[DOCS] sweep: automation-mode.md contradicts SKILL.md on open_ids for --only / --min-severity runs

[tronghieu]

What type of documentation issue is this?

Outdated content

Documentation location

src/bmad_loop/data/skills/bmad-loop-sweep/automation-mode.md, the "Validation rules the orchestrator enforces" list (lines 54-59 at v0.12.0, unchanged on main).

What's the issue?

#651 added --only and --min-severity. SKILL.md was updated for them. automation-mode.md was not. The two files now give the triage agent opposite instructions for a subset sweep.

SKILL.md:

If the invocation carries --only DW-1,DW-2,..., those ids are the complete triage universe for this session. Read the full ledger, but verify and partition exactly those named open entries; do not add other raw-open entries to open_ids or any result category. The orchestrator has already applied the operator's named-subset or severity-floor selector and validates your result against this exact scope.

automation-mode.md:

Validation rules the orchestrator enforces (a violation fails the whole result and burns a retry):

• open_ids must list exactly the ledger's status: open entries — the orchestrator parses the ledger itself and compares.
• Every open id appears in exactly ONE of already_resolved / bundles / blocked / skip / decisions. No misses, no duplicates, no invented ids.

The engine agrees with SKILL.md. In sweep.py at v0.12.0, the run loop builds selected_ids from selection.selected and passes it to _cycle. That set reaches validate_triage as expected_open_ids. A subset sweep is validated against the selected ids, not against every open entry.

So this is not a runtime bug. The risk is the agent. automation-mode.md is the mandatory contract, and it says a violation burns a retry. An agent that obeys it on a --only run lists every open entry in open_ids. Validation then fails with "open_ids do not match the ledger's open entries", and the retry is burned.

An automated PR reviewer read the pair as unsatisfiable while reviewing a 0.12.0 upgrade. A triage agent can read it the same way.

Suggested improvement

Scope both bullets to the session's triage universe. For example:

• open_ids must list exactly the open entries in this session's triage universe: every status: open entry, or only the named ids when the invocation carries --only. The orchestrator computes the same set and compares.
• Every id in that universe appears in exactly ONE of already_resolved / bundles / blocked / skip / decisions. No misses, no duplicates, no invented ids.

bmad-loop Version (if applicable)

0.12.0