fix(ship): agent-mode qa-verify no longer runs the host test suite; arch debate runs serial

CHANGELOG.md

@@ -4,6 +4,11 @@ All notable changes to forge will be documented in this file. Format follows [Ke
 
 ## [Unreleased]
 
+### Fixed
+
+- **`forge ship --agent-mode` could run the host project's entire native test suite itself and hang for many minutes with zero output.** The `qa-verify` checkpoint's Phase 2 called `runQATestSuite`, which shells out to `npm test` / `go test ./...` / `pytest` on the target repo — with no agent-mode guard. On a large host project that is a multi-minute blocking run with its own output stream, and forge's `procspawn` timeout does not reliably reap a killed runner's orphaned worker processes on Windows (jest/vitest workers keep the stdout pipe open, so the read never sees EOF). When the pipeline reached `qa-verify` without pausing first — e.g. after `checkArch` short-circuits on an already-present `arch.md` from an earlier answered turn — the result was `forge ship --agent-mode` silently executing the caller's full test suite and appearing to hang. In agent mode the host agent *is* the QA agent (the checkpoint's own description: "QA agent: probe MCP server tools or run native test suite"), so `qa-verify` now emits an advisory telling the host agent to run and report the suite instead of running it itself. Non-agent-mode runs are unchanged.
+- **`checkArch`'s parallel role debate still fanned out six goroutines against the agent bridge even though the bridge can only surface one pending turn per run.** The 1.10.4 mutex stopped the data race but not the pile-up: five of the six goroutines do throwaway work that is redone on the next replay, while serialising behind `Bridge.mu` as the first goroutine holds it across `savePending`'s file I/O — a contributing shape behind the "hangs with zero output mid-arch-debate" reports. In agent mode (`pipe.Bridge() != nil`) the debate now runs sequentially and stops at the first owed turn, matching `RunWithOptions`'s existing `serial` intent; the parallel path is kept for real-provider runs. New regression test `TestRunParallelArchDebate_AgentModeSerialNoFanOut`.
+
 ## [1.10.4] — 2026-09-05 — LLM-pipeline auto-fallback, a critical wrong-feature resume, an arch-debate data race, and dry-run leaks
 
 ### Fixed

internal/cli/cmdship/arch.go

@@ -57,9 +57,21 @@ func defaultArchRoles() []archRoleDebate {
 	}
 }
 
-// runParallelArchDebate concurrently invokes all arch reviewer roles and
-// collects their concerns. Results are appended to the arch document as a
-// "## Reviewer Concerns" section. A nil pipe is a no-op.
+// runParallelArchDebate invokes all arch reviewer roles and collects their
+// concerns. Results are appended to the arch document as a "## Reviewer
+// Concerns" section. A nil pipe is a no-op.
+//
+// Concurrency: the roles run in parallel goroutines only when the reasoning
+// plane is a real provider. In agent mode (pipe.Bridge() != nil) they run
+// sequentially and stop at the first pause. Fanning out against the bridge
+// there is pointless and unsafe: the bridge can only surface one pending turn
+// per run, so every goroutine after the first miss does throwaway work that is
+// redone on the next replay anyway — and it serialises six goroutines behind
+// Bridge.mu while the first holds it across file I/O in savePending(), the
+// exact shape behind the "forge ship --agent-mode hangs with zero output
+// mid-arch-debate" reports (the 1.10.4 mutex stopped the data race but not the
+// pile-up). Sequential + early-exit removes the hazard at the source and
+// matches RunWithOptions's own `serial` intent for agent mode.
 func runParallelArchDebate(pipe *LLMPipe, description, archDoc string, maxTokens int) string {
 	if pipe == nil {
 		return ""
@@ -70,27 +82,42 @@ func runParallelArchDebate(pipe *LLMPipe, description, archDoc string, maxTokens
 		concern string
 	}
 	results := make([]result, len(roles))
-	var wg sync.WaitGroup
-
-	for i, role := range roles {
-		wg.Add(1)
-		go func(idx int, r archRoleDebate) {
-			defer wg.Done()
-			concern, err := pipe.InvokeDebateRound(
-				"arch-parallel-debate",
-				description,
-				r.persona,
-				archDoc,
-				"",
-				maxTokens,
-			)
-			if err != nil || strings.TrimSpace(concern) == "" {
-				concern = "(no concerns raised)"
+
+	askRole := func(idx int, r archRoleDebate) {
+		concern, err := pipe.InvokeDebateRound(
+			"arch-parallel-debate",
+			description,
+			r.persona,
+			archDoc,
+			"",
+			maxTokens,
+		)
+		if err != nil || strings.TrimSpace(concern) == "" {
+			concern = "(no concerns raised)"
+		}
+		results[idx] = result{name: r.name, concern: concern}
+	}
+
+	if bridge := pipe.Bridge(); bridge != nil {
+		// Agent mode: sequential, and stop as soon as a turn is owed.
+		for i, role := range roles {
+			if bridge.Paused() {
+				results[i] = result{name: role.name, concern: "(no concerns raised)"}
+				continue
 			}
-			results[idx] = result{name: r.name, concern: concern}
-		}(i, role)
+			askRole(i, role)
+		}
+	} else {
+		var wg sync.WaitGroup
+		for i, role := range roles {
+			wg.Add(1)
+			go func(idx int, r archRoleDebate) {
+				defer wg.Done()
+				askRole(idx, r)
+			}(i, role)
+		}
+		wg.Wait()
 	}
-	wg.Wait()
 
 	var sb strings.Builder
 	sb.WriteString("\n\n## Reviewer Concerns (parallel debate)\n\n")

internal/cli/cmdship/rfc005_p1p2_test.go

@@ -19,6 +19,9 @@ import (
 	"path/filepath"
 	"strings"
 	"testing"
+	"time"
+
+	"github.com/teragrid/forge/internal/agentbridge"
 )
 
