| name | gh-address-cr |
| description | Use when a GitHub Pull Request has unresolved review threads, pending reviews, stale/outdated threads, local findings ingestion, or needs mandatory final-gate proof in one PR-scoped session. |
gh-address-cr
Use this skill as the thin adapter and behavioral policy layer for the
gh-address-cr runtime CLI.
The runtime owns session state, intake routing, leases, GitHub side effects,
reply evidence, and completion truth.
Primary Commands
Use the runtime help and manifest as the authoritative command inventory:
gh-address-cr --help
gh-address-cr agent manifest
Start from one of two canonical modes:
gh-address-cr review <owner/repo> <pr_number>
gh-address-cr address <owner/repo> <pr_number> --lean
review / 完整审查: ingest explicit structured findings and handle them
together with GitHub review threads.
address / 处理评审: handle existing GitHub review threads without
starting a new review producer.
If the PR number is unknown, run
gh-address-cr active-pr [--repo <owner/repo>] [--head <branch>].
Compose a review producer only when the user explicitly names or supplies one:
Use $gh-address-cr review PR #123 with $engineering:code-review as the findings producer.
Require a JSON array with title, body, path, and line; use [] when
clean. Do not ingest narrative Markdown. Read
references/mode-producer-matrix.md for the exact intake command.
Packaging And Runtime Boundary
This file is part of the packaged gh-address-cr skill. All paths in this document are relative to the installed skill root.
Repository tests, CI, and release metadata are outside the packaged skill
payload.
- Runtime entrypoints:
gh-address-cr and python3 -m gh_address_cr
- Runtime version: inspect
gh-address-cr version and compare it with
runtime-requirements.json
- Protocol compatibility: inspect
gh-address-cr adapter check-runtime
If the runtime or required version is unavailable, fail before session
mutation. Do not copy runtime state-machine logic into the skill.
Execution Ladder
- Run the selected public main entrypoint.
- Read only the machine summary fields
status, reason_code, waiting_on,
next_action, commands, and counts.
- Prefer the returned
commands templates over reconstructing commands.
- Follow
references/status-action-map.md until the runtime accepts evidence.
- Submit decisions through
gh-address-cr agent resolve; publish accepted
GitHub-thread evidence through gh-address-cr agent publish.
- Run
gh-address-cr final-gate <owner/repo> <pr_number> last.
If review returns BLOCKED, inspect the loop request artifact, apply fix,
clarify, defer, or reject through runtime evidence, then rerun the same
review command.
GitHub review comment reply tasks are incomplete until the runtime accepts the
response and publishes both reply and resolve side effects. Use
references/agent-protocol.md for item, batch, stale, decline, evidence, and
lease-recovery command shapes.
When exactly one cached PR session exists, PR-scoped commands may omit the
target. For NO_ACTIVE_PR_SCOPE or AMBIGUOUS_PR_SCOPE, pass the target
explicitly instead of guessing.
Completion And Telemetry
Record measured timing on every --validation when known:
gh-address-cr agent resolve <owner/repo> <pr_number> <item_id> \
--commit <sha> --files <paths> --summary "..." --why "..." \
--validation "unit-tests=passed@4200ms"
gh-address-cr final-gate <owner/repo> <pr_number>
Completion requires a freshly passing final gate. Include its exact
completion_summary_line in the final response. Coverage is complete,
partial, runtime-only, or unavailable. Telemetry degradation is
diagnostic, not review-resolution failure; explain abnormal diagnostics. Read
references/completion-contract.md before claiming completion.
For cross-invocation correlation, keep one stable session identifier:
export GH_ADDRESS_CR_CONVERSATION_ID="<stable-session-id>"
Process telemetry is exported by the runtime through its configured Honeycomb
relay. It is fail-open and can be disabled with DISABLE_TELEMETRY=1 or
DO_NOT_TRACK=1.
Never include raw prompts, tokens, usernames, machine identifiers, or
unnecessary absolute paths.
Common Mistakes
- Do not infer state from prose or logs; follow machine fields and returned
commands.
- Do not post GitHub replies or resolve threads directly.
- Do not treat
STALE or outdated threads as clean.
- Do not invent P0-P4 severity. Preserve explicit producer or reviewer
evidence, and provide
--severity-note for an intentional override.
- Do not omit measured validation timing when it is known.
- Do not create workspace findings files when
findings --input - is enough.
- Do not treat
agents/openai.yaml as a behavioral contract.
review-to-findings does not accept arbitrary Markdown. It only accepts fenced
finding blocks, and rejects plain narrative Markdown review output. Local
findings require terminal notes and validation evidence.
GitHub review threads require durable reply evidence from the current
authenticated login.
Reference Surface
Read only the reference required by the current runtime state:
- For blocked or waiting states:
references/status-action-map.md
- For resolve, batch, evidence, or lease details:
references/agent-protocol.md
- Before reporting completion:
references/completion-contract.md
- For producer or input routing:
references/mode-producer-matrix.md
- Before deciding
fix, clarify, defer, or reject:
references/cr-triage-checklist.md
- To understand runtime-owned side-effect evidence:
references/evidence-ledger.md
- When the skill itself blocks progress:
references/feedback.md; filing a
qualifying tooling-bug is an expected automatic step.