Carbon Monitor

Agentic AI pull request reviews, grounded by Carbon MCP.

Carbon Monitor is two things. It is a working reviewer for pull requests in
carbon-design-system/carbon and carbon-design-system/carbon-ai-chat (the
repositories in GITHUB_AI_AGENT_REPOS): each open pull request goes through an AI
CLI agent (IBM Bob by default, or Claude), every Carbon-specific claim (component
props, tokens, icons, accessibility patterns) is verified against the Carbon MCP
server, and the result lands on the PR as one review of inline comments and a
summary with a recommendation. A clean PR gets one short verdict comment
instead. It is also a harness for running any agent on a schedule in IBM Cloud, with cost
tracking, telemetry and a Slack report, that you can point at your own
repository or your own job. A skill adapts it for you. It asks up to ten
questions and writes the SPEC you build from, and you keep the Cloud deploy,
cost caps, telemetry and Slack report as they are.

Quick start · Adapt it with a skill · How a review works · Deploy · Architecture · API · Contributing

· Apache-2.0 · Built for the Carbon Design System

---

Supported agents

Agent	CLI	Select with	Key from .env	Cost unit
IBM Bob	Bob Shell CLI (bob run), the default	GITHUB_AI_AGENT_CLI=bob	BOB_API_KEY	Bobcoins
Claude	Claude Code CLI in print mode (claude -p)	GITHUB_AI_AGENT_CLI=claude	ANTHROPIC_API_KEY	US dollars

Both CLIs are in the container image, and one variable picks the reviewer. The
rubric, the prompt, the output parser, the Carbon MCP checks, the schedule and
the console are the same for both. The summary comment says which one ran
(Reviewed by: IBM Bob or Reviewed by: Claude), and cost is recorded in the
agent's own unit. Each agent runs in its own private configuration directory, so
REVIEW_CONCURRENCY applies to both. Another CLI is one adapter file; see
Add an agent.

What it does

Stage	What happens
Fetch	Queries GitHub for open PRs, excluding drafts, bot-authored PRs (Dependabot, Renovate), PRs older than 21 days and PRs already labelled AIReviewed, in each repository in GITHUB_AI_AGENT_REPOS. Up to 10 PRs per repository per run, carbon first, then carbon-ai-chat. Both limits are configurable and apply per repository.
Bundle	Builds a temporary workspace per PR: PR_REVIEW_REQUEST.md (description, changed files, the unified diff, and a list of files left out), pr.json / files.json / diff.patch, the agent rules (.bob/rules/01-output.md for Bob, CLAUDE.md for Claude), and the Carbon Builder skill. A Closes #N, Fixes #N or Resolves #N reference pulls in the linked issue (and its parent) as acceptance criteria, capped at 5,000 tokens. The diff is assembled from each file's patch as whole files fitted to the input budget, tests and docs after source, so no PR is skipped for its size; what does not fit is listed under Files Not Reviewed.
Review	Runs the selected agent in the bundle: bob run --trust --accept-license [--team-id …] "<prompt>" or claude -p … "<prompt>". The agent writes a visible catalogue of every suspicious item per file, classified as Category 1 (Carbon API, needs MCP) or Category 2 (visible in the diff).
Verify	Resolves every catalogue item. Category 1 items are checked with the Carbon Builder skill and Carbon MCP tools (code_search, docs_search, get_charts, labs_search, code_audit); Category 2 items from the diff alone. The parser flags runs that skipped the catalogue or hid their resolutions.
Post	Findings of minor severity or above that map to a diff line go out as inline comments in one GitHub review. The rest, plus the summary and recommendation, go in one PR comment, with any nits in a collapsed list; nits are never inline. A clean PR gets one short verdict comment and no review. A partial review says so in that comment and lists the files it did not cover. If no file fits the budget, no agent runs and the PR gets a Not Reviewed comment. The AIReviewed label is added.
Report	Scheduled runs post a digest to Slack when SLACK_WEBHOOK_URL is set. Review events are written to Elasticsearch for the Kibana dashboard.

