Stabilize swarm activation control lifecycle

docs/logs/engineering-log.md

@@ -124,6 +124,20 @@
 - Result: the focused test passed normal/race at `-count=100`, the complete
   harness package passed normal/race, and `./scripts/test-regression.sh` passed
   with 85.6% total coverage and zero uncovered functions.
+
+## 2026-07-30 (Swarm Activation Control Lifecycle Race — Issue #1046)
+
+- Symptom: hosted fast CI intermittently reported `agent_swarm` missing from an
+  unrestricted run immediately after test-local activation.
+- Cause: the control reused an exhausted scripted provider, so terminal cleanup
+  could clear the activation before the fixture inspected definitions.
+- Planned fix: use a dedicated provider that blocks until after the definition
+  assertion, then release it and wait for normal terminal cleanup.
+- Verification contract: focused normal/race stress, harness normal/race, full
+  regression, and GitHub required checks.
+- Result: the focused test passed normal/race at `-count=100`, the complete
+  harness package passed normal/race, and `./scripts/test-regression.sh` passed
+  with 85.6% total coverage and zero uncovered functions.
 ## 2026-07-30 (Worktree Containment CI Synchronization — Issue #1039)
 
 - Symptom: GitHub Actions fast run 30551198514 received

docs/logs/long-term-thinking-log.md

@@ -148,6 +148,19 @@
   pinned states remain temporary exceptions, subscriber cancellation still
   triggers cleanup, and all focused through full gates are green.
 
+## 2026-07-30 (Swarm Activation Control Lifecycle Race — Issue #1046)
+
+- Command intent: clear the hosted baseline flake blocking the cron/callback
+  merge chain.
+- User intent: prove all accepted gates are green before merging to `main`.
+- Success definition: the test deterministically exercises both a denied member
+  and a live unrestricted control, then cleans up normally.
+- Guardrails: issue-first isolated worktree, test-only lifecycle control, no
+  production activation or Runner change.
+- Outcome: the dedicated provider holds the actual control run live through
+  activation inspection, then releases it to normal completion; focused,
+  package, race, and coverage gates are green.
+
 ## 2026-07-30 (Workflow Subscription Cancellation Test — Issue #1035)
 
 - Command intent: clear the race-gate blocker exposed while verifying the

docs/plans/2026-07-30-issue-1046-swarm-activation-control-impact-map.md

@@ -0,0 +1,67 @@
+# Cross-Surface Impact Map: Issue #1046 Swarm Activation Control
+
+## Task
+
+- Task / issue: stabilize the unrestricted swarm activation control, #1046.
+- Plan: `2026-07-30-issue-1046-swarm-activation-control-plan.md`.
+- Owner: Codex.
+- Status: implemented and fully verified locally; hosted checks pending.
+
+## Current Ownership, Callers, and Data Flow
+
+- Entry: `TestAgentSwarmDeniedForMemberRuns`.
+- Production source of truth: `ActivationTracker` owns per-run activations;
+  terminal Runner paths call `Cleanup(runID)`.
+- Test flow: the member run exhausts `capturingProvider.turns`; the following
+  unrestricted run can complete immediately, racing activation inspection.
+- Search conclusion: adjacent swarm tests do not reuse an exhausted provider
+  for a live activation control.
+
+## Config, API, CLI, and Tools
+
+- Config/env/defaults: none.
+- API/CLI/wire/tool behavior: unchanged.
+- Tool contract: keep `agent_swarm` hidden for denied members and visible after
+  activation for unrestricted runs.
+
+## Persistence and Compatibility
+
+- State: test-local release channel only.
+- Schemas/migrations/caches: none.
+- Compatibility: no runtime change.
+
+## Lifecycle, Security, and Reliability
+
+- Concurrency: block the control provider until after definitions are read.
+- Cleanup: release and wait for terminal status; Runner cleanup remains intact.
+- Auth/privacy/secrets: none.
+
+## Product and Integration Surfaces
+
+- Harness production, API, TUI, web, macOS GUI, and providers: no code change.
+- Automation: fast and race-enabled CI become repeatable.
+- UX/accessibility: none.
+
+## Deployment and Operations
+
+- Deployment/migration/flags: none.
+- Rollback: revert if the control no longer runs through normal Runner paths.
+- Observability/operator docs: none.
+
+## Regression Tests
+
+- Red: hosted missing-definition failure after terminal cleanup.
+- Green: focused normal/race `-count=100`.
+- Controls: denied member remains hidden and uncallable; live unrestricted run
+  sees the activated definition; released control reaches terminal state.
+- Full: harness normal/race and `./scripts/test-regression.sh`.
+
+## Documentation and Handoff
+
+- Update engineering log, long-term log, plan, impact map, and plans index.
+- No public docs.
+
+## Warning Check
+
+- Every runtime/product surface is explicitly unaffected because the repair is
+  confined to the test provider lifecycle.

