PR code review agent for any GitHub repository. The target repo is named at
onboarding and stored in work/CONFIG.md. It runs on the Claude Code, pi,
Bob and Copilot harnesses (docs/logging.md → Harness adapters),
uses the GitHub CLI (gh) to fetch open pull requests and delivers a structured
review to the chat UI and the GitHub PR thread. Optionally — opt-in at
onboarding — it also nudges reviewers on Slack when PRs wait too long for human
review (the PR Shepherd role).
Six kinds of scheduled run exist. Five start with the deterministic
pre-flight script scripts/preflight.sh, and the
health monitor with scripts/monitor.sh. Both scripts
only detect: they make no GitHub writes and compute the run's worklist. The
one exception is the monitor's opt-in direct DM and dashboard publish (below).
With work, the agent performs all of it per the docs/ procedures.
An empty worklist starts no session at all. Every schedule except the weekly
audit runs its script as its platform precheck
(scripts/precheck.sh): the gate's exit code decides
whether the model is woken, so idle ticks cost nothing, and a started run
receives the worklist the gate already computed. Bookkeeping that nobody waits
on — prunes, self-heals, status resets —
rides along with the next run that has real work
(docs/worklist.md → The schedule gate). Every
scheduled session carries a title that says what it does — Review PR #1457,
Triage CI on PR #1457, Audit agent health 2026-10-09
(docs/runbook.md → Session titles).
Review heartbeat — every 5 minutes, around the clock by default. An
optional active window checks every 15 minutes in the quiet hours outside it.
preflight.sh review lists open non-draft PRs in one REST call and decides per
PR what is due:
- Never-reviewed PRs get a first review automatically. Already-reviewed PRs get
a re-review only on an explicit trigger — the
rereview_label, the optionalrereview_delta_label, or GitHub's "Re-request review" whenrereview_triggerenables it. New commits alone just flip the tracking row toawaiting_label. - Same-HEAD PRs are skipped via
work/REVIEWS.mdplus the remote dedup-marker check, with self-heal. Closed and merged PRs are verified per PR and queued for pruning. PRs carrying the optionalurgent_labeljump the queue and get a rapid, criticals-only review that approves when it finds no critical, before the full one. - The agent then follows
docs/review.md+docs/skills.md: context and diff, clone, every configured review skill, HEAD freshness re-verified right before posting, one GitHub review (summary + inline comments, signed withbot_display_name, carrying the hidden dedup marker), label bookkeeping, and tracking inwork/REVIEWS.mdandwork/reviews/pr-<number>.md. Reviews are concise and assume agent-written, agent-read code unless memory names another audience. Re-review scope follows the trigger: the label asks for the whole PR again, a review request, the delta label or an on-demand ask for the delta only. - The same heartbeat answers comments and PR descriptions addressed to the bot:
questions get a reply, explicit review feedback is recorded to memory and
confirmed in the reply, and "please re-review" is served on demand
(
docs/mentions.md). - With the admin's opt-in (
auto_merge, off by default), the same heartbeat merges a PR a person labeled withauto_merge_labelonce the review approved it as a quick check and every gate holds — clean mergeable state, green checks, size cap, no.github/or sensitive path (docs/auto-merge.md). - With the admin's opt-in (
agent_fixes, off by default), a person'sagent_fix_labelasks for one round in which the agent fixes its own blocking findings and pushes one commit to the PR branch — never to a fork, never past a commit pushed meanwhile; such a PR never auto-merges (docs/agent-fixes.md).
Shepherd sweep — hourly on working days and hours; runs only when Slack
notifications are enabled. preflight.sh shepherd classifies every open
non-draft PR from independent reviews and applies the age gate, cooldown and
escalation ladder, flagging merge-conflicted PRs for an author-directed rebase
nudge. The agent sends each due nudge to the shared Slack channel and records
it immediately after the send — send-then-record, roster-only mentions
(docs/shepherd.md). A reviewer nudge carries the
review's triage — estimated minutes, what a person must judge, what was
checked — and shepherd_scope: needs_human skips PRs the review called a
quick check.
