ci: inherit gadfly's default swarm (slim caller, re-pin @b02b11d)

steve/gadfly#10 moved the curated swarm (3 cloud + Claude Code sonnet/opus/
opus:max, 5-lens suite) into the reusable workflow's input defaults. Drop this
repo's explicit `with:` swarm block and inherit it — only the consumer-specific
allow-list remains. Re-pin to the post-#10 gadfly commit (@b02b11d).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.gitea/workflows/adversarial-review.yml

@@ -1,11 +1,8 @@
-# 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 executus PRs with 3 ollama-cloud models (3-lens suite). Gadfly
-# is a simple system — findings are advisory; always double-check before acting.
+# Gadfly adversarial review — subscribes to steve/gadfly's reusable workflow and
+# INHERITS its default swarm (3 cloud models + Claude Code sonnet/opus/opus:max,
+# 5-lens suite). This stub holds only the triggers, the actor gate, secret
+# forwarding, and the allow-list; the swarm config lives centrally in gadfly's
+# review-reusable.yml. Advisory only — never blocks a merge.
 
 name: Adversarial Review (Gadfly)
 
@@ -33,60 +30,23 @@ 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.
+    # the allowed_users input below (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
-    # Full fleet: 3 cloud (lens fan-out) + M1/M5 Macs via foreman. The slow local
-    # lanes dominate wall time, so allow plenty of headroom.
-    timeout-minutes: 90
-    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 }}
-          # Local Macs, reached through their foreman queues (native Ollama on the
-          # wire). GADFLY_ENDPOINT_M5 registers provider "m5",
-          # each a foreman-preset Ollama client at the secret's URL, of the form:
-          #   foreman|https://<foreman-host>|<token>
-          # Needs an image with foreman provider-type support (this one). If a Mac
-          # is offline that model's comment shows an error and the others still post.
-          # (Gitea secrets aren't auto-exposed — map each explicitly.)
-          GADFLY_ENDPOINT_M5: ${{ secrets.GADFLY_ENDPOINT_M5 }}
-          # Full fleet: 3 cloud + M1 Pro + M5 Max. The Macs are back so the
-          # gadfly-reports scoreboard can quantify whether they earn their keep
-          # (they previously took 26–29 min for ZERO real findings — now measured).
-          # Cloud concurrency lives in the LENSES: one cloud model at a time
-          # (ollama-cloud=1) with its 3 lenses concurrent (LENS ollama-cloud=3) so
-          # its comment lands sooner; each Mac runs one model, lenses serial (its
-          # foreman queue serializes anyway). All three provider lanes run parallel.
-          GADFLY_MODELS: "minimax-m3:cloud,glm-5.2:cloud,glm-5.1:cloud,kimi-k2.7-code:cloud,deepseek-v4-pro:cloud,nemotron-3-super:cloud,gpt-oss:120b-cloud,qwen3-coder:480b-cloud,gemma4:cloud,m5/qwen3.6:35b-mlx"
-          GADFLY_PROVIDER_CONCURRENCY: "ollama-cloud=3,m5=1"
-          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 so the slow local models stay 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 }}
+    # Pinned to an immutable gadfly commit (not @main): a push to gadfly can't
+    # silently change the code that runs with our forwarded secrets.
+    uses: steve/gadfly/.gitea/workflows/review-reusable.yml@b02b11d69139843665da4cdbf776bc0b3583490d
+    # Least privilege: forward ONLY the secrets the swarm uses (GITEA_TOKEN is auto).
+    secrets:
+      OLLAMA_CLOUD_API_KEY:    ${{ secrets.OLLAMA_CLOUD_API_KEY }}
+      CLAUDE_CODE_OAUTH_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
+      GADFLY_FINDINGS_URL:     ${{ secrets.GADFLY_FINDINGS_URL }}
+      GADFLY_FINDINGS_TOKEN:   ${{ secrets.GADFLY_FINDINGS_TOKEN }}
+    with:
+      # Inherit the default swarm from gadfly's review-reusable.yml; only the
+      # consumer-specific allow-list is set here.
+      allowed_users: "steve,fizi,dazed"

run/critic.go

@@ -114,11 +114,11 @@ func (b *criticBinding) maxStepsOption(base int) agent.Option {
 	})
 }
 
