+				map[string]any{"type": "image_url", "image_url": map[string]any{"url": "data:image/png;base64,AQID"}},
+			}},
+			map[string]any{"role": "assistant", "content": "checking", "tool_calls": []any{
+				map[string]any{"id": "call_1", "type": "function", "function": map[string]any{
+					"name": "get_weather", "arguments": `{"city":"Boston"}`,
+				}},
+			}},
+			map[string]any{"role": "tool", "content": "72F", "tool_call_id": "call_1"},
+			map[string]any{"role": "tool", "content": "ERROR: boom", "tool_call_id": "call_2"},
+			map[string]any{"role": "user", "content": "thanks"},
+		},
+		"tools": []any{
+			map[string]any{"type": "function", "function": map[string]any{
+				"name":        "get_weather",
+				"description": "Get current weather",
+				"parameters":  map[string]any{"type": "object", "properties": map[string]any{"city": map[string]any{"type": "string"}}},
+			}},
+		},
+		"tool_choice":           "auto",
+		"temperature":           0.5,
+		"top_p":                 0.9,
+		"max_completion_tokens": float64(256),
+		"stop":                  []any{"END"},
+		"reasoning_effort":      "high",
+		"response_format": map[string]any{"type": "json_schema", "json_schema": map[string]any{
+			"name":   "verdict",
+			"schema": map[string]any{"type": "object", "properties": map[string]any{"ok": map[string]any{"type": "boolean"}}},
+		}},
+	}
+	if !reflect.DeepEqual(rec.body, want) {
+		got, _ := json.MarshalIndent(rec.body, "", "  ")
+		exp, _ := json.MarshalIndent(want, "", "  ")
+		t.Errorf("request body mismatch\ngot:\n%s\nwant:\n%s", got, exp)
+	}
+}
+
+func TestToolChoiceForms(t *testing.T) {
+	tests := []struct {
+		choice string
+		want   any // nil = key absent
+	}{
+		{"", nil},
+		{"auto", "auto"},
+		{"none", "none"},
+		{"required", "required"},
+		{"get_weather", map[string]any{"type": "function", "function": map[string]any{"name": "get_weather"}}},
+	}
+	for _, tt := range tests {
+		t.Run("choice="+tt.choice, func(t *testing.T) {
+			srv, rec := newServer(t, http.StatusOK, textResponse)
+			m := testModel(t, srv, nil)
+			req := llm.Request{
+				Messages:   []llm.Message{llm.UserText("hi")},
+				Tools:      []llm.Tool{{Name: "get_weather"}},
+				ToolChoice: tt.choice,
+			}
+			if _, err := m.Generate(context.Background(), req); err != nil {
+				t.Fatalf("Generate: %v", err)
+			}
+			got, present := rec.body["tool_choice"]
+			if tt.want == nil {
+				if present {
+					t.Errorf("tool_choice present, want omitted: %v", got)
+				}
+				return
+			}
+			if !reflect.DeepEqual(got, tt.want) {
+				t.Errorf("tool_choice = %v, want %v", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestMaxTokensField(t *testing.T) {
+	t.Run("default uses max_completion_tokens", func(t *testing.T) {
+		srv, rec := newServer(t, http.StatusOK, textResponse)
+		m := testModel(t, srv, nil)
+		req := llm.Request{Messages: []llm.Message{llm.UserText("hi")}, MaxTokens: 64}
+		if _, err := m.Generate(context.Background(), req); err != nil {
+			t.Fatalf("Generate: %v", err)
+		}
+		if got := rec.body["max_completion_tokens"]; got != float64(64) {
+			t.Errorf("max_completion_tokens = %v, want 64", got)
+		}
+		if _, present := rec.body["max_tokens"]; present {
+			t.Error("max_tokens present, want omitted")
+		}
+	})
+	t.Run("WithLegacyMaxTokens uses max_tokens", func(t *testing.T) {
+		srv, rec := newServer(t, http.StatusOK, textResponse)
+		m := testModel(t, srv, []Option{WithLegacyMaxTokens()})
+		req := llm.Request{Messages: []llm.Message{llm.UserText("hi")}, MaxTokens: 64}
+		if _, err := m.Generate(context.Background(), req); err != nil {
+			t.Fatalf("Generate: %v", err)
+		}
+		if got := rec.body["max_tokens"]; got != float64(64) {
+			t.Errorf("max_tokens = %v, want 64", got)
+		}
+		if _, present := rec.body["max_completion_tokens"]; present {
+			t.Error("max_completion_tokens present, want omitted")
+		}
+	})
+	t.Run("zero omits both", func(t *testing.T) {
+		srv, rec := newServer(t, http.StatusOK, textResponse)
+		m := testModel(t, srv, nil)
+		req := llm.Request{Messages: []llm.Message{llm.UserText("hi")}}
+		if _, err := m.Generate(context.Background(), req); err != nil {
+			t.Fatalf("Generate: %v", err)
+		}
+		if _, present := rec.body["max_tokens"]; present {
+			t.Error("max_tokens present, want omitted")
+		}
+		if _, present := rec.body["max_completion_tokens"]; present {
+			t.Error("max_completion_tokens present, want omitted")
+		}
+	})
+}
+
+func TestSchemaNameDefault(t *testing.T) {
+	srv, rec := newServer(t, http.StatusOK, textResponse)
+	m := testModel(t, srv, nil)
+	req := llm.Request{
+		Messages: []llm.Message{llm.UserText("hi")},
+		Schema:   json.RawMessage(`{"type":"object"}`),
+	}
+	if _, err := m.Generate(context.Background(), req); err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	rf, ok := rec.body["response_format"].(map[string]any)