Adapt it with a skill

The agent is IBM Bob or Claude. Carbon Monitor is the harness around it:
everything that runs it on a schedule in IBM Cloud, pays for it, logs it, and
reports on it. The
review logic is the one part that is specific to the job, and it sits behind
named seams.

Harness, kept as is	Task layer, replaced per job
Code Engine deployment, image, secret sync	Work-item source (GitHub today)
Daily schedule with budget auto-pause	Workspace, prompt and output parser
Agent runner and MCP token refresh	Write-back and the done marker
Cost ledger, Elasticsearch events, Kibana	Event fields, digest lines, console labels
Slack transport, console shell, sign-in	Health tiles for the external system

Three ways to make it yours:

Path	What changes
Another repository	Configuration, the reviewing account's access, and optionally the rubric and the MCP server
Another job	The task layer, against any system: Jira, ServiceNow, Box, a database, a feed
Another agent	Bob and Claude are built in and switch with GITHUB_AI_AGENT_CLI. Another CLI is one adapter file in src/agents/, plus its install and key; Codex, Gemini CLI and Copilot CLI are mapped

The carbon-monitor-adapt-harness skill walks you through any of them. It
reads the harness map, asks at most ten questions, and writes a SPEC you
implement from. To use it:

1. From a clone of this repository, install the skill into your coding agent.
   No skill support? Say "Read skills/carbon-monitor-adapt-harness/SKILL.md
   and use it as your instructions" instead.

   npx skills add https://github.com/carbon-design-system/Carbon-Monitor/tree/main/skills/carbon-monitor-adapt-harness -a claude-code   # or -a bob, cursor, codex, github-copilot, gemini-cli
   

2. Open your agent in the clone and say "Adapt this harness to my use case"
   (in Claude Code, /carbon-monitor-adapt-harness). Answer the questions;
   "default" accepts a default. It writes specs/<name>/SPEC.md.

3. Start a fresh session and say "Implement specs/<name>/SPEC.md, one task
   at a time, verifying each against its acceptance criteria."

The map is docs/HARNESS.md. The skill is in skills/carbon-monitor-adapt-harness; per-agent
swap notes and two finished example SPECs are in docs/adapt-harness.

What a review looks like

The summary comment on a reviewed PR, followed by the AIReviewed label.

A PR with nothing above nit gets a much shorter comment: a verdict line, the
agent's summary and, when there are nits, a collapsed list of them with title and
location. It carries no recommendation section, findings table or token usage, and
the PR gets no inline comments. Nits are never inline on any PR. If the review's
reliability is in doubt, the PR gets the full summary comment with a reliability
note instead.

Every review ends in one of three recommendations:

Value	Meaning
Consider Revising	One or more blocking findings, or two or more major findings
Suggested Improvements	Any major or minor finding, or review reliability was uncertain
Looks Good	nit-level findings only, or no findings

Each finding carries a severity:

Severity	Meaning
blocking	Breaks user-facing behaviour or violates an accessibility requirement
major	Likely to cause a visible rendering difference or runtime error
minor	Maintainability risk or subtle behaviour difference — not immediately breaking
nit	Convention / hygiene — no functional consequence

The inline, summary and Looks Good comment formats, the partial-review note, and
the reliability checks behind the recommendation, are in
docs/REVIEW_OUTPUT.md.

See it running

The operator console after a clean single-PR run.

Carbon MCP connection: sign in once with IBMid, then the container renews its own token.	Scheduled reviews: one daily run, off until an operator turns it on.

The Slack digest posted after a scheduled run.

The console sign-in page, shown when AUTH_REQUIRED is on.

Who can run this

Reading, testing and adapting the code needs nothing special: npm ci and
npm run test:all run anywhere with Node 22.15 or later. Running a review does
need IBM access. Bob Shell requires the IBM office network or VPN; outside it,
requests fail with Cloudflare errors. Claude talks to the Anthropic API and
needs no IBM network for inference. Carbon MCP authenticates with an IBMid
for either agent, and its token expires, so the console renews it after one
Sign in with IBMid.
On the GitHub side, it needs a token for the reviewing GitHub account (i.e.
machine user), which must hold Triage on every repository it reviews; see docs/GITHUB_ACCESS.md.

