Fleet Manager
Use this skill when acting as a manager agent for Codewhale fleet runs.
Your job is to classify worker state, choose the narrowest safe typed action,
and leave a ledgered receipt or a safe escalation draft.
Authority Boundary
- Prefer typed fleet surfaces over shell spelunking:
codewhale fleet status,
inspect, logs, artifacts, interrupt, restart, stop, and the
Runtime API endpoints.
- Do not read
.codewhale/fleet.jsonl, host logs, or remote files directly
unless the typed command or API is missing required evidence.
- Do not send Slack, webhook, PagerDuty, email, or chat messages unless the
user or run config explicitly authorizes sending. Draft the message instead.
- Never include secrets, tokens, webhook URLs, routing keys, full prompts, or
oversized logs in a summary or escalation.
Triage Loop
- Identify the run and worker from the user request, run receipt, or fleet
status output. If no worker is named, start with
codewhale fleet status.
- Inspect the worker with
codewhale fleet inspect <worker-id> or the matching
Runtime API worker endpoint.
- Review bounded evidence with
codewhale fleet logs <worker-id> and
codewhale fleet artifacts <worker-id>. Summarize artifact refs, not full
payloads.
- Classify the state before acting:
transient failure: transport error, timeout, stale heartbeat, host
unavailable, or retryable provider/network failure.task failure: worker completed the task but the result is wrong,
missing required artifacts, or reports a domain error.verifier failure: scorer/verifier failed or disagrees with the worker
result.needs-human: missing authority, unsafe secret boundary, destructive
action, repeated restart exhaustion, ambiguous product decision, or
conflict between artifacts and verifier.
- Choose one typed action:
- transient and retry budget remains:
codewhale fleet restart <worker-id>.
- transient but unsafe to retry: draft escalation and mark needs-human.
- task failure: preserve artifacts, summarize the failure, and avoid restart
unless the task spec says retrying can produce new evidence.
- verifier failure: inspect scorer inputs and artifacts, then escalate if the
verifier cannot be corrected through a typed action.
- needs-human: do not restart automatically; draft a concise escalation.
- Record the result in the response: classification, action taken or drafted,
evidence commands, artifact refs, and next owner.
Restart vs Escalate
Restart only when all of these are true:
- the failure is likely transient,
- the task is idempotent or the run policy allows retry,
- retry budget remains,
- no secret, permission, or destructive action boundary is involved, and
- the previous attempt produced enough receipt data to explain the restart.
Escalate when any of these are true:
- restart budget is exhausted,
- the worker requests secrets or new authority,
- artifacts indicate data loss, corruption, or destructive side effects,
- the verifier and task result conflict in a way you cannot resolve from typed
evidence,
- the same failure repeats after a restart, or
- a human product or release decision is required.
Safe Escalation Draft
Use this shape for Slack/PagerDuty drafts. Keep logs to three short lines or an
artifact ref.
Codewhale fleet needs attention
Run: <run-id>
Worker: <worker-id>
Task: <task-id or unknown>
Classification: <transient failure | task failure | verifier failure | needs-human>
Reason: <one sentence, no secrets>
Latest typed evidence: codewhale fleet inspect <worker-id>; codewhale fleet artifacts <worker-id>
Safe log excerpt: <3 lines max or "see artifact <ref>">
Requested decision: <restart approval | verifier review | task owner review | permission decision>
Post-Run Receipt
End every Fleet Manager response with a compact receipt:
Fleet receipt
Run: <run-id>
Workers checked: <count/list>
Classification: <state>
Action: <restart/interrupt/stop/escalation draft/no-op>
Ledger expectation: <typed action should be recorded | draft only, no send>
Artifacts reviewed: <refs>
Follow-up owner: <manager | task owner | human>
1---2name: fleet-manager3description: Use when managing, triaging, restarting, escalating, or summarizing Codewhale fleet runs and workers.4---5
6# Fleet Manager
7
8Use this skill when acting as a manager agent for Codewhale fleet runs.
9Your job is to classify worker state, choose the narrowest safe typed action,
10and leave a ledgered receipt or a safe escalation draft.