+	if !ok {
+		t.Fatalf("response_format missing: %v", rec.body)
+	}
+	js, ok := rf["json_schema"].(map[string]any)
+	if !ok {
+		t.Fatalf("json_schema missing: %v", rf)
+	}
+	if js["name"] != "response" {
+		t.Errorf("schema name = %v, want %q", js["name"], "response")
+	}
+}
+
+func TestGenerateTextResponse(t *testing.T) {
+	srv, _ := newServer(t, http.StatusOK, textResponse)
+	m := testModel(t, srv, nil)
+	resp, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if got := resp.Text(); got != "hello" {
+		t.Errorf("Text = %q, want %q", got, "hello")
+	}
+	if resp.FinishReason != llm.FinishStop {
+		t.Errorf("FinishReason = %v, want %v", resp.FinishReason, llm.FinishStop)
+	}
+	if resp.Usage != (llm.Usage{InputTokens: 19, OutputTokens: 10}) {
+		t.Errorf("Usage = %+v, want {19 10}", resp.Usage)
+	}
+	if resp.Model != "openai/gpt-test" {
+		t.Errorf("Model = %q, want %q", resp.Model, "openai/gpt-test")
+	}
+	if len(resp.ToolCalls) != 0 {
+		t.Errorf("ToolCalls = %v, want none", resp.ToolCalls)
+	}
+	if resp.Raw == nil {
+		t.Error("Raw is nil, want wire response")
+	}
+}
+
+func TestGenerateToolCallResponse(t *testing.T) {
+	const body = `{
+	  "id": "chatcmpl-2", "object": "chat.completion", "created": 1, "model": "gpt-test",
+	  "choices": [{
+	    "index": 0,
+	    "message": {"role": "assistant", "content": null, "tool_calls": [
+	      {"id": "call_9", "type": "function", "function": {"name": "get_weather", "arguments": "{\"city\":\"Boston\"}"}},
+	      {"id": "", "type": "function", "function": {"name": "get_time", "arguments": "{}"}}
+	    ]},
+	    "finish_reason": "stop"
+	  }],
+	  "usage": {"prompt_tokens": 3, "completion_tokens": 4, "total_tokens": 7}
+	}`
+	srv, _ := newServer(t, http.StatusOK, body)
+	m := testModel(t, srv, nil)
+	resp, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if len(resp.ToolCalls) != 2 {
+		t.Fatalf("ToolCalls = %d, want 2", len(resp.ToolCalls))
+	}
+	tc := resp.ToolCalls[0]
+	if tc.ID != "call_9" || tc.Name != "get_weather" || string(tc.Arguments) != `{"city":"Boston"}` {
+		t.Errorf("ToolCalls[0] = %+v", tc)
+	}
+	if resp.ToolCalls[1].ID != "call_1" {
+		t.Errorf("synthesized ID = %q, want %q", resp.ToolCalls[1].ID, "call_1")
+	}
+	// finish_reason "stop" with tool_calls present: presence wins.
+	if resp.FinishReason != llm.FinishToolCalls {
+		t.Errorf("FinishReason = %v, want %v", resp.FinishReason, llm.FinishToolCalls)
+	}
+	if len(resp.Parts) != 0 {
+		t.Errorf("Parts = %v, want none", resp.Parts)
+	}
+}
+
+func TestFinishReasonMapping(t *testing.T) {
+	tests := []struct {
+		wire string
+		want llm.FinishReason
+	}{
+		{"stop", llm.FinishStop},
+		{"length", llm.FinishLength},
+		{"tool_calls", llm.FinishToolCalls},
+		{"content_filter", llm.FinishContentFilter},
+		{"function_call", llm.FinishOther},
+		{"weird_new_reason", llm.FinishOther},
+	}
+	for _, tt := range tests {
+		t.Run(tt.wire, func(t *testing.T) {
+			body := `{"choices":[{"index":0,"message":{"role":"assistant","content":"x"},"finish_reason":"` + tt.wire + `"}],"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2}}`
+			srv, _ := newServer(t, http.StatusOK, body)
+			m := testModel(t, srv, nil)
+			resp, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+			if err != nil {
+				t.Fatalf("Generate: %v", err)
+			}
+			if resp.FinishReason != tt.want {
+				t.Errorf("FinishReason = %v, want %v", resp.FinishReason, tt.want)
+			}
+		})
+	}
+}
+
+func TestAPIErrorMapping(t *testing.T) {
+	t.Run("429 rate limit is transient", func(t *testing.T) {
+		const body = `{"error":{"message":"Rate limit reached","type":"rate_limit_error","param":null,"code":"rate_limit_exceeded"}}`
+		srv, _ := newServer(t, http.StatusTooManyRequests, body)
+		m := testModel(t, srv, nil)
+		_, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+		apiErr, ok := errors.AsType[*llm.APIError](err)
+		if !ok {
+			t.Fatalf("err = %v (%T), want *llm.APIError", err, err)
+		}
+		if apiErr.Status != http.StatusTooManyRequests {
+			t.Errorf("Status = %d, want 429", apiErr.Status)
+		}
+		if apiErr.Code != "rate_limit_exceeded" {
+			t.Errorf("Code = %q, want %q", apiErr.Code, "rate_limit_exceeded")
+		}
+		if apiErr.Message != "Rate limit reached" {
+			t.Errorf("Message = %q", apiErr.Message)
+		}
+		if apiErr.Provider != "openai" || apiErr.Model != "gpt-test" {
+			t.Errorf("Provider/Model = %q/%q", apiErr.Provider, apiErr.Model)
+		}
+		if got := llm.Classify(err); got != llm.ClassTransient {
+			t.Errorf("Classify = %v, want transient", got)
+		}
+	})
+	t.Run("401 code null falls back to type, permanent", func(t *testing.T) {
+		const body = `{"error":{"message":"Incorrect API key provided","type":"authentication_error","param":null,"code":null}}`
+		srv, _ := newServer(t, http.StatusUnauthorized, body)
+		m := testModel(t, srv, nil)