Quick start

Prerequisite	Notes
Node.js 24 or later	Required by Bob Shell; the code itself runs on >= 22.15.0
GitHub token	For the reviewing GitHub account (i.e. machine user), which needs the Triage role on every repository in GITHUB_AI_AGENT_REPOS, plus Pull requests and Issues write and Contents read. See docs/GITHUB_ACCESS.md.
Bob Shell CLI	Installed and on PATH, when the agent is bob (the default)
Claude Code CLI	Installed and on PATH, when the agent is claude: npm install -g @anthropic-ai/claude-code. Claude Code 2.1.259 or later (the image pins 2.1.302)
Carbon MCP	Configured in Bob as a streamable-HTTP connection; Claude reads the same credentials (see below)

git clone https://github.com/carbon-design-system/Carbon-Monitor.git
cd Carbon-Monitor
npm ci
cp .env.example .env

Set the two required variables in .env:

GITHUB_AI_AGENT_TOKEN=ghp_your_github_token_here
BOB_API_KEY=your_bob_api_key_here

To review with Claude instead, set the agent and its key (an API key from the
Claude Console; BOB_API_KEY is then not needed):

GITHUB_AI_AGENT_CLI=claude
ANTHROPIC_API_KEY=your_anthropic_api_key_here

One app reviews a list of repositories under GITHUB_AI_AGENT_OWNER. The default
list is carbon; to review carbon-ai-chat as well, set:

GITHUB_AI_AGENT_REPOS=carbon,carbon-ai-chat

List order is batch order, so every carbon PR is queued before any
carbon-ai-chat PR, in scheduled runs, console runs and the budget. The profile
that reviews a PR follows its repository: carbon uses the core profile,
carbon-ai-chat the caic profile, which runs on Bob only. GITHUB_AI_AGENT_REPO
still works as a one-repository fallback when GITHUB_AI_AGENT_REPOS is unset.

Start the console and API, then review one PR from the CLI:

npm run build:client && npm run api   # console + API on http://localhost:3000
npm start -- --pr <number>            # review a single PR
npm start -- --pr carbon-ai-chat#<number>   # name the repository when more than one is listed

npm start with no flag reviews every PR that passes the Fetch filter, in each
listed repository. A bare number is looked up in every listed repository; if it
is open in more than one, or in none, the run stops before any review, and
repo#number says which one you mean.

Configuration

The variables most deployments set. Every variable, with its default, is in
docs/CONFIGURATION.md.

Variable	Required	Default	Description
GITHUB_AI_AGENT_TOKEN	Yes	—	GitHub PAT. Needs Pull requests + Issues write and Contents read, and Triage on every listed repo — see GitHub access
GITHUB_AI_AGENT_CLI	No	bob	Worker agent: bob or claude. Another CLI needs an adapter, see Add an agent
BOB_API_KEY	With bob	—	Bob Shell API key. Create with Scope Inference in the Bob web portal — browser SSO cannot complete in a container
ANTHROPIC_API_KEY	With claude	—	Claude Code API key from the Claude Console. Required when GITHUB_AI_AGENT_CLI=claude; leave blank for Bob
GITHUB_AI_AGENT_OWNER	No	carbon-design-system	Repository owner
GITHUB_AI_AGENT_REPOS	No	carbon	Comma-separated repository names under the owner, in batch order, such as carbon,carbon-ai-chat. Falls back to GITHUB_AI_AGENT_REPO when unset
GITHUB_AI_AGENT_MAX_PRS	No	10	Max PRs reviewed per repository per run
GITHUB_AI_AGENT_DAYS_BACK	No	21	How far back to look for PRs (days), per repository
CLAUDE_MODEL	No	claude-sonnet-5-5	Claude model
CLAUDE_EFFORT	No	high	Claude reasoning effort
CLAUDE_MAX_BUDGET_USD	No	5	Per-review spend cap in USD, passed as --max-budget-usd. 0 disables it. Spend can pass the cap slightly
CLAUDE_MAX_TURNS	No	60	Per-review turn cap, passed as --max-turns. 0 disables it
CLAUDE_ALLOWED_TOOLS	No	read-only set	Read(./**),Glob,Grep,Skill(carbon-builder),mcp__carbon-mcp
CLAUDE_DISALLOWED_TOOLS	No	edit and shell set	Bash,Edit,Write,MultiEdit,NotebookEdit,WebFetch,WebSearch,Agent. Three more deny rules are always added, see docs/CONFIGURATION.md
CLAUDE_MONTHLY_BUDGET_USD	No	unset	Monthly budget in USD for scheduled reviews. Unset means not metered
CLAUDE_BUDGET_LOW_PCT	No	10	Scheduled reviews pause below this percentage of the monthly budget

