OpenClaw repair
Parse <selector>, --issue <finding-id>, [--all-findings], [--yes] from "$ARGUMENTS".
Every fix enters here: a repair exists only as a row in findings-catalog.md carrying an id, severity,
detection, fix, verification and risk class. An unknown symptom gets a catalog entry — with its
documentation citation — first, and a fix second.
Process
Resolve the id to a row — Skill fleet-diagnostics, then ${CLAUDE_PLUGIN_ROOT}/skills/fleet-diagnostics/references/findings-catalog.md. Id
forms: fleet.* catalog ids · upstream checkIds (fs.*, gateway.*, tools.exec.*, plugins.*,
security.exposure.*) verbatim from the live --json · <instance>/<check-id>. No row → stop.
Resolve the targets as a mutation:
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/fleet.py" resolve "<selector>" --mutation --table
Exit 3 refuses an empty selector, all, an alien instance or manage:false, with the reason per
name; a legacy layout is refused too (red line legacy-instance-mutation).
Load the skill that owns the family — auth → provider-auth; config and model chain →
config-surgery; memory → memory-ops; secrets → secrets-infisical; shared assets → shared-assets;
exposure → security-audit. Upgrades belong to /openclaw-ops:update, never here.
Re-detect before repairing — run the row's detect step now through
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/ocexec.py" <instance> --json -- <args>. A finding that no longer reproduces is resolved, not repaired.
Build the plan at the row's risk class (gate.make_plan) with all eight blocks —
TARGET · PRECHECK · CHANGE · BACKUP · IMPACT · VALIDATE · ROLLBACK · APPLY (plus
IRREVERSIBLE · CONFIRM for R3/R4). Block contents belong to their owners: the eight-block contract
to fleet-model, the deletion count and the .bak-ring snapshot rule to config-surgery. ROLLBACK is
executable, prose does not validate, and anything replacing an executable artefact pins first (gate.pin):
cp <snapshot-path> <config-path> && python3 "${CLAUDE_PLUGIN_ROOT}/scripts/ocexec.py" <instance> --yes --plan-id <plan-id> -- gateway restart.
Show the plan and stop. Mint its id (scripts/lib/gate.py plan mint repair <instance> --risk <R>)
and print it with the plan: the applying ocexec.py call passes it as --plan-id, the door checks it
against the registry and burns it — one plan, one mutation, so a rollback needs its own id.
--yes belongs to a later turn; R4 also needs the typed phrase from gate.confirm_phrase.
Apply, then verify with the row's verify command — one whose result differs before and after;
"looks fine now" is not one. Re-run detect: it must come back empty, and zero-retry families
(gate.retry_policy) stop on the first failure.
--all-findings
Batch over every open finding. Broken → attempted-repair, so the policy is continue-and-report
(gate.batch_policy): one row per instance and finding — repaired / failed / refused — and the
batch finishes. R3/R4 rows leave for their own plan; a healthy-instance batch is fail-fast behind the canary.
Never
Improvise a fix for an id with no row · pass --accept-capabilities, printed for a human · run the automatic repair
flags of lint or security audit by default (red line doctor-fix-or-security-fix) · restart a zombie instead of triaging it.
1---2name: repair3description: 'The only repair funnel — dispatch one catalog finding id through its eight-block plan, apply on a later turn, verify Arguments: <selector> --issue <finding-id> [--all-findings] [--yes].'4metadata:5 source: https://skillmd.com/skills/agents-store/repair6---78# OpenClaw repair910Parse `<selector>`, `--issue <finding-id>`, `[--all-findings]`, `[--yes]` from "$ARGUMENTS".1112Every fix enters here: a repair exists only as a row in `findings-catalog.md` carrying an id, severity,13detection, fix, verification and risk class. An unknown symptom gets a catalog entry — with its14documentation citation — first, and a fix second.1516## Process17181. **Resolve the id to a row** — `Skill` `fleet-diagnostics`, then `${CLAUDE_PLUGIN_ROOT}/skills/fleet-diagnostics/references/findings-catalog.md`. Id19 forms: `fleet.*` catalog ids · upstream `checkId`s (`fs.*`, `gateway.*`, `tools.exec.*`, `plugins.*`,20 `security.exposure.*`) verbatim from the live `--json` · `<instance>/<check-id>`. **No row → stop.**21222. **Resolve the targets as a mutation**:23 ```bash24 python3 "${CLAUDE_PLUGIN_ROOT}/scripts/fleet.py" resolve "<selector>" --mutation --table25 ```26 Exit 3 refuses an empty selector, `all`, an alien instance or `manage:false`, with the reason per27 name; a legacy layout is refused too (red line `legacy-instance-mutation`).28293. **Load the skill that owns the family** — auth → `provider-auth`; config and model chain →30 `config-surgery`; memory → `memory-ops`; secrets → `secrets-infisical`; shared assets → `shared-assets`;31 exposure → `security-audit`. Upgrades belong to `/openclaw-ops:update`, never here.32334. **Re-detect before repairing** — run the row's detect step now through34 `python3 "${CLAUDE_PLUGIN_ROOT}/scripts/ocexec.py" <instance> --json -- <args>`. A finding that no longer reproduces is resolved, not repaired.35365. **Build the plan** at the row's risk class (`gate.make_plan`) with all eight blocks —37 **TARGET · PRECHECK · CHANGE · BACKUP · IMPACT · VALIDATE · ROLLBACK · APPLY** (plus38 **IRREVERSIBLE · CONFIRM** for R3/R4). Block contents belong to their owners: the eight-block contract39 to `fleet-model`, the deletion count and the `.bak`-ring snapshot rule to `config-surgery`. ROLLBACK is40 executable, prose does not validate, and anything replacing an executable artefact pins first (`gate.pin`):41 `cp <snapshot-path> <config-path> && python3 "${CLAUDE_PLUGIN_ROOT}/scripts/ocexec.py" <instance> --yes --plan-id <plan-id> -- gateway restart`.42436. **Show the plan and stop.** Mint its id (`scripts/lib/gate.py plan mint repair <instance> --risk <R>`)44 and print it with the plan: the applying `ocexec.py` call passes it as `--plan-id`, the door checks it45 against the registry and **burns it** — one plan, one mutation, so a rollback needs its own id.46 `--yes` belongs to a later turn; R4 also needs the typed phrase from `gate.confirm_phrase`.47487. **Apply, then verify** with the row's verify command — one whose result differs before and after;49 "looks fine now" is not one. Re-run detect: it must come back empty, and zero-retry families50 (`gate.retry_policy`) stop on the **first** failure.5152## `--all-findings`5354Batch over every open finding. Broken → attempted-repair, so the policy is **continue-and-report**55(`gate.batch_policy`): one row per instance and finding — `repaired` / `failed` / `refused` — and the56batch finishes. R3/R4 rows leave for their own plan; a healthy-instance batch is fail-fast behind the canary.5758## Never5960Improvise a fix for an id with no row · pass `--accept-capabilities`, printed for a human · run the automatic repair61flags of lint or security audit by default (red line `doctor-fix-or-security-fix`) · restart a zombie instead of triaging it.