Multi-skill review (same fan-out as `/repo-review-full`) that prints the
aggregated BLOCK/WARN/NIT report instead of posting inline comments. Routes every
host harness through Pi and works against an open PR or local branch. Requires
the tinaudio-synth-setter-skills plugin.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Multi-skill review (same fan-out as `/repo-review-full`) that prints the
aggregated BLOCK/WARN/NIT report instead of posting inline comments. Routes every
host harness through Pi and works against an open PR or local branch. Requires
the tinaudio-synth-setter-skills plugin.
repo-review-full-no-comments — Multi-Skill Review Without Posting
Same analysis as /repo-review-full, with two differences:
The final delivery step renders findings in chat instead of submitting a
GitHub review.
A PR is not required. If no PR exists for the current branch, the
orchestrator reviews the local branch vs. the default branch.
The foreground dry run posts nothing. On an existing PR, deferred second passes
may later post only new Codex-verified findings through detached follow-up; this
preserves the sub-ten-minute response without dropping slow independent review.
The review implementation is Pi-native. Claude Code and Codex invoke the same
headless Pi entrypoint instead of maintaining separate nested-agent harnesses.
What you (the main agent) do
Capture the target argument: if the command was invoked with an explicit
<N>, keep it; otherwise the orchestrator resolves PR-or-local-branch mode
itself.
If SYNTH_SETTER_PI_REVIEW is not 1, follow
agent/_shared/pi-review-host-contract.md with
repo-review-full-no-comments as the selected skill. Relay the command's
output verbatim and stop; the child Pi session owns the review.
If SYNTH_SETTER_PI_REVIEW=1, do not invoke the launcher again. Execute the
orchestrator brief in this Pi session and use Tintin's pr-review-worker
Agent for the flat Step 4 fan-out. Follow the allocation, fallback, merge,
and transcript-audit rules in the shared analysis exactly.
The agent returns the full rendered Markdown report ending in a final
Sentinel: <path> line. Print exactly what the orchestrator returned,
verbatim — that trailing line already surfaces the sentinel path, so do not
append any narration of your own. Do not re-run the pipeline.
Orchestrator agent brief
You are the orchestrator for a multi-skill dry-run PR review that posts
nothing to GitHub. Complete every step in order and do not stop early. "You"
throughout the steps below and in the shared analysis file means you, this
orchestrator agent.
Model policy. Use the shared dynamic routing table and supply every
worker's model and thinking level explicitly.
Step 1: Resolve the target (PR or local branch)
Pick whichever mode applies. Steps 3–6 work the same once the metadata is in
hand.
PR mode. Use this when an explicit <N> was passed, or when
gh pr view --json number resolves a PR for the current branch. If no <N>
was passed, first resolve it from the current branch
(N=$(gh pr view --json number -q .number)) and use that value wherever <N>
appears below — never run the command with the literal <N> placeholder:
This is the exact call from agent/skills/_shared/repo-review-full-analysis.md
Step 1 — use that file's guidance for parsing it.
Local-branch mode. Use this when no <N> was passed AND
gh pr view --json number fails / returns nothing for the current branch.
Derive the same fields from local git:
base_ref=$(gh repo view --json defaultBranchRef -q .defaultBranchRef.name)
base_sha=$(git merge-base HEAD "origin/${base_ref}")
head_sha=$(git rev-parse HEAD)
head_ref=$(git rev-parse --abbrev-ref HEAD)
# File list with status (for tdd-refactor R-detection in Step 3)
git diff --name-status "${base_sha}..${head_sha}"
Build a synthetic metadata object equivalent to the gh pr view JSON:
number: null — no PR yet; use the branch name in any user-facing text.
headRefOid: head_sha
baseRefOid:
Notes
This skill's foreground result is side-effect-free on GitHub. For an existing
PR, detached follow-up may post one review containing only new Codex-verified
findings from passes deferred at the response deadline. It rechecks the exact
head immediately before posting. Local-branch mode remains fully side-effect-free.
A non-PASS report starts a remediation loop, not a license to retry the same
review. Follow the non-PASS Summary instruction before another review or a
handoff. This is advisory so investigation and deliberately uncommitted
experiments remain possible.
Like /repo-review-full, this skill depends on the
tinaudio-synth-setter-skills plugin being enabled. If a sub-skill
invocation fails, surface the error — don't silently skip.
Do not dedupe findings across skills. Keep each skill's signal independent,
same as /repo-review-full.
mergeable, mergeStateStatus, statusCheckRollup: not available —
Step 2 handles this.
If there are zero changed files between base_sha and head_sha, set
is_zero_diff=true, skip Steps
2–6 and go straight to Step 7's PASS short form: write the sentinel file
and return the rendered PASS report ending in Sentinel: <path> (note
PASS — no diff vs ${base_ref} in the report body). Do not early-return a
bare string — the pre-PR gate needs the sentinel on disk.
Step 2: Inspect PR health
PR mode. Follow agent/skills/_shared/repo-review-full-analysis.md Step 2
exactly as written (merge conflicts + failing checks).
Local-branch mode. Skip the GitHub-side mergeable and
statusCheckRollup checks — they don't exist pre-PR. Do not synthesize
PR-health BLOCKs and do not emit a ## PR health section in Step 6's JSON.
Instead, Step 6 prepends a one-line caveat to the review_body so the user
knows merge/CI health was not verified.