Bob Shell. Install with the official script,
curl -fsSL https://bob.ibm.com/download/bobshell.sh | bash (IBM network or
VPN), then run bob trust /private/tmp and bob trust /path/to/Carbon-Monitor
so Bob can use MCP tools in the review bundles.

Claude Code. Install with npm install -g @anthropic-ai/claude-code
(Node 22 or later; Claude Code 2.1.259 or later, the image pins 2.1.302), create an API key in the Claude Console, and put it in
.env as ANTHROPIC_API_KEY. The key is read from the environment, so there is
no login step. Claude reviews read-only: file reads, search and the Carbon tools
are allowed, and shell, edits, web access and subagents are denied.

Carbon MCP. Add a carbon-mcp entry of type streamable-http pointing at
https://mcp.carbondesignsystem.com/mcp to ~/.bob/settings/mcp.json, then
check that /mcp list inside bob reports carbon-mcp - Ready. That file is
the one place the credentials live for both agents: for each Claude review the
harness copies them into the bundle's own MCP file.

Carbon Builder skill. npm ci (or npm install) places it in .bob/skills/carbon-builder/
if it is missing. Otherwise follow the
Carbon Builder skill setup guide.

The full mcp.json example is in docs/CONFIGURATION.md;
token permissions are in docs/GITHUB_ACCESS.md.

How a review works

POST /api/review/trigger          202 Accepted at once
  └─► reviewPRs()                 src/index.js
        ├─► GitHub: select open PRs, fetch changed files and patches + linked issue
        ├─► buildReviewBundle()   src/reviewBundle.js
        ├─► runAgent()            src/agentRunner.js, through src/agents/<agent>.js
        │     bob run --trust --accept-license [--team-id …] "<prompt>"
        │     claude -p --output-format stream-json … "<prompt>"
        │     └─► Carbon MCP: code_search · docs_search · get_charts · labs_search · code_audit
        ├─► parseReviewOutput()   src/reviewParser.js
        ├─► post one review of inline comments, then the summary (or verdict) comment, apply the AIReviewed label
        └─► Slack digest (scheduled runs only, when SLACK_WEBHOOK_URL is set)

Only one batch runs at a time, across all listed repositories, with up to
REVIEW_CONCURRENCY reviews (default 2) in flight inside it, for either agent. Two agents reviewing the same PR would post the
same findings twice, so a second trigger while one is in flight gets 409 with
the running batch's details. The guard is in-process, and the Code Engine app
is pinned to one instance, so one lock covers the deployment. A review takes
minutes, longer than Code Engine's 300-second request ceiling, so the trigger
returns 202 straight away and callers poll GET /api/review/status. The full
contract, including request bodies, validation and authentication, is in
docs/API.md.

