Skill

echo-review

Review a GitHub PR and post a PENDING review with inline comments. Applies a universal best-practices floor and, if present in the target repo, team-mined patterns from .echoreview/patterns.md. Use when the user types /echo-review <pr> or asks you to review a PR for them.

Invocation

How this skill is triggered — by the user, by Claude, or both

Slash command

/echoreview:review

User invocable

Model invocable

Inline context

Default effort

Context Preview

The summary Claude sees in its skill listing — used to decide when to auto-load this skill

Review a single GitHub PR and post a **PENDING** review back to GitHub. The human always clicks Submit — never finalize a review automatically.

Supporting Files

references/comment-template.mdreferences/re-review.mdreferences/reasoning-examples.mdreferences/reasoning-pass.mdreferences/universal-best-practices.mdscripts/build-diff-map.shscripts/extract-context.shscripts/fetch-comments.shscripts/submit-review.sh

SKILL.md

167 lines · ~2.3k tokens

Stats

LanguageShell

Parent stars0

MaintenanceGood

Last CommitMay 23, 2026

Actions

View Source View Plugin View on GitHub View README

Stats

Actions

echo-review

Review a single GitHub PR and post a PENDING review back to GitHub. The human always clicks Submit — never finalize a review automatically.

This skill walks four phases with two user checkpoints (after Phase 1 and before Phase 4 submission). Skip neither.

Inputs

/echo-review <pr> where <pr> is one of:

A full PR URL: https://github.com/OWNER/REPO/pull/N
An OWNER/REPO#N reference
A bare number N (only valid when the current working directory is a clone of the target repo)

If the user gave no argument, ask for one. Don't guess.

Phase 0 — Setup (silent, do this before Phase 1)

Resolve the input to OWNER, REPO, NUMBER:
- URL form: parse OWNER, REPO, N from the URL path.
- OWNER/REPO#N form: split on / and #.
- Bare number: run gh repo view --json owner,name --jq '"\(.owner.login)/\(.name)"' from the current working directory to fill in OWNER/REPO; if that fails (not in a repo), tell the user and stop.
Verify gh auth status succeeds. If not, instruct the user to run gh auth login and stop.
Set WORK_DIR=/tmp/echoreview-${NUMBER} and mkdir -p it.

Phase 1 — Context extract

Run the two extract scripts in parallel from ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/:

extract-context.sh "$OWNER" "$REPO" "$NUMBER" — writes metadata.json, diff.patch, files.txt, file-types.json to $WORK_DIR.
fetch-comments.sh "$OWNER" "$REPO" "$NUMBER" — writes existing-comments.json, existing-reviews.json, review-iteration.txt, previous-comments.md to $WORK_DIR.

After both complete:

Read metadata.json for the PR title, author, base/head refs, and stats.
Read file-types.json and tally counts per type field.
Read review-iteration.txt. If 2, this is a re-review — load references/re-review.md and read previous-comments.md.

Checkpoint 1 — confirm before proceeding

Print a summary to the user that looks like:

PR #<N>: <title>
  Author: @<login>
  Base ← Head: <base> ← <head>
  Files changed: <count> (<type breakdown>)
  Prior comments: <count> (<iteration label: first review | re-review>)

Continue with the review? (y/N)

Pause and wait for the user's explicit y/N. If anything other than y / yes / Y, stop and report no review was created.

Phase 2 — Pattern review

Load references/universal-best-practices.md.
From the user's current working directory, check whether .echoreview/patterns.md exists. If so, load it. If not, proceed with the universal floor only — this is the expected v0.1 path until echo-extract ships.
Read $WORK_DIR/diff.patch.
Apply both rule sets against the diff. For each finding, capture:
- file (post-rename, new-side path)
- line (new-side line number)
- severity per universal-best-practices.md (or the rule's declared severity from patterns.md)
- source — one of universal, patterns
- pattern_id — only set for patterns-driven findings (e.g. ECHO-ARC-001); used as metadata, never printed in the comment
- comment — drafted per references/comment-template.md. For findings tied to a patterns.md rule, the rule's verbatim evidence quotes are the voice guide for this comment — match their register, length, vocabulary, emoji usage, code-block usage, capitalization, punctuation, and phrasing patterns. Do not impose a house style on top of them.
Write the structured findings to $WORK_DIR/findings.json as a JSON array.

Phase 3 — Reasoning pass

Check the skip conditions from references/reasoning-pass.md. If any matches, skip this phase entirely — do not surface any reasoning-pass findings and do not penalize the PR for being lockfile-only / docs-only / generated / a pure version bump.

Otherwise:

Load references/reasoning-pass.md and references/reasoning-examples.md.
Run the open-ended contradiction-finding pass over $WORK_DIR/diff.patch.
For each candidate finding, assign an explicit numeric confidence (0–100). Drop any finding below 80. Drop any finding whose severity would be suggestion or note — Phase 3 is BLOCKER/WARNING only.
Drop any finding you would have to hedge with "possibly", "could be", "might".
Append surviving findings to $WORK_DIR/findings.json with source: "reasoning".