Steps 3–6: Run the shared analysis pipeline
Read and follow agent/skills/_shared/repo-review-full-analysis.md Steps 3
through 6 using the metadata from Step 1 above. The shared file owns skill
selection, the parallel fan-out, finding aggregation, and the findings-JSON
construction; it is mode-agnostic. When it says "the calling skill," that is
repo-review-full-no-comments:
Use repo-review-full-no-comments as the calling-skill name in any
[<calling-skill>:block] prefixes inside the ## PR health bullets (PR mode
only — local-branch mode has no PR-health bullets).
Before writing findings, create an invocation-isolated path:
Capture the exact printed path and write this invocation's findings JSON only
there. Shell variables do not persist across tool calls, so substitute that
exact path in every later command; never use a shared fixed filename. Bash
tool calls do not share an EXIT trap, so remove the path on every
controlled success or failure; an interrupted run leaves only an isolated
file in the platform temporary directory and cannot contaminate another run.
For pr_number in the JSON: use the resolved PR number in PR mode, or
null in local-branch mode. Do not place branch names in pr_number; keep
branch/target details in review_body and, if the shared JSON builder
supports it, a separate metadata field such as target.
Phrase the review_body lead-in to reflect what was reviewed and that
nothing was posted. In PR mode, write:
Multi-skill dry-run review of PR #<N> — <K> parallel passes (<list of skills>). Findings below; no inline comments posted.
In local-branch mode, write:
Multi-skill dry-run review of branch <head_ref> vs <base_ref> (no PR yet) — <K> parallel passes (<list of skills>). Findings below; no inline comments posted. Merge/CI health not checked — rerun after opening the PR for that.
Step 7: Render the findings — write the sentinel file and return the report
A terminal failure after target resolution bypasses this ordinary renderer.
Follow the shared Terminal failure delivery section with
--mode no-comments; the helper writes the canonical blocking HEAD sentinel
and returns the exact failure report before exiting nonzero.
Do NOT invoke post_review.py. Do NOT call any gh api .../reviews or
gh pr review command. The foreground step has zero GitHub side effects;
repo-review-follow-up.md owns the narrowly scoped late-finding exception.
Transform the JSON payload at the exact printed findings path into a Markdown
report. The
report is both written to a sentinel file and returned as your final
message (the main agent prints it for the user). The retained
pre-pr-review-gate.sh parser validates this filename after its local
PreToolUse registration is restored.
Render through the deterministic helper. Do not hand-write the sentinel
path, reconstruct Markdown, or embed the payload in a generated shell program.
The helper validates the isolated payload, derives Git/progress state, writes
the canonical sentinel atomically, removes only that payload, and prints the
exact foreground deliverable:
For the Step 1 zero-diff path, which intentionally has no findings payload,
invoke the same helper with --zero-diff --target <target> --reviewed-head <head-sha> and omit --payload, --skill-count, --next-step, and
--remove-payload.
Execute the applicable form once and return its stdout verbatim. The result ends with the
absolute or repository-relative canonical Sentinel: <path> line. The layout
below documents helper output; it is not an instruction to generate another
report.
Write the report to $REVIEW_PATH using this layout:
# repo-review-full-no-comments — <target><review_bodyverbatim, includingthe `## PRhealth` sectionifpresent>## Inline findings (would be posted by `/repo-review-full`)### `<path>`-**L<line>** — <body>-**L<line>** — <body>### `<otherpath>`-**L<line>** — <body>## Summary- B BLOCK, W WARN, N NIT across K skills
- PR-health flags: <Mmerge-conflict / Ffailing-check> (omit if zero or in local-branch mode)
- Reviewed at: <full-sha-from-git-rev-parse-HEAD>-<next-steptip>
For <target> in the header, use PR #<N> in PR mode or
branch <head_ref> in local-branch mode.
For <next-step tip> in the Summary section, use (substitute
<REVIEW_PATH> with the actual path computed above — do not emit the literal
placeholder):
PR mode: Run /repo-review-full <N> to post these as inline review comments.
Local-branch mode: Open a PR, then run /repo-review-full to post these as inline review comments if desired. Review sentinel: <REVIEW_PATH>.
If there are any findings or PR-health flags, append: After remediation and relevant checks, commit and push coherent progress before re-running this review or ending the session. If that is unsafe or impossible, state the blocker instead of retrying unchanged.
Rules for the rendering:
Group inline findings by path, then list each finding as **L<line>** — <body>.
Use the same body text you put into the JSON (**[<short-tag>:<severity>]** <description>).
Preserve the PR-health and ## Nits bullets from review_body verbatim —
they are important for human reviewers and easy to lose if you re-summarize.
NITs stay in the body here for the same reason they do in PR mode: the
pre-PR gate must never see a :nit] tag it could mistake for an obligation.
Pi PASS report. When Pi has no findings or PR-health flags, preserve the
complete ## Pi review audit section from review_body between the PASS
line and ## Summary; a successful review must not discard its model,
attempt, agent-id, or transcript evidence. The fixed short form below is
only for zero-diff reviews, which skip worker allocation and audit rows.
PASS short form. If is_zero_diff == true, still write the sentinel
file. A non-zero diff with no findings keeps the complete Pi audit above
instead of using this short form. The gate's size guard rejects files under
200 bytes, and the header + PASS line + Reviewed at: line are
~130 bytes — pad with a one-line context summary so the total is ≥200
bytes. Use this exact template (substitute <target> and <sha>):
The sentinel file is the gate's contract; your returned report is the human
deliverable. Always produce both.
On ordinary success, --remove-payload removes the exact isolated findings
file. Terminal failure delivery remains responsible for its own cleanup.
Return value. Reply with the full Markdown report (the exact content you
wrote to the sentinel) followed by a final line: Sentinel: <REVIEW_PATH>.
The main agent prints the report and surfaces the path. Return the rendered
report as data — do not summarize or re-narrate it.
