| name | linksim-policy-audit |
| description | Audit recurring LinkSim failures, corrections, memory drift, duplicated code or UI, and workflow friction as Steward, then suggest evidence-backed policy wording without editing anything. Use for manual or quota-gated policy reviews, recurring operational problems, inconsistent agent behavior, or proposed improvements to AGENTS.md, skills, prompts, permissions, workflows, and infrastructure safeguards. |
LinkSim Policy Audit
Audit read-only and return suggestions only. Follow AGENTS.md and treat the
current checked-in policy as authoritative until a reviewed Forge change is
merged.
Gather evidence
- Define the audit window and sources. Prefer issues, review threads, CI runs,
repeated corrections, incident reports, and checked-in policy.
- Separate recurring evidence from one-off preference. Cite concrete events,
files, symbols, or artifact links; do not infer a pattern from one ambiguous
occurrence.
- Check whether current policy already covers the behavior and whether the
failure was policy absence, ambiguity, drift, non-compliance, or tooling.
- Search for overlapping rules, skills, scripts, code, and UI before proposing
anything new. Recommend consolidation where it reduces conflicting sources.
- Include quota impact, security, least privilege, provenance, maintenance,
and failure modes where relevant.
- When evidence supports a durable lesson, suggest a compact knowledge
registry entry with affected agents, evidence, and proposed wording. Never
edit the knowledge registry; an explicitly approved Forge task is required.
If evidence does not support a change, report No policy suggestion with the
searched evidence. Do not create work merely because the audit ran.
Form a reviewable suggestion
For each supported improvement, provide:
- evidence and recurrence count;
- current policy and observed gap;
- exact proposed wording or deterministic-tool change;
- expected benefit and token or maintenance cost;
- compatibility, security, and authority implications;
- alternatives considered and why this is narrower;
- validation and rollback approach;
- files or systems that an approved Forge task would change.
Keep facts, inferences, and recommendations visibly distinct. Never edit
AGENTS.md, skills, prompts, permissions, workflows, infrastructure, the
knowledge registry, or memory.
Never implement the recommendation. Human acceptance followed by an explicit
Forge task is required.
Publication gate
For a scheduled audit, first confirm the shared quota guard permits a model
call using the thresholds and durable issue in
config/steward-policy-audit.json. If that configuration is disabled, its
executor is unconfigured, telemetry is stale, or the guard is not normal, do
not start a model call. Comment only when at least one evidence-backed
improvement exists. Append through the deterministic publisher to the single
configured documentation + pending-discussion issue; do not rewrite its
body or create repeated issues. Publication must be idempotent, signed, and
provenance-validated. Resolve scheduled publication identity, delivery
identity, visible attribution, and permitted issue authority from
config/ai-agents.json. Before publication, the deterministic publisher must
compare its authenticated GitHub identity with Steward's configured
githubIdentity, confirm the requested action is
publish-signed-suggestion, confirm that action is allowed and not prohibited,
and stop before publication on any mismatch. The publisher fails closed on any
mismatch. Use scripts/ai-provenance.mjs to validate the registry and rendered
artifact. The artifact must display only the
registry-approved attribution for the configured Steward publisher. Do not
hard-code, inherit, or claim another agent's delivery identity or authority.
End every output with Suggestion only — no policy change applied. followed by
the signature and marker returned by node scripts/ai-provenance.mjs footer;
append both returned fields and never prescribe or hand-build either one. Fail
closed if attribution or provenance cannot be generated. Render the complete
comment to a file, then require
node scripts/ai-provenance.mjs validate-artifact --agent Steward to pass
before publication.