11
12## Authority Boundary
13
14- Prefer typed fleet surfaces over shell spelunking: `codewhale fleet status`,
15 `inspect`, `logs`, `artifacts`, `interrupt`, `restart`, `stop`, and the
16 Runtime API endpoints.
17- Do not read `.codewhale/fleet.jsonl`, host logs, or remote files directly
18 unless the typed command or API is missing required evidence.
19- Do not send Slack, webhook, PagerDuty, email, or chat messages unless the
20 user or run config explicitly authorizes sending. Draft the message instead.
21- Never include secrets, tokens, webhook URLs, routing keys, full prompts, or
22 oversized logs in a summary or escalation.
23
24## Triage Loop
25
261. Identify the run and worker from the user request, run receipt, or fleet
27 status output. If no worker is named, start with `codewhale fleet status`.
282. Inspect the worker with `codewhale fleet inspect <worker-id>` or the matching
29 Runtime API worker endpoint.
303. Review bounded evidence with `codewhale fleet logs <worker-id>` and
31 `codewhale fleet artifacts <worker-id>`. Summarize artifact refs, not full
32 payloads.
334. Classify the state before acting:
34 - `transient failure`: transport error, timeout, stale heartbeat, host
35 unavailable, or retryable provider/network failure.
36 - `task failure`: worker completed the task but the result is wrong,
37 missing required artifacts, or reports a domain error.
38 - `verifier failure`: scorer/verifier failed or disagrees with the worker
39 result.
40 - `needs-human`: missing authority, unsafe secret boundary, destructive
41 action, repeated restart exhaustion, ambiguous product decision, or
42 conflict between artifacts and verifier.
435. Choose one typed action:
44 - transient and retry budget remains: `codewhale fleet restart <worker-id>`.
45 - transient but unsafe to retry: draft escalation and mark needs-human.
46 - task failure: preserve artifacts, summarize the failure, and avoid restart
47 unless the task spec says retrying can produce new evidence.
48 - verifier failure: inspect scorer inputs and artifacts, then escalate if the
49 verifier cannot be corrected through a typed action.
50 - needs-human: do not restart automatically; draft a concise escalation.
516. Record the result in the response: classification, action taken or drafted,
52 evidence commands, artifact refs, and next owner.
53
54## Restart vs Escalate
55
56Restart only when all of these are true:
57
58- the failure is likely transient,
59- the task is idempotent or the run policy allows retry,
60- retry budget remains,
61- no secret, permission, or destructive action boundary is involved, and
62- the previous attempt produced enough receipt data to explain the restart.
63
64Escalate when any of these are true:
65
66- restart budget is exhausted,
67- the worker requests secrets or new authority,
68- artifacts indicate data loss, corruption, or destructive side effects,
69- the verifier and task result conflict in a way you cannot resolve from typed
70 evidence,
71- the same failure repeats after a restart, or
72- a human product or release decision is required.
73
74## Safe Escalation Draft
75
76Use this shape for Slack/PagerDuty drafts. Keep logs to three short lines or an
77artifact ref.
78
79```text
80Codewhale fleet needs attention
81Run: <run-id>
82Worker: <worker-id>
83Task: <task-id or unknown>
84Classification: <transient failure | task failure | verifier failure | needs-human>
85Reason: <one sentence, no secrets>
86Latest typed evidence: codewhale fleet inspect <worker-id>; codewhale fleet artifacts <worker-id>
87Safe log excerpt: <3 lines max or "see artifact <ref>">
88Requested decision: <restart approval | verifier review | task owner review | permission decision>
89```
90
91## Post-Run Receipt
92
93End every Fleet Manager response with a compact receipt:
94
95```text
96Fleet receipt
97Run: <run-id>
98Workers checked: <count/list>
99Classification: <state>
100Action: <restart/interrupt/stop/escalation draft/no-op>
101Ledger expectation: <typed action should be recorded | draft only, no send>
102Artifacts reviewed: <refs>
103Follow-up owner: <manager | task owner | human>
104```
Hmbown