Summary
When the architect runs pre_check_batch while the task is in rework_required, the reducer correctly throws TASK_WORKFLOW_CODER_MUTATION_REQUIRED. Guardrails then file that error as a Stage A attribution failure and print a remediation that names a human-only command:
[guardrails] Stage A write failed for task 1.1: TASK_WORKFLOW_CODER_MUTATION_REQUIRED - pre_check_batch result was NOT attributed. Run /swarm recover 1.1.
The diagnosis is wrong (the result was attributed; the state simply requires a coder change first) and the advice is not actionable by the architect.
Version
opencode-swarm@7.179.1 on OpenCode 1.18.21, Windows 11. Byte-identical on the 7.180.0 release branch.
Reproduction
- Task in
rework_required (see the companion issue on rework_required).
- Architect calls
pre_check_batch for the task's files.
- The CRITICAL-WARN above is logged, twice in our run, once per attempt.
Where it happens
src/hooks/gate-evidence.ts (around lines 647-655): the throw is intentional and correct for this state.src/hooks/guardrails/index.ts (around lines 201-204 and 1322-1329): the catch labels the code as an attribution miss and adds the /swarm recover advice.
Why it matters
In our run the architect followed the advice path as far as it could: it re-ran pre_check_batch, then stashed the finished deliverables and sent a coder with instructions to make no change, in an attempt to satisfy "a coder mutation is required". That burned about 7 minutes and a third of the run's model spend before the wall clock ended it. A message that said "this task needs a real code change from a coder, or is blocked" would have stopped that.
Expected
- Label
TASK_WORKFLOW_CODER_MUTATION_REQUIRED for what it is: the workflow needs an accepted coder mutation before Stage A can pass again.
- Do not recommend
/swarm recover to an agent that cannot run it. If the state is unrecoverable for the architect, say so.
Summary
When the architect runs
pre_check_batchwhile the task is inrework_required, the reducer correctly throwsTASK_WORKFLOW_CODER_MUTATION_REQUIRED. Guardrails then file that error as a Stage A attribution failure and print a remediation that names a human-only command:The diagnosis is wrong (the result was attributed; the state simply requires a coder change first) and the advice is not actionable by the architect.
Version
opencode-swarm@7.179.1on OpenCode 1.18.21, Windows 11. Byte-identical on the 7.180.0 release branch.Reproduction
rework_required(see the companion issue onrework_required).pre_check_batchfor the task's files.Where it happens
src/hooks/gate-evidence.ts(around lines 647-655): the throw is intentional and correct for this state.src/hooks/guardrails/index.ts(around lines 201-204 and 1322-1329): the catch labels the code as an attribution miss and adds the/swarm recoveradvice.Why it matters
In our run the architect followed the advice path as far as it could: it re-ran
pre_check_batch, then stashed the finished deliverables and sent a coder with instructions to make no change, in an attempt to satisfy "a coder mutation is required". That burned about 7 minutes and a third of the run's model spend before the wall clock ended it. A message that said "this task needs a real code change from a coder, or is blocked" would have stopped that.Expected
TASK_WORKFLOW_CODER_MUTATION_REQUIREDfor what it is: the workflow needs an accepted coder mutation before Stage A can pass again./swarm recoverto an agent that cannot run it. If the state is unrecoverable for the architect, say so.