Deploy

Target	How
Local	npm run build:client && npm run api serves the console and API on one port: PORT, then API_PORT, default 3000.
Container	npm run build:client, then podman build --platform linux/amd64 -t carbon-monitor . and podman run -d -p 8080:8080 --env-file .env carbon-monitor, or docker compose up -d with the included docker-compose.yml, which keeps /workspace in a named volume so a restart does not force a new IBMid sign-in. The image must be Debian-based: Bob Shell's node-pty prebuild needs glibc, so Alpine cannot run it.
IBM Cloud Code Engine	./deploy_ce.sh reads credentials and the deployment target (REGION, RESOURCE_GROUP, PROJECT, NAMESPACE) from .env, syncs the secret, builds, pushes and deploys. The app runs at min-scale 1 and max-scale 1: the Carbon MCP refresh timer needs an instance that never scales to zero, and the rotating refresh token needs exactly one writer.

After the first deploy, open the console and press Sign in with IBMid once.
The deployment topology, ordering rules and image layout are in
docs/ARCHITECTURE.md §8.

./deploy_ce.sh passes GITHUB_AI_AGENT_CLI from .env (default bob) and the
CLAUDE_* settings through, and syncs ANTHROPIC_API_KEY into the Code Engine
secret beside BOB_API_KEY. It refuses to deploy without the key of the agent
you selected: BOB_API_KEY for bob, ANTHROPIC_API_KEY for claude. Moving a
deployment to Claude is GITHUB_AI_AGENT_CLI=claude plus the key, then a
redeploy. The image holds both CLIs, so it is the same image.

Reviewing more than one repository

One app, one Carbon MCP sign-in, one batch and one Bobcoins budget cover every
repository in GITHUB_AI_AGENT_REPOS. To add carbon-ai-chat:

1. Give the reviewing account Triage on the new repository, so it can add
   AIReviewed there (docs/GITHUB_ACCESS.md).
2. In .env, set GITHUB_AI_AGENT_REPOS=carbon,carbon-ai-chat. Four concurrent
   reviews want REVIEW_CONCURRENCY=4 and MEMORY=8G. Before deploying, confirm
   in the Code Engine documentation that CPU=1 with MEMORY=8G is a supported
   combination; if it is not, set CPU=2. The sign-in carries over.
3. Add the profile field to Elasticsearch by hand, once. Events now carry the
   review profile, and the mapping is not dynamic, so an unmapped field is stored
   but cannot be searched. Run
   PUT <events index>/_mapping { "properties": { "profile": { "type": "keyword" } } }
   on the current index and add the same property to
   carbon-monitor-events-template for future indices. node scripts/init-index.js
   upserts the template with it but never changes an existing index. Only events
   emitted after the change are searchable by profile.
4. Deploy with ./deploy_ce.sh, then run one explicit carbon-ai-chat PR from the
   console and check its [AI agent review - CAIC] header and AIReviewed label.

To go back, set GITHUB_AI_AGENT_REPOS=carbon and redeploy.

Console sign-in

Sign-in is off in code but on in .env.example and in the deploy_ce.sh
default. With AUTH_REQUIRED=true, the console shows a sign-in page that
accepts CONSOLE_USERNAME / CONSOLE_PASSWORD, scripts authenticate with
Authorization: Bearer <API_SECRET>, and SESSION_SIGNING_KEY should be set so
a restart does not sign everyone out. With it off, anyone with the app's URL can
read status and trigger a review. See
docs/ARCHITECTURE.md §5.

Operate

Feature	What it does
Operator console	Status tiles for the agent CLI, Carbon MCP, GitHub (every listed repository), profiles and Elasticsearch that say what is wrong, a Reviews panel to run the latest PRs or a named list (type PR numbers; the console finds the repository and asks only when a number is open in more than one), and a live log of the current or last run.
Scheduled reviews	One daily run, configured in Settings → Scheduled reviews. Off on a fresh install. npm start also runs a single pass and exits, so an external scheduler can drive it.
Slack digest	Set SLACK_WEBHOOK_URL to a Slack incoming webhook. The digest fires after scheduled runs only; runs started from the console do not post. One digest per batch, grouped by repository in batch order; a digest over Slack's 50-block limit is split by repository, then into numbered parts, and no result is dropped. A partial review shows as partial review: N of M files on its line.
Kibana dashboard	Review metrics from Elasticsearch, when ES_URL is set. Setup is in kibana/README.md.
Cost tracking	Bob: the Bobcoins balance and per-review spend; scheduled reviews pause themselves, and alert Slack, when the balance falls below BOB_CREDITS_LOW_PCT (default 10%). Claude: per-review spend in dollars and the month so far; with CLAUDE_MONTHLY_BUDGET_USD set, scheduled reviews pause below CLAUDE_BUDGET_LOW_PCT (default 10%) of it, and without it they are not metered.

