forge-harness

Routes a proposed skill, plugin, or agent to its correct home — forge-harness (FH) meta-skill, project-local agent, or drop — by applying a 4-criteria meta-skill bar followed by a project-local value test.

Reduces token waste in Claude Code sessions across two axes — context footprint (auto-generates .claudeignore, guides over-read files and /clear timing) and command output (routes to a command-output proxy/hook such as rtk to trim verbose CLI stdout). In hub environments, regularly audits bloated CLAUDE.md/MEMORY.md/memory/*.md files and proposes compression. Usable standalone without a hub clone.

Mines improvement insights from long-form AI dialogue logs and long texts — an external or pasted corpus, not the current session (current-session wrap-up belongs to harvest-loop). First operation is sycophancy removal — separates speakers, strips the counterpart's agreement/restatement spans, converts the user's own utterances into falsifiable propositions, and labels each proposition induced (frame first introduced by the counterpart) or independent (originated with the user). Ships a drop counter so removed material is visible, never silently lost. Triggered by "mine insights from this chat log", "이 대화록에서 통찰 캐줘", "strip the agreement and show me what's actually mine", "동조 걷어내고 알맹이만", "what did I actually contribute in this thread".

Collects the latest AI/harness engineering trends from HackerNews and arxiv, then synthesizes them into actionable insights directly relevant to forge-harness operations. Triggered by "AI trends", "latest trends", "this week's AI updates", "what's on the frontier", "harness trends", "frontier-digest".

Scans a project's harness structure (.claude/ directory, CLAUDE.md, rules, agents) to diagnose complexity, drift, missing files, and broken references, then suggests improvements. `--lint` flag adds language-pattern scan (self-marketing, cushion words, version brags). Triggered by "harness check", "structure diagnosis", "CLAUDE.md audit".

Checks a submitted PR against the environment's baseline assets (CLAUDE.md, memory, naming, asset classification) and attaches a review comment with a merge recommendation. 5 steps — diff read, 8-area consistency check, self-catch, comment, merge recommendation.

Reads the current work context and plans the optimal agent dispatch. Clarifies direction with 1-2 questions when unclear; infers and proceeds immediately when execution path is unclear. Runs an automatic recording gate after each Wave completes. Triggered by "compose agents", "which agent should I use?", "run in parallel", or "agent-composer".

Recruits cross-family verifier sidecars (codex, agy, local 4090 over Tailscale) for adversarial verification of load-bearing changes, maximizing model-family diversity against the orchestrator. Mechanically discovers the available sidecar panel, recruits at least one cross-family verifier when present, and degrades gracefully when none are. The governor (Claude) keeps the terminal verdict; sidecar findings must be source-grounded before acceptance. Opt-in via one-time consent, stored in the UAP; fires only on load-bearing changes. Triggered by "recruit a cross-family check", "decorrelate this verification", "use the idle sidecars to verify", "auto-decorrelation".

Autonomously runs external user reaction simulations, internal audits, ideation scans, artifact validation, and quality reviews. Profiles the target artifact first, then derives task-appropriate personas, dispatches them as parallel agents, classifies findings into M/S/R tiers, and completes the pipeline through to commit automatically.

Given a task description, searches internal and external open-source ecosystems (including Codex marketplace and Claude Code marketplace) to find and recommend suitable plugins with installation guidance. Recommendation is quality-validation based (marketplace-listed + performance-validated), not source-origin based. Activates on "recommend a plugin", "what tool should I use?", "is there a plugin for this?", "recommend a tool". Also checks for duplicate installations.

Renders the FH hub map on demand — the door menu, a starter set of skills, and the most-used trigger phrases — without requiring a greeting. State-aware; composes live candidates from the session card and tracks.

Splits an over-loaded always-loaded context asset — a SKILL.md, CLAUDE.md, or memory index — into a lean always-loaded layer + an on-demand layer, using a governance-semantic criterion (not length, but when the content is needed), connected by imperative pointers. Based on paper §9.5 Protocol-Priority Split pattern. Diagnoses, classifies, splits, and verifies in one pass. Renamed from skill-splitter (old name still routes here). Triggers: "SKILL.md too large", "split this skill", "skill is bloated", "skill file too long", "CLAUDE.md 너무 커".

A skill that immediately updates the baseline and reflects it in the next session when the user raises a counter-argument to an AI recommendation. Triggered by "is that right?", "re-examine this", "something seems off here". Explicit /verify-bidirectional call also possible.