+		_, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+		apiErr, ok := errors.AsType[*llm.APIError](err)
+		if !ok {
+			t.Fatalf("err = %v (%T), want *llm.APIError", err, err)
+		}
+		if apiErr.Status != http.StatusUnauthorized || apiErr.Code != "authentication_error" {
+			t.Errorf("Status/Code = %d/%q", apiErr.Status, apiErr.Code)
+		}
+		if got := llm.Classify(err); got != llm.ClassPermanent {
+			t.Errorf("Classify = %v, want permanent", got)
+		}
+	})
+	t.Run("non-JSON body becomes message", func(t *testing.T) {
+		srv, _ := newServer(t, http.StatusServiceUnavailable, "upstream exploded\n")
+		m := testModel(t, srv, nil)
+		_, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+		apiErr, ok := errors.AsType[*llm.APIError](err)
+		if !ok {
+			t.Fatalf("err = %v (%T), want *llm.APIError", err, err)
+		}
+		if apiErr.Status != http.StatusServiceUnavailable || apiErr.Message != "upstream exploded" {
+			t.Errorf("Status/Message = %d/%q", apiErr.Status, apiErr.Message)
+		}
+	})
+}
+
+func TestMissingAPIKey(t *testing.T) {
+	t.Setenv("OPENAI_API_KEY", "")
+	srv, rec := newServer(t, http.StatusOK, textResponse)
+	m, err := New(WithBaseURL(srv.URL)).Model("gpt-test")
+	if err != nil {
+		t.Fatalf("Model: %v", err)
+	}
+	_, err = m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	apiErr, ok := errors.AsType[*llm.APIError](err)
+	if !ok {
+		t.Fatalf("err = %v (%T), want *llm.APIError", err, err)
+	}
+	if apiErr.Status != http.StatusUnauthorized || apiErr.Code != "missing_api_key" {
+		t.Errorf("Status/Code = %d/%q, want 401/missing_api_key", apiErr.Status, apiErr.Code)
+	}
+	if rec.hits != 0 {
+		t.Errorf("server hit %d times, want 0", rec.hits)
+	}
+}
+
+func TestEnvAPIKeyReadAtConstruction(t *testing.T) {
+	srv, rec := newServer(t, http.StatusOK, textResponse)
+	t.Setenv("OPENAI_API_KEY", "env-secret")
+	p := New(WithBaseURL(srv.URL))
+	t.Setenv("OPENAI_API_KEY", "changed-later") // must not affect p
+	m, err := p.Model("gpt-test")
+	if err != nil {
+		t.Fatalf("Model: %v", err)
+	}
+	if _, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}}); err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if got := rec.header.Get("Authorization"); got != "Bearer env-secret" {
+		t.Errorf("Authorization = %q, want %q", got, "Bearer env-secret")
+	}
+}
+
+func TestAuthAndContentTypeHeaders(t *testing.T) {
+	srv, rec := newServer(t, http.StatusOK, textResponse)
+	m := testModel(t, srv, nil)
+	if _, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}}); err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if got := rec.header.Get("Authorization"); got != "Bearer test-key" {
+		t.Errorf("Authorization = %q, want %q", got, "Bearer test-key")
+	}
+	if got := rec.header.Get("Content-Type"); got != "application/json" {
+		t.Errorf("Content-Type = %q, want application/json", got)
+	}
+	if rec.path != "/chat/completions" {
+		t.Errorf("path = %q, want /chat/completions", rec.path)
+	}
+}
+
+func TestCompatEndpointNameAndBaseURL(t *testing.T) {
+	srv, rec := newServer(t, http.StatusOK, textResponse)
+	p := New(WithName("groq"), WithAPIKey("k"), WithBaseURL(srv.URL+"/openai/v1/"))
+	if p.Name() != "groq" {
+		t.Errorf("Name = %q, want groq", p.Name())
+	}
+	m, err := p.Model("llama-3.3-70b")
+	if err != nil {
+		t.Fatalf("Model: %v", err)
+	}
+	resp, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if rec.path != "/openai/v1/chat/completions" {
+		t.Errorf("path = %q, want /openai/v1/chat/completions (trailing slash trimmed)", rec.path)
+	}
+	if resp.Model != "groq/llama-3.3-70b" {
+		t.Errorf("Model = %q, want groq/llama-3.3-70b", resp.Model)
+	}
+	if rec.body["model"] != "llama-3.3-70b" {
+		t.Errorf("wire model = %v, want llama-3.3-70b (verbatim)", rec.body["model"])
+	}
+}
+
+func TestCapabilityEnforcement(t *testing.T) {
+	img := func(mime string, n int) llm.Part { return llm.Image(mime, make([]byte, n)) }
+	tests := []struct {
+		name string
+		caps *llm.Capabilities // nil = provider defaults
+		msg  llm.Message
+	}{
+		{
+			name: "images unsupported",
+			caps: &llm.Capabilities{SupportsTools: true, SupportsStreaming: true},
+			msg:  llm.UserParts(img("image/png", 4)),
+		},
+		{
+			name: "too many images",
+			caps: &llm.Capabilities{MaxImagesPerReq: 1},
+			msg:  llm.UserParts(img("image/png", 4), img("image/png", 4)),
+		},
+		{
+			name: "disallowed MIME under defaults",
+			msg:  llm.UserParts(img("image/bmp", 4)),
+		},
+		{
+			name: "image too large",
+			caps: &llm.Capabilities{MaxImagesPerReq: 4, MaxImageBytes: 2},
+			msg:  llm.UserParts(img("image/png", 3)),
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			srv, rec := newServer(t, http.StatusOK, textResponse)
+			var mopts []llm.ModelOption
+			if tt.caps != nil {
+				mopts = append(mopts, llm.WithCapabilities(*tt.caps))
+			}
+			m := testModel(t, srv, nil, mopts...)
