devboy-tools

Walk a project from "no values provisioned" to "doctor --secrets is green" — eight idempotent steps with resume support via setup-state.toml. Wraps the secret framework (ADR-023 §3.8) for AI agents and headless onboarding.

Bootstrap devboy from scratch — install the CLI if missing, register the MCP server, run `devboy onboard` for the active agent, optionally bootstrap the secret framework, verify with `doctor`. First-run skill for both manual installs and the Claude Code / Codex plugin.

First-run wizard for the devboy secret framework — walk a fresh project from "no secret manifest, no router, no daemon" to "every required secret provisioned and verified". Idempotent eight-step flow per ADR-023 §3.8 with state at ~/.devboy/secrets/setup-state.toml so the user can resume or skip.

Analyse the user's Claude Code (or other agent) logs and auto-configure the layered-pipeline compression profiles for their tools, models, and workflow.

Diagnose and fix a broken devboy-tools setup — corrupt config, missing tokens, keychain trouble, wrong paths, plugin install failures.

Enumerate and introspect the active tool bundle — names, categories, schemas, how to invoke each tool from the CLI.

Create a well-structured issue in the configured tracker using Feature, Defect, or Task templates.

Fetch and summarise issues from the configured tracker — filter, paginate, then drill into a single issue when needed.

Connect two issues with a typed relationship — blocks, blocked-by, related, duplicates, or parent/subtask.

Full cycle for shipping an issue — read it, branch, implement, push, open an MR, and link back to the original ticket.

Change an issue's state, labels, assignees, or priority with a partial update, leaving unrelated fields untouched.

Apply reviewer feedback on an MR — fix the code, run local checks, commit, push, and reply to each discussion.

Strict but calibrated code review of a single MR or PR — checklist, inline comments, one overall summary.

Dry-run a review pass over your own MR before asking a human reviewer — local checklist, no inline posting.

Graphic monthly/weekly digest of Claude Code session patterns — biome aquarium (whale/shark/dolphin/fish/shrimp/plankton), 8-class archetype, rhythm, stack palette, DORA radar (CFR + lead time + pushes), friction (compacts/pivots/subagents). Backend installs on first run via curl.

Summarise today's trace activity, merged MRs, and closed issues into a single report.

Extract the lesson from a session that failed, then succeeded, and propose where to codify it.

Fan out parallel QA sub-agents against the `devboy` CLI — each one hunts a specific class of regression (exit codes, stdout hygiene, error propagation, schema drift, …) and reports findings into a shared bug log.

Aggregate the last N days of traces, MRs, and CI runs to surface patterns worth fixing.

Run a command, trace every step, verify the expected outcome, and retry on failure.

Find meetings by keyword, participant, host, or date range and surface a short, rankable hit list.

Extract deduplicated action items from a meeting, confirm with the user, then create issues in the configured tracker.

Fetch a meeting transcript — full, paginated, or filtered to a single speaker or phrase — without leaking raw PII into long-term storage.

Search messenger history for messages mentioning a keyword, optionally scoped to one chat or a date window.

Catch the user up on a channel or DM by summarising messages over a time window, chunk by chunk, into grouped bullet points.

Post a short, structured notification — subject, bullets, optional link — to a chat or channel the user explicitly names.

Analyze higher-level patterns in Claude Code usage. Outputs (a) graphic monthly/weekly digest with metaphors — aquarium of biomes (🐋🦈🐬🐟🦐🦠), archetypes (⚙️🔬🌐🛠📝🔍🏗💬), rhythm, stack palette, DORA radar (CFR + lead time + pushes), friction (compacts/pivots/subagents); (b) per-session parquet bundles for further analysis (biome, archetype, rhythm, growth, milestones, idle gaps, subagent pyramids, compact patterns, parallelism, burst classes, topics). Use when the user asks about weekly/monthly reports, session biomes, productivity profiles, "what kind of work was done", when a session became a whale, DORA metrics, or wants drill-down view of a specific session.

Rebuild the Claude Code agent-usage research dataset (sessions, loops, tool calls, gold-selection events, correlations) from ~/.claude/projects/*.jsonl logs into /tmp/claude_analysis/*.parquet. Use when the user wants to bootstrap analytics on a new machine, refresh after new sessions, validate reproducibility, or onboard the research pipeline after cloning this repo.