samples mode cannot replay call-workflow, dispatch-workflow, or replace-label

[dsyme]

Summary

Three configured safe outputs cannot be exercised deterministically with --use-samples:

• call-workflow compiles with a sample, but replay returns Tool not found: call_workflow.
• dispatch-workflow compiles with a sample, but replay returns Tool not found: dispatch_workflow.
• replace-label runs no engine and performs no output when no sample is present, but adding a sample is rejected at compile time: samples: no MCP tool schema found for "replace_label".

The handlers exist in actions/setup/js/safe_output_handler_manager.cjs, but sample validation/replay does not expose schemas/tools for them consistently.

Reproductions

Call workflow:

safe-outputs:
  call-workflow:
    workflows: [worker]
    samples:
      - workflow_name: worker
        inputs: { sentinel: sample }

Dispatch workflow:

safe-outputs:
  dispatch-workflow:
    workflows: [worker]
    samples:
      - workflow_name: worker
        inputs: { sentinel: sample }

Both compile, then fail in Replay safe-outputs samples (deterministic) with JSON-RPC -32601 Tool not found.

Replace label:

safe-outputs:
  replace-label:
    allowed-transitions:
      - from: old
        to: new
    samples:
      - item_number: "${{ github.event.issue.number }}"
        label_to_remove: old
        label_to_add: new

Compilation fails with:

samples: no MCP tool schema found for "replace_label" (yaml key "replace-label"). Available tools come from pkg/workflow/js/safe_outputs_tools.json

Expected behavior

Every safe output that supports workflow configuration should either support deterministic sample replay with a schema/tool, or fail configuration with a clear, consistent unsupported-samples diagnostic. call-workflow, dispatch-workflow, and replace-label should be replayable so suites using --use-samples do not silently skip the operation or fail at runtime.

E2E evidence

gh-aw ref: main at 1bdfdfa5a0f9; samples: true; AI engine not called.

• test-copilot-call-workflow: https://github.com/githubnext/gh-aw-test/actions/runs/32550237271 — Tool not found: call_workflow (call_workflow)
• test-copilot-dispatch-workflow: https://github.com/githubnext/gh-aw-test/actions/runs/32550237779 — Tool not found: dispatch_workflow (dispatch_workflow)
• test-copilot-replace-label: https://github.com/githubnext/gh-aw-test/actions/runs/32549825512 — all jobs skipped and fixture label unchanged; adding the sample above reproduces the compiler rejection locally
• Parent E2E run: https://github.com/githubnext/gh-aw-test/actions/runs/32548557621

No matching open issue was found using searches for apply_samples, the exact Tool not found messages, and the exact missing-schema diagnostic.

[dsyme]

Reopening because the replace-label portion still reproduces on gh-aw main at commit 818e3a3 with --use-samples. test-copilot-replace-label has no usable sample because adding the canonical sample still fails compilation with samples: no MCP tool schema found for "replace_label"; without a sample, the sampled E2E leaves the fixture label unchanged. Parent: https://github.com/githubnext/gh-aw-test/actions/runs/32926231185. Recorded child https://github.com/githubnext/gh-aw-test/actions/runs/32927235241 is a skipped sibling issue event, not an executed replacement.

[dsyme]

replace-label still reproduces on gh-aw main bde2c79aa0d6 with --use-samples (AI not called). Parent: https://github.com/githubnext/gh-aw-test/actions/runs/33586188587 Child: https://github.com/githubnext/gh-aw-test/actions/runs/33587225609

At 2026-09-02T03:32:03Z, GH_AW_SAMPLES is []; apply_samples exits with nothing to replay. The run succeeds with zero safe-output messages, and fixture issue #13528 retains copilot-remove-label-test.