+			_, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{tt.msg}})
+			if !errors.Is(err, llm.ErrUnsupported) {
+				t.Fatalf("err = %v, want ErrUnsupported", err)
+			}
+			if got := llm.Classify(err); got != llm.ClassPermanent {
+				t.Errorf("Classify = %v, want permanent", got)
+			}
+			if rec.hits != 0 {
+				t.Errorf("server hit %d times, want 0 (must refuse before sending)", rec.hits)
+			}
+		})
+	}
+
+	t.Run("streaming unsupported", func(t *testing.T) {
+		srv, rec := newServer(t, http.StatusOK, textResponse)
+		m := testModel(t, srv, nil, llm.WithCapabilities(llm.Capabilities{SupportsTools: true}))
+		_, err := m.Stream(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+		if !errors.Is(err, llm.ErrUnsupported) {
+			t.Fatalf("err = %v, want ErrUnsupported", err)
+		}
+		if rec.hits != 0 {
+			t.Errorf("server hit %d times, want 0", rec.hits)
+		}
+	})
+}
+
+func TestModelCapabilitiesOverride(t *testing.T) {
+	p := New(WithAPIKey("k"))
+	def, err := p.Model("a")
+	if err != nil {
+		t.Fatalf("Model: %v", err)
+	}
+	if caps := def.Capabilities(); !caps.SupportsTools || caps.MaxImagesPerReq != 100 || caps.MaxImageBytes != 20<<20 {
+		t.Errorf("default caps = %+v", caps)
+	}
+	custom := llm.Capabilities{SupportsStreaming: true, ContextWindow: 8192}
+	ovr, err := p.Model("b", llm.WithCapabilities(custom))
+	if err != nil {
+		t.Fatalf("Model: %v", err)
+	}
+	if got := ovr.Capabilities(); !reflect.DeepEqual(got, custom) {
+		t.Errorf("override caps = %+v, want %+v", got, custom)
+	}
+}
+
+func TestTransportErrorIsNotAPIError(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {}))
+	url := srv.URL
+	srv.Close() // guarantee connection refused
+	p := New(WithAPIKey("k"), WithBaseURL(url))
+	m, err := p.Model("gpt-test")
+	if err != nil {
+		t.Fatalf("Model: %v", err)
+	}
+	_, err = m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err == nil {
+		t.Fatal("Generate succeeded against closed server")
+	}
+	if _, ok := errors.AsType[*llm.APIError](err); ok {
+		t.Errorf("transport error wrapped in APIError: %v", err)
+	}
+	if !strings.Contains(err.Error(), "openai: do request") {
+		t.Errorf("err = %v, want openai: do request context", err)
+	}
+	if got := llm.Classify(err); got != llm.ClassTransient {
+		t.Errorf("Classify = %v, want transient (net error must stay visible)", got)
+	}
+}
+
+func TestDecodeErrorWrapped(t *testing.T) {
+	srv, _ := newServer(t, http.StatusOK, "{not json")
+	m := testModel(t, srv, nil)
+	_, err := m.Generate(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err == nil || !strings.Contains(err.Error(), "openai: decode response") {
+		t.Errorf("err = %v, want decode response context", err)
+	}
+	if _, ok := errors.AsType[*llm.APIError](err); ok {
+		t.Errorf("decode error wrapped in APIError: %v", err)
+	}
+}
diff --git a/provider/openai/stream.go b/provider/openai/stream.go
new file mode 100644
index 0000000..a6a7e76
--- /dev/null
+++ b/provider/openai/stream.go
@@ -0,0 +1,183 @@
+package openai
+
+import (
+	"bufio"
+	"encoding/json"
+	"fmt"
+	"io"
+	"strings"
+	"sync"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// stream consumes the data-only SSE stream of chat.completion.chunk events.
+//
+// Delivery contract: TextDelta events as content fragments arrive; ToolCall
+// events only once fully assembled (fragments are buffered internally and
+// flushed at stream end — simplest correct handling of interleaved parallel
+// calls); exactly one final Response event; then io.EOF.
+type stream struct {
+	m    *model
+	body io.ReadCloser
+	sc   *bufio.Scanner
+
+	closeOnce sync.Once
+	closeErr  error
+
+	queue []llm.StreamEvent
+	done  bool // finalize ran; drain queue then io.EOF
+
+	text    strings.Builder
+	calls   []*toolCallAcc // first-appearance order
+	byIndex map[int]*toolCallAcc
+	finish  string
+	usage   llm.Usage
+}
+
+// toolCallAcc accumulates one tool call's fragments. The id and name arrive
+// on the first fragment for an index; arguments arrive as string pieces to
+// concatenate.
+type toolCallAcc struct {
+	id   string
+	name string
+	args strings.Builder
+}
+
+// Next implements llm.Stream.
+func (s *stream) Next() (llm.StreamEvent, error) {
+	for {
+		if len(s.queue) > 0 {
+			ev := s.queue[0]
+			s.queue = s.queue[1:]
+			return ev, nil
+		}
+		if s.done {
+			return llm.StreamEvent{}, io.EOF
+		}
+		if !s.sc.Scan() {
+			if err := s.sc.Err(); err != nil {
+				return llm.StreamEvent{}, fmt.Errorf("openai: read stream: %w", err)
+			}
+			// Why: some compat servers close the body without a [DONE]
+			// sentinel; a clean EOF still finalizes with what arrived.
+			s.finalize()
+			continue
+		}
+		line := strings.TrimSpace(s.sc.Text())
+		if !strings.HasPrefix(line, "data:") {
+			continue // SSE comments, event:/id: fields, blank separators
+		}
+		payload := strings.TrimSpace(strings.TrimPrefix(line, "data:"))
+		if payload == "" {
+			continue
+		}
+		if payload == "[DONE]" {
+			s.finalize()
+			continue
+		}
+		if err := s.handleChunk([]byte(payload)); err != nil {
+			return llm.StreamEvent{}, err
+		}
+	}
+}
+
+// handleChunk folds one chat.completion.chunk into the stream state,
+// queueing any events it produces.
+func (s *stream) handleChunk(data []byte) error {
+	var chunk streamChunk
+	if err := json.Unmarshal(data, &chunk); err != nil {
+		return fmt.Errorf("openai: decode stream chunk: %w", err)
+	}
+	if chunk.Error != nil {
+		// Mid-stream error event on an otherwise-200 stream. Status stays 0:
+		// there is no failing HTTP status to report.
+		apiErr := &llm.APIError{
+			Provider: s.m.p.name,
+			Model:    s.m.id,
+			Code:     chunk.Error.Code,
+			Message:  chunk.Error.Message,
+		}
+		if apiErr.Code == "" {
+			apiErr.Code = chunk.Error.Type
+		}
+		return apiErr
+	}
+	if chunk.Usage != nil {
+		s.usage = llm.Usage{
+			InputTokens:  chunk.Usage.PromptTokens,
+			OutputTokens: chunk.Usage.CompletionTokens,
+		}
+	}
+	// Why the guard: the include_usage chunk arrives with an EMPTY choices
+	// array; indexing choices[0] unconditionally would panic on it.
+	if len(chunk.Choices) == 0 {
+		return nil
+	}
+	choice := chunk.Choices[0]
+	if choice.FinishReason != "" {
+		s.finish = choice.FinishReason
+	}
+	if choice.Delta.Content != "" {
+		s.text.WriteString(choice.Delta.Content)
+		s.queue = append(s.queue, llm.StreamEvent{TextDelta: choice.Delta.Content})
+	}
+	for _, tc := range choice.Delta.ToolCalls {
+		acc := s.byIndex[tc.Index]
+		if acc == nil {
+			if s.byIndex == nil {
+				s.byIndex = make(map[int]*toolCallAcc)
+			}
+			acc = &toolCallAcc{}
+			s.byIndex[tc.Index] = acc
+			s.calls = append(s.calls, acc)
+		}
+		if tc.ID != "" {
+			acc.id = tc.ID
+		}
+		if tc.Function.Name != "" {
+			acc.name = tc.Function.Name
+		}
+		acc.args.WriteString(tc.Function.Arguments)
+	}
+	return nil
+}
+
+// finalize assembles the buffered tool calls and the final Response, queues
+// them (ToolCall events first, Response last), and marks the stream done.
