ci(gadfly): trim the weakest reviewers from the swarm

Drop the four lowest-graded reviewers — m5/qwen3.6:35b-mlx, gemma4:cloud,
gpt-oss:120b-cloud, kimi-k2.7-code:cloud. Removing m5/qwen3.6 takes the last
local Mac out, so this is now a cloud-only fleet of 6 ollama-cloud models;
GADFLY_ENDPOINT_M5 and the m5 concurrency entry are gone and the per-job timeout
drops to 45m. README/CLAUDE.md kept in sync.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

.gitea/workflows/adversarial-review.yml

@@ -0,0 +1,82 @@
+# Gadfly — agentic adversarial PR reviewer (https://gitea.stevedudenhoeffer.com/steve/gadfly).
+#
+# Runs the published Gadfly image (pinned to an immutable :sha- tag — act_runner
+# caches :latest, and this build is what carries foreman provider-type support)
+# as a specialist swarm and posts
+# ONE consolidated review comment as gitea-actions. Advisory only — never blocks a
+# merge. This reviews majordomo PRs with 6 ollama-cloud models (3-lens suite).
+# Gadfly is a simple system — findings are advisory; always double-check before
+# acting.
+
+name: Adversarial Review (Gadfly)
+
+on:
+  pull_request:
+    types: [opened, reopened, ready_for_review]
+  issue_comment:
+    types: [created]
+  workflow_dispatch:
+    inputs:
+      pr_number:
+        description: "PR number to review"
+        required: true
+
+permissions:
+  contents: read
+  issues: write
+  pull-requests: write
+
+concurrency:
+  group: gadfly-${{ github.event.issue.number || github.event.pull_request.number || github.event.inputs.pr_number }}
+  cancel-in-progress: true
+
+jobs:
+  review:
+    # Security: only trusted users may trigger a secret-bearing run via a PR
+    # comment (pull_request + workflow_dispatch are already trusted). Mirrors
+    # GADFLY_ALLOWED_USERS, the in-container belt-and-suspenders check.
+    if: >-
+      github.event_name != 'issue_comment'
+      || (github.event.issue.pull_request
+          && (github.actor == 'steve'
+              || github.actor == 'fizi'
+              || github.actor == 'dazed'))
+    runs-on: ubuntu-latest
+    # Fleet: 6 ollama-cloud models (lens fan-out), no local Macs. (Trimmed the
+    # weakest reviewers by grade — m5/qwen3.6, gemma4, gpt-oss, kimi-k2.7 — plus
+    # the earlier M1 drop.) Plenty of headroom for the cloud lanes.
+    timeout-minutes: 45
+    steps:
+      - uses: docker://gitea.stevedudenhoeffer.com/steve/gadfly:sha-d7f364d
+        env:
+          GITEA_API: ${{ github.server_url }}/api/v1/repos/${{ github.repository }}
+          GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
+          OLLAMA_CLOUD_API_KEY: ${{ secrets.OLLAMA_CLOUD_API_KEY }}
+          # Cloud-only fleet (no local Macs). Cloud concurrency lives in the
+          # LENSES: models run a few at a time (ollama-cloud=3) with their 3 lenses
+          # concurrent (LENS ollama-cloud=3) so comments land sooner.
+          GADFLY_MODELS: "minimax-m3:cloud,glm-5.2:cloud,glm-5.1:cloud,deepseek-v4-pro:cloud,nemotron-3-super:cloud,qwen3-coder:480b-cloud"
+          GADFLY_PROVIDER_CONCURRENCY: "ollama-cloud=3"
+          GADFLY_PROVIDER_LENS_CONCURRENCY: "ollama-cloud=3"
+          # Default => the 3-lens suite (security, correctness, error-handling).
+          # Set the repo var GADFLY_SPECIALISTS to override (csv / "all" / "auto").
+          GADFLY_SPECIALISTS: ${{ vars.GADFLY_SPECIALISTS || 'security,correctness,error-handling' }}
+          # Per-lens deadline + bounded steps to keep each reviewer's run sane.
+          GADFLY_TIMEOUT_SECS: "600"
+          GADFLY_MAX_STEPS: "14"
+          # Allow-list for the comment trigger (mirrors the job-level if: guard).
+          GADFLY_ALLOWED_USERS: "steve,fizi,dazed"
+          # --- findings telemetry: POST runs + findings to the gadfly-reports store ---
+          # Advisory & off unless GADFLY_FINDINGS_URL is set; failures only log to
+          # stderr and never affect the review. GADFLY_REPO / GADFLY_PR are derived
+          # in-container; the URL + token are user-scope secrets.
+          GADFLY_FINDINGS_URL: ${{ secrets.GADFLY_FINDINGS_URL }}
+          GADFLY_FINDINGS_TOKEN: ${{ secrets.GADFLY_FINDINGS_TOKEN }}
+          # --- event context (leave as-is) ---
+          EVENT_NAME: ${{ github.event_name }}
+          PR: ${{ github.event.pull_request.number || github.event.issue.number || github.event.inputs.pr_number }}
+          PR_BRANCH: ${{ github.head_ref }}
+          IS_DRAFT: ${{ github.event.pull_request.draft }}
+          COMMENT_BODY: ${{ github.event.comment.body }}
+          COMMENT_ID: ${{ github.event.comment.id }}
+          ACTOR: ${{ github.actor }}

