agentleFS
ryanportfolioryanportfolio/Harness-Firmware/.claude/skills/astra-fullreview/SKILL.md

astra-fullreview

Skill1 starsChanged 21 days agoMITView on GitHub

Full multi-agent Codex review on gpt-6-astra at medium reasoning. Same verified lifecycle as codex-fullreview. Use for /astra-fullreview or 'full Astra review with sub-reviewers'.

Files
SKILL.md

Astra full review

Read .claude/skills/codex-fullreview/SKILL.md and follow its complete execution contract: local preflight, the required .agents/skills/impartial-review/SKILL.md check (when it is absent, stop and point to astra-review, not codex-review), exact scope, unique run directory, source manifest, the author brief (written to $RUN/brief.md and named in the prompt unless the user asks for a fully blind review), plain codex exec launch that reads .agents/skills/impartial-review/SKILL.md by file and acts as its Manager, process completion, report identity, the sub-reviewer spawn count, finding verification, and honest attribution. This entrypoint changes only the defaults:

SettingAstra default
Modelgpt-6-astra, for the Manager and every sub-reviewer
Effortmedium, regardless of diff size
Run directory prefix.tmp/astra-fullreview-

Skip the newest-Sol check in codex-fullreview Step 1; this entrypoint never swaps Astra for Sol. Honor an explicit user model/effort choice. Confirm supported local options before inference; a model identifier in this file is not proof of availability. $ARGUMENTS carries scope as in codex-fullreview.

Example for a branch diff after creating a fresh $RUN directory (POSIX shell):

mkdir -p .tmp
RUN=$(mktemp -d .tmp/astra-fullreview-XXXXXXXX)
git fetch origin main
BASE=$(git merge-base origin/main HEAD); HEAD_SHA=$(git rev-parse HEAD)
codex exec -s read-only -m gpt-6-astra -c model_reasoning_effort=medium -o "$RUN/report.md" "Read .agents/skills/impartial-review/SKILL.md completely and act as its Manager on the branch diff: base $BASE, head $HEAD_SHA. Read the diff with git diff $BASE $HEAD_SHA. Spawn fresh-context sub-reviewers with spawn_agent, fork_turns \"none\", model gpt-6-astra, reasoning_effort medium, as the skill directs; each is a leaf reviewer. Start no codex exec or other reviewer CLI process; if agent tools are missing, say so in the report. Author brief: $RUN/brief.md; give it only to the intent reviewer, as the skill directs; every other sub-reviewer gets the diff without it. No file edits, no Git writes, no publication. Verify the sub-reviewers' findings and write the final verified report. First line: Scope: base $BASE, head $HEAD_SHA. Second line: Sub-reviewers: <number spawned>." < /dev/null > "$RUN/run.log" 2>&1

Commit and uncommitted prompts (with the Astra model and effort in place of Sol and high), the root-commit base, the file-named skill invocation and prompt escaping, the session-file checks for skill load and spawned children, the scope-line check, and attribution follow codex-fullreview. Zero spawned children is reported as "no sub-reviewers ran", never hidden.

On PowerShell use a GUID-named directory and supported stdin/output redirection as described in codex-fullreview. Distinct prefixes do not prevent collisions between two Astra invocations; every run must still be unique.

If Astra is unavailable, report that outcome. Do not silently inherit another model or attribute a fallback to Astra. A rejected Astra id is reported, never retried on another model; any other failure gets the one automatic retry from codex-fullreview. A retry uses a new directory and reports its observed model/effort, or explicitly marks model resolution unverified. Keep lifecycle fixes in codex-fullreview rather than forking them here.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

Reviews can't be read right now.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.