+func (s *stream) finalize() {
+	if s.done {
+		return
+	}
+	s.done = true
+	resp := &llm.Response{Model: s.m.p.name + "/" + s.m.id, Usage: s.usage}
+	if s.text.Len() > 0 {
+		resp.Parts = []llm.Part{llm.TextPart{Text: s.text.String()}}
+	}
+	for i, acc := range s.calls {
+		id := acc.id
+		if id == "" {
+			// Why: ToolResult.ID must echo ToolCall.ID; synthesize for
+			// compat servers that stream calls without ids.
+			id = fmt.Sprintf("call_%d", i)
+		}
+		resp.ToolCalls = append(resp.ToolCalls, llm.ToolCall{
+			ID:        id,
+			Name:      acc.name,
+			Arguments: json.RawMessage(acc.args.String()),
+		})
+	}
+	resp.FinishReason = mapFinish(s.finish, len(resp.ToolCalls) > 0)
+	for i := range resp.ToolCalls {
+		tc := resp.ToolCalls[i] // copy so the event doesn't alias the slice
+		s.queue = append(s.queue, llm.StreamEvent{ToolCall: &tc})
+	}
+	s.queue = append(s.queue, llm.StreamEvent{Response: resp})
+}
+
+// Close implements llm.Stream. Closing the body unblocks any in-flight read
+// and aborts the HTTP stream; safe to call at any time, including twice.
+func (s *stream) Close() error {
+	s.closeOnce.Do(func() { s.closeErr = s.body.Close() })
+	return s.closeErr
+}
diff --git a/provider/openai/stream_test.go b/provider/openai/stream_test.go
new file mode 100644
index 0000000..23944e2
--- /dev/null
+++ b/provider/openai/stream_test.go
@@ -0,0 +1,267 @@
+package openai
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// sseServer streams each payload as one "data: <payload>" SSE event and
+// records the request like newServer.
+func sseServer(t *testing.T, payloads ...string) (*httptest.Server, *recorded) {
+	t.Helper()
+	rec := &recorded{}
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		rec.hits++
+		rec.header = r.Header.Clone()
+		rec.path = r.URL.Path
+		raw, err := io.ReadAll(r.Body)
+		if err != nil {
+			t.Errorf("read request body: %v", err)
+		}
+		if len(raw) > 0 {
+			if err := json.Unmarshal(raw, &rec.body); err != nil {
+				t.Errorf("request body is not JSON: %v\n%s", err, raw)
+			}
+		}
+		w.Header().Set("Content-Type", "text/event-stream")
+		for _, p := range payloads {
+			io.WriteString(w, "data: "+p+"\n\n")
+		}
+	}))
+	t.Cleanup(srv.Close)
+	return srv, rec
+}
+
+// collect drains a stream to io.EOF, failing the test on any other error.
+func collect(t *testing.T, s llm.Stream) []llm.StreamEvent {
+	t.Helper()
+	var events []llm.StreamEvent
+	for {
+		ev, err := s.Next()
+		if err == io.EOF {
+			return events
+		}
+		if err != nil {
+			t.Fatalf("Next: %v", err)
+		}
+		events = append(events, ev)
+	}
+}
+
+func TestStreamText(t *testing.T) {
+	srv, rec := sseServer(t,
+		`{"id":"c1","object":"chat.completion.chunk","created":1,"model":"gpt-test","choices":[{"index":0,"delta":{"role":"assistant","content":"","refusal":null},"finish_reason":null}]}`,
+		`{"choices":[{"index":0,"delta":{"content":"Hel"},"finish_reason":null}],"obfuscation":"xK9q"}`,
+		`{"choices":[{"index":0,"delta":{"content":"lo"},"finish_reason":null}]}`,
+		`{"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}`,
+		`{"choices":[],"usage":{"prompt_tokens":5,"completion_tokens":2,"total_tokens":7}}`,
+		`[DONE]`,
+	)
+	m := testModel(t, srv, nil)
+	s, err := m.Stream(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err != nil {
+		t.Fatalf("Stream: %v", err)
+	}
+	defer s.Close()
+	events := collect(t, s)
+
+	// Request shape: stream flag, usage opt-in, SSE accept header.
+	if rec.body["stream"] != true {
+		t.Errorf("stream = %v, want true", rec.body["stream"])
+	}
+	so, _ := rec.body["stream_options"].(map[string]any)
+	if so == nil || so["include_usage"] != true {
+		t.Errorf("stream_options = %v, want include_usage true", rec.body["stream_options"])
+	}
+	if got := rec.header.Get("Accept"); got != "text/event-stream" {
+		t.Errorf("Accept = %q, want text/event-stream", got)
+	}
+
+	if len(events) != 3 {
+		t.Fatalf("got %d events, want 3: %+v", len(events), events)
+	}
+	if events[0].TextDelta != "Hel" || events[1].TextDelta != "lo" {
+		t.Errorf("deltas = %q, %q, want Hel, lo", events[0].TextDelta, events[1].TextDelta)
+	}
+	final := events[2].Response
+	if final == nil {
+		t.Fatal("last event has no Response")
+	}
+	if got := final.Text(); got != "Hello" {
+		t.Errorf("final text = %q, want Hello", got)
+	}
+	if final.FinishReason != llm.FinishStop {
+		t.Errorf("FinishReason = %v, want stop", final.FinishReason)
+	}
+	if final.Usage != (llm.Usage{InputTokens: 5, OutputTokens: 2}) {
+		t.Errorf("Usage = %+v, want {5 2}", final.Usage)
+	}
+	if final.Model != "openai/gpt-test" {
+		t.Errorf("Model = %q, want openai/gpt-test", final.Model)
+	}
+
+	// Next after EOF keeps returning EOF; Close is idempotent.
+	if _, err := s.Next(); err != io.EOF {
+		t.Errorf("Next after EOF = %v, want io.EOF", err)
+	}
+	if err := s.Close(); err != nil {
+		t.Errorf("first Close: %v", err)
+	}
+	if err := s.Close(); err != nil {
+		t.Errorf("second Close: %v", err)
+	}
+}
+
+func TestStreamParallelToolCalls(t *testing.T) {
+	// Two interleaved calls with distinct indexes; id/name only on the first
+	// fragment of each; arguments split across fragments.
