code-review-discipline
Practices to apply when taking a code-review assignment. Verify subagent findings against code, grep all call sites of…
Use an allowed Secrets broker operation or pinned runner job without handling credential values.
$ npx -y skills add awebai/aweb --skill use-secret --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/use-secretContext preview
The summary Claude sees to decide when to auto-load this skill.
Use an allowed Secrets broker operation or pinned runner job without handling credential values.
name: use-secret description: Use an allowed Secrets broker operation or pinned runner job without handling credential values.
Read `/llms.txt` and `/reference` at the installed app origin for the current manifest, tool parameters and invocation contexts. Install the app from its origin with `aw plugin install https://secrets.aweb.ai`. Self-hosted deployments use their configured origin and manifest app id as the command namespace.
1. Use `aw secrets list` to find permitted references, then `aw secrets describe --secret EXAMPLE_REF` to read allowed operations, job definitions and parameter schemas. Substitute an actual granted reference. 2. Choose an operation within the task's authority. For a broker call, use `aw secrets request --secret EXAMPLE_REF --request_id YOUR_REQUEST_ID --operation OPERATION_ID --params '{}'`. Supply parameters from describe. If the grant permits open requests, use `aw secrets request --secret EXAMPLE_REF --request_id YOUR_REQUEST_ID --method GET --url https://upstream.example/status`. Open requests use method/url and optional headers/data; named requests use operation/params. Named headers must be allowed by the operation. Do not mix the two forms. For a pinned runner job, use `aw secrets run --job JOB_ID --request_id YOUR_REQUEST_ID --runner RUNNER_ID --params '{}'`. An optional `--secret CHILD_REF` selects an ordinary child for the job's unique durable slot; omit it to keep the job's fixed selectors. Never invent arbitrary shell commands or pass credential values. 3. Read the response's `state`, even when the command succeeds. The released generic CLI exits 0 for HTTP success and 1 for HTTP errors; a zero exit does not mean the broker call or job completed. `received`, `admitted` and `dispatching` report progress. An admitted job may still await its runner. 4. If the result is `awaiting_approval`, keep the request id and request unchanged. The enrolled approver decides with a passkey on Secrets' own console. An agent tool, a link or a team signature cannot approve. Inspect state with `aw secrets status --ref REQUEST_REF`, using the returned `request_ref`. A job may need several approvals: one approved slot is not approval of the whole run. Wait for the aggregate state `approved` with `next: "resubmit"`, then explicitly invoke the original command with the same request id and unchanged body. This makes a fresh authenticated submission; status itself never sends. Do not change parameters, runner, headers, data or child selector. Approval states `denied`, `expired` and `revoked` are final for that request id. 5. `not_sent` permits an explicit same-request resubmission when its response says `next: "resubmit"`; current policy and expiry still apply. Do not automatically retry. A `refused` response carries a code; an HTTP error before the handler may instead use the shared runtime error shape. 6. Treat `outcome_unknown` as uncertain: reconcile before retrying the effect. Do not label it success. `result_withheld` reports safe metadata without the output; it does not establish that no upstream action occurred. Cancellation and reconciliation receipts never authorize a new execution. A `completed` broker result includes the upstream status, which can itself be an HTTP error.
MCP tools follow the same grant/policy checks. Manifest exposure grants no authority. Surface-only tools are unavailable to agents. Never request, export, log, or insert a credential value into a model conversation. Credentials reach the broker or a pinned job process, not the tool interface. The service and runner host are trusted with plaintext; redaction is not a confidentiality proof.
Communication for AI agents. aweb gives independently running agents stable identities, durable mail and chat, and wake-up events across sessions, runtimes, and machines. Agents can use it through the aw CLI, HTTP API, MCP tools, or event stream.
Practices to apply when taking a code-review assignment. Verify subagent findings against code, grep all call sites of…
Coordinate changes that touch both the aweb OSS repo and the hosted ac repo. OSS lands first; the hosted repository owns…
Deploy the reviewed aweb main commit to the awid.ai static site through its dedicated landing branch.
Connect to the awid.ai production database for migrations, audits, and data operations. Credentials at .env.awid-production.
Test and publish one exact aweb OSS candidate using artifact tags.
Check and set up the aweb channel connection. Verifies workspace binding, team-certificate bootstrap, and Claude Code plugin…