Phase 4 — Submit

Skip gate — bail before composing a payload if nothing is actionable

Before running build-diff-map.sh, compute inline_findings_count:

New findings in $WORK_DIR/findings.json (each one becomes an inline comment if it maps to the diff).
Plus, in re-review mode (review-iteration.txt == 2), the count of prior items classified as NOT_FIXED or PARTIALLY_FIXED — these become fresh inline comments per references/re-review.md.

FIXED and WONTFIX_ACKNOWLEDGED items do not count — they only ever appear in the resolution table, never as inline comments.

If inline_findings_count == 0:

Do not run build-diff-map.sh.
Do not run submit-review.sh.
Do not show Checkpoint 2.
Print the no-actionable-items status block to the Claude Code session per references/re-review.md (different shape for first-review vs re-review). The resolution table, if any, prints to chat — not to the PR.
Exit cleanly. The PR stays clean.

Posting a "nothing to add" PENDING review is exactly the noise the "no PR summary comment" rule in DESIGN.md forbids.

Otherwise — compose and submit

Run ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/build-diff-map.sh "$OWNER" "$REPO" "$NUMBER" to produce $WORK_DIR/diff-map.json.
For each finding in $WORK_DIR/findings.json:
- Look up "<file>:<line>" in diff-map.json.
- If present, record the comment with path, position, body.
- If missing, demote to the review body under a section ## Additional findings (outside diff range):.
Sort comments by severity (blocker → warning → suggestion → note), then by file:line ascending. No count cap.
If review-iteration.txt is 2, prepend the resolution table from references/re-review.md to the review body. (The resolution table only contextualizes re-flagged items; it never appears on its own — that case is handled by the skip gate above.)
Write $WORK_DIR/review-comments.json (the in-diff comment array) and $WORK_DIR/submission-payload.json with this shape:

{
  "body": "<resolution table + out-of-diff findings, or empty string>",
  "comments": [
    { "path": "<file>", "position": <int>, "body": "<comment body>" }
  ]
}

Do not include an event field. On GitHub's API, omitting event produces a PENDING review; setting it to APPROVE / REQUEST_CHANGES / COMMENT submits the review immediately. submit-review.sh refuses to POST any payload that has an event field set.

Checkpoint 2 — confirm before submission

Print to the user:

The compiled findings list, grouped by severity, each shown as <severity> <file>:<line> <comment> (truncate the comment to one line for the summary).
A count line: <N> inline, <M> in body (out-of-diff).
The question: Submit PENDING review on PR #<N>? (y/N)

Pause and wait for the user's explicit y/N. Anything other than y / yes / Y stops the run; tell the user the payload is on disk at $WORK_DIR/submission-payload.json if they want to inspect.

On y:

Run ${CLAUDE_PLUGIN_ROOT}/skills/review/scripts/submit-review.sh "$OWNER" "$REPO" "$NUMBER". It POSTs the payload.
Print the resulting PR URL (read from metadata.json .html_url) so the user can click through and Submit on GitHub.

Output to the user (final message)

Short — three lines:

Posted PENDING review on PR #<N>: <html_url>
<inline_count> inline comments, <body_count> in body.
Click "Submit review" on GitHub when you're ready.

Do not write a celebration, a next-steps list, or a summary of what the PR was about.

Working directory

All intermediates live in /tmp/echoreview-${NUMBER}/. Never write inside the user's repo. The user owns cleanup — don't auto-delete.

What this skill must not do

Do not change the event to COMMENT, APPROVE, or REQUEST_CHANGES. PENDING only.
Do not skip either checkpoint.
Do not invent skip conditions beyond those listed in references/reasoning-pass.md.
Do not include pattern IDs in posted comment bodies, impersonate named reviewers, or harass the PR author (see references/comment-template.md for the three guardrails).
Do not cap the number of comments. Trust the severity declarations and the reasoning-pass threshold.

echo-review

Invocation

Context Preview

Supporting Files

SKILL.md

echo-review

Invocation

Context Preview

Supporting Files

SKILL.md

echo-review

Inputs

Phase 0 — Setup (silent, do this before Phase 1)

Phase 1 — Context extract

Checkpoint 1 — confirm before proceeding

Phase 2 — Pattern review

Phase 3 — Reasoning pass

Phase 4 — Submit

Skip gate — bail before composing a payload if nothing is actionable

Otherwise — compose and submit

Checkpoint 2 — confirm before submission

Output to the user (final message)

Working directory

What this skill must not do

Similar Skills

echo-review

Inputs

Phase 0 — Setup (silent, do this before Phase 1)

Phase 1 — Context extract

Checkpoint 1 — confirm before proceeding

Phase 2 — Pattern review

Phase 3 — Reasoning pass

Phase 4 — Submit

Skip gate — bail before composing a payload if nothing is actionable

Otherwise — compose and submit

Checkpoint 2 — confirm before submission

Output to the user (final message)

Working directory

What this skill must not do

Similar Skills