+	srv, _ := sseServer(t,
+		`{"choices":[{"index":0,"delta":{"role":"assistant","content":null,"tool_calls":[{"index":0,"id":"call_a","type":"function","function":{"name":"get_weather","arguments":""}}]},"finish_reason":null}]}`,
+		`{"choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"city\":"}}]},"finish_reason":null}]}`,
+		`{"choices":[{"index":0,"delta":{"tool_calls":[{"index":1,"id":"call_b","type":"function","function":{"name":"get_time","arguments":"{\"tz\":"}}]},"finish_reason":null}]}`,
+		`{"choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\"Boston\"}"}}]},"finish_reason":null}]}`,
+		`{"choices":[{"index":0,"delta":{"tool_calls":[{"index":1,"function":{"arguments":"\"EST\"}"}}]},"finish_reason":null}]}`,
+		`{"choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}`,
+		`{"choices":[],"usage":{"prompt_tokens":11,"completion_tokens":9,"total_tokens":20}}`,
+		`[DONE]`,
+	)
+	m := testModel(t, srv, nil)
+	s, err := m.Stream(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err != nil {
+		t.Fatalf("Stream: %v", err)
+	}
+	defer s.Close()
+	events := collect(t, s)
+
+	if len(events) != 3 {
+		t.Fatalf("got %d events, want 3 (two tool calls + response): %+v", len(events), events)
+	}
+	a, b := events[0].ToolCall, events[1].ToolCall
+	if a == nil || b == nil {
+		t.Fatalf("events 0/1 are not tool calls: %+v", events)
+	}
+	if a.ID != "call_a" || a.Name != "get_weather" || string(a.Arguments) != `{"city":"Boston"}` {
+		t.Errorf("first call = %+v", a)
+	}
+	if b.ID != "call_b" || b.Name != "get_time" || string(b.Arguments) != `{"tz":"EST"}` {
+		t.Errorf("second call = %+v", b)
+	}
+	final := events[2].Response
+	if final == nil {
+		t.Fatal("last event has no Response")
+	}
+	if len(final.ToolCalls) != 2 {
+		t.Fatalf("final ToolCalls = %d, want 2", len(final.ToolCalls))
+	}
+	if final.ToolCalls[0].ID != "call_a" || final.ToolCalls[1].ID != "call_b" {
+		t.Errorf("final ToolCalls order = %q, %q", final.ToolCalls[0].ID, final.ToolCalls[1].ID)
+	}
+	if final.FinishReason != llm.FinishToolCalls {
+		t.Errorf("FinishReason = %v, want tool_calls", final.FinishReason)
+	}
+	if final.Usage != (llm.Usage{InputTokens: 11, OutputTokens: 9}) {
+		t.Errorf("Usage = %+v, want {11 9}", final.Usage)
+	}
+}
+
+func TestStreamMidStreamError(t *testing.T) {
+	srv, _ := sseServer(t,
+		`{"choices":[{"index":0,"delta":{"content":"par"},"finish_reason":null}]}`,
+		`{"error":{"message":"The server had an error while processing your request","type":"server_error","param":null,"code":null}}`,
+	)
+	m := testModel(t, srv, nil)
+	s, err := m.Stream(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err != nil {
+		t.Fatalf("Stream: %v", err)
+	}
+	defer s.Close()
+
+	ev, err := s.Next()
+	if err != nil || ev.TextDelta != "par" {
+		t.Fatalf("first event = %+v, %v; want TextDelta par", ev, err)
+	}
+	_, err = s.Next()
+	apiErr, ok := errors.AsType[*llm.APIError](err)
+	if !ok {
+		t.Fatalf("err = %v (%T), want *llm.APIError", err, err)
+	}
+	if apiErr.Code != "server_error" {
+		t.Errorf("Code = %q, want server_error", apiErr.Code)
+	}
+	if apiErr.Message != "The server had an error while processing your request" {
+		t.Errorf("Message = %q", apiErr.Message)
+	}
+	if apiErr.Status != 0 {
+		t.Errorf("Status = %d, want 0 (the HTTP stream was 200)", apiErr.Status)
+	}
+}
+
+func TestStreamHTTPError(t *testing.T) {
+	srv, _ := newServer(t, http.StatusTooManyRequests,
+		`{"error":{"message":"Rate limit reached","type":"rate_limit_error","param":null,"code":"rate_limit_exceeded"}}`)
+	m := testModel(t, srv, nil)
+	_, err := m.Stream(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	apiErr, ok := errors.AsType[*llm.APIError](err)
+	if !ok {
+		t.Fatalf("err = %v (%T), want *llm.APIError from Stream itself", err, err)
+	}
+	if apiErr.Status != http.StatusTooManyRequests || apiErr.Code != "rate_limit_exceeded" {
+		t.Errorf("Status/Code = %d/%q", apiErr.Status, apiErr.Code)
+	}
+	if got := llm.Classify(err); got != llm.ClassTransient {
+		t.Errorf("Classify = %v, want transient", got)
+	}
+}
+
+func TestStreamWithoutDoneSentinel(t *testing.T) {
+	// Why: some compat servers close the connection without "data: [DONE]";
+	// a clean EOF must still produce the final Response.
+	srv, _ := sseServer(t,
+		`{"choices":[{"index":0,"delta":{"content":"ok"},"finish_reason":null}]}`,
+		`{"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}`,
+	)
+	m := testModel(t, srv, nil)
+	s, err := m.Stream(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err != nil {
+		t.Fatalf("Stream: %v", err)
+	}
+	defer s.Close()
+	events := collect(t, s)
+	if len(events) != 2 {
+		t.Fatalf("got %d events, want 2: %+v", len(events), events)
+	}
+	final := events[1].Response
+	if final == nil || final.Text() != "ok" || final.FinishReason != llm.FinishStop {
+		t.Errorf("final = %+v", final)
+	}
+}
+
+func TestStreamCloseEarly(t *testing.T) {
+	srv, _ := sseServer(t,
+		`{"choices":[{"index":0,"delta":{"content":"a"},"finish_reason":null}]}`,
+		`{"choices":[{"index":0,"delta":{"content":"b"},"finish_reason":null}]}`,
+		`{"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}`,
+		`[DONE]`,
+	)
+	m := testModel(t, srv, nil)
+	s, err := m.Stream(context.Background(), llm.Request{Messages: []llm.Message{llm.UserText("hi")}})
+	if err != nil {
+		t.Fatalf("Stream: %v", err)
+	}
+	if _, err := s.Next(); err != nil {
+		t.Fatalf("Next: %v", err)
+	}
+	if err := s.Close(); err != nil {
+		t.Errorf("Close mid-stream: %v", err)
+	}
+	if err := s.Close(); err != nil {
+		t.Errorf("Close again: %v", err)
+	}
+}
diff --git a/provider/openai/wire.go b/provider/openai/wire.go
new file mode 100644
index 0000000..5553ccc
--- /dev/null
+++ b/provider/openai/wire.go
@@ -0,0 +1,321 @@
+package openai
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"strings"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+)
+
+// --- request wire shapes ---
+
+type chatRequest struct {
+	Model    string        `json:"model"`
+	Messages []wireMessage `json:"messages"`
+	Tools    []wireTool    `json:"tools,omitempty"`
+	// ToolChoice is "auto"/"none"/"required" (string) or a named-function
+	// object; any avoids two fields for one wire key.
