laicluse-agent-fieldkit

Single-occupancy lock for a git worktree or standalone working directory: a vendor-neutral, cross-platform CLI so only one coding agent mutates a repository checkout at a time. claim/release/check use an atomic exclusive-create lockfile keyed by the nearest git worktree root, falling back to directory realpath outside git, with pid-based liveness that self-heals a dead holder and respects a live one. No flock, no native dependency. bonsai and the discipline hooks consume it.

Git workflow skills, repo-aware push policy, and commit/push hooks for agent sessions and direct CLI commits.

Local-first Pattern Memory for reusable implementation recipes. Stores personal patterns as Markdown and keeps private precedents out of public artifacts.

Marketplace-wide utilities for the laicluse-agent-fieldkit marketplace. Currently ships /whats-new (namespaced as /laicluse-agent-fieldkit:whats-new when needed), a one-stop reader for the post-update CHANGELOG section of any installed plugin.

Dispatch a rover at a task and stay back while it decides in the field: a SURVEY/DRIVE/INSPECT/STOW phase machine with a decide framework, pride and trim quality gates, verify evidence discipline, prepare for queuing missions across repos, and a stop communique. Asks the autonomous layer at startup whether it needs to keep itself alive. Depends on gurus for the INSPECT panel review.

An opinionated baseline for software craft: naming, testing, and programming doctrine in the tradition of Beck, Martin, and Fowler. These are chosen conventions among defensible options, not universal law. Skills: programming-philosophy, testing-philosophy, naming-is-hard.

Reach outside your own head when you are unsure: research what others did before you commit (inspire), and verify your own claims against external sources when challenged (ground). Two lifelines for the moment your own certainty runs out. Skills: inspire, ground.

Copy core content of the last answer to the macOS clipboard via the clipboard-copy helper (pbcopy for plain, pbcopy-html for Slack rich text).

Worktree lifecycle as a cross-platform CLI: create a worktree + branch off the freshest default, set it up from a .bonsai file list, and tear it down behind a hard safety gate that keeps non-integrated work and warns on orphaned commits or interference. Skills: bonsai (create+setup), setup, prune.

Guardrail hooks plus /duh, /just-a-question, and /not-your-monkey correction skills for Claude Code and Codex.

Find and converge parallel paths in any artefact: code, prose, design systems. Methodology guide that disciplines how a duplication audit runs (Type-4 clone framing, verifier-burden LLM pass, allow-list scoping, drift hypothesis, triage, two-fates discipline, contrarian second-pass, method-as-artefact) without prescribing what duplication looks like in your codebase. One orchestrator (drydry:drydry) with quick and audit modes; the calling session formulates the audit checklist from the codebase in front of it. Agent-only sub-skills handle sweep, checklist seeds, triage, research, upstream/framework checks, and instruction-file audits.

Opinionated review panels that challenge code, decisions, and prose from multiple perspectives. Ships gurus:software, gurus:council, gurus:writers, and the gurus:gurus orchestrator; Claude also receives the gurus:sonnet-max panel subagent.

Bring another coding agent in as a peer to review work just done or just discussed. Claude hands work to Codex; Codex hands work to Claude. Intervision is peer consultation among equals, not supervision from above.

Speech mode for macOS: the agent speaks every response aloud through the `say` command, translating screen content into spoken language instead of reading it out verbatim. /saysay toggles it on, /saysay off exits. Ships a per-user phonetic dictionary plus serialized, ducked playback so parallel sessions speak one after another.

Drill N layers deep into a question or goal (default 10). Claude asks and answers "why?" itself, then analyses the chain for a better direction. Use when tackling unclear decisions, vague goals, root-cause analysis, or self-improvement on something that is not working as desired.

Reference for plugin naming, skill resolution, Claude/Codex adapter sync, marketplace aliases, runtime caches, and plugin:skill invocation.

Keep an autonomous mission running across turns. A startup capability probe decides whether keep-alive machinery is needed: a cron heartbeat with exponential backoff, auto-stop, and wake/restore for an interactive session; nothing for a persistent process. The rover plugin's decision framework calls this layer at dispatch.

Routes user feedback about agent behavior to the strongest durable target: hooks, skills, project code, or instruction files.

Cuss at your coding agent to capture a one-line friction note, then fix the real problem later. The cuss commands (/fuck, /fucking, /fucked, /shit, /crap, /wtf, /bullshit) append one cheap line to a global log and move on; safeword commands (/safeword, /pineapple, /pineapplejuice, /pinapplejuice, /flugelhorn, /banana) interrupt current work and fix one visible friction point now without logging it for later. /anger-management:repair cools down (~22 min, a background headless agent investigates), reads the full capture history, scores confidence, chooses the mitigation layer and target scope, and only applies or routes a fix when confidence is high enough. Lower-confidence findings leave the breadcrumb trail open for a future pass. Modelled on Stripe's bad-day button, course-correction-before-escalation, and the constructed-emotion view of anger; sometimes the irritation is real, sometimes it is playful. Vocabulary avoids slurs and identity-targeted abuse. State is stored locally; the optional background investigation uses the configured runner.