docs/plans/2026-07-30-issue-1046-swarm-activation-control-plan.md

@@ -0,0 +1,60 @@
+# Plan: Keep the Swarm Activation Control Run Live
+
+## Context
+
+- Governing GitHub issue: #1046.
+- Problem: `TestAgentSwarmDeniedForMemberRuns` reuses an exhausted scripted
+  provider for its unrestricted control. The control run can complete and clear
+  its activation between `Activate` and `filteredToolsForRun`.
+- Red evidence: hosted fast CI failed at `runner_swarm_test.go:242` with
+  `agent_swarm missing from unrestricted run definitions after activation`.
+- Constraint: preserve production activation cleanup and both sides of the
+  denied/unrestricted contract.
+
+## Scope
+
+- In scope: a dedicated, test-local blocking provider for the control run.
+- Out of scope: production Runner lifecycle, activation tracking, agent swarm,
+  permissions, or deferred-tool filtering.
+
+## Documentation Contract
+
+- Feature status: test-only bug repair.
+- Public docs affected: none.
+- Evidence: engineering and long-term logs, plan, impact map, plans index.
+
+## Test Plan
+
+- Red: retain the exact hosted lifecycle-race failure as regression evidence.
+- Green: focused normal and race tests at `-count=100`.
+- Package: `internal/harness` normal and race.
+- Full: `./scripts/test-regression.sh` and GitHub required checks.
+
+## Cross-Surface Impact Map
+
+- See `2026-07-30-issue-1046-swarm-activation-control-impact-map.md`.
+
+## Implementation Checklist
+
+- [x] Create contract-complete bug #1046.
+- [x] Capture hosted failure and architecture search evidence.
+- [x] Write plan and impact map before code.
+- [x] Hold the unrestricted control run live through its assertion.
+- [x] Run focused stress, package, and full local gates.
+- [ ] Pass hosted required checks.
+- [ ] Merge through a closing PR.
+
+## Verification
+
+- Focused normal and race tests passed at `-count=100`.
+- The complete `internal/harness` package passed normal and race tests.
+- `./scripts/test-regression.sh` passed normal, race, and the 85.6% coverage
+  gate with zero uncovered functions.
+
+## Risks and Mitigations
+
+- Risk: the control provider could leak when an assertion fails.
+- Mitigation: register release and shutdown cleanup before starting the run.
+- Risk: the control could stop proving a real unrestricted run.
+- Mitigation: start a separate Runner normally, activate its actual run ID, and
+  assert through `filteredToolsForRun`.

docs/plans/INDEX.md

@@ -20,6 +20,8 @@
 - `2026-07-30-issue-1042-default-registry-repeat-impact-map.md` — Cross-surface impact map for Issue #1042.
 - `2026-07-30-issue-1048-pinned-retention-quota-plan.md` — Issue #1048 planned exclusion of subscriber-pinned runs from the drainable terminal retention quota.
 - `2026-07-30-issue-1048-pinned-retention-quota-impact-map.md` — Cross-surface impact map for Issue #1048.
+- `2026-07-30-issue-1046-swarm-activation-control-plan.md` — Issue #1046 planned live unrestricted control for the swarm activation regression.
+- `2026-07-30-issue-1046-swarm-activation-control-impact-map.md` — Cross-surface impact map for Issue #1046.
 - `2026-07-30-issue-1035-workflow-subscription-cancel-plan.md` — Issue #1035 planned deterministic workflow subscription-cancellation regression repair.
 - `2026-07-30-issue-1035-workflow-subscription-cancel-impact-map.md` — Cross-surface impact map for Issue #1035.
 - `2026-07-30-issue-1026-feedback-direct-publish-plan.md` — Issue #1026 attached-image, publish-by-default GitHub feedback intake (implemented and verified; promotion pending).

internal/harness/runner_swarm_test.go

@@ -227,21 +227,43 @@ func TestAgentSwarmDeniedForMemberRuns(t *testing.T) {
 	}
 
 	// Control: an unrestricted run sees the activated definition.
-	run2, err := runner.StartRun(RunRequest{Prompt: "parent"})
+	releaseControl := make(chan struct{})
+	var releaseControlOnce sync.Once
+	release := func() {
+		releaseControlOnce.Do(func() { close(releaseControl) })
+	}
+	t.Cleanup(release)
+	controlProvider := &funcProvider{fn: func(ctx context.Context, _ CompletionRequest) (CompletionResult, error) {
+		select {
+		case <-releaseControl:
+			return CompletionResult{Content: "done"}, nil
+		case <-ctx.Done():
+			return CompletionResult{}, ctx.Err()
+		}
+	}}
+	controlRunner := NewRunner(controlProvider, registry, RunnerConfig{
+		DefaultModel:        "test",
+		DefaultSystemPrompt: "sys",
+		Activations:         activations,
+	})
+	run2, err := controlRunner.StartRun(RunRequest{Prompt: "parent"})
 	if err != nil {
 		t.Fatal(err)
 	}
 	activations.Activate(run2.ID, "agent_swarm")
 	seen := false
-	for _, def := range runner.filteredToolsForRun(run2.ID) {
+	for _, def := range controlRunner.filteredToolsForRun(run2.ID) {
 		if def.Name == "agent_swarm" {
 			seen = true
 		}
 	}
 	if !seen {
 		t.Fatal("agent_swarm missing from unrestricted run definitions after activation")
 	}
-	waitForStatus(t, runner, run2.ID, RunStatusCompleted, RunStatusFailed)
+	release()
+	if got := waitForStatus(t, controlRunner, run2.ID, RunStatusCompleted, RunStatusFailed); got != RunStatusCompleted {
+		t.Fatalf("control run status = %q, want completed", got)
+	}
 }
 
 func TestAgentSwarmApprovalFlowSurfacesMutatingCall(t *testing.T) {