ARCH-008: dependency.py workflow completion check rejects completed_with_failures

[Muizzkolapo]

Summary

workflow/parallel/dependency.py:167 checks upstream workflow completion using strict equality:

return all(details.get("status") == "completed" for details in status_data.values())

This returns False when any upstream action has completed_with_failures status, blocking downstream execution. Since batch actions never reach completed_with_failures (see ARCH-007), this creates an asymmetry: online workflows with partial failures are blocked, batch workflows with the same failures proceed freely.

Impact

• Online: An upstream action that processes 99/100 items successfully gets completed_with_failures. Downstream workflows that depend on it are blocked — even though 99 good records are available.
• Batch: The same scenario gets completed, and downstream proceeds.
• Net effect: Online workflows are more fragile than batch for the same failure rate. This is backwards — online should have the advantage of immediate visibility, not be penalized for it.

Suggested Fix

return all(
    details.get("status") in {"completed", "completed_with_failures"}
    for details in status_data.values()
)

This matches the pattern already used in state.py:128-133 (is_workflow_complete()) and state.py:136-138 (is_workflow_done()).

Related

• ARCH-007 (ARCH-007: Batch completion path never detects partial item failures (completed_with_failures is online-only) #89) — batch path never sets completed_with_failures (masks this bug for batch)
• ARCH-004 (ARCH-004: Formalize action state machine to prevent mode-specific state leaks #86) — formalized state machine would make terminal/complete sets consistent

Files Affected

• agent_actions/workflow/parallel/dependency.py:167

Acceptance Criteria

• Upstream actions with completed_with_failures do not block downstream workflow execution
• Behavior is symmetric between online and batch for the same failure pattern

[Muizzkolapo]

Heads-up — bundle with Phase 2 PR

This is a one-line fix that should ship alongside the Phase 2 completed_with_failures work currently in progress. Without it, online actions that complete with partial failures will block downstream workflow execution at dependency.py:167, while batch actions (which never reach this status — see #89) proceed freely.

# dependency.py:167 — change from:
return all(details.get("status") == "completed" for details in status_data.values())

# to:
return all(
    details.get("status") in {"completed", "completed_with_failures"}
    for details in status_data.values()
)

This matches the pattern already in state.py:128-133 and state.py:136-138.

[Muizzkolapo]

Fixed in PR #92 — dependency.py:167 now uses in {"completed", "completed_with_failures"}.