-// steerOptions returns the agent RunOptions that drain the critic's steer
-// messages into the loop. Empty when there is no critic.
-func (b *criticBinding) steerOptions() []agent.RunOption {
+// drainSteer returns the critic's queued steer messages (nil-safe), so the
+// executor can merge them with the session steer mailbox into one WithSteer.
+func (b *criticBinding) drainSteer() []llm.Message {
 	if b == nil {
 		return nil
 	}
-	return []agent.RunOption{agent.WithSteer(b.h.Steer)}
+	return b.h.Steer()
 }

run/executor.go

@@ -101,6 +101,10 @@ type Result struct {
 	Steps  []tool.Step
 	Usage  llm.Usage
 	Err    error
+	// PostRunResult carries artifacts produced by a SessionToolFactory's PostRun
+	// hook (rendered images, files). nil when no factory was set or PostRun
+	// returned nil. The host delivers these (e.g. mort's chat API / Discord).
+	PostRunResult *tool.PostRunResult
 }
 
 // Run executes ra with the given invocation + input and returns the Result. It
@@ -176,6 +180,12 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 		inv.RunState = stateAcc
 	}
 
+	// Steer mailbox: lets session tools (via inv.AttachImages) feed multimodal
+	// messages into the running conversation before its next step. Created BEFORE
+	// the toolbox build so any tool's handler captures the live AttachImages seam.
+	mailbox := &steerMailbox{}
+	inv.AttachImages = (&runSession{mailbox: mailbox}).AttachImages
+
 	// Build the toolbox from the agent's low-level tools.
 	toolbox, err := e.cfg.Registry.Build(ra.LowLevelTools, inv, tool.Visibility("private"), nil)
 	if err != nil {
@@ -192,6 +202,34 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 		return res
 	}
 
+	// Per-invocation ExtraTools + a SessionToolFactory's per-run tools, added on
+	// top of the agent's palette. The factory closes over the live session (the
+	// AttachImages mailbox); its PostRun hook (held for after the run) produces
+	// artifacts attached to res.PostRunResult, and its Cleanup is deferred. All
+	// nil-safe.
+	for _, t := range inv.ExtraTools {
+		if err := toolbox.Add(t); err != nil {
+			res.Err = fmt.Errorf("add extra tool: %w", err)
+			e.finishAudit(ctx, rec, "error", res, started, res.Err)
+			return res
+		}
+	}
+	var postRun func(ctx context.Context, transcript []llm.Message, output string, runErr error) *tool.PostRunResult
+	if inv.SessionToolFactory != nil {
+		st := inv.SessionToolFactory(&runSession{mailbox: mailbox})
+		if st.Cleanup != nil {
+			defer safeCleanup(st.Cleanup) // panic-isolated, like runPostRun
+		}
+		for _, t := range st.Tools {
+			if err := toolbox.Add(t); err != nil {
+				res.Err = fmt.Errorf("add session tool: %w", err)
+				e.finishAudit(ctx, rec, "error", res, started, res.Err)
+				return res
+			}
+		}
+		postRun = st.PostRun
+	}
+
 	// Run context: bound by MaxRuntime, detached from the caller's deadline so a
 	// lane/queue wait doesn't eat the run budget (mort's V10 lesson). Caller
 	// cancellation still propagates via MergeCancellation. Created BEFORE the
@@ -279,7 +317,10 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 	}
 
 	ag := agent.New(model, e.systemPrompt(ra), opts...)
-	runRes, runErr := ag.Run(runCtx, input, critic.steerOptions()...)
+	// One WithSteer drains BOTH the session mailbox (a tool's AttachImages) and
+	// the critic's nudges before each step.
+	steer := func() []llm.Message { return append(mailbox.drain(), critic.drainSteer()...) }
+	runRes, runErr := ag.Run(runCtx, input, agent.WithSteer(steer))
 
 	status := statusFor(runCtx, runErr)
 	if runRes != nil {
@@ -289,6 +330,19 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 	res.Steps = emitter.snapshot()
 	res.Err = runErr
 
