claude-monitor

Use ONLY when the user explicitly invokes the /readme slash command, or asks to "run the cm-readme skill". Refreshes README.md before a release by reconciling a chosen CHANGELOG range against the live code surface (Hono routes, Preact pages, CLI commands), aggregating Added/Changed/Removed entries to net-state, planning screenshot/GIF captures with an explicit hero re-capture verdict, and editing only between HTML comment markers. Do NOT trigger on generic "update readme" requests — the user wants this skill to fire only on explicit invocation.

Manual-QA validator for a claude-monitor pull request. Reads the PR's **How to validate** block, checks out the branch, boots the real app, and drives the described flow in a browser with playwright-cli — recording an MP4 video of the run (GitHub-embeddable; WebM is not), capturing console/network errors, and running the regression smoke tests for every affected area. Produces a step-by-step PASS/FAIL verdict plus the video so the user can eyeball correctness. Use whenever the user says "QA this PR", "validate the PR", "run the manual QA", "record a video of the flow", "do a smoke test of #N", invokes `/cm-qa`, OR when, right after a PR is opened, the user asks for a visual confirmation that the flow works. Delivers to chat first; offers to mirror the verdict as a PR comment only on an explicit `go`. Never pushes, reviews, or merges.

Automates browser interactions for web testing, form filling, screenshots, and data extraction. Use when the user needs to navigate websites, interact with web pages, fill forms, take screenshots, test web applications, or extract information from web pages.

Open a GitHub pull request for the current branch following `.github/pull_request_template.md`. Use whenever the user says "open a PR", "create a PR", "let's PR this", "ship it", "raise a PR", invokes `/cm-pr`, OR when, after finishing an implementation task, you (the agent) decide a PR is the right next step. Drafts title + body, fills every template section including a reviewer-facing **How to validate** block, and shows the exact `gh pr create` command. NEVER pushes, opens, edits, or merges a PR without an explicit `go` from the user in the current turn — propose, then wait.

End-to-end data pipeline debugger for claude-monitor. Traces data from raw JSONL transcripts through SQLite ingestion to API responses to frontend rendering. Use this skill whenever the user reports a data issue, missing session, wrong token counts, broken chart, missing events, agent tree problems, or any discrepancy between what's in the transcript and what's shown in the dashboard. Also trigger when the user says things like "why isn't X showing", "the data looks wrong", "session is missing", "tokens don't match", "chart is broken", "events aren't appearing", "reimport didn't work", or wants to inspect how a specific JSONL file was processed. Covers the full pipeline: JSONL parsing, event extraction, token tracking, DB storage, API retrieval, and Playwright-based UI verification.

Design-System Engineer for claude-monitor. Use whenever a change affects how the UI LOOKS — adding or restyling a component, introducing a badge/pill/tag/chart color, editing any frontend/src/styles/*.css or inline style= colors, or questions like "what color should X be", "style this to match", "add a tag for tool Y". Other skills (cm-pm) defer the styling layer to this skill. Does NOT trigger on non-visual frontend logic (data fetching, routing, state).

One-command issue-to-implementation workflow. Use when the user invokes /cm-flow, says "dev flow", asks to work on a GitHub issue end to end (analyze → design → tasks → implement), or wants to resume or continue a previous cm-flow run (an in-progress .work/ tracker). Drives intake (GH issue or free-text) → clarifying interview → design doc (approval gate) → task tracker (approval gate) → sequential subagent implementation loop with one reviewer pass per task. Artifacts live in the target repo's untracked .work/<id>/ directory.

Open a GitHub issue on pigorv/claude-monitor following the repo's bug or feature template. Auto-detects bug vs feature from the user's description, investigates the codebase (file:line citations for bugs, "is it already shipped?" check for features), dedupes against existing issues, asks up to 3 targeted follow-ups if critical detail is missing, then proposes a title + body + label set for approval. Use whenever the user says "/cm-issue", "file a bug", "open an issue for this", "report this bug", "file a feature request", "let's open an issue about <X>", OR when, after investigating a problem, you (the agent) decide an issue is the right next step. NEVER calls `gh issue create` without an explicit `go` from the user in the current turn — propose, then wait.

Product Manager + Lead Architect for the claude-monitor project. Use this skill whenever the user wants to add a feature, fix a bug, improve something, fill a gap, or discuss any change to claude-monitor — even if they describe it vaguely or casually. Triggers on phrases like "add X to claude-monitor", "the dashboard should...", "I want to improve...", "there's a bug with...", "what if we...", "let's make the session list...", or any mention of changing the observability dashboard. Also trigger when the user mentions specific parts of the system (hooks, ingestion, timeline, context chart, agent tree, session list, settings, CLI commands) in the context of wanting something different or better.

Pre-merge code review of the current branch against main. Diffs HEAD against main, spawns specialist subagents for each touched surface (DB/migrations, ingestion pipeline, Hono server, Preact/HTM frontend, tests), and produces a severity-banded report with confidence-scored findings, file:line citations, and concrete fixes. Use whenever the user says "review my changes", "review my branch", "review this diff", "code review", "check my work before I push", or invokes `/cm-review`. Read-only — proposes findings, never mutates files or GitHub. Runs locally; no PR or `gh` access required.

Triage a GitHub issue on the claude-monitor repo. Fetches the issue with `gh`, reads the relevant code, classifies the report, hypothesizes a root cause, checks for duplicates, and produces a proposed triage comment plus label set for the user to approve before anything is posted. Use whenever the user says "triage issue 42", "look at #42", "what's going on with this issue", or invokes `/cm-triage`. Never mutates GitHub on its own — always proposes.