CLAUDE.md

@@ -135,6 +135,27 @@ CI: `.gitea/workflows/ci.yaml` (Gitea Actions, mirrors foreman). README.md
 must match reality in the same commit that changes behavior — no
 aspirational docs; unbuilt features are marked pending in the matrix.
 
+## Adversarial review loop (Gadfly)
+
+Ship work through PRs and let Gadfly review it before merge:
+
+- **Push to a PR, never straight to `main`.** Branch, push, open a PR.
+  `.gitea/workflows/adversarial-review.yml` runs Gadfly (the standalone
+  agentic adversarial reviewer) — a fleet of 6 ollama-cloud models, each
+  running the 3-lens suite (security, correctness, error-handling). Advisory
+  only; it never blocks the merge.
+- **Wait for Gadfly to finish, then read its output.** Don't merge while the
+  review is still running. Each model posts one consolidated comment; weigh
+  every finding on its merits and fix the real ones (Gadfly is a simple
+  system — findings are advisory, so confirm before acting).
+- **Grade the findings back to the Gadfly MCP.** For each finding, call
+  `mcp__gadfly__record_finding_grade`: `is_real=true` + a `severity`
+  (trivial|small|medium|high|critical) for a genuine problem, or
+  `is_real=false` for a false positive; add `notes`/`usefulness` when
+  useful. Use `mcp__gadfly__list_findings` (`only_ungraded=true`) to find
+  what still needs grading and `mcp__gadfly__scoreboard` for the per-model
+  rollup. This telemetry is how we measure whether each model earns its keep.
+
 ## Out of scope (anti-creep)
 
 No persistent store (health is in-memory behind the registry), no

provider/llamaswap/image.go

@@ -51,6 +51,9 @@ func (m *imageModel) Generate(ctx context.Context, req imagegen.Request, opts ..
 	if strings.TrimSpace(req.Prompt) == "" {
 		return nil, fmt.Errorf("%w: image generation requires a prompt", llm.ErrUnsupported)
 	}
+	if req.N < 0 {
+		return nil, fmt.Errorf("%w: image count N must be >= 0, got %d", llm.ErrUnsupported, req.N)
+	}
 
 	wire := imageRequest{
 		Model:          m.id,
@@ -61,7 +64,7 @@ func (m *imageModel) Generate(ctx context.Context, req imagegen.Request, opts ..
 	}
 
 	var resp imageResponse
-	if err := m.p.doJSON(ctx, http.MethodPost, "/v1/images/generations", &wire, &resp); err != nil {
+	if err := m.p.doJSON(ctx, http.MethodPost, "/v1/images/generations", m.id, &wire, &resp); err != nil {
 		return nil, err
 	}
 

provider/llamaswap/llamaswap.go

@@ -38,6 +38,11 @@ import (
 // DefaultName is the registry name used when WithName is not given.
 const DefaultName = "llama-swap"
 