Spend caps and budgets

Per review, the agent CLI enforces the caps, so they hold whichever agent runs:

Cap	IBM Bob	Claude
Cost per review	BOB_MAX_COST in Bobcoins, passed as --max-cost	CLAUDE_MAX_BUDGET_USD (default 5), passed as --max-budget-usd
Turns per review	BOB_MAX_TURNS, passed as --max-turns	CLAUDE_MAX_TURNS (default 60), passed as --max-turns
Wall time	the harness timeout	the harness timeout

Both CLIs check the cost cap between turns, so a run can pass it by one turn's cost. Leave headroom. A run stopped by a cap is recorded with its spend and reported as a cost or turn limit in the log, the ledger and telemetry.

Per month, the two agents differ:

• Bob. The account's Bobcoins allowance is enforced by the Bob gateway. When it is spent, reviews are refused whatever the harness does. The harness pauses scheduled reviews earlier, below BOB_CREDITS_LOW_PCT of the balance, and tells Slack.
• Claude. There is no account-side stop the harness can see. CLAUDE_MONTHLY_BUDGET_USD is the harness's own gate over recorded spend (Elasticsearch, with the local ledger as a fallback). Scheduled reviews pause below CLAUDE_BUDGET_LOW_PCT of it. The gate covers scheduled runs only, and reviews in flight are not counted yet, so spend can pass the budget by up to REVIEW_CONCURRENCY × CLAUDE_MAX_BUDGET_USD. Unset means not metered. For a hard account-level stop, set a monthly spend limit on the workspace that owns the API key in the Anthropic Console. Details in docs/CONFIGURATION.md.

Documentation

Document	Covers
docs/ARCHITECTURE.md	Container image, Carbon MCP token refresh, authentication, Code Engine deployment, module map
docs/HARNESS.md	Harness versus task layer, every extension point, the agent adapters, the names that change
skills/carbon-monitor-adapt-harness	The skill that interviews you and writes a SPEC for your own repository, job or agent
docs/GITHUB_ACCESS.md	The reviewing GitHub account (i.e. machine user), its token permissions, and why it needs Triage
docs/CONFIGURATION.md	Environment variables, Bob Shell, Claude Code, Carbon MCP and Carbon Builder setup
docs/API.md	REST endpoints and the full review trigger contract
docs/REVIEW_OUTPUT.md	Inline, summary and Looks Good comment formats, partial reviews, recommendations, reliability checks
docs/INLINE_COMMENTS.md	How findings are mapped to diff lines
docs/TOKEN_USAGE_TRACKING.md	Token usage estimates in the summary comment
docs/MCP_TOKEN_REFRESH.md	Manual Carbon MCP token refresh, the fallback to IBMid sign-in
kibana/README.md	Kibana dashboard setup
CONTRIBUTING.md	Setup, tests and pull request guidelines
SECURITY.md	Reporting a vulnerability

Troubleshooting

Bob not found in PATH. Install it with
curl -fsSL https://bob.ibm.com/download/bobshell.sh | bash, then confirm with
which bob.

Carbon MCP not connected. Run /mcp list inside bob; it should show
carbon-mcp - Ready. If not, check ~/.bob/settings/mcp.json, sign in with
IBMid from the console, or call POST /api/mcp/refresh.

No PRs to review. Expected when every open PR already has the AIReviewed
label, or none was opened within the DAYS_BACK window.

