Merge pull request 'run: fold inv.Images into the initial user message (multimodal opening turn)' (#16) from feat/initial-images into main

Every reviewer flagged that runAgent appended llm.Text(input) unconditionally, so
an image-only run (blank prompt) emitted an empty TextPart — inconsistent with the
sibling runSession.AttachImages which guards it. Mirror that guard
(strings.TrimSpace(input) != ""). Also:
- copy opts before appending (variadic backing array can have spare capacity; avoid
  aliasing a caller's slice).
- reword the doc comment to drop the mort-agentexec reference (executus is a
  standalone lib; a consumer name doesn't belong in its godoc).

Tests: image+text are co-located in ONE user message; an image-only run emits no
blank TextPart.

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).
+# Gadfly adversarial review — subscribes to steve/gadfly's reusable workflow and
-#
+# INHERITS its default swarm. This stub holds only the triggers, the actor gate,
-# Runs the published Gadfly image (pinned to an immutable :sha- tag — act_runner
+# secret forwarding, and the allow-list; the swarm config (models, lenses,
-# caches :latest, and this build is what carries foreman provider-type support)
+# concurrency, timeouts) lives centrally in gadfly's review-reusable.yml so it is
-# as a specialist swarm and posts
+# tuned in ONE place. Advisory only — never blocks a merge.
-# 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.
 
 name: Adversarial Review (Gadfly)
 
@@ -32,59 +29,26 @@ concurrency:
 jobs:
   review:
     # Security: only trusted users may trigger a secret-bearing run via a PR
-    # comment (pull_request + workflow_dispatch are already trusted). Mirrors
+    # comment (pull_request + workflow_dispatch are already trusted). Mirrors the
-    # GADFLY_ALLOWED_USERS, the in-container belt-and-suspenders check.
+    # allowed_users input below (the in-container belt-and-suspenders check) — both
+    # lists must stay in sync; a workflow if: can't read a workflow_call input.
     if: >-
       github.event_name != 'issue_comment'
       || (github.event.issue.pull_request
           && (github.actor == 'steve'
               || github.actor == 'fizi'
               || github.actor == 'dazed'))
-    runs-on: ubuntu-latest
+    # Tracks gadfly's v1 release tag — a curated pointer re-moved on each release
-    # Full fleet: 3 cloud (lens fan-out) + M1/M5 Macs via foreman. The slow local
+    # (unlike @main, which moves on every push). Central swarm tuning propagates
-    # lanes dominate wall time, so allow plenty of headroom.
+    # here automatically; the tradeoff vs a full sha pin is that v1 is mutable.
-    timeout-minutes: 90
+    uses: steve/gadfly/.gitea/workflows/review-reusable.yml@v1
-    steps:
+    # Least privilege: forward only the review secrets (not `secrets: inherit`,
-      - uses: docker://gitea.stevedudenhoeffer.com/steve/gadfly:sha-d7f364d
+    # which would expose every repo secret). GITEA_TOKEN is the automatic token.
-        env:
+    secrets:
-          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
+      CLAUDE_CODE_OAUTH_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
-                    # 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.)
-          # 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,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 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) ---
+    with:
-          EVENT_NAME: ${{ github.event_name }}
+      # Consumer-specific allow-list; everything else is inherited.
-          PR: ${{ github.event.pull_request.number || github.event.issue.number || github.event.inputs.pr_number }}
+      allowed_users: "steve,fizi,dazed"
-          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 }}

run/executor.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"strings"
 	"time"
 
 	"gitea.stevedudenhoeffer.com/steve/majordomo/agent"
@@ -320,7 +321,7 @@ func (e *Executor) Run(ctx context.Context, ra RunnableAgent, inv tool.Invocatio
 	// 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))
+	runRes, runErr := runAgent(runCtx, ag, input, inv.Images, agent.WithSteer(steer))
 