+// maxResponseBytes caps the JSON body read on the success path. Generous
+// enough for a multi-image b64 payload, bounded so a hostile/buggy upstream
+// can't make a decode allocate without limit.
+const maxResponseBytes = 64 << 20
+
 // Provider is a llama-swap client. It satisfies llm.Provider (chat, delegated
 // to provider/openai) and imagegen.Provider (image generation), and exposes
 // llama-swap's management endpoints as concrete methods.
@@ -136,7 +141,7 @@ func (p *Provider) ListModels(ctx context.Context) ([]ModelInfo, error) {
 	var out struct {
 		Data []ModelInfo `json:"data"`
 	}
-	if err := p.doJSON(ctx, http.MethodGet, "/v1/models", nil, &out); err != nil {
+	if err := p.doJSON(ctx, http.MethodGet, "/v1/models", "", nil, &out); err != nil {
 		return nil, err
 	}
 	return out.Data, nil
@@ -148,7 +153,7 @@ func (p *Provider) ListModels(ctx context.Context) ([]ModelInfo, error) {
 // would have to guess.
 func (p *Provider) Running(ctx context.Context) (json.RawMessage, error) {
 	var out json.RawMessage
-	if err := p.doJSON(ctx, http.MethodGet, "/running", nil, &out); err != nil {
+	if err := p.doJSON(ctx, http.MethodGet, "/running", "", nil, &out); err != nil {
 		return nil, err
 	}
 	return out, nil
@@ -160,18 +165,30 @@ func (p *Provider) Running(ctx context.Context) (json.RawMessage, error) {
 func (p *Provider) Unload(ctx context.Context, model string) error {
 	path := "/api/models/unload"
 	if model != "" {
+		// Why reject rather than percent-escape: llama-swap model ids legitimately
+		// contain ":" (e.g. "qwen3:14b"), which is path-legal and must reach the
+		// server verbatim; only path-structure characters are dangerous (they'd
+		// redirect the request to another endpoint), and those never appear in a
+		// real model id.
+		if strings.ContainsAny(model, "/?#") {
+			return fmt.Errorf("llama-swap: invalid model id %q for unload (contains a path separator)", model)
+		}
 		path += "/" + model
 	}
-	return p.doJSON(ctx, http.MethodPost, path, nil, nil)
+	return p.doJSON(ctx, http.MethodPost, path, "", nil, nil)
 }
 
 // --- shared HTTP helper for management + image endpoints ---
 
 // doJSON performs a request to a llama-swap endpoint relative to baseURL,
-// optionally encoding body and decoding into out (either may be nil). Transport
+// optionally encoding body and decoding into out (either may be nil). model
-// failures are wrapped raw so llm.Classify still sees the underlying net error;
+// labels the failing target in any *llm.APIError ("" for endpoints that aren't
-// non-2xx responses become *llm.APIError.
+// model-specific). Transport failures are wrapped raw so llm.Classify still
-func (p *Provider) doJSON(ctx context.Context, method, path string, body, out any) error {
+// sees the underlying net error; non-2xx responses become *llm.APIError.
+func (p *Provider) doJSON(ctx context.Context, method, path, model string, body, out any) error {
+	if p.baseURL == "" {
+		return fmt.Errorf("llama-swap provider %q: no base URL configured (set one via WithBaseURL or an LLM_* env DSN)", p.name)
+	}
 	var rdr io.Reader
 	if body != nil {
 		b, err := json.Marshal(body)
@@ -196,12 +213,15 @@ func (p *Provider) doJSON(ctx context.Context, method, path string, body, out an
 	}
 	defer resp.Body.Close()
 	if resp.StatusCode/100 != 2 {
-		return p.apiError(resp, "")
+		return p.apiError(resp, model)
 	}
 	if out != nil {
-		if err := json.NewDecoder(resp.Body).Decode(out); err != nil {
+		if err := json.NewDecoder(io.LimitReader(resp.Body, maxResponseBytes)).Decode(out); err != nil {
 			return fmt.Errorf("llama-swap: decode response: %w", err)
 		}
+	} else {
+		// Drain (bounded) so the connection can be reused.
+		_, _ = io.Copy(io.Discard, io.LimitReader(resp.Body, maxResponseBytes))
 	}
 	return nil
 }

provider/llamaswap/llamaswap_test.go

@@ -127,6 +127,24 @@ func TestUnload(t *testing.T) {
 	if gotPath != "/api/models/unload" {
 		t.Errorf("unload-all path = %q", gotPath)
 	}
+
+	// A model id with a path separator is rejected before any request.
+	if err := p.Unload(context.Background(), "../admin"); err == nil {
+		t.Error("expected error for model id with path separator")
+	}
+}
+
+func TestManagementNoBaseURL(t *testing.T) {
+	p := New() // no base URL
+	if _, err := p.ListModels(context.Background()); err == nil {
+		t.Error("ListModels: expected error for missing base URL")
+	}
+	if _, err := p.Running(context.Background()); err == nil {
+		t.Error("Running: expected error for missing base URL")
+	}
+	if err := p.Unload(context.Background(), "m"); err == nil {
+		t.Error("Unload: expected error for missing base URL")
+	}
 }
 
 func TestRunningRaw(t *testing.T) {