+	// PostRun: hand the SessionToolFactory's hook the full transcript (populated
+	// even on partial results) so it can produce artifacts. Best-effort +
+	// panic-isolated — a PostRun failure never fails an otherwise-successful run.
+	if postRun != nil {
+		var transcript []llm.Message
+		if runRes != nil {
+			transcript = runRes.Messages
+		}
+		// Detach from the caller's ctx: a finished/cancelled caller must not abort
+		// artifact production (the hook owns its own bounding, per its contract).
+		res.PostRunResult = runPostRun(detach(ctx), postRun, transcript, res.Output, runErr)
+	}
+
 	e.finishAudit(ctx, rec, status, res, started, runErr)
 	if e.cfg.Ports.Budget != nil {
 		e.cfg.Ports.Budget.Commit(detach(ctx), inv.CallerID, time.Since(started).Seconds())

run/session.go

@@ -0,0 +1,88 @@
+package run
+
+import (
+	"context"
+	"log/slog"
+	"strings"
+	"sync"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+// runPostRun invokes a SessionToolFactory's PostRun hook with panic isolation:
+// a PostRun panic (or a slow artifact build that the hook mishandles) must not
+// fail an otherwise-successful run — artifacts are best-effort, the agent's text
+// output is the source of truth.
+func runPostRun(ctx context.Context,
+	hook func(context.Context, []llm.Message, string, error) *tool.PostRunResult,
+	transcript []llm.Message, output string, runErr error) (prr *tool.PostRunResult) {
+	defer func() {
+		if r := recover(); r != nil {
+			slog.Error("run: PostRun hook panicked; no artifacts produced", "panic", r)
+			prr = nil
+		}
+	}()
+	return hook(ctx, transcript, output, runErr)
+}
+
+// steerMailbox is a thread-safe queue of messages a session tool (via
+// tool.Invocation.AttachImages) wants injected into the agent loop before its
+// next step — the same WithSteer mechanism the critic uses for nudges, exposed
+// to ordinary tools so they can show the model content (e.g. a rendered
+// preview) it must SEE, not just be told about.
+type steerMailbox struct {
+	mu   sync.Mutex
+	msgs []llm.Message
+}
+
+func (m *steerMailbox) push(msg llm.Message) {
+	m.mu.Lock()
+	m.msgs = append(m.msgs, msg)
+	m.mu.Unlock()
+}
+
+// drain returns and clears the queued messages (nil when empty).
+func (m *steerMailbox) drain() []llm.Message {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if len(m.msgs) == 0 {
+		return nil
+	}
+	out := m.msgs
+	m.msgs = nil
+	return out
+}
+
+// runSession implements tool.AgentSession over a steer mailbox: AttachImages
+// queues a user-role multimodal message the agent loop injects before its next
+// step. Replaces legacy agentkit's Agent.AttachImages — majordomo's *agent.Agent
+// is immutable mid-run, so mutation flows through the run-scoped steer mailbox.
+type runSession struct{ mailbox *steerMailbox }
+
+func (s *runSession) AttachImages(text string, images ...llm.ImagePart) {
+	parts := make([]llm.Part, 0, len(images)+1)
+	if strings.TrimSpace(text) != "" {
+		parts = append(parts, llm.Text(text))
+	}
+	for _, img := range images {
+		parts = append(parts, img)
+	}
+	if len(parts) == 0 {
+		return
+	}
+	s.mailbox.push(llm.UserParts(parts...))
+}
+
+// safeCleanup runs a SessionTools.Cleanup with panic isolation, so a misbehaving
+// teardown (temp-dir removal, handle close) can't clobber an otherwise-successful
+// run via the executor's top-level recover.
+func safeCleanup(fn func()) {
+	defer func() {
+		if r := recover(); r != nil {
+			slog.Error("run: session Cleanup panicked", "panic", r)
+		}
+	}()
+	fn()
+}

