artifact_exists can never pass for a dot-directory path: the worker's journaled artifacts list omits them

[khaliqgant]

Summary

An artifact_exists gate can never pass for a file under a dot-directory, because the worker's journaled artifacts list omits those paths entirely. The gate is correct by its own rules — it reads the journal — but the journal does not describe what the agent actually wrote.

.workflow-artifacts/ is the conventional artifact directory in AgentWorkforce/relay (every flow spec in that repo writes there), so in practice artifact_exists is unusable for the thing it exists to check.

Evidence

Run 01M2ZQP6P9ME3PRBWDXQA32MT3, CLI 2.0.22, macOS arm64, flows run --local-agent.

The run died at shadow-rust.gate with the file it was gating sitting on disk:

FAILED [step_failed] Step "shadow-rust.gate" (deterministic) completionReason: retries_exhausted exit=1.

$ ls -la .workflow-artifacts/migrate-native-delivery/phase-0-seam-20260920c/reviews/shadow-rust.md
-rw-r--r--  1 khaliqgant  staff  25257 Sep 20 08:57 ...

The gate was { type: 'artifact_exists', path: '.workflow-artifacts/migrate-native-delivery/phase-0-seam-20260920c/reviews/shadow-rust.md' }. shadow-rust journaled zero artifacts.

The preceding step makes the pattern unambiguous. implement-rust journaled 6,837 artifacts:

top-level prefixes: {'target': 6832, 'crates': 5}

target/ and crates/ only — and nothing under .workflow-artifacts/, despite that step provably having written evidence/mutation-proof.md there:

$ stat -f "%Sm %N" -t "%H:%M" .workflow-artifacts/.../evidence/mutation-proof.md
08:48 .workflow-artifacts/.../evidence/mutation-proof.md

So the exclusion is not .gitignore-driven: target/ is gitignored too and is included in full. The distinguishing property of the missing path is the leading dot.

Why it matters

artifact_exists is documented as the journal-honest way to check "the agent wrote X", and it is the gate an author naturally reaches for on a review or report step. Silently excluding a whole class of paths means:

• the gate fails on a file that exists, with no indication why;
• the failure is a bare retries_exhausted exit=1 (the lowered gate decides in-process, so there is no command output to inspect either);
• agent-written evidence is invisible in the journal generally, not just to this gate — any consumer reading output.artifacts to find what a step produced will miss it.

What to change

Either include dot-directories in the worker's artifact scan, or — if they are excluded deliberately, which is defensible for .git/, .relayflowd/ and friends — make the exclusion explicit and inspectable:

• document which prefixes are excluded;
• have flows check refuse an artifact_exists gate whose path falls inside an excluded prefix, with a message naming the exclusion, rather than compiling a gate that cannot pass;
• consider an opt-in (artifacts.include) so a flow can declare its own artifact directory.

A deny-list of exactly .git/ and the runtime's own state directories would cover the real hazard without swallowing every author-chosen artifact path.

Acceptance

• Either an agent writing .workflow-artifacts/x/y.md has that path in its journaled artifacts, or flows check refuses the artifact_exists gate naming it, before the run starts.
• A test pins a dot-directory artifact path either way.

Out of scope

• Whether artifact_exists should consult the disk (it should not; journal-honest replay is the right design).
• The unrelated subprocess_gate stdio loss (A failing subprocess_gate journals empty stdout/stderr: the lowering uses stdio:'inherit' and the daemon's stdio is captured nowhere #511), though the two combine badly: neither gate mechanism can tell an author why it failed.