forge-studio

Use when the user describes a multi-step feature or refactor and you need to decide whether to handle it solo, dispatch a parallel `/fan-out`, or run a `/worktree-team` planner→generator→reviewer pipeline. Outputs a routing recommendation with the reasoning behind it.

Use whenever the user asks to apply the same operation to many independent files or components — bulk migration, batch refactor across controllers, parallel exploration of unrelated subsystems. Dispatches one subagent per file with a shared prompt template, then collects the results.

Use to toggle the safe-mode flag at .claude/safe-mode. The flag is written automatically by consecutive-failure-guard once tool failures hit FORGE_SAFE_MODE_THRESHOLD (default 5), and block-destructive.sh denies every Bash/Write/Edit while it exists. /safe-mode off clears the flag, logs the exit, and prompts /postmortem; /safe-mode on enters manually; /safe-mode status reports current state.

Run the 5-stage Token Transformation Pipeline (TRAE §5.2.2) over the current session state — Collection, Ranking, Compression, Budgeting, Assembly. Emits a concrete next-action recommendation (/compact, /lean-md, or /progress-log + fresh session) instead of a generic warning.

Audit or scaffold a CLAUDE.md against the Karpathy 4-section structure (Think Before Coding · Simplicity First · Surgical Changes · Goal-Driven Execution). Companion to /lean-md — structure audit, not size trim.

Print every policy enforcement point declared in plugins/diagnostics/registry/policies.json — id, verdict, plugin, hook event, severity, bypass — grouped by verdict. Single discoverable index of what the harness blocks, anchors, nudges, or logs at runtime.

Audit the project against the R.E.S.T. framework (TRAE) — Reliability, Efficiency, Security, Traceability. Reads ledger entries, hook state, and artifact presence across long-session + policy-gateway + context-engine + traces. Single PASS/WARN/FAIL table. Cross-cut meta-surface.

Read the SessionStart timing log written by the diagnostics time-hook wrapper and report per-hook duration plus a cold-vs-warm split across recent sessions. Surfaces which plugin's bootstrap is dominating session-open latency.

Score every SKILL.md in the marketplace against staleness signals (last-edit age, eval coverage, citation freshness, frontmatter compliance) and emit a ranked report. Read-only — never modifies a skill.

Expand the

Produce `.claude/forward-briefing.md` — a forward-framed view of the last 5 progress entries. Re-presents accumulated blockers as open questions and surfaces next steps first, so the resuming session starts in a problem-solving posture rather than a failure-replay posture.

Generate an executable init.sh that bootstraps the dev environment (install, build, test, run commands) so fresh sessions can recreate state in one command.

Append the current session's net outcomes to claude-progress.txt (durable, session-to-session log). Records completions, in-progress work, blockers, and next-step recommendations. Emits a ledger entry for unified audit.

Resume work from the long-session artifacts (claude-progress.txt + spec.md + features.json). Reads the append-only log, the living spec, and the testable-requirements list, then briefs the current session.

Use when the user asks "how do I run things in parallel", "should I use worktrees", or wants the playbook for multi-session and parallel execution — worktrees, fan-out, writer/reviewer splits, headless mode, queue-and-collect. Reference-style passive skill that surfaces examples inline.

Use when the user asks about "thinking modes", "effort levels", "how hard should Claude think", or wants to pick between low/medium/high/xhigh/max for a task — explains the cost/quality tradeoff per level and recommends one based on the work shape.

Use to apply an assessed self-evolution proposal — snapshots the prior resource version under `.claude/lineage/versions/`, writes the new content, and appends a commit entry to the ledger. Refuses to run unless the most recent `assess` verdict is `pass` and the user has explicitly approved.

Initialize .claude/spec.md from the latest plan's

Use to reverse a prior self-evolution commit — restores a snapshot from `.claude/lineage/versions/<slug>/<version>` and appends a rollback entry to the ledger. The rollback itself is logged so history stays append-only and auditable.

Remove signs of AI-generated writing from text. Use when editing or reviewing text to make it sound more natural and human-written. Based on Wikipedia's comprehensive "Signs of AI writing" guide. Detects and fixes patterns including: inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary words, passive voice, negative parallelisms, and filler phrases.

Use when the user says "create a stacked branch", "start a new branch on top of the current one", or "add a layer to the stack" — creates a new git branch whose parent is the currently checked-out branch and registers it in the stack graph so subsequent stack-submit and stack-restack know the lineage.

