Compiler silently drops jobs.agent.continue-on-error

[HardMax71]

jobs.agent.continue-on-error: true passes compilation without warnings but is absent from the generated agent job. This is silent configuration loss, independently of whether error tolerance should be supported for that built-in job.

Reproduction

Reproduced with gh aw version v0.86.2. Save .github/workflows/ignored-error-policy.md:

---
on: workflow_dispatch
permissions:
  contents: read
engine: copilot
network:
  allowed: [defaults]
jobs:
  agent:
    continue-on-error: true
---

Report the repository name. Do not change files or post to GitHub.

Run gh aw compile ignored-error-policy --no-check-update. It reports one successful compilation and zero warnings. Inspect the job-level fields under jobs.agent in the generated lock file: continue-on-error is missing. Existing step-level occurrences do not satisfy the requested job-level setting. No agent execution or provider credential is needed to reproduce this.

Source analysis

Cross-checked source at main commit ebfaf2ac2b0c637e1ff858168544f08b983c05ee (source inspection, not a build of that revision):

• The frontmatter schema accepts boolean jobs.*.continue-on-error.
• Built-in job augmentation tracks needs, conditions, permissions, and timeout presence, but not this field. Its empty-augmentation check therefore ignores a configuration containing only continue-on-error.
• Custom-job handling already has extractCustomJobContinueOnError and stores the value in Job.ContinueOnError.

Proposed agentic plan

1. In pkg/workflow/compiler_builtin_job_augmentation.go, explicitly handle the presence of continue-on-error. For supported built-in jobs, reuse Job.ContinueOnError and its renderer. Include field presence in the empty-augmentation guard so explicit false is handled too.
2. For intentionally unsupported built-in jobs, reject the field with an actionable compile error. Preserve activation/security prerequisites and current failure defaults; do not broadly enable error tolerance on generated jobs.
3. Add compiler regressions for agent true, explicit false, omitted and imported configuration, and unaffected custom-job handling. Assert parsed job-level YAML, not a global text match that could match step settings.
4. Document supported built-in jobs in docs/src/content/docs/reference/steps-jobs.md. Follow the developer validation flow: make build, make fmt, and final make agent-report-progress with compiler/workflow drift checks.

Acceptance: the reproducer either emits the requested native job-level property or fails clearly because it is unsupported; it must not succeed silently while dropping the setting.

Related behavior and scope

The motivating downstream run retried 402 Insufficient Balance four times and failed its optional reviewer. A native, provider-independent policy would be preferable to balance preflight calls or local error-parsing scripts.

#39581 addresses Copilot-specific quota retries; #49415 addresses detection error tolerance. Neither covers this compiler omission. #59316 deliberately fails incomplete runs, so this report does not ask to make report_incomplete silently successful. Native job-level continue-on-error is not equivalent to a neutral/skipped review, nor does it selectively recognize billing errors; that broader policy is separate.

Searched existing issues and PRs again before filing and found no exact duplicate. Suggested maintainer labels: bug, workflow.

Analysis and reproduction prepared with a coding agent, following the community issue-plan contribution process. No PR submitted.