+	ToolChoice          any                `json:"tool_choice,omitempty"`
+	Temperature         *float64           `json:"temperature,omitempty"`
+	TopP                *float64           `json:"top_p,omitempty"`
+	MaxCompletionTokens int                `json:"max_completion_tokens,omitempty"`
+	MaxTokens           int                `json:"max_tokens,omitempty"`
+	Stop                []string           `json:"stop,omitempty"`
+	ReasoningEffort     string             `json:"reasoning_effort,omitempty"`
+	ResponseFormat      *wireRespFormat    `json:"response_format,omitempty"`
+	Stream              bool               `json:"stream,omitempty"`
+	StreamOptions       *wireStreamOptions `json:"stream_options,omitempty"`
+}
+
+type wireMessage struct {
+	Role string `json:"role"`
+	// Content is a string for text-only turns, a part array for multimodal
+	// turns, or nil (wire null) for assistant turns that only call tools.
+	Content    any            `json:"content"`
+	ToolCalls  []wireToolCall `json:"tool_calls,omitempty"`
+	ToolCallID string         `json:"tool_call_id,omitempty"`
+}
+
+type wireTextPart struct {
+	Type string `json:"type"`
+	Text string `json:"text"`
+}
+
+type wireImagePart struct {
+	Type     string       `json:"type"`
+	ImageURL wireImageURL `json:"image_url"`
+}
+
+type wireImageURL struct {
+	URL string `json:"url"`
+}
+
+type wireToolCall struct {
+	ID       string           `json:"id"`
+	Type     string           `json:"type"`
+	Function wireFunctionCall `json:"function"`
+}
+
+type wireFunctionCall struct {
+	Name string `json:"name"`
+	// Arguments is a JSON-encoded STRING per the wire format, not an object.
+	Arguments string `json:"arguments"`
+}
+
+type wireTool struct {
+	Type     string           `json:"type"`
+	Function wireToolFunction `json:"function"`
+}
+
+type wireToolFunction struct {
+	Name        string          `json:"name"`
+	Description string          `json:"description,omitempty"`
+	Parameters  json.RawMessage `json:"parameters,omitempty"`
+}
+
+type wireNamedToolChoice struct {
+	Type     string       `json:"type"`
+	Function wireToolName `json:"function"`
+}
+
+type wireToolName struct {
+	Name string `json:"name"`
+}
+
+type wireRespFormat struct {
+	Type       string          `json:"type"`
+	JSONSchema *wireJSONSchema `json:"json_schema,omitempty"`
+}
+
+// wireJSONSchema omits the strict flag on purpose: strict mode imposes
+// schema rewrites (every property required, additionalProperties:false at
+// every level) that belong to the caller, not the transport.
+type wireJSONSchema struct {
+	Name   string          `json:"name"`
+	Schema json.RawMessage `json:"schema"`
+}
+
+type wireStreamOptions struct {
+	IncludeUsage bool `json:"include_usage"`
+}
+
+// --- response wire shapes (loose: unknown fields ignored) ---
+
+type chatResponse struct {
+	ID      string       `json:"id"`
+	Object  string       `json:"object"`
+	Created int64        `json:"created"`
+	Model   string       `json:"model"`
+	Choices []chatChoice `json:"choices"`
+	Usage   *wireUsage   `json:"usage"`
+}
+
+type chatChoice struct {
+	Index        int             `json:"index"`
+	Message      wireRespMessage `json:"message"`
+	FinishReason string          `json:"finish_reason"`
+}
+
+type wireRespMessage struct {
+	Role      string         `json:"role"`
+	Content   string         `json:"content"` // null decodes to ""
+	Refusal   string         `json:"refusal"` // tolerated, unused
+	ToolCalls []wireToolCall `json:"tool_calls"`
+}
+
+type wireUsage struct {
+	PromptTokens     int `json:"prompt_tokens"`
+	CompletionTokens int `json:"completion_tokens"`
+	TotalTokens      int `json:"total_tokens"`
+}
+
+type errorEnvelope struct {
+	Error wireError `json:"error"`
+}
+
+type wireError struct {
+	Message string `json:"message"`
+	Type    string `json:"type"`
+	Code    string `json:"code"` // null decodes to ""
+}
+
+// --- streaming wire shapes ---
+
+type streamChunk struct {
+	Choices []streamChoice `json:"choices"`
+	Usage   *wireUsage     `json:"usage"`
+	Error   *wireError     `json:"error"` // mid-stream error event
+}
+
+type streamChoice struct {
+	Index        int         `json:"index"`
+	Delta        streamDelta `json:"delta"`
+	FinishReason string      `json:"finish_reason"` // null decodes to ""
+}
+
+type streamDelta struct {
+	Content   string                `json:"content"` // null decodes to ""
+	ToolCalls []streamToolCallDelta `json:"tool_calls"`
+}
+
+// streamToolCallDelta is one tool-call fragment. The id and name appear only
+// on a call's first fragment; later fragments carry just index + an
+// arguments substring. Accumulation keys on Index, never ID.
+type streamToolCallDelta struct {
+	Index    int              `json:"index"`
+	ID       string           `json:"id"`
+	Function wireFunctionCall `json:"function"`
+}
+
+// --- mapping: llm.Request -> chatRequest ---
+
+// buildRequest translates the canonical request to the wire shape. The
+// capability check has already passed by the time this runs.
+func (m *model) buildRequest(req llm.Request, stream bool) *chatRequest {
+	out := &chatRequest{
+		Model:           m.id,
+		Temperature:     req.Temperature,
+		TopP:            req.TopP,
+		Stop:            req.StopSequences,
+		ReasoningEffort: req.ReasoningEffort,
+	}
+
+	// Fold Request.System and every RoleSystem message into one leading
+	// system message, System field first. Why: the canonical contract allows
+	// system content in both places; OpenAI wants one system mechanism.