run/session_test.go

@@ -0,0 +1,94 @@
+package run_test
+
+import (
+	"context"
+	"testing"
+
+	"gitea.stevedudenhoeffer.com/steve/majordomo/llm"
+	"gitea.stevedudenhoeffer.com/steve/majordomo/provider/fake"
+
+	"gitea.stevedudenhoeffer.com/steve/executus/run"
+	"gitea.stevedudenhoeffer.com/steve/executus/tool"
+)
+
+// TestSessionToolFactoryPostRun: a SessionToolFactory's PostRun hook produces an
+// artifact (from the run output + transcript) that lands on Result.PostRunResult,
+// and its Cleanup is deferred.
+func TestSessionToolFactoryPostRun(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m", fake.Reply("hello artifacts"))
+	m, _ := fp.Model("m")
+
+	cleanupCalled := false
+	inv := tool.Invocation{
+		RunID: "r1",
+		SessionToolFactory: func(_ tool.AgentSession) tool.SessionTools {
+			return tool.SessionTools{
+				PostRun: func(_ context.Context, transcript []llm.Message, output string, _ error) *tool.PostRunResult {
+					return &tool.PostRunResult{
+						Artifacts: []tool.Artifact{{Name: "out.txt", MimeType: "text/plain", Data: []byte(output)}},
+						Metadata:  map[string]any{"transcript_len": len(transcript)},
+					}
+				},
+				Cleanup: func() { cleanupCalled = true },
+			}
+		},
+	}
+	ex := run.New(run.Config{
+		Registry: tool.NewRegistry(),
+		Models:   func(ctx context.Context, _ string) (context.Context, llm.Model, error) { return ctx, m, nil },
+	})
+	res := ex.Run(context.Background(), run.RunnableAgent{ModelTier: "m"}, inv, "go")
+	if res.Err != nil {
+		t.Fatalf("run error: %v", res.Err)
+	}
+	if res.PostRunResult == nil {
+		t.Fatal("Result.PostRunResult is nil — PostRun hook not invoked / not attached")
+	}
+	if n := len(res.PostRunResult.Artifacts); n != 1 {
+		t.Fatalf("artifacts = %d, want 1", n)
+	}
+	a := res.PostRunResult.Artifacts[0]
+	if a.Name != "out.txt" || string(a.Data) != "hello artifacts" {
+		t.Errorf("artifact = {%q, %q}", a.Name, string(a.Data))
+	}
+	if tl, _ := res.PostRunResult.Metadata["transcript_len"].(int); tl < 1 {
+		t.Errorf("transcript not passed to PostRun (len=%d)", tl)
+	}
+	if !cleanupCalled {
+		t.Error("Cleanup was not deferred/called")
+	}
+}
+
+// TestSessionToolFactoryAddsTool: tools the factory returns join the run's
+// toolbox and are callable by the model.
+func TestSessionToolFactoryAddsTool(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m",
+		fake.ReplyWith(llm.Response{ToolCalls: []llm.ToolCall{{ID: "c1", Name: "render", Arguments: []byte(`{}`)}}}),
+		fake.Reply("rendered"),
+	)
+	m, _ := fp.Model("m")
+
+	toolCalled := false
+	renderTool := llm.DefineTool("render", "render a preview",
+		func(_ context.Context, _ struct{}) (any, error) { toolCalled = true; return "ok", nil })
+	inv := tool.Invocation{
+		RunID: "r2",
+		SessionToolFactory: func(_ tool.AgentSession) tool.SessionTools {
+			return tool.SessionTools{Tools: []llm.Tool{renderTool}}
+		},
+	}
+	ex := run.New(run.Config{
+		Registry: tool.NewRegistry(),
+		Models:   func(ctx context.Context, _ string) (context.Context, llm.Model, error) { return ctx, m, nil },
+	})
+	res := ex.Run(context.Background(),
+		run.RunnableAgent{ModelTier: "m", MaxIterations: 5}, inv, "go")
+	if res.Err != nil {
+		t.Fatalf("run error: %v", res.Err)
+	}
+	if !toolCalled {
+		t.Error("session-factory tool was not added to the toolbox / not called")
+	}
+}