When two skills, agents, or independent research tracks produce conflicting verdicts on the same output, reads the conflict as a signal rather than an error and harvests new skill candidates or insight deltas — including a Dual-Track Grounding conflict (open-frontier track vs internally-grounded recall track disagreeing). Routes skills born from contention to fh-meta (meta-layer), the commons plugin (project-agnostic), or field harvest (domain-specific). Triggered by "two skills conflict", "they produce different conclusions", "contention-layer", "contention harvest", "open vs grounded contradiction", "research tracks disagree".

Scans git history of field projects to detect patterns worth backporting to the meta-harness (Mode A), and logs session work to the knowledge hub when a session ends (Mode B). Triggers on "save this pattern", "make this reusable across projects", "can we automate this repetition?", "push this to FH", "reverse absorption", "harvest pattern", "session log", "log today's work", "sync to hub".

An 8-step self-evolution pipeline that runs after field sessions end (field-harvest → contention-layer → challenger + persona-innovator → synthesizer → Critic agent → harness-doctor → verify-bidirectional → curator), absorbing session learnings back into the FH ecosystem. Runs automatically at session end in the main dev environment; proposes execution first for external users. Triggered by "session harvest", "learning absorption", "fh evolution", or "harvest-loop". (The phrase "run the pipeline" is ceded to pipeline-conductor — use that for end-to-end verification sweeps.)

Ingests a video's content (YouTube and similar) for agent context, routing by capability, task need, and video length — a natively multimodal engine (Gemini via agy or the Gemini API) for direct-URL and long video, Claude vision over ffmpeg-extracted frames when no such engine is present, yt-dlp captions when the transcript alone suffices, and the logged-in claude-in-chrome session for auth-gated videos. The governor cross-checks any multimodal comprehensive-read. Triggered by "ingest this video", "what does this video show", "get the transcript from this YouTube video", "video-ingest".

Multi-perspective synthesis structure — Innovator (propose) → Devil-Advocate (challenge) → Mediator (synthesize) 3-layer execution. Outputs conditional verdicts without binary win/loss. Activates on "deliberation", "battle this out", "weigh the pros and cons", "review from multiple angles", "which side is right?". Optional deep-insight persona jurors for domain-specific views. Designed for design decisions, skill proposals, and architectural choices.

Reviews a technical proposal from the perspective of organizational decision-makers (CTO, technical lead, QA lead, conference reviewers, etc.) and generates an HTML presentation deck. Outputs approval gate results per persona and connects to sim-conductor for improvement suggestions.

Run when setting up a new project for the first time or onboarding after installing FH (first setup, initial configuration, onboarding start, configure project, help me set up). Performs environment detection → gap diagnosis → item-by-item suggestions → user approval → execution → acceleration baseline setup in sequence. Use --dry-run to output diagnosis report only (bg dispatch compatible).

Generates structured prompts to send to each agent in an agent dispatch plan. Triggered by "write the instructions", "what do I say to the agent?", "write the prompt for me", "meta-prompt-builder". Bridges agent-composer (which agents) and prompt content (what to say). Uses Goal/Context/Constraints/Done When structure.

All-angle verification meta-skill for near-complete artifacts. Turns vague design anxiety into structured challenger waves using fh-commons:quench-challenger, then drives defense and convergence until root weaknesses, residual risks, and added complexity are explicit. Covers standard attack/defense rounds, optional Meta-Aware Adversary mode for AI-specific risks such as hallucination, context collapse, prompt injection, and tool lock-in, and Wave-P3 re-attack after an upstream gate declares PASS. Built-in outputs emphasize attack-plus-prescription pairs and can feed fh-meta:persona-innovator after convergence. Triggered by: "quench this", "devil's judgment", "all-angle review", "end-to-end verification", "steel quench", "deep pre-completion inspection", "did it really pass?".

Input-tracing grounding audit for artifacts such as test cases, analysis reports, and design docs. Extracts proper nouns, numbers, citations, version claims, and branching conditions, then back-traces each to declared local files by grep or to external sources by fetch-and-support checks. Marks missing anchors as Phantom Claims and cited-but-unsupporting anchors as Unsupported. A claim is grounded only by non-model evidence: a local hit or literal source span, never another model's agreement. Renamed from source-grounding-audit; old-name references still route here. Triggered by: "phantom detection", "phantom claim", "source back-trace", "where did this come from", "verify source", "does the source support this claim", "grounding audit", "source grounding audit", "citation support check".

Scans git-tracked (public) files for operator-private tokens that should live only in gitignored files — real usernames, absolute home paths, companion-store names, company asset names. Reports file:line + matched token + severity, so a public/private split stays clean before publish. Triggered by "public surface audit", "did I leak anything", "check tracked files for private tokens", "private token scan", "public-surface-audit".

