barkpark

Open, log, checkpoint, and close a persistent Barkpark session record so work survives context limits, machine switches, and new conversations. Invoke at the START of any substantial work session ("start a session", "track this session", "begin session tracking") — and when invoked MID-conversation, the open captures the FULL active thread so far (first checkpoint distills the entire conversation, trail backfilled, transcript covers turn one onward), after every milestone during the session (a paper published, a task closed, an epic wave sealed, a successful git push), before context is likely to run out or before a long/risky operation ("checkpoint", "save progress", "write a checkpoint"), and at the END of a work session ("wrap up", "end of session", "close out", "hand off to another machine", "I'm done for now"). Produces a `type:session` document on the Barkpark server (via `bp session open|log|publish`) carrying metadata, a milestone event trail, synthesis blocks (current task, progress, key files, decis

Turn THIS Claude Code session into a Personal Dev Fleet listener (worker) — it stays alive and executes orders that arrive as bp tasks on the ledger. Invoke when the user says "become a listener", "start listening for orders", "be a fleet worker", "join the fleet as <name>", "listen for tasks", or "worker mode". The session claims assignee-routed tasks with a fenced claim, does the work IN-TURN, stamps evidence, closes on the claim epoch, and returns to listening. Part of the Personal Dev Fleet (paper /papers/personal-dev-fleet-worker-protocol).

Pick up a Barkpark session record left by a previous conversation or a different machine and load it back into context — metadata, milestone events, and the synthesis blocks (current task, progress, key files, decisions, next steps, learnings). Invoke when the user says "resume session <slug>", "continue where I left off", "pick up the other session", "load my last session", "what was I doing", or hands you a `session-...` slug. Warns loudly if the cwd, git repo, or branch don't match what the session recorded — a mismatch usually means you're on the wrong machine or checkout for this handoff. Pairs with the `session` skill, which produced the record.

Turn THIS Claude Code session into the Personal Dev Fleet orchestrator — it takes a wish, decomposes it into fence-disjoint orders, files them as bp tasks routed to listeners, tracks completion, and grades/reports. Invoke when the user says "be the orchestrator", "orchestrate the fleet", "conduct the fleet", "dispatch this to the workers", "run a fleet campaign", "decompose this wish for the fleet", or hands you a goal to distribute across fleet listeners. Pairs with the fleet-listener skill. Part of the Personal Dev Fleet (paper /papers/personal-dev-fleet-mvp).

Run a five-scale, experiment-gated Barkpark epic for enormous numeric inventories and builder-heavy gruntwork. Use when the user asks for a legendary cycle, a 5x epic cycle, a massive document or Paper repair swarm, or a numerically sharded quality campaign across Studio, TUI, email, CLI, or other readers.

Run one evidence-gated Barkpark epic wave in Codex, from strategy through multi-wave agent fleets, verification, high-effort build, review, and debrief. Use when the user asks to run, continue, perfect, or design an epic cycle/wave.

Use when the user asks "how good is the codebase", "what's the code quality", "what should we improve / refactor", "where's our technical debt", "score the repo", "audit the codebase", "build the codebase knowledge graph", "what is each file for / how important / how useful", "what intentions / goals does the code serve", "what open work is highest-leverage", or "publish the codebase into Barkpark". Four arcs: ASSESS (13-dimension quality scorecard + ROI improvement plan), ENRICH (per-file importance, reach+why, intentions, ownership, git history), PUBLISH (one Barkpark paper per file, typed references, interconnected graph), RELATE (co-change + tasks as a third graph layer, plus triage of open work by predicted impact). Programmatic scans measure for free; agents are spent only on judgment a parser can't make, and only on what changed. Lives in tooling/.

Use when you're about to work a task and want the exact file set to read FIRST instead of crawling the tree — "what files are involved in X", "scope this task", "what does the auth/search/sheets/onboarding feature touch", "give me a context pack for <task>", "which files advance <intention>", "where do I start on <feature>". Maps a task description or an intention id to the intention(s) it advances, then emits a SCOPED CONTEXT PACK: every file advancing it with its one-line role/why + reach + dependencies BOTH directions + blast-radius. Exploration → lookup. Lives in tooling/scope/.