openharness

TRIGGER when: asked to delegate work, execute or parallelize a plan, "run this plan", "delegate this", or after /prd or plan creation. Decomposes work by dependency, launches worker sub-agents in parallel waves, validates completion, and reports results while preserving failure isolation and recursion limits.

Dispatcher for the harness wiki knowledge base — routes the first token of $ARGUMENTS to one of three subcommands: ingest, query, or lint. The corpus lives at .oh/skills/wiki/corpus/ (entity pages) and corpus/raw/ (immutable snapshots), owned by this skill and gitignored-by-default (curated entries are whitelisted into git with `git add -f`). Canonical schema: .oh/skills/wiki/references/schema.md. Full per-subcommand procedures live in references/{ingest,query,lint}.md. Always logs per the Memory Improvement Protocol. TRIGGER when: "add to wiki", "capture this page", "snapshot this source", "ingest <url|path>", or promoting a sub-agent draft -> ingest; "what does the wiki say about X", "find wiki entries for X", "look up X in the wiki" -> query; "lint the wiki", "regenerate the wiki index", "find stale/orphaned wiki entries" -> lint.

Turn a source scenario into a publishable blog post. First supported playbook: Loom/demo-to-blog, where a raw notes file plus Loom screenshots become a polished site post after Advisor briefing, photo/text alignment audit, site convention audit, narrative/fact-check audit, source-faithful image handling (exact URLs or local files downloaded from them), optional social/channel-promotion artifact generation, and build verification. Manual-invoke only because it writes content and may spawn delegates. TRIGGER when: /blog <scenario> invoked; asked to create a blog post from a Loom/video/demo/transcript/screenshots; asked to turn .claude/specs/<name>/ demo.md into an openharness-web blog post.

Start, inspect, or stop T3 Code (`npx t3`) in the Open Harness sandbox. Use this for the browser-based T3 Code harness on port 3773, including tmux launch, pairing URL discovery, logs, status, and shutdown. T3 Code wraps an already-authenticated Claude Code, Codex, or OpenCode backend. TRIGGER when: user asks to run T3 Code, start `npx t3`, open the T3 browser UI, get the T3 pairing URL, check T3 Code status/logs, or stop T3 Code.

Scientific session-closing retrospective: scan the current conversation, turn each signal into a falsifiable hypothesis, cite session evidence for AND against it, assign a verdict (supported / refuted / inconclusive) and a confidence level, then promote only supported, sufficiently-confident hypotheses into the harness memory tiers (.oh/memory/MEMORY.md, .oh/context/IDENTITY.md) behind a propose-then-confirm gate. Reflects on six learning/knowledge subsystems through the lens of this session — continual learning, context compression, reinforcement learning, wiki, docs, and memory scaffolding — and points at the deep-dive lint/audit skills rather than running them. Operationalizes the Memory Improvement Protocol (.oh/skills/retro/references/memory-protocol.md) as an explicit, evidence-driven, session-closing skill rather than a per-run afterthought. Always appends a log entry. TRIGGER when: /retro invoked, or session closing with decisions, surprises, or failures worth preserving.

Self-improvement loop: each hourly run implements the oldest open issue labeled `autopilot` that has no open PR. When the queue is empty it runs first-principles `/harness-audit` research, files its own `autopilot` ticket from the top-ranked finding, and builds that. Decomposes via the pm agent, then runs `/ship-spec --issue`, which owns the build end-to-end (the two compacts bracketing implement, a worktree Advisor, `/delegate` workers running `.oh/scripts/ralph.sh`, the `/eval` gate, and the `/pr-audit` promotable undraft) through to a ready-for-review PR whose description states why this item was selected. Harness-infra only (skills/rules/docs/scripts/crons/wiki) — never sandbox application code. Runs in its own per-run tmux session. Caps: 6 open autopilot PRs created per UTC day AND 10 total open; never auto-merges. TRIGGER when: the hourly .oh/crons/autopilot.md fires, or invoked manually on demand (e.g. /autopilot --dry-run to preview the next selection).

Open Harness git workflow: issues, branches, commits, PR titles/bodies, changelog discipline, worktrees, branch catch-up, stacked PRs, releases, and post-push CI checks. TRIGGER when: any chat mentions git, GitHub, branches, commits, pushes, pulls, PRs, issues, worktrees, merge conflicts, dirty workspaces, changelog entries, release branch/tag workflow, or project git conventions.

Explain the Open Harness architecture, layout, and conventions. Use this skill when the user asks about project structure, layout, git workflow, rules, or how things are organized. Returns specific file paths and `path:line` citations. TRIGGER when: asked about harness structure, layout, conventions, rules, process, or project organization.

