[sangilish]

Summary

• Honor explicit requires_tool gates before a stateful cascade in ordinary single-agent execution, matching the routed path.
• Keep controller-specific skip reasons and the existing called-tool advertisement check.
• Add nine focused regression/control cases for same-phase and downstream consumers, supported execution, real failures, non-stateful parity, controller diagnostics, and zero-coverage protection.

Refs adcontextprotocol/adcp#8506.

Boundary

This is not a blanket conversion of missing-tool cascades into passing evidence. A genuinely failed prerequisite still fails its consumers when their declared tools are available, an earlier real failure still fails the storyboard, and an entirely tool-gated run cannot pass without executed coverage.

The paired storyboard source change must declare sync_creatives on the dependent provenance audit reads. The runner change alone does not repair the existing controller-only declarations. This also does not claim to resolve the broader fixture, hosted-diagnostic and coverage issues in adcontextprotocol/adcp#8576.

Validation

• npm run build:lib
• npm run typecheck
• Nine focused cascade, skip-counting, requires, capability and routed suites: 349 tests passed.
• Prettier check passed.
• ESLint: zero errors; 24 warnings, also present on the unmodified runner at the base revision.
• Changeset status, PR title validation and git diff --check passed.

The full monorepo suite and hosted grader were not run. Includes an SDK patch changeset; no protocol wire shape or generated release artifact is changed.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-30T03:18:56.771444Z We couldn’t complete this review. Comment @codex review to try again.	ba0c26b	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[aao-secretariat]

Ladon verdict: Approve

Approve — clean patch with no blocking findings.

Checked: drops the routingContext && guard at runner.ts:6066 so a cascade-tripped stateful step honors its explicit requires_tool gate in single-agent runs, matching the routed and non-cascade paths (runner.ts:8668, same skip-reason mapping). The missingReason change preserves the controller-specific reason via step.requires_tool === missingTool. Boundaries hold — real failures still cascade when the declared tool is advertised, earlier real failures still fail the run, and zero executed coverage still fails. Nine regression tests cover same-phase/downstream/supported/failure/non-stateful/controller/zero-coverage branches. Changeset is a patch and appropriate — no wire or export change.

No critical/high/medium findings. gated_paths is false, high_risk is false, no author team gate. review_decision is REVIEW_REQUIRED but the gated-paths hard gate (row 2) only fires when gated_paths is true, which it is not. Falls through to row 9.