harden-record-author

name	harden-record-author
description	Work findings from an inspection-report@1 under scope and safety discipline and emit a hardening-record@1. Validates input with inspect-lint before reading findings, classifies each finding as agent/operator/split (honouring operator_prerequisites), halts at plan and review gates, and never pushes, merges, rotates credentials, or executes irreversible actions. Use for "/harden", "work the inspection backlog", or "remediate what /inspect found". Do NOT use to discover findings (inspection-report-author), judge report quality (inspection-report-audit), or ship a backlog class:improvement task (ship-tasks — distinct "harden a task" path).
license	Apache-2.0
metadata	{"version":"1.0.0","module":"skill","stage":"harden","cyberos-template":"hardening-record@1","cyberos-rubric-target":"hardening_record_rubric@1.0"}
allowed_memory_scopes	{"read":["project:*","module:*","company:locked-decisions","memories:decisions","memories:projects"],"write":["project:*","memories:projects"]}
allowed_mcp_tools	["kb.read","kb.search","memory.search","memory.write_memory","audit.append","chat.notify"]
escalation	{"to_persona_on_legal":"cuo-clo","to_persona_on_security":"cuo-cseco","to_persona_on_compliance":"cuo-clo","to_human_on_irreversible":true}
invocation_modes	["standalone","chained"]
expects	{"schema_ref":"./envelopes/input.json","required_fields":["report_path"],"optional_fields":["finding_id","want_commits","caller_persona","trace_id","chain_to"]}
produces	{"schema_ref":"./envelopes/output.json","output_kind":"artefact","human_summary_ref":"./HUMAN_SUMMARY.md"}
depends_on_contracts	[{"id":"nats-subjects","version":"v1","purpose":"wire_protocol_emission","pin_path":"cyberos/skill/contracts/nats-subjects/"}]
exposable_as	{"internal":true,"agent_plugin":true,"mcp_tool":false,"partner_connector":false}
audit	{"emit_to":"genie.action_log","row_kind":"artefact_write","payload_hash_field":"artefact_hash","explanation_pane":"required"}
confidence_band	{"default":0.7,"defer_below":0.5,"cite_sources":"required"}
untrusted_inputs	{"wrap_in_marker":"untrusted_content","injection_scan":"required","on_marker_hit":"surface_to_human"}
self_audit	{"invariants_ref":"./INVARIANTS.md","check_at":["on_node_boundary","on_completion"]}
determinism	{"reproducible":false,"fixity_notes":"Planner (harden-plan.mjs) is deterministic; remediation judgement is not."}
chains_from	inspection-report-audit
chains_to	harden-record-audit
next_skill_recommendation	harden-record-audit
emitted_source_freshness_tier	30
untrusted_content_wrapping	required

harden-record-author

Purpose

Close findings from an inspection-report@1 without exceeding each finding's declared scope, and leave a hardening-record@1 a re-inspection can reconcile by fingerprint.

Never discovers findings, never re-inspects, never decides a finding is not worth fixing. Distinct from ship-tasks "harden a task" (class: improvement).

Contract

Rules in references/harden-prompt.md (HRD-* families). Non-negotiable:

1. Linter exits 0 before any finding is read (HRD-IN-1).
2. Only declared-scope files change, plus named collateral (HRD-SCOPE).
3. Nothing irreversible; no credential rotation; no push/merge/deploy (HRD-SAFE).
4. Both human gates reached; silence is not approval (HRD-HITL).

Procedure

Validate → resolve next action → tools/harden-plan.mjs → plan HITL → work one finding per cycle → review HITL → emit record → chain to harden-record-audit.

Actor classification

tools/harden-plan.mjs classifies agent / operator / split from remediation, rollback, and operator_prerequisites. A non-none prerequisite must not classify as pure agent (shopass INS-F-0002 regression).

Anti-scope

No out-of-scope writes, no PR open, no task status mutation, no live credential tests.