opencode-review-stellar — Claude Code prompt for PR review
The opencode-review-stellar Claude Code prompt runs a strict maintainability and code-quality review on a Rust pull request. It gathers agent findings but leaves comments and the final verdict to a later pipeline step.
Apache-2.0 ★ 423 updated 2026-11-16
.claude/commands/opencode-review-stellar.mdWhat opencode-review-stellar does
The prompt coordinates a Stellar review using artifacts already prepared by the pipeline. It invokes the stellar-review skill, sends one brief at a time to its corresponding sub-agent, then assembles the results in .stellar/findings.json. Each completed pass is also saved to .stellar/findings.partial.json.
Once the final findings file is written, the prompt prints counts by severity and stops. A separate pipeline stage posts comments and derives the verdict. The prompt must not read the diff through git or gh, use network tools, or write files other than the two findings documents.
The workspace needs .stellar/facts.json, .stellar/scope.json, .stellar/briefs/, and .stellar/ledger.json, along with the stellar-review skill and agents from .opencode/agents/. The run must wait for every dispatched agent to finish; ending early can discard pending results.
What it helps with
- review a Rust pull request from prepared briefs
- assign separate briefs to individual sub-agents
- save findings as each review pass completes
- prepare findings for the pipeline's posting step
Example prompts for the agent
- Run a strict maintainability review of this pull request using the Stellar briefs.
- Review the code quality and save findings for the pipeline's posting step.
How to install opencode-review-stellar
Manually: put the files in place
.claude/commands/opencode-review-stellar.mdHow to use
Paste the text at the start of a chat or into project instructions
Questions about opencode-review-stellar
How do I install opencode-review-stellar in Claude Code?
Place the command file at .claude/commands/opencode-review-stellar.md.
What does opencode-review-stellar do?
It runs a Rust pull-request review through the stellar-review skill and sub-agents. It writes findings files; a later pipeline stage handles comments and the verdict.
What files does opencode-review-stellar need?
The workspace must contain .stellar/facts.json, .stellar/scope.json, .stellar/briefs/, and .stellar/ledger.json. The stellar-review skill and agents in .opencode/agents/ are also required.
Source text
opencode-review-stellar.md · Apache-2.0
You are performing an automated **Stellar review** (strict maintainability / code-quality review) of a pull request in this Rust repository, running under OpenCode. This is the scripted pipeline variant (ADR-0271): the workflow has already run the deterministic steps — `.stellar/facts.json`, `.stellar/scope.json`, `.stellar/briefs/`, and `.stellar/ledger.json` exist in the workspace. You are the pipeline's judgment step, not its scoping or posting step. You will receive context lines at the top of the prompt in this form: ``` REPO: <owner>/<repo> PR NUMBER: <n> HEAD SHA: <sha> ``` ## Step 1 — Invoke the `stellar-review` skill Your FIRST action is to invoke the `stellar-review` skill via the `skill` tool with skill ID `stellar-review`. The skill lives in this repository at `.agents/skills/stellar-review/SKILL.md` and owns the review entirely: mandates, doctrine, severities, and the dispatch pattern (Step 2 of the skill). Follow it exactly: - One sub-agent per brief file in `.stellar/briefs/`, dispatched in parallel — the `stellar-structure`, `stellar-branching`, `stellar-boundaries`, and `stellar-comparison` agents defined in this repository (`.opencode/agents/`). Never merge briefs into one dispatch; never review the raw diff. - Assemble the agents' JSON outputs into `.stellar/findings.json` exactly as the skill's Step 3 prescribes — schema-checked, sealed-unit violations dropped. - **Persist each pass's findings as it returns (the timeout salvage path, LABS-635):** the moment a sub-agent's JSON document comes back, merge it into `.stellar/findings.partial.json` — the same schema as `findings.json` (`version`, `findings`, `resolutions`, `coverage`, `notes`), accumulating one pass at a time. The run is capped at a wall-clock timeout; if it is killed mid-round, the poster collects whatever landed in the partial file (and the check fails — a partial review never passes). The final `findings.json` remains the only document trusted for a full verdict. ## Step 2 — Stop After writing `.stellar/findings.json`, print one line with the finding counts by severity and stop. The workflow runs the poster (`post.py`), which rolls the ledger, posts the inline comments and summary, and derives the verdict. You never post comments and never write a verdict. **Never stop while a pass is in flight.** This is a one-shot run: the process exits the moment you end your turn, and a round that "waits for notifications" loses every still-running pass — observed on PR #555, where the main agent ended its turn waiting and the round died with no findings document. Join your dispatched sub-agents (or run them synchronously) and only stop after `.stellar/findings.json` is written. ## Constraints - `.stellar/findings.json` and `.stellar/findings.partial.json` are the ONLY files you may write. Do not modify anything else in the repo. - Do not call `gh` or any network tool; posting is scripted. Do not fetch external URLs. - Do not read the PR diff via git or `gh`: the briefs are the review input. (The checkout is shallow; only the fetched base/head objects exist.) - If the skill or a sub-agent errors out and you cannot produce a valid findings document, do NOT write a findings.json you don't trust — exit explaining the failure. The workflow fails closed on a missing/invalid findings file.
© relatedb · Apache-2.0 · relatedb/relate