Wraps /goal with a tiered safety + orchestration ladder. core (default): a token budget gate (pre-run estimate), mid-run budget thresholds, and an automatic post-run quality verification via pipeline-conductor — closing /goal's two gaps (Haiku evaluates completion, pipeline-conductor evaluates correctness). pro: adds context-doctor token reduction and agent-composer goal decomposition. max: adds internal-first capability-gap fill (LOCAL_SKILL_REGISTRY skill-bus scan → plugin-recommender external search → auto_project_mapping map/cluster) and cross-ecosystem-synergy-detection pre-validation. The Phase-1 budget verdict auto-recommends the mode. Triggered by "goal with quality gate", "safe goal", "goal-quench", "orchestrate this goal", or before running /goal on high-stakes tasks.

Detects "stale-but-confident" memory entries — facts that were once verified but have silently drifted. Scans memory/*.md for entries past their staleness threshold and proposes re-verification or archival. Runs automatically as part of harvest-loop Step 0-c and on explicit invocation.

A universal gate-reinforcement meta-skill that replaces the "single-pass = done" pattern with a converging loop of up to N rounds. Can be applied to any gate, checkpoint, or verification step. Only declares "truly passed" after FAIL→FIX→re-verify repeats until convergence (all items pass). Escalates to structural redesign if not converged within N rounds. Triggers on: "convergence-loop", "how many rounds do we need", "suspicious of single-pass", "not sure if it really passed", or equivalent phrasing.

Estimates token cost before a multi-step task and outputs a Green/Yellow/Red gate verdict. Tracks actual vs. estimated after completion for calibration. Triggers on "token budget", "how much will this cost", "will this be expensive", "estimate tokens", before long multi-agent tasks.

Fetches and normalizes multi-version public-domain corpora into a single verified-axiom store for verbatim-relay grounding, recording per-version license and provenance.

Expands a named persona seed into a tiered, judgment-mapped cast — tiering each persona by a domain safety rule, mapping each to a decision-lens in the user's vocabulary, then proposing additional voices with sourced anchors.

Detects MCP tool failure patterns and trips a circuit breaker to stop cascading retries. Proposes fallback alternatives and resets when the tool recovers. Triggers on "MCP failing", "tool keeps erroring", "circuit-breaker", repeated tool call failures.

Implements a predict-verify loop for harness edits. Every SKILL.md, rules/*.md, or CLAUDE.md change is paired with a predicted impact stored in a manifest. The next session verifies predictions against actual outcomes, and a validation gate accepts edits only when improvement is measurable. Runs automatically on harvest-loop Step 0-c and on explicit invocation.

DEPRECATED — merged into harness-doctor --lint mode (2026-06-02). Language pattern detection (self-marketing, cushion words, version labels) is now Step 3-L of harness-doctor. Use /harness-doctor --lint instead.

DEPRECATED — merged into agent-composer Step 3-a (2026-06-02). Context Card injection (N≤2 standard / N≥3 Registry mode) + coordination-overhead budget + Focus Mode are now part of agent-composer Step 3. Invoke /agent-composer for parallel dispatch.

Chains the four core FH verification pipelines (harvest-loop → steel-quench → phantom-quench → sim-conductor) into a single gated sweep. Accepts a scope (single skill, specific asset, full harness) and aggregates results into one structured report. Supports --quick mode (steps 2+3 only) and --full mode (all four steps). Triggered by "run the full pipeline", "chain all verifications", "end-to-end sweep", "pipeline-conductor", or "verify everything".

Detects harness regressions by running standard prompt probes after rule/skill changes and comparing outputs against saved baselines. Triggers on "prompt regression", "did my changes break anything", "regression check", "test harness changes".

Evaluates whether a repository meets marketplace listing criteria via a 5-point check and outputs a listing suitability verdict. Checks README completeness, zero-config readiness, maintenance signals, duplication, and public safety.

Audits SKILL.md files in plugins/ for return-path gate compliance. Detects fire-and-forget chains where a caller dispatches a downstream skill but does not gate its next step on the callee's verdict. Reports each chain as CLOSED or OPEN and outputs specific fix suggestions. Triggered by "chain audit", "return-path check", "fire-and-forget detection", "chain closure audit", "skill chain compliance".

Automatically discovers useful combinations (synergy pairs) across multiple installed plugins and skills, and presents them as a ranked table. Proactively suggests undiscovered synergies when new projects or skills are registered in the registry. Activates on phrases like "do my installed tools work well together?", "they seem to work in isolation", "find synergies".

Scans the existing environment before and after meta-harness plugin install to diagnose conflicts, duplicates, and silent overwrite risks. Reports potential conflict areas with existing CLAUDE.md, skills, hooks, and audit practices. Activates on "install conflict", "any overlaps?", "is it safe to install?".