 // ── snapshot.go ──────────────────────────────────────────────────────────────
@@ -593,6 +596,58 @@ func TestRunParallelArchDebate_EmptyDocNoPanic(t *testing.T) {
 	_ = runParallelArchDebate(nil, "feat", "", 100)
 }
 
+// TestRunParallelArchDebate_AgentModeSerialNoFanOut is a regression test for the
+// recurring "forge ship --agent-mode hangs with zero output mid-arch-debate"
+// reports (docs/plans/FORGE_SHIP_ISSUES_2026-09-04.md ISSUE 5). The 1.10.4 fix
+// added a mutex to Bridge, which stopped the data race but not the underlying
+// hazard: six goroutines still fanned out against a bridge that can only ever
+// surface one pending turn, piling up behind Bridge.mu while the first holds it
+// across savePending()'s file I/O. In agent mode the debate must instead run
+// sequentially and stop at the first owed turn.
+func TestRunParallelArchDebate_AgentModeSerialNoFanOut(t *testing.T) {
+	t.Parallel()
+	root := t.TempDir()
+	bridge, err := agentbridge.Open(root, agentbridge.DefaultSession)
+	if err != nil {
+		t.Fatalf("open bridge: %v", err)
+	}
+	pipe := newLLMPipeAgent(root, bridge)
+
+	done := make(chan string, 1)
+	go func() { done <- runParallelArchDebate(pipe, "feat", "# Arch Doc", 300) }()
+
+	var out string
+	select {
+	case out = <-done:
+	case <-time.After(10 * time.Second):
+		t.Fatal("runParallelArchDebate hung in agent mode — it must run sequentially and return at the first owed turn")
+	}
+
+	// Every role still appears in the appended section (unanswered ones as
+	// placeholders), so the arch document shape is unchanged.
+	if !strings.Contains(out, "## Reviewer Concerns (parallel debate)") {
+		t.Fatalf("missing Reviewer Concerns section:\n%s", out)
+	}
+	for _, r := range defaultArchRoles() {
+		if !strings.Contains(out, "### "+r.name) {
+			t.Errorf("role %q missing from debate output", r.name)
+		}
+	}
+
+	// Exactly one turn is owed — the first role's — not six, and the bridge
+	// state is coherent (not corrupted by concurrent writes).
+	st := bridge.Stats()
+	if st.Pending == nil {
+		t.Fatal("expected one pending turn after the agent-mode debate, got none")
+	}
+	if st.Pending.Operation != "arch-parallel-debate" {
+		t.Fatalf("pending turn operation = %q, want arch-parallel-debate", st.Pending.Operation)
+	}
+	if !bridge.Paused() {
+		t.Fatal("bridge must be paused once the first role's turn is owed")
+	}
+}
+
 // ── DAG parallel pipeline ──────────────────────────────────────────────────
 //
 // Test design:

internal/cli/cmdship/ship.go

@@ -1987,6 +1987,26 @@ func checkQAVerify(root, description, specName string, pipe *LLMPipe) Checkpoint
 	}
 
 	// ── Phase 2: automated test suite ───────────────────────────────────────
+	// In agent mode the host agent *is* the QA agent — the checkpoint's own
+	// description is "QA agent: probe MCP server tools or run native test
+	// suite". forge must not shell out to `npm test` / `go test ./...` /
+	// `pytest` itself here: on a large host project that is a multi-minute
+	// blocking run with its own output stream, and forge's procspawn timeout
+	// does not reliably reap a killed runner's orphaned worker processes on
+	// Windows (jest/vitest workers keep the stdout pipe open) — the shape
+	// behind "forge ship --agent-mode hangs for 15 min with zero output"
+	// reports where the pipeline reached qa-verify without pausing. Emit an
+	// advisory instead and let the host agent run and report the suite.
+	if pipe != nil && pipe.Bridge() != nil {
+		cp.Status = "warning"
+		cp.Detail = "QA-Verify (agent mode): native test suite not run by forge — " +
+			"the host agent runs and reports it (4-stage testing pipeline, stage 1/2)"
+		if auditRes.SpecFound && len(auditRes.Gaps) > 0 {
+			cp.Detail += fmt.Sprintf("; %d spec audit warning(s)", len(auditRes.Gaps))
+		}
+		return cp
+	}
+
 	cp.Status, cp.Detail = runQATestSuite(root)
 
 	// Append spec audit warnings to detail regardless of runner.