A comment says "Partial review". The PR did not fit the diff budget
(GITHUB_AI_AGENT_MAX_INPUT_TOKENS) or the file cap (GITHUB_AI_AGENT_MAX_FILES,
default 300), so the comment lists the files that were not reviewed. Raising
either limit covers more; see
Partial reviews. If not even one line of
any file fits, no agent runs: the PR gets a Not Reviewed comment and the
AIReviewed label.

Claude: no API key. With GITHUB_AI_AGENT_CLI=claude and
no ANTHROPIC_API_KEY, the log and GET /api/ready warnings say
ANTHROPIC_API_KEY is not set, and deploy_ce.sh stops before it deploys. The
process stays up so the console can show the problem. Set the key in .env.

Claude: /api/ready returns 503. The reason names what is missing, for
example Agent CLI 'claude' is not available on PATH. Install the CLI, or use
the image. A missing key is only a warning on a 200, and the agent tile turns
red with ANTHROPIC_API_KEY is not set.

Claude: permission_denials warning in the log. The log line reads
claude was denied N tool call(s). Claude tried a tool that
CLAUDE_DISALLOWED_TOOLS (or the always-added deny rules) forbids, or one that
CLAUDE_ALLOWED_TOOLS does not list. The review still ran; check the named tools
and widen the allow list only if the review needs them.

Claude: carbon-mcp warning at session start. A log line such as
carbon-mcp is failed in the claude session means the Carbon MCP server did not
connect, and Carbon findings fall back to model knowledge. The harness writes the
bearer into <bundle>/claude-config/claude-mcp.json from the credentials in
~/.bob/settings/mcp.json, the file the MCP refresh timer maintains. With no
bearer there, the log says No Carbon MCP credentials found. Sign in with IBMid
from the console, or call POST /api/mcp/refresh, as for Bob.

Claude: carbon-builder skill was not discovered warning. The session's
skill list lacks carbon-builder, so Carbon findings will not be verified by the
skill. The skills: N count in the session line includes the CLI's own bundled
skills (19 in a live run), so only a missing carbon-builder warns. Check that
the bundle holds .claude/skills/carbon-builder/SKILL.md, which is copied from
.bob/skills/carbon-builder in the repository.

Claude: review stopped by a limit. A run that ends with error_max_budget_usd
or error_max_turns fails as a cost or turn limit, records its spend in the
ledger, and emits agent_limit_reached. Raise CLAUDE_MAX_BUDGET_USD or
CLAUDE_MAX_TURNS, or set them to 0 to remove the cap.

Claude on an arm64 Mac, in the amd64 image. claude --version aborts under
qemu, so the agent tile and /api/ready fail when you run the amd64 container
locally. Amd64 Code Engine is not affected. Run npm start on the host, or build
an arm64 image, to test locally.

Keep artifacts for debugging. Run GITHUB_AI_AGENT_KEEP_ARTIFACTS=true npm start.
The bundle path is printed; it holds agent-output.txt (the agent's full response), agent-events.ndjson
(its raw event stream) and agent-stderr.txt. The bundle also holds the
agent's credential file (.env for Bob, claude-config/claude-mcp.json for
Claude, both mode 0600), so delete it when you are done.

Development

Adding review rubric checks

The review prompt in src/reviewPrompt.js (buildReviewPrompt) contains the full numbered rubric the agent follows. Adding a new check means appending a numbered rule to the prompt — no other files need to change.

Changing the comment format

Summary comment rendering is in formatSummaryComment, the clean-PR comment in formatLooksGoodComment and inline comment rendering in formatInlineComment.

Adjusting finding filters

reviewParser.js contains filterUnverifiedCarbonFindings and computeRecommendation. The recommendation logic is in computeRecommendation.

---

Build with us

• Issues: report bugs and propose features in GitHub issues.
• Pull requests: welcome. Read CONTRIBUTING.md first.
• Security: report vulnerabilities privately, as described in SECURITY.md.
• License: Apache-2.0. See LICENSE.