Use when a parent branch was squash-merged into its base and the child branch now shows duplicate commits — replays only the child's own commits onto the new base using rebase --onto, then retargets the child's PR and force-pushes.

Use when branches in a stacked-PR chain have diverged from their parents and need rebasing — runs the full stack rebase using --update-refs, then force-pushes each moved branch and reconciles its PR body.

Use when you want a read-only view of the current stacked-PR chain — which branches are in the stack, their PR numbers and states (OPEN, MERGED, CLOSED), and which branches have diverged from their parent and need a restack. Renders the full tree without touching any branch or remote.

Use when you want to open or refresh pull requests for every branch in a stacked-PR chain — creates missing PRs with the correct base branch and refreshes the body of existing ones, then force-pushes any branch whose tip has moved.

Run comparative benchmarks for a skill — with-skill vs without-skill, N iterations each. Emits per-iteration benchmark.json with pass_rate/time/token mean+stddev+delta, and per-run grading.json with per-assertion evidence.

Use when you need to answer "if I change <symbol>, what actually breaks?" — joins the static caller graph (codegraph MCP) with recent execution traces to separate code-reachable callers from runtime-exercised ones. Surfaces three sets: static-known-and-exercised (real blast radius), static-known-but-dormant (callable but unused this session), and runtime-only (called but not in graph — likely dynamic dispatch or graph drift).

Analyze recent execution traces to find recurring patterns, failures, and optimization opportunities.

Audit forge-studio SKILL.md frontmatter for the SSL overlay (scheduling / structural / logical). Reports skills missing a measurable success criterion. Read-only.

Compute seven harness-level quality dimensions (arXiv:2605.18747 §5.2.1, arXiv:2605.26112 §4.2) from existing Forge Studio artifacts and render a Markdown scorecard. Use when you want a snapshot of how well the harness is performing across trajectory efficiency, verification strength, recovery ability, state consistency, safety compliance, replayability, and memory hygiene.

Use when the user references prior work, asks "what did we decide about X", or starts a topic that may have been explored before — loads `.claude/memory/index.md`, picks the matching tier-2 topic file, and brings the relevant facts into the current turn. Performs a tier-aware retrieval (pointer → topic → optional transcript) instead of dumping everything.

Use whenever the user shares (or you discover) something that should survive session boundaries — architectural decisions, recurring user preferences, non-obvious constraints, "always do it this way" rules. Writes a tier-2 topic file in `.claude/memory/topics/`, adds a tier-1 pointer to the index, and snapshots any prior version to the lineage ledger.

Use when the user asks to "audit my session", "check for sycophancy", "review behavioral compliance", or otherwise wants a self-discipline pass — scans the current transcript for sycophancy, unnecessary apologies, scope creep, focus drift, and filler language, then reports violations against the rules in `behavioral-core/hooks/rules.d/`.

Use when two or more streams of work must not interleave edits — concurrent feature branches, long-running research alongside implementation, or any time `/fan-out` would race because items share mutable state. Bootstraps N (max 5) parallel agents each in an isolated git worktree with a role-scoped CLAUDE.md and optional path ownership, so each role can only write to its declared subtree.

Scan the marketplace for documentation drift, registration gaps, convention violations, stale memory, and HARNESS_SPEC invariant compliance. Reports only — no writes.

Internal helper — parses a plan file's `## Convergence` section and evaluates its `criterion` shell command against current state. Returns a structured report with `met`, `evidence`, and `gap` fields. Called by /verify and /status; not for direct user invocation.

Manually invoke the agentic workflow for the current task. Reads the active plan and dispatches the right pattern (single-agent, pipeline, fan-out, TDD loop).

Use when the user names a paper PDF in `docs/papers/` and asks to research, audit, or apply its findings to this marketplace — produces a single research brief that grounds the paper's claims in the existing plugin/hook/skill/agent surface, prefers reuse before proposing new, and verifies every Claude Code CLI mechanic the paper touches against current docs.

Use whenever a task is about to be marked done — runs the listed verification commands (tests, lint, type-check, behavioral spot-check), captures the actual output, and compares against `.claude/features.json` or the `/contract` criteria. Refuses to mark done unless every gate produced evidence; closes the trust-then-verify gap that produces "I think it works" claims.

Use when a long session has compacted or handed off, before any Edit, to confirm Claude's belief about file contents still matches disk. Diffs current sha256 signatures against the last-recorded snapshot in `.claude/state/belief.jsonl` for the N most-recently-edited files (default 5).