home-dev-lab

upgrade-canary

Invoke ONLY when you suspect Claude Code may have upgraded, when you are about to cut a plugin release, or when the user says "run the upgrade canary", "did the update break the workflows", "re-check the runtime surface", or "what changed in the runtime". Re-verify both runtimes the toolkit depends on: the interactive `claude` binary and the SDK-bundled one. Run the full matrix only when versions changed since the last pass, when forced, or when the last verdict failed. Maintainer tool only; requires this repo's `toolkit/` tree plus local Claude Code auth.

toolkit-scaffold

Invoke ONLY when you are about to turn a job description into a new build-clean `.workflow.ts` skeleton through the toolkit path, or when the user says "scaffold a workflow", "start a new workflow", "generate a workflow skeleton", "set up a .workflow.ts", "wire these patterns together", or "use the toolkit to build a workflow". Pick the patterns, write the small JSON spec, run `workflow-toolbox scaffold`, then fill in placeholders and build plus check the result. Not for deep one-off authoring guidance, debugging a failed run, or re-verifying the runtime after an upgrade.

workflow-debugger

Invoke when you are diagnosing a workflow run that errored, returned a wrong or partial result, or seems stuck, or when the user asks "why did my workflow fail", "debug my workflow run", "the workflow errored / died", "can I resume this run", or "what happened in run wf_…". Read the on-disk run journal first, use per-agent transcripts only as fallback, classify the failure mode, and say honestly whether `resumeFromRunId` will replay cached work or save nothing. Not for authoring or restructuring a workflow script, or for checking runtime drift after a Claude Code upgrade.

what-next

Invoke when you have just finished a chunk of work or need the next recommendation grounded in the tracked backlog, or when the user types /what-next or asks "what's next", "what should I tackle now", or "what comes after this". Read the project's Planka board through .claude/planka.json when the `planka` MCP is reachable, otherwise fall back to .claude/progress.md, plus the current conversation and any existing knowledge index such as `MEMORY.md`. Read-only: propose a next step, never mutate the board, `progress.md`, or task state.

pilot-wave

Invoke when you are about to drive tracked cards end-to-end through delegated execution and need to compose the environment brief, select the cards, and spawn the adopted `pilot` or `pilot-orchestrator`, or when the user says "run these cards", "pilot this wave", "spawn a pilot for this ticket", or "set up the orchestrator for these". Keep the main session as the arbiter and wake-up relay, require adopted project copies first, and if they are missing, propose `adopt` before spawning. Not for a single mechanical edit.

adopt

Invoke ONLY when you are about to install, check, refresh, or propose adopted editable copies from workflow-toolbox into a config or project directory and need the versioned, fingerprinted path, or when the user explicitly asks to "adopt the delegation rules", "install the workflow-toolbox rules as editable files", "install project copies of the pilot agents", "adopt the pilot watchdog into this project", or "check my adopted workflow-toolbox rules/agents for updates". Write ONLY on request; never run automatically. Not for authoring workflows or composing a pilot wave.

deep-grounding

Invoke when you are about to assert a checkable claim or start a non-trivial analysis, design, diagnosis, or recommendation and need to ground it against the real sources before committing. Ask "where could the evidence for this be?", read the relevant sources, follow their leads recursively, and report conclusions with evidence tiers plus any residual unknowns. Use this **PROACTIVELY** when you catch yourself answering from memory, from a single source or angle, or when a result beats your expectation and could skip verification.

independent-analysis

Invoke when you are about to commit to a non-trivial design, plan, decision, claim, architecture, root-cause theory, or change set and want it stress-tested by fresh, conversation-blind analysis, or when the user says "are you sure?", "what am I missing?", "did we forget anything?", or "make sure no angle is forgotten". Run the bundled independent-analysis workflow so missing angles surface and refute-first verifiers kill plausible-but-wrong ones. Use this **PROACTIVELY** when you may be moving fast or confirming a prior.

workflow-composer

Invoke ONLY when you are about to build, generate, repair, or restructure a workflow for Claude Code's Workflow tool and need authoring guidance or code changes, or when the user says "I want a workflow that…", "turn this process into a workflow", or "set up a multi-agent pipeline", or needs help picking orchestration patterns or tuning model, effort, or agentType routing. Write or repair the workflow artifact through the `@workflow-toolbox` path for repeatable work or the raw single-file path for one-offs. Not for debugging a specific failed run or re-verifying runtime drift after an upgrade.

planka-tracking

Invoke ONLY when the current project is not yet managed by Planka and you are about to onboard it to board-based tracking, or when the user says "onboard this project to Planka", "set up the board here", "start tracking tasks on Planka", or "track my tasks on a kanban board". Create the project, board, lists, and labels through the `planka` MCP, migrate any existing .claude/progress.md zero-loss into cards, and write .claude/planka.json. Requires a reachable `planka` MCP server; do not run without it.

lesson-harvest

Invoke **PROACTIVELY** whenever you are integrating a closed pilot or executor report — nobody will ask you to — to get its "## Lessons for the memory" section without rereading the whole file. Extract candidate lesson items verbatim and return a distinct verdict for three cases: the section says "None.", the section is missing entirely, or the section carries content. Detection and extraction only: never write to a knowledge base, invent wording, classify a fiche, or decide whether to persist the lesson.

stale-card-sweep

Invoke **PROACTIVELY** whenever you are about to close a task-tracker card — nobody will ask you to — to check whether its diff already covered the defect another open card still describes. Also invoke when the user says "sweep for stale cards", "check what this closes", or "did this diff subsume anything else". Shortlist candidates from the changed files, then judge each one against the diff. Advisory only: flag candidate cards with a comment; never close, move, or rewrite them. Not for card creation or age-triggered dormant-card sweeps.

delegation-chain

Load this BEFORE spawning a delegate, arming a watcher, or reasoning about what wakes a session. Consolidates the downward spawn chain (anonymous/named/isolated, and what each shape costs an observer or a worktree), the upward report chain (what actually reaches a spawner and what silently doesn't), who wakes whom (and what wakes nothing), and the watchers/guards this plugin ships with an explicit "does NOT cover" column for each. Not a task to run — a reference to consult before a delegation decision, so the same question isn't re-answered from scratch every time.

adopt-rules

Invoke ONLY when the user explicitly asks to install / adopt editable copies from workflow-toolbox into their config, or to check adopted copies for updates. Two managed sets — the cross-cutting RULE files (the delegation ladder) and the pilot AGENT-definition copies (pilot / pilot-watchdog / pilot-orchestrator, whose project copies enable the watchdog observer pairing that plugin-installed agents can't). E.g. "adopt the delegation rules", "install the workflow-toolbox rules as editable files", "install project copies of the pilot agents", "adopt the pilot watchdog into this project", "check my adopted workflow-toolbox rules/agents for updates". Writes versioned, fingerprinted, editable files ONLY on explicit request — never automatically. Re-invoke to detect stale copies (installed version behind the plugin) and refresh them. Not for authoring workflows (workflow-composer) or composing a pilot wave (pilot-wave).