Health monitor — every 15 minutes; a tick with no alert starts no
session. monitor.sh reads the local logs for model and GitHub outages,
classifier refusals, a silent review schedule, a failing backup, harness drift
and stalled or anomalous reviews, and keeps 90 days of hourly technical history
and incidents. With the opt-ins it sends its alerts itself as a templated
Slack DM, so they arrive during a model outage, and publishes a health
dashboard (docs/monitor.md,
docs/health.md).
Weekly audit — Friday morning by default, on while audit_report is
enabled (the default). preflight.sh audit computes 7-day statistics and
deterministic health checks — auth, missed heartbeats, error logs, state
consistency, verdict and findings shifts, skill freshness, 👍/👎 reactions.
The agent adds the judgment checks — schedules, memory-rule compliance, nudge
integrity, lessons from 👎-flagged findings — and sends a traffic-light report
to Slack when enabled, and to the chat UI always
(docs/audit.md).
Codebase survey — weekly in a quiet hour, gated by survey. A review reads
a diff; a survey reads one area of the repository as it stands, traces its
inputs to their dangerous sinks, and reports only the dangerous findings a diff
structurally cannot show — security holes, data loss, reachable failures. The area is chosen deterministically (never
surveyed first, then the oldest pass), the pass is capped before it starts, and
the history accumulates in one artifact with a stable URL
(docs/survey.md).
Weekly trends — every audit also appends its measured week to
work/audit/ and republishes one accumulated artifact: volume, verdicts,
findings and their acceptance, latency, cost, stalls and errors, plus the
repository's own health — how much of what merged was reviewed, PR size, the
wait for a first human review — against the previous week and the 4-week
average. The history is append-only and the step makes no model call
(docs/trends.md).
Model benchmark — the 1st of the month by default; runs only when the
benchmark key is enabled. The agent replays ≥5 synthetic review fixtures
with known seeded defects through its full pipeline, measures time and tokens
per review, scores each output against the fixture's manifest, and republishes
an accumulated report, kept with full provenance so review quality stays
comparable across model, harness and definition versions. Unrecorded trial
runs score feature branches during development, several at once
(docs/benchmark.md).
The definition is split so the always-loaded part stays small.
CLAUDE.md holds only the run types and the rule to read
docs/runbook.md once preflight or the monitor reports
work; the runbook holds the run procedures and the hard invariants; every other
procedure lives in its own docs/ file, read only when the matching work
happens — a review run reads exactly the read_set preflight computes for its
worklist. An idle heartbeat loads nothing.
The definition is versioned (VERSION +
CHANGELOG.md). At updates, on demand, and before any
self-modification the agent checks version freshness, applies the changelog's
upgrade steps, and warns the operator when the instance is outdated
(docs/persistence.md → Definition version &
upgrade).
Feedback persists. What the user says in chat, or in a PR comment addressed
to the bot, goes to work/MEMORY.md (global) or to that PR's
## PR-local overrides (PR-specific), so later runs respect it and never
re-flag a dismissed finding. The agent also learns passively: generalizable
insights from human reviews and author replies are recorded as observed
insights, and the weekly audit consolidates them — merge, promote to rules,
drop stale — so memory stays useful and bounded
(docs/preferences.md).
Bringing up a new code-guardian agent takes three steps — or one, on a
platform whose starter kit catalog offers this repository: creating the
agent from the Code Guardian kit grants the connections, registers the
schedules and points the first turn at ONBOARDING.md,
which then asks for the repository to review.
Bringing it up by hand:
-
Create the agent on the platform, with GitHub — and optionally Slack — connections granted (see Connections).
-
Grab the link to
ONBOARDING.md— from the repo or fork you actually deploy from, for examplehttps://github.com/<your-org>/code-guardian/blob/main/ONBOARDING.md. The agent derives its definition repo from this URL, host included, so a GitHub Enterprise URL works the same; it is stored asdefinition_repo, so a fork's agent stays pinned to the fork and never resets itself to upstream. -
Tell the agent, in its first message:
Here is a file — read it and set yourself up according to it: https://github.com//code-guardian/blob/main/ONBOARDING.md
That is a complete initialization. The agent checks out its own definition,
wires up work/, walks you through a short configuration dialog (bot name,
review marker, review skills, Slack — each value lands in work/CONFIG.md),
registers the schedules (the review heartbeat, the Friday audit, the 15-minute
health monitor, plus the hourly shepherd sweep, the monthly benchmark and the
weekly survey when their keys are enabled), and marks itself onboarded so it
never repeats the process.
One onboarding question is: do you want Slack notifications? — the PR Shepherd reviewer nudging, and the switch for every other Slack message.
- No, or no answer → the agent never touches Slack. Reviews, reports and alerts still land in the chat UI and on GitHub.
- Yes → the agent builds with you the reviewer roster
(
work/DEVELOPERS.md), the only set of people it may ever @-mention: the GitHub logins of an org team you pick (else the repo's top contributors), seed expertise from their recent PRs, and each member's Slack member ID, which you paste (in Slack: profile → ⋮ → Copy member ID).
The decision is stored in work/CONFIG.md and can be changed later by telling
the agent; it flips the flag, and builds the roster on first enable.
Independently of this opt-in, anyone in the connected channel — or, in a
GitHub comment @-mentioning the bot, an account mention_authors admits — can
ask the agent to review a specific PR, equivalent to adding the re-review
label, including restarting a stuck review (docs/review-on-demand.md →
On-demand review). Any other change request from a channel is declined
and reported to the operator; under
definition_issues: enabled it is also filed as an anonymous tracking issue on
the definition repo, with the link in the reply (docs/runbook.md →
Instruction sources & trust boundary).
The agent reads no environment variables of its own: onboarding asks for
the target repository and the optional state-backup repository and stores both
in work/CONFIG.md (github_repo, work_repo), which every scheduled run
reads from a fresh shell. GH_HOST is the only ambient value it honours, and
only to name the default GitHub host.
The definition is project-agnostic: everything specific to one deployment lives
in work/CONFIG.md, which onboarding fills in interactively, auto-detecting
what it can and asking for the rest. Per-key semantics are in
docs/config.md; summary:
| Key | Filled at onboarding by | Purpose |
|---|---|---|
github_repo |
Step 0 answer | Target-repo reference ([host/]owner/repo) — the only source the runtime reads. |
work_repo |
Step 0 answer (optional) | Repo backing work/ up ([host/]owner/repo); omitted = local-only persistence (docs/persistence.md). |
definition_repo |
derived from the ONBOARDING.md URL, host included | The repo this definition came from (fork-aware) — outer-repo origin, target of definition PRs, review-footer link. |
definition_branch |
derived from the ONBOARDING.md URL, else main |
Branch of definition_repo this instance runs from. A per-agent deployment choice; definition PRs are still based on main. |
bot_login |
auto-detected via GraphQL viewer (else gh api user), confirmed |
GitHub account people mention and request — "independent reviewer" classification. A GitHub App token can post beside a user bot_login; an App bot_login uses the labels (docs/github-access.md → Triggers by identity). |
bot_display_name |
asked (default Code Guardian) |
Name the agent signs reviews with. Cosmetic only. |
review_marker |
asked (default code-guardian:review) |
Prefix of the hidden dedup marker in every posted review. Immutable once the first review is posted. |
rereview_label |
asked (default code-guardian:rereview) |
PR label that requests a complete re-review. Without a trigger, new commits are not re-reviewed. The agent removes the label once the re-review is posted. |
rereview_trigger |
asked with rereview_label (default label) |
How re-reviews are requested: label, review-request (needs the bot as a collaborator), or both. A served review request clears itself. A GitHub App bot_login is label-only. |
rereview_delta_label |
asked (default off) | PR label that requests a delta re-review — the changes since the last review, as a review request does. The delta trigger of a GitHub App bot_login. The agent removes it once the re-review is posted. |
urgent_label |
asked with the labels (default off) | Optional human-managed label marking a PR urgent: its due reviews jump the queue and run rapid-first (a rapid pass without a critical finding approves), and with Slack enabled a newly urgent PR gets one immediate mention-free channel alert (docs/review-urgent.md → Urgent PRs). |
review_progress |
asked (default disabled) |
Publishes each review's progress to the PR as a check run (GitHub App) or a commit status (user token) on the reviewed SHA (docs/review-bookkeeping.md → Progress signal on GitHub). Always success when it finishes, so it never gates a merge; the context is the instance's review_marker. |
ci_triage |
asked (default disabled) |
After a review posts, a failing check on the reviewed SHA gets one comment with the probable cause and the smallest fix (docs/ci-triage.md). Reads and explains only — it never restarts a job, changes a label, or changes a verdict. |
review_dispatch |
asked (default enabled) |
A heartbeat that finds several PRs keeps its first and starts every other PR at once in a one-time session of its own (docs/worklist.md → Dispatch); disabled keeps them in the run, one after another. |
mention_replies |
asked (default enabled) |
GitHub comments addressed to the bot are answered every heartbeat — replies, feedback recorded to memory, review requests served (docs/mentions.md). |
mention_authors |
asked (default collaborators) |
Whose GitHub mentions are handled: repository owners, members and collaborators, or anyone (docs/config.md). |
project_profile |
asked (default enabled) |
Generated map of the reviewed repository (work/PROFILE.md), kept current by a structural fingerprint and handed to every review and skill subagent — orientation only, never evidence (docs/profile.md). |
guide_label |
asked with the review guide row (default off) | PR label that asks for the review guide whatever the review's triage; it stays on the PR (docs/artifact.md → Review guide). |
## Review skills table |
defaulted to the public set (issue-fit + review-guide + doc-drift + typescript-engineering + react-ui-engineering), operator-adjustable, every row validated | Per-PR review skills: name, per-skill source ([host/]owner/repo, or harness), trigger (always or an extension list), and the review-section heading. The definition holds the mechanics; this table defines what runs when and from where. Onboarding ends with the offer of one review skill made for the repository from an interview (docs/tailored-skill.md). |
## Watch rules table |
not filled — added later in chat when a team asks | Instance-local "when a PR does X, give a heads-up in Y" rules, delivered to vetted targets: chat UI, a Slack channel, or a PR comment (docs/watches.md). Keeps team-specific triggers out of this public definition. |
slack_notifications |
asked (default disabled) |
Gates all Slack activity (shepherd nudging, watch notifications). |
audit_report |
defaulted to enabled |
Weekly health check and report (Slack when enabled, chat UI otherwise). |
definition_issues |
asked (default disabled) |
Whether the agent files anonymous tracking issues on the definition repo — audit-found definition bugs and refused channel requests; scripts/definition-issue.sh blocks any draft that names instance data (docs/runbook.md → Definition-repo issues). |
benchmark |
asked (default off) | Monthly self-benchmark of the review pipeline on ≥5 synthetic fixtures with known defects, time and tokens measured per review (docs/benchmark.md). |
benchmark_judge |
asked with benchmark (default off) |
Pinned model id for the LLM-judged quality scores, run through the claude CLI (or pi on the same model as the fallback) whatever model the reviews use; off = deterministic scoring only. |
merge_ready_nudge |
asked with Slack (default disabled) |
One Slack line when a PR is approved, conflict-free, green and carries no open critical of the agent's own — once per approval, to the author (docs/shepherd.md → Ready to land). |
shepherd_scope |
asked with Slack (default all) |
needs_human limits reviewer nudges to PRs whose current review did not call a quick check (docs/shepherd.md → Scope and brief). |
human_review_paths |
asked with auto-merge or Slack (default none) | Globs of files that always make a PR's triage needs-human (docs/review-mechanics.md → Summary body format). |
auto_merge, auto_merge_label, auto_merge_max_lines, auto_merge_method |
asked of the repo admin (default off) | Merges a PR a person labeled, that the review approved as a quick check and that passes every gate — clean mergeable state, green checks, size cap, no .github/ or sensitive path (docs/auto-merge.md). |
agent_fixes, agent_fix_label |
asked of the repo admin (default off) | A person's label asks for one round in which the agent fixes its own blocking findings and pushes one commit to the PR branch, never to a fork; such a PR never auto-merges (docs/agent-fixes.md). |
survey |
asked (default disabled) |
Weekly security pass over one area of the repository — security holes, data loss and reachable failures, traced from input to sink; no suggestion, no refactor (docs/survey.md). One area per run, capped, read-only: it never changes code and never posts on a PR. |
survey_report |
asked with survey (default dam) |
Surface for the accumulated survey artifact, updated in place at a stable URL: dam or off. |
survey_report_findings |
asked with survey_report: dam (default disabled) |
Whether the public survey artifact shows the findings themselves, each code location linked to the commit the pass read, or stays an index of areas and passes. |
survey_interval_days |
not set (= 7) |
The floor between two survey passes, so a drifting cron never surveys twice in one interval. |
audit_trend |
defaulted to dam |
Surface for the weekly trend artifact, updated in place at a stable URL: dam or off. |
benchmark_report |
asked with benchmark (default dam) |
Surface for the accumulated report artifact, updated in place at a stable URL: dam or off. |
## Benchmark model prices table |
offered with benchmark (optional) |
Per-MTok USD prices that price the est $ columns of the benchmark report and the weekly trends (docs/benchmark.md → Model prices). |
active_hours, active_days, review_interval_active, review_interval_quiet |
asked (default 24/7: 00-23, Mon-Sun, 5 min; 15 min quiet under a window) |
The heartbeat's two cadences and the window between them; each interval bounds how long a new PR waits for its run. They are the source of truth for the registered crons (ONBOARDING.md Step 6a) — an edited key takes effect once the schedules are re-registered. |
stall_alert_threshold |
not set (= 4) |
Stalled reviews within 24 h that record one history incident (no alert), at most once per UTC day; 0/off disables. |
review_anomaly_factor |
not set (= 4) |
A review whose cost, time, context, repeated calls, failures or largest tool result reaches max(floor, this × its model's median) records one history incident (no alert); 0/off disables. |
model_alert_threshold |
not set (= 2) |
Runs ended on an API failure within one outage before the monitor alerts; 0/off disables. |
github_alert_threshold |
not set (= 2) |
Gate passes in a row that could not list the open PRs before the monitor alerts; 0/off disables. |
classifier_alert_threshold |
not set (= 20) |
Auto-mode classifier refusals within 24 h that record one history incident per UTC day (no alert); 0/off disables. |
log_level |
not set (= info) |
Verbosity of the structured events log work/logs/events-*.jsonl (docs/logging.md); debug also records successful external tool calls. |
escalation_owner |
asked (only when Slack enabled) | Roster member @-mentioned at nudge level 4, and the fallback DM target of the monitor's alerts. |
monitor_owner |
asked with the escalation owner (default: not set) | Roster member the monitor's alert DMs go to; not set = escalation_owner. |
monitor_slack_direct |
asked with the escalation owner (default disabled) |
The monitor sends its alerts itself as a templated DM to the monitor owner, so they arrive during a model outage (docs/monitor.md → Direct delivery). |
health_dashboard |
asked (default off) |
dam publishes the health dashboard — service status, 90 days of hourly history, incidents — updated in place (docs/health.md). |
-
Platform: the agent assumes the DAM agent infrastructure —
$HOMEat/home/agenton a persistent/workspacevolume, the platform's outbound auth proxy for GitHub tokens, and themcp__platform-outbound__*tools for schedules, one-time review sessions, session titles and Slack. The monitor's opt-ins call that server from a script, at theplatform-outboundURL of~/.mcp.json(scripts/lib/mcp.sh). The benchmark runs its reviews in platform sub-agents through the platform's Driver REST routes beside that URL (docs/benchmark.md→ Isolation). Running elsewhere requires adapting those assumptions. -
GitHub hosts: every repo reference is
[<host>/]<owner>/<repo>, so the target repo, this definition, the skill sources and thework_repobackup may each live on a different host —github.comor a GitHub Enterprise instance. Each host in play must be authenticated separately (gh auth login --hostname <host>, operator-only) and reachable from the pod; a non-github.comtarget host is persisted asGH_HOSTat onboarding. -
GitHub identity: the agent posts reviews and comments as the account behind its token — a user token or a GitHub App installation token, both supported (
docs/github-access.md). Use a dedicated machine or bot account, not a personal one, that is a collaborator on the target repo with permission to review PRs. GitHub ignores review requests and approvals from a PR's own author, so the bot account must not open the PRs it reviews. When it does — one GitHub App that opens and reviews PRs — the agent posts its verdict as a comment (docs/github-access.md→ Own pull request). -
Token scopes — the single place these are specified:
Scope Required? What needs it repoyes PRs, reviews, comments, labels and issues on the target repo; push to the work_repobackup and the definition reporead:orgoptional Onboarding only: lists your org's teams to seed the reviewer roster. Without it onboarding falls back to the repo's top contributors; no scheduled run uses it The audit's
token_scopescheck asserts the required ones only. A missing scope is operator-only to fix: the agent reports it and never works around it. A fine-grained or GitHub App token has permissions in place of scopes — the list and the access check that reads them are indocs/github-access.md→ Feature map. -
External services: artifacts and reports publish to the platform's DAM Artifact Library (best-effort, behind the owner's experimental flag), whose
visibility:"public"share URL is reachable by anyone holding it. The HTML passes a redaction pass before the publish (docs/artifact.md).
- A GitHub connection must be granted so
ghcan authenticate (the Envoy sidecar injects the token, a user's or a GitHub App's, on outbound GitHub requests). The same token serves the target repo, thework_repobackup and the definition repo. - A Slack connection is optional, needed only when Slack notifications
are enabled, so
mcp__platform-outbound__send_channel_messagecan reach the shared channel. With notifications disabled — the default — the agent never calls Slack and the full review pipeline runs unaffected.
work/MEMORY.md, work/REVIEWS.md and work/reviews/ hold the agent's
learned preferences and review history. work/CONFIG.md and — with Slack
enabled — work/DEVELOPERS.md (reviewer roster) and work/SHEPHERD.md (nudge
ledger) live alongside them. They sit on the /workspace PVC at runtime
(mounted as /home/agent/work/), so they survive pod restarts. How work/ is
seeded depends on the work_repo key:
- Set —
work/is a plain data directory backed up to that repo after every run, via a disposable pod-local clone, giving durable, versioned, cross-pod history. Naming an existing backup at onboarding continues that agent: state and logs return verified, and the accumulated reports are published again (docs/persistence.md→ After a restore). - Omitted —
REVIEWS.mdandreviews/are reconstructed from the agent's marker-carrying reviews on the target repo;MEMORY.md, which is not derivable from PRs, starts from the seed template inONBOARDING.md(Step 3b).
work/ is not tracked by this definition repo: the allowlist .gitignore
hides everything under it, so the two never collide and a definition update
never touches live runtime state (docs/persistence.md → Two stores: shared
live state + durable backup). Treat work/ as confidential — it may hold
private data (roster Slack IDs, preferences, logs), and it leaves the agent
only via this backup or the configured output surfaces (docs/runbook.md →
Hard invariants).
CLAUDE.md— the bootstrap loaded on every run: repo resolution, run types, and the rule to read the runbook once preflight or the monitor reports work.docs/runbook.md— the operating manual read only then: run procedures, trust boundary, hard invariants, and the map ofdocs/.docs/worklist.md— the schedule gate, the entry command and the worklist contract, read by a run without a gated worklist.docs/— every other procedure, read on demand (docs/runbook.md→ Map ofdocs/).scripts/preflight.sh— deterministic pre-flight for every run type but the monitor; detects work, never acts on GitHub.scripts/monitor.sh— the monitor's detector: the health alerts, from local files only (docs/monitor.md), plus the opt-in direct DM and dashboard publish (docs/health.md).scripts/precheck.sh— the schedule gate: one pass of either detector before the session starts, an exit code that skips an idle fire, and the worklist path a started run reads.scripts/review-pr.sh— the mechanical half of one PR review: guards, lock, context, clone, post, holds, merges and agent fixes, one subcommand each (its header lists them). The agent decides what the review says.scripts/verify-onboarding.sh— one-shot, read-only post-onboarding check (ONBOARDING Step 7): the checkout andwork/against the templates, and with--livethe environment a run needs. PrintsFAIL … — fix: …lines.scripts/log.sh+scripts/harness/— the structured events log (work/logs/events-*.jsonl, 14-day retention) and the per-harness adapters — on Claude Code the hooks that auto-capture failed tool calls and keep unused tools out of every request, on pi an extension that runs the same hooks, plus the subagent fan-out and the benchmark's token snapshot.- The other
scripts/— profile, dispatch, backup, survey, benchmark, trends, health, definition-issue — each state their contract in their header and theirdocs/home. scripts/tests/+.github/workflows/ci.yml— deterministic stub tests (gh/curl faked, offline) and the CI that runs them plus thedocs/self-modification.md§9 sweeps on every PR..agents/skills/— skills bundled with the definition:issue-fit, a review skill installed through a## Review skillsrow whosesourceis the instance'sdefinition_repo— does the diff deliver what the linked issue asked;review-guide, a guide skill installed the same way — for a PR whose review needs a person, a visual page linked from the review: why a person, the grounded risks, the review's verified findings, the project map with the change marked, then one-line sections that open to a picture each (why the PR exists, asked vs delivered, behavior before → after, the dependency chain, decisions, fit with the project's records, problems met, how it is proven, churn, impact, rollout and rollback, limits); the maps it draws stay inwork/diagrams/as project knowledge (docs/artifact.md → Review guide);review-remediation, the author side — install it in whichever agent answers the reviews, to fix every blocking finding at every location and run the checks the review carries. It tells reviewers apart by login and review marker, so a GitHub App login that posts for several agents works too.pr-openis its first-push counterpart: it self-reviews the branch and writes a PR body with the scope, a guide for a human reviewer and the person's decisions in a hiddenauthor-decisionsblock, which the reviewer reads. Only protocol skills live here; a domain skill (language, framework, OWASP or WCAG checklist) lives in its team's skills repository or in the reviewed repository (docs/skills.md→ Where a skill lives).ONBOARDING.md— the first-run setup runbook; it also carries thework/MEMORY.mdandwork/REVIEWS.mdseed templates (Step 3b).kit.yaml— the starter kit this definition is offered as: the connections it asks for, the schedules it registers, the bundled skills and the seed.VERSION+CHANGELOG.md— definition semver, bumped with every change, and the per-version upgrade steps instances apply.
Apache License 2.0 — see LICENSE.