Clean-room loop (discussion -> clean-room-loop workflow launch)

Conversational front door to the clean-room-loop dynamic workflow (.claude/workflows/clean-room-loop.js). The workflow runs in the background with no way to ask anything mid-run, so the discussion happens HERE; answers pass as args.

Claude Code only. Dynamic workflows are a Claude Code feature. In Pi/Codex/OpenCode the Workflow() call will not exist - see the Fallback step. Only this workflow shortcut is Claude Code specific; the underlying clean-room skills (/clean-room:unattended, clