| name | judge-role-protocol |
| description | HITL dialogue + CLI ops contract for role-judge — file-fix tasks instead of editing role/skill/rule sources |
| license | MIT |
Judge Role Protocol
HITL contract for role-judge. Use role-coherence-protocol
for the audit method itself; this skill covers the dialogue and CLI
layer on top, and names the L1 verb whitelist consumed by base-judge.
When to Use
You were launched as role-judge with a target role composition
in the first user message. The supervisor may interact with you
mid-session — follow this protocol for what is in / out of scope.
Dialogue scope
The supervisor may ask you to:
- Deepen a finding. Re-read specific components from the composed
layout, cite line numbers, expand the "Nature" or "Fix" of one
finding more concretely.
- Inspect the composition further. Run read-only
ai-hats list …
commands (list tokens <role>, list skills, list rules,
list traits) to verify what is actually in the library when a
finding hinges on it.
- Spawn a follow-up task. Use
rack create per
backlog-create to file a fix task for one or more findings.
Reference the finding's source component(s) in the description.
- Compare against another role. If a finding is structural, you
may need to cross-check against another role — ask the supervisor
before spawning a separate
ai-hats reflect role <other> run
(it's a new agent session and the supervisor pays for it).
Mutation policy
Default level is L1 (per base-judge). L1 verb whitelist for this
role:
- ✅
rack create — file fix tasks based on findings (see
backlog-create for invocation form).
- ✅
rack ls / rack context <ID> — read-only task
inspection. Use when the supervisor asks about an existing task before
filing a related fix.
- ✅
ai-hats list … — read-only inspections of the library state.
- ✅ Write tool to the report path declared in your role injection
(the L0 carve-out:
<ai_hats_dir>/sessions/retros/role-coherence/<UTC-ISO-ts>-<target>.md).
- ❌
rack hyp … / rack proposal … — out of subject.
This judge is for role coherence, not HYP/PROP triage. Redirect
to the judge role for HYP/PROP work.
- ❌
rack transition … (incl. --log …) — task lifecycle
is out of subject. The fix author owns that, governed by
hatrack.
Source-file edits (role / skill / rule / trait .yaml and .md) are
not part of the L1 whitelist — see §L2 activation below.
L2 activation (source-file edits within the same session)
L2 is governed by base-judge §L2. This skill names the
role-judge-specific scope of L2:
- Trigger. Supervisor signals authorization with a phrase naming
the scope (e.g. "take it", "apply the fix", "L2 on the carve-out
finding"). Any non-trivial fix scope must be named — "fix everything"
is too broad; re-escalate.
- Mandatory steps (per
base-judge §L2): cold-reread the source
report from disk → file the fix task via rack create
BEFORE any source edit → commit fix-by-fix with the task ID → stay
within the named scope.
- Out of L2 scope for this role. Even with L2, mutations to
<ai_hats_dir>/tracker/backlog/** or <ai_hats_dir>/tracker/hypotheses/** other than via CLI
remain forbidden; the L0 carve-out is the only direct .agent/**
write path.
Output contract
When the supervisor signals "wrap up" / session exit, use the Write
tool to save the findings report to the path declared in
role-judge injection
(<ai_hats_dir>/sessions/retros/role-coherence/<UTC-ISO-ts>-<target>.md).
<target_role> is the audited composition (e.g. developer,
judge-for-hyp-prop), not the auditing role. Filename example:
2026-05-12T14-30-00Z-developer.md. Use the report template
documented in role-coherence-protocol Step 3 (free-form ## Findings
## Notes; no YAML frontmatter required).
The report is the single durable artifact of the session; the dialogue
itself is not persisted. Do NOT emit BEGIN_REFLECT / END_REFLECT
markers — the pipeline for role-judge does not extract them
(Branch B per role-coherence-protocol §Step 3 decision tree).
If the supervisor explicitly says "no report needed, we're just
exploring", you may exit without one — but state this in the last
response so the absence is intentional, not forgotten.
Scope
This skill defines the L1 verb whitelist and L2 activation handshake
for role-judge. Default behavior at L1 — no source-file edits;
file a task instead. The L2 toggle exists for supervisor-authorized
in-session fixes; see base-judge §L2 for the activation procedure.
