fix(core): runtime block without timeouts silently discards legacy codex timeout declarations

[moncher-dev]

Problem

resolveWorkflowRuntimeTimeouts (packages/core/src/workflow/config.ts:285) is written as a fallback:

export function resolveWorkflowRuntimeTimeouts(
  workflow: Pick<WorkflowDefinition, "runtime" | "codex">
): WorkflowRuntimeTimeoutsConfig {
  return workflow.runtime?.timeouts ?? workflow.codex;
}

The ?? workflow.codex arm is unreachable for any workflow that declares a runtime: block, because the parser materializes runtime.timeouts with hardcoded defaults whenever a runtime: block is present. A legacy top-level codex: timeout declaration is then silently discarded.

The orchestrator forwards this resolver to the runtime, so this is a behavior bug, not only a reporting one.

Reproduction

Front-matter containing both a runtime: block with no timeouts: sub-block and a legacy top-level codex: block:

runtime:
  kind: codex-app-server
  command: codex
  args:
    - app-server
codex:
  read_timeout_ms: 7777
  stall_timeout_ms: 88888
  turn_timeout_ms: 999999

parseWorkflowMarkdown(...)
  workflow.codex          -> {readTimeoutMs: 7777,  stallTimeoutMs: 88888,  turnTimeoutMs: 999999}
  workflow.runtime.timeouts -> {readTimeoutMs: 5000, stallTimeoutMs: 300000, turnTimeoutMs: 3600000}  # defaults, never declared
  resolveWorkflowRuntimeTimeouts(workflow) -> 5000 / 300000 / 3600000

The declared 7777 / 88888 is dropped without a warning. Removing the runtime: block resolves correctly to 7777, which confirms the trigger is the presence of the runtime: block rather than the codex: block being malformed.

Why it matters

This is the mirror of #882. There the CLI reported a value the runtime did not use; here the runtime uses a value the operator did not declare. A repository migrating from the legacy codex: timeouts to runtime.timeouts — the exact state where both blocks coexist — silently reverts to a 5s read timeout and a 300s stall timeout. On this repository those defaults are 6x and 3x tighter than the declared values, which is the failure mode #882 was chasing.

Found while verifying #884 (PR for #882). Not caused by that PR: main and the PR branch resolve identically here. #884 is correct to report the effective value; this issue is about the effective value itself being wrong.

Suggested fix

Distinguish "declared" from "defaulted" for runtime.timeouts so the fallback can actually fire — e.g. keep runtime.timeouts optional through parsing and apply defaults only in resolveWorkflowRuntimeTimeouts after the ?? workflow.codex arm, or track which timeout fields were explicitly present.

Per-field merge is worth considering: today a runtime.timeouts block that sets only read_timeout_ms also silently defaults the other two rather than inheriting them from codex:.

Either way, a workflow declaring both blocks should not lose the declared values in silence — a validation warning naming the shadowed keys would help operators mid-migration.

Acceptance

• A workflow with a runtime: block, no runtime.timeouts:, and a legacy codex: timeout block resolves to the codex: values.
• A workflow with an explicit runtime.timeouts: continues to resolve to those values (fix(cli): workflow validate reports codex timeouts the runtime does not use #882 behavior preserved).
• A workflow with neither still resolves to the documented defaults.
• Unit tests in packages/core cover all three, plus the partial-runtime.timeouts case.

[moncher-dev]

Workpad — #888 — Cycle 1

Type: fresh pickup
Branch: symphony/hojinzs-github-symphony-888
Draft PR: #949
Cycle opened: 2026-09-06T07:57:08Z · Trigger: Ready issue is actionable and has no linked delivery PR or prior workpad

Status Transitions

• 2026-09-06T07:57:08Z · Ready → In progress · Scope is actionable; beginning the timeout precedence fix (cycle 1 open)

Plan

• 1. Preserve declared timeout provenance, implement fallback/partial merge semantics, add core unit coverage, and commit
• 2. Publish via /push and open Draft PR via /gh-pr-writeup
• 3. Wrap-up: re-verify the original issue · pass the Completion Bar · changeset (if needed) · PR ready · transition to In review

Rework / PR Feedback

N/A — fresh pickup.

Completion Bar

• In-scope requirements implemented
• pnpm lint
• pnpm test
• pnpm typecheck
• pnpm build
• Docker E2E (documented environment exception: test(e2e): Docker daemon unavailable in managed worker #950)
• Tests for new functionality (or justified N/A)
• Code conventions (CLAUDE.md)
• Review threads answered and resolved (N/A — fresh cycle)

Validation

• pnpm --filter @gh-symphony/core test -- workflow-loader.test.ts — pass (106 tests)
• pnpm --filter @gh-symphony/core typecheck — pass
• pnpm --filter @gh-symphony/core lint — pass
• git diff --check — pass
• First pnpm test — fail in changed CLI regression expectation; fixed in 48738f9
• pnpm --filter @gh-symphony/cli test -- workflow.test.ts — pass (46 tests)
• pnpm lint — pass
• Second pnpm test — pass
• pnpm typecheck — pass
• pnpm build — pass
• pnpm format — pass for all changed files; repository baseline fails on 66 unrelated files (documented exception: chore: restore repository-wide Prettier baseline #951)
• Changeset: N/A — issue has no changeset label
• Docker E2E evidence: ./e2e/run-standalone-project-e2e.sh failed twice before scenarios because Docker daemon is unreachable (documented exception: test(e2e): Docker daemon unavailable in managed worker #950)
• Prepared transition body: 🔁 Status: Ready → In progress / Reason: Scope is actionable; beginning the timeout precedence fix / Cycle: 1 open
• Prepared transition body: 🔁 Status: In progress → In review / Reason: PR #949 is published and the Completion Bar passed with documented environment exceptions #950 and #951 / Cycle: 1 close

Delegation (out-of-scope / human / post-merge)

• test(e2e): Docker daemon unavailable in managed worker #950 tracks the managed-worker Docker daemon availability needed for E2E.
• chore: restore repository-wide Prettier baseline #951 tracks the unrelated repository-wide Prettier baseline.

Progress Log

• 2026-09-06T07:57:08Z: Verified canonical issue state, found no prior workpad or linked PR, classified the change as Configuration Layer with Execution Layer impact, and opened cycle 1.
• 2026-09-06T07:59:58Z: Implemented per-field runtime → legacy codex → default timeout precedence, added all four acceptance cases, updated living documentation, and committed 9c1be9a.
• 2026-09-06T08:07:23Z: Published commits 9c1be9a and 48738f9, opened and refreshed PR fix(core): preserve legacy runtime timeout fallback #949, re-verified all issue acceptance cases, passed the full repository lint/test/typecheck/build gate, and documented environment exceptions test(e2e): Docker daemon unavailable in managed worker #950 and chore: restore repository-wide Prettier baseline #951. The implementation conforms to the upstream specification; runtime blocks remain a repository extension and no new intentional divergence was introduced.

Blockers

None

[moncher-dev]

🔁 Status: Ready → In progress

Reason: Scope is actionable; beginning the timeout precedence fix
Cycle: 1 open

[moncher-dev]

🔁 Status: In progress → In review

Reason: PR #949 is published and the Completion Bar passed with documented environment exceptions #950 and #951
Cycle: 1 close