Context-as-environment decomposition (Layer B of the RLM integration). Treat a large artifact (file / dir / log) as a REPL/filesystem the root agent greps and slices instead of ingesting — partition it into addressable chunks via query-context.mjs --map, RECURSE sub-agent calls over the relevant chunks under a bounded depth/children/step budget, AGGREGATE (piping competing per-chunk answers through /weigh), then PERSIST the recursion trace. The anti-context-rot move: ADDRESS context, don't dump the whole artifact into the window. Reuses .oh/scripts/ralph.sh (the recursion loop) and .oh/worktrees/ (isolated branches) BY REFERENCE — never edits either. Manual-invoke (spawns agents, burns tokens). TRIGGER when: /rlm invoked, or asked to "answer a question over a huge file/log", "decompose a large artifact", "recurse over chunks", "address context instead of ingesting it", "beat context rot on a long input", "RLM <file> <query>".

Compose /prd → wiki alignment + DeepWiki comparison → 2 critics → /ralph → gh issue → branch → draft PR → implementation/eval/wiki revision/CI → ready-for-review PR into a single end-to-end invocation. Produces a fully-scaffolded task, opens a draft PR for observability while work is pending, then marks the PR ready after the implementation, eval, wiki, and CI gates pass. TRIGGER when: asked to scaffold a spec end-to-end, "ship a spec", "set up a task PR", or after planning a feature and ready to formalize.

Manage .oh/worktrees/ lifecycle: create worktree, list worktrees, remove worktree, clean worktrees, stale worktrees audit, isolate work, project clone. TRIGGER when: any git worktree operation, branch isolation needed, stale worktrees review, project clone under .oh/worktrees/project/ (e.g. "clone <owner>/<repo> to worktrees", "add <repo> to .oh/worktrees", "clone this repo into worktrees"), worktree cleanup. A leading-slash harness dir like "/worktrees" still means the repo-relative .oh/worktrees/ — never a literal filesystem-root path.

Per-unit verdict gate (the build⇄audit gate composed by `/spec execute`) — decide whether ONE implementation satisfies its task graph and is promotable, then emit a single PASS/FAIL verdict. Composes (never forks) the existing checks: prd.json task-graph conformance + /eval regression floor + /pr-audit promotable classification (subsumes /ci-status) + /agent-browser for UI stories. Distinct from the audit-FAMILY survey skills (/pr-audit = whole open-PR queue, /context-audit = context budget, /harness-audit = whole harness): this is one impl vs. one spec → one verdict. TRIGGER when: an implementation is complete and needs a go/no-go verdict before retro; the build⇄audit gate of `/spec execute` runs; "audit this task", "is <slug> promotable", "verify the implementation against its prd.json".

Score every eval probe (.oh/evals/probes/*.sh) and capability benchmark task (.oh/evals/capability/tasks/CB-*.md + repo-orientation) across seven anti-Goodhart failure modes and emit a KEEP/GROOM/CUT matrix. Read-only by default — like /skill-lint, it only reads and prints; it never mutates a probe or capability file. Deterministic file/grep/git/stat scoring, no LLM judgment. TRIGGER when: asked to lint the eval suite, groom the probes, check for Goodharted / stale / duplicate / always-SKIP probes, audit the capability benchmark tasks, "eval lint", or before/after adding probes.

Rank past prompts by session outcome and mine the markers that produce the best sessions. Runs the deterministic mine-traces.mjs engine over Claude + Pi JSONL traces, scores each session by a friction + ground-truth outcome proxy, ranks the initiating prompts, then synthesizes falsifiable prompt markers STRATIFIED by session type and proposes harness memory/identity improvements behind a /retro-style propose-then-confirm gate. A cross-session, data-driven cousin of /retro. TRIGGER when: /prompt-miner invoked, or asked to "mine prompts", "rank prompts by outcome", "what prompt patterns work best", "mine session traces", "find good prompt markers", "analyze prompt quality".

Assess sandbox host resources (memory, swap, disk, CPU, Docker usage) and readiness to start or build a stack. Ranks reclaim levers by safety×yield, runs the zero-risk ones, and confirms before any destructive removal. TRIGGER when: asked for a health check, "do we have enough memory/disk", before starting a heavy stack or docker build, "system health", "free up space", "reclaim resources", or when a build/run is at risk of OOM / disk-full.

Score the default-loaded context budget across 4 dimensions and emit KEEP/TRIM/DEMOTE/CUT verdicts per file. Optional Tier-2 ablation harness removes a target file, runs a fixed probe suite, and measures behavior degradation — the only provably safe gate for cutting load-bearing content. TRIGGER when: asked to audit context window, check default context load, "what's in my context", evaluate rules for signal vs noise, or before/after any change to .oh/context/ or CLAUDE.md.