+	var sys []string
+	if req.System != "" {
+		sys = append(sys, req.System)
+	}
+	for _, msg := range req.Messages {
+		if msg.Role == llm.RoleSystem {
+			if t := msg.Text(); t != "" {
+				sys = append(sys, t)
+			}
+		}
+	}
+	if joined := strings.Join(sys, "\n\n"); joined != "" {
+		out.Messages = append(out.Messages, wireMessage{Role: "system", Content: joined})
+	}
+
+	for _, msg := range req.Messages {
+		switch msg.Role {
+		case llm.RoleSystem:
+			// Folded above; excluded from the normal message list.
+		case llm.RoleUser:
+			out.Messages = append(out.Messages, wireMessage{Role: "user", Content: contentValue(msg.Parts)})
+		case llm.RoleAssistant:
+			wm := wireMessage{Role: "assistant"}
+			if text := msg.Text(); text != "" {
+				wm.Content = text
+			}
+			for _, tc := range msg.ToolCalls {
+				args := string(tc.Arguments)
+				if args == "" {
+					// Why: arguments must be a valid JSON document string;
+					// an empty string is not one.
+					args = "{}"
+				}
+				wm.ToolCalls = append(wm.ToolCalls, wireToolCall{
+					ID:       tc.ID,
+					Type:     "function",
+					Function: wireFunctionCall{Name: tc.Name, Arguments: args},
+				})
+			}
+			out.Messages = append(out.Messages, wm)
+		case llm.RoleTool:
+			// One wire message per result: the API pairs each tool output
+			// with its call via tool_call_id, one message each.
+			for _, tr := range msg.ToolResults {
+				content := tr.Content
+				if tr.IsError {
+					content = "ERROR: " + content
+				}
+				out.Messages = append(out.Messages, wireMessage{
+					Role:       "tool",
+					Content:    content,
+					ToolCallID: tr.ID,
+				})
+			}
+		}
+	}
+
+	for _, t := range req.Tools {
+		out.Tools = append(out.Tools, wireTool{
+			Type:     "function",
+			Function: wireToolFunction{Name: t.Name, Description: t.Description, Parameters: t.Parameters},
+		})
+	}
+
+	switch req.ToolChoice {
+	case "":
+		// Omit: provider default ("auto" when tools are present).
+	case "auto", "none", "required":
+		out.ToolChoice = req.ToolChoice
+	default:
+		// Any other value names the one tool the model must call.
+		out.ToolChoice = wireNamedToolChoice{Type: "function", Function: wireToolName{Name: req.ToolChoice}}
+	}
+
+	if req.MaxTokens > 0 {
+		if m.p.legacyMaxTokens {
+			out.MaxTokens = req.MaxTokens
+		} else {
+			out.MaxCompletionTokens = req.MaxTokens
+		}
+	}
+
+	if len(req.Schema) > 0 {
+		name := req.SchemaName
+		if name == "" {
+			name = "response"
+		}
+		out.ResponseFormat = &wireRespFormat{
+			Type:       "json_schema",
+			JSONSchema: &wireJSONSchema{Name: name, Schema: req.Schema},
+		}
+	}
+
+	if stream {
+		out.Stream = true
+		// Why: without include_usage the stream never reports token counts;
+		// the usage arrives in one extra chunk with an empty choices array.
+		out.StreamOptions = &wireStreamOptions{IncludeUsage: true}
+	}
+
+	return out
+}
+
+// contentValue renders message parts as the wire content value: a plain
+// string when text-only (maximum compat), a part array when images are
+// present.
+func contentValue(parts []llm.Part) any {
+	multimodal := false
+	for _, p := range parts {
+		if _, ok := p.(llm.ImagePart); ok {
+			multimodal = true
+			break
+		}
+	}
+	if !multimodal {
+		var b strings.Builder
+		for _, p := range parts {
+			if t, ok := p.(llm.TextPart); ok {
+				b.WriteString(t.Text)
+			}
+		}
+		return b.String()
+	}
+	out := make([]any, 0, len(parts))
+	for _, p := range parts {
+		switch v := p.(type) {
+		case llm.TextPart:
+			out = append(out, wireTextPart{Type: "text", Text: v.Text})
+		case llm.ImagePart:
+			url := "data:" + v.MIME + ";base64," + base64.StdEncoding.EncodeToString(v.Data)
+			out = append(out, wireImagePart{Type: "image_url", ImageURL: wireImageURL{URL: url}})
+		}
+	}
+	return out
+}
diff --git a/registry.go b/registry.go
index 430982f..2992b01 100644
--- a/registry.go
+++ b/registry.go
@@ -2,6 +2,7 @@ package majordomo
 
 import (
 	"fmt"
+	"net/http"
 	"os"
 	"strings"
 	"sync"
@@ -73,11 +74,12 @@ func (c ChainConfig) classify(err error) llm.ErrorClass {
 }
 
 type registryConfig struct {
-	health    health.Config
-	chain     ChainConfig
-	envLookup func(string) string
-	environ   func() []string
-	skipEnv   bool
+	health     health.Config
+	chain      ChainConfig
+	envLookup  func(string) string
+	environ    func() []string
+	skipEnv    bool
+	httpClient *http.Client
 }
 
 // RegistryOption configures New.
@@ -115,6 +117,14 @@ func WithoutEnvProviders() RegistryOption {
 	return func(rc *registryConfig) { rc.skipEnv = true }
 }
 
+// WithHTTPClient sets the HTTP client used by built-in providers and
+// env-DSN scheme factories created by this registry (proxies, custom TLS,
+// test servers). Providers registered explicitly via RegisterProvider keep
+// whatever client they were built with.
+func WithHTTPClient(c *http.Client) RegistryOption {
+	return func(rc *registryConfig) { rc.httpClient = c }
+}
+
 // New creates a Registry with all built-in providers and scheme factories
 // registered, then loads LLM_* env-DSN providers from the process
 // environment (unless WithoutEnvProviders is given). Malformed LLM_* entries
@@ -139,7 +149,7 @@ func New(opts ...RegistryOption) *Registry {
 		envLookup: cfg.envLookup,
 	}
 
-	registerBuiltins(r)
+	registerBuiltins(r, cfg.httpClient)
 
 	if !cfg.skipEnv {
 		env := make(map[string]string)