 	status := statusFor(runCtx, runErr)
 	if runRes != nil {
@@ -440,3 +441,28 @@ func detach(ctx context.Context) context.Context {
 	_ = cancel // bounded by the timeout; nothing to cancel early
 	return c
 }
+
+// runAgent dispatches the majordomo agent loop. majordomo's Run takes a text-only
+// input arg, so when the invocation carries images they're folded into the first
+// user message (text + image parts) via WithHistory and Run is called with an
+// empty input — the model then sees a multimodal opening turn. The image-less path
+// passes the prompt straight through.
+//
+// The text part is omitted when input is blank (image-only run), matching
+// runSession.AttachImages so no empty TextPart is sent.
+func runAgent(ctx context.Context, ag *agent.Agent, input string, images []llm.ImagePart, opts ...agent.RunOption) (*agent.Result, error) {
+	if len(images) == 0 {
+		return ag.Run(ctx, input, opts...)
+	}
+	parts := make([]llm.Part, 0, len(images)+1)
+	if strings.TrimSpace(input) != "" {
+		parts = append(parts, llm.Text(input))
+	}
+	for _, img := range images {
+		parts = append(parts, img)
+	}
+	// Copy opts before appending so a caller-supplied backing array is never
+	// mutated/aliased (the variadic slice can have spare capacity).
+	opts = append(opts[:len(opts):len(opts)], agent.WithHistory([]llm.Message{llm.UserParts(parts...)}))
+	return ag.Run(ctx, "", opts...)
+}

run/images_test.go

@@ -0,0 +1,121 @@
+package run_test
+
+import (
+	"context"
+	"strings"
+	"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"
+)
+
+// TestExecutorFoldsInitialImages: when the invocation carries Images, they're
+// folded into the first user message (alongside the prompt text) instead of being
+// dropped — majordomo's Run input arg is text-only, so the executor seeds the
+// multimodal opening turn via history.
+func TestExecutorFoldsInitialImages(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m", fake.Reply("saw the image"))
+	m, _ := fp.Model("m")
+
+	img := llm.ImagePart{MIME: "image/png", Data: []byte("PNGDATA")}
+	inv := tool.Invocation{RunID: "r1", Images: []llm.ImagePart{img}}
+	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, "describe this")
+	if res.Err != nil {
+		t.Fatalf("run error: %v", res.Err)
+	}
+
+	calls := fp.Calls()
+	if len(calls) == 0 {
+		t.Fatal("no model calls recorded")
+	}
+	// The text + image must be CO-LOCATED in a single user message (not split
+	// across two), so the model reads them as one multimodal turn.
+	coLocated := false
+	for _, msg := range calls[0].Request.Messages {
+		sawImage, sawText := false, false
+		for _, p := range msg.Parts {
+			switch pp := p.(type) {
+			case llm.ImagePart:
+				if string(pp.Data) == "PNGDATA" {
+					sawImage = true
+				}
+			case llm.TextPart:
+				if strings.Contains(pp.Text, "describe this") {
+					sawText = true
+				}
+			}
+		}
+		if sawImage && sawText {
+			coLocated = true
+		}
+	}
+	if !coLocated {
+		t.Error("image + prompt text were not folded into the SAME user message")
+	}
+}
+
+// TestExecutorImageOnlyNoBlankText: an image-only run (blank prompt) must NOT emit
+// an empty TextPart — the message carries just the image, matching
+// runSession.AttachImages's guard.
+func TestExecutorImageOnlyNoBlankText(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m", fake.Reply("saw it"))
+	m, _ := fp.Model("m")
+
+	inv := tool.Invocation{RunID: "r3", Images: []llm.ImagePart{{MIME: "image/png", Data: []byte("IMG")}}}
+	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, "   ")
+	if res.Err != nil {
+		t.Fatalf("run error: %v", res.Err)
+	}
+	for _, msg := range fp.Calls()[0].Request.Messages {
+		for _, p := range msg.Parts {
+			if tp, ok := p.(llm.TextPart); ok && strings.TrimSpace(tp.Text) == "" {
+				t.Error("image-only run emitted a blank TextPart")
+			}
+		}
+	}
+}
+
+// TestExecutorTextOnlyUnchanged: with no Images, the prompt flows through as the
+// text input (regression guard that the fold path didn't break the common case).
+func TestExecutorTextOnlyUnchanged(t *testing.T) {
+	fp := fake.New("fake")
+	fp.Enqueue("m", fake.Reply("ok"))
+	m, _ := fp.Model("m")
+
+	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"}, tool.Invocation{RunID: "r2"}, "plain prompt")
+	if res.Err != nil {
+		t.Fatalf("run error: %v", res.Err)
+	}
+	calls := fp.Calls()
+	if len(calls) == 0 {
+		t.Fatal("no model calls recorded")
+	}
+	sawText := false
+	for _, msg := range calls[0].Request.Messages {
+		for _, p := range msg.Parts {
+			if tp, ok := p.(llm.TextPart); ok && strings.Contains(tp.Text, "plain prompt") {
+				sawText = true
+			}
+		}
+	}
+	if !sawText {
+		t.Error("text-only prompt did not reach the model")
+	}
+}