Audit all open PRs in one bulk query and triage them into actionable buckets: ready-to-merge, needs-review, CI-failing, conflicting/behind, plus stale/convention/duplicate issue-reference flags. Draft PRs are split out as a separate work-in-progress class (promotable / WIP / limbo), checked first and never mixed into the actionable buckets. Read-only by default; --deep escalates flagged PRs to parallel diff reviewers, --proof writes each PR's verdict + evidence back as a comment for human reviewers, and opt-in flags (--label-apply/--close-stale) apply triage actions after confirmation. TRIGGER when: asked to audit open PRs, "check the open PRs", "what PRs are stuck", "triage the PR queue", PR backlog review, or before a merge sweep.

Open a URL in the headless agent-browser, with a preflight health check. Verifies agent-browser is installed and Chromium launches before navigating. Errors out with a diagnostic trace if anything fails. TRIGGER when: asked to open a page, browse a URL, take a screenshot, or test a site with agent-browser.

The approve (approve|deny) decision gate of the canonical workflow — read a spec's critique findings and decide go/no-go BEFORE any GitHub-side state exists, then emit a single APPROVED/DENIED verdict. Auto-gates on critic SEVERITY (any unmitigated high → deny), with an optional human override for borderline findings. This is the decision half of the critique→approve pair; the critics that produce the findings are the separate `/critique` node. TRIGGER when: a plan/PRD has been critiqued and needs a commitment decision before implementation; "approve this spec", "should we build <slug>", "gate <slug> on its critique".

Progress-ceiling verdict gate (part of `/spec execute`'s improve tail) — decide whether ONE landed change was actually BENEFICIAL (moved or held the capability-benchmark ceiling without breaking the regression floor, and is worth its complexity), then emit a single BENEFICIAL/NOT-BENEFICIAL verdict. Composes (never forks) the existing instruments: /eval (the regression floor — probes stay green) + the capability-benchmark ceiling delta (.oh/evals/capability/RESULTS.md vs. the counterfactual). Machinery added with no benchmark movement is NOT-BENEFICIAL by definition. Distinct from /audit (per-unit promotability = floor) — this is the ceiling: did the harness get BETTER, not just not-broken. TRIGGER when: a change has landed and the loop needs a benefit-vs-counterfactual verdict before the cycle repeats; the improve tail of `/spec execute` runs; "was <change> beneficial", "score the capability benchmark", "benchmark this cycle".

Write a Conventional Commit message in caveman style — terse subject, fragment body, full technical accuracy. Subject line stays a valid Conventional Commit (type: description). TRIGGER when: /caveman-commit invoked, or asked for a caveman commit message.

Rewrite a target file (memory note, doc, scratch plan) in compressed caveman form to shrink its token footprint, preserving every technical fact, code block, and link. Shows a diff and confirms before overwriting. TRIGGER when: /caveman-compress <file> invoked, or asked to compress a doc/memory file.

Produce PR review comments in caveman style — terse, fragment-based, one finding per line, severity-tagged. Compresses prose, never the cited code/symbol/line refs. TRIGGER when: /caveman-review invoked, or asked for compressed PR review notes.

Ultra-compressed output mode. Cuts response tokens ~65–75% by speaking in telegraphic, fragment-based phrasing while preserving full technical accuracy. Persists across the session until explicitly disabled. TRIGGER when: /caveman invoked, or user says "caveman mode", "talk like caveman", "compress your output", "fewer tokens", "be terse"; disable on "stop caveman" / "normal mode".

Estimate caveman token savings — compare a normal-prose baseline against its compressed form and report tokens-in / tokens-out / % reduction. Heuristic estimate (≈4 chars/token), not exact billing. TRIGGER when: /caveman-stats invoked, or asked how much caveman saved.

Check the CI pipeline status for the current branch after pushing changes. Reports pass/fail with failure details. Use this after every push to confirm your changes are truly done — CI must be green. TRIGGER when: after git push, after committing changes, when asked to check CI, or when verifying that work is complete.

Start or explain a Cloudflared tunnel for a sandbox app port. Cloudflared is the default public sharing method for Open Harness previews; this skill replaces generic sharing guidance with a portable pointer to the installed cloudflared CLI and tmux process convention. TRIGGER when: asked to share a local app publicly, expose a sandbox port, make localhost reachable from another machine, open a preview URL, or run cloudflared.

The critique node of the canonical workflow — run two adversarial critics in parallel (implementer lens + user lens) against a freshly-planned spec, write their findings to .oh/tasks/<slug>/critique.md, and hand off to the approve gate. This is the EVIDENCE half of the critique→approve pair; it produces findings, it does not decide (the /approve gate decides). Runs on local artifacts only, before any GitHub-side state. TRIGGER when: a PRD/plan exists and needs adversarial review before the commitment gate; "critique this spec", "run the critics on <slug>", "review prd before we build".

Detect and report three classes of silent harness drift: framework drift (origin↔upstream), branch-behind/append-file drift, and cron-staleness drift. Never auto-remediates — prints a recommended command for each finding without executing it. TRIGGER when: asked to check for drift, before a session-critical commit, after a long session gap, when MEMORY.md or log.md may have been concurrently modified, or on suspicion that a recently-merged cron is not running.

Run the context fitness-function probe suite (.oh/evals/probes/*.sh) against real state and write the .oh/evals/RESULTS.md benchmark. Each probe is a deterministic 3-state oracle (PASS/REGRESSION/SKIPPED); a green→red transition is surfaced as a REGRESSION naming the lesson it closes. Tier-B behavioral evals are out of scope. TRIGGER when: asked to run evals, check the probe suite, "run /eval", verify a lesson's probe is green, benchmark the harness, or before/after editing a rule/skill that a probe guards.

Spawn 4 parallel sub-agents (PM, Implementer, Critic, Explorer) to audit the harness for improvements. Synthesizes findings into tier-ranked actionable list. Outputs recommended next 3 actions. TRIGGER when: asked to audit the harness, find improvements, review system health, "what should we fix", or periodically via heartbeat.

One-shot draft PRD sketch from a fuzzy scenario. Writes a single artifact to .claude/specs/<slug>/spec.md — a path that is gitignored by default (`.gitignore:51`), so the spec never enters git history. Includes a mermaid diagram. Output is purpose-built as input for `/ship-spec --plan .claude/specs/<slug>/spec.md`, which bypasses /prd's clarifying questions when a plan is provided. No clarifying questions in /imagine itself — generate directly from the <scenario> argument. TRIGGER when: /imagine invoked, or asked to "sketch a spec", "imagine a feature", "draft a quick PRD", "what if we built X".

Adaptive pre-work clarifier. Picks 2–4 task-specific questions via AskUserQuestion, echoes a brief scope summary, then proceeds. Refuses for trivial tasks. TRIGGER when: /interview invoked explicitly, OR before non-trivial work where scope/intent/constraints are ambiguous AND the user hasn't already specified them.

Interact with the Post Bridge social media post management API. Use when the user wants to publish posts, upload media, manage social accounts, schedule content, or retrieve analytics via the Post Bridge platform. Triggers on: post to instagram, upload media to post bridge, schedule a social post, list social accounts, create a post, publish video, post bridge api.

Generate a Product Requirements Document (PRD) for a new feature. Use when planning a feature, starting a new project, or when asked to create a PRD. Triggers on: create a prd, write prd for, plan this feature, requirements for, spec out.

Convert PRDs to prd.json format for the Ralph autonomous agent system. Use when you have an existing PRD and need to convert it to Ralph's JSON format. Triggers on: convert this prd, turn this into ralph format, create prd.json from this, ralph json.

Cut a CalVer release: compute version, pre-flight, push release branch + tag, poll CI, verify GitHub Release + GHCR image. TRIGGER when: asked to release, version, tag, ship, cut a release, or push a new version.

Render an artifact (or in-context material) as a bespoke, self-contained HTML file for one-shot human consumption. Writes to .oh/memory/<UTC-date>/<slug>.html. Output is gitignored — these are consumption artifacts, not source. TRIGGER when: asked to render HTML, generate an HTML report, visualize an audit/council/lint/digest, "make this readable", "make a dashboard for", or as a follow-up to /harness-audit, /strategic-proposal, /skill-lint.

Author and refine Claude Code skills (.claude/skills/*/SKILL.md) the harness way — the evolving pointer to skill best practices (replaces the deprecated skill-builder agent). Covers skill-vs-command-vs-agent, frontmatter, progressive disclosure, argument wiring, side-effect handling, and a validation checklist; defers the full field schema to references/skill-authoring.md. TRIGGER when: asked to create/build/scaffold a new skill, fix/review/audit an existing skill, "make a /command", convert an agent to a skill, or check a skill against best practices.

Score all skills for staleness using 5 dimensions (freshness, usage, integrity, format, dependencies). Reports CURRENT/STALE/BROKEN/DELETE verdicts with specific recommendations. TRIGGER when: asked to audit skills, check skill health, "are my skills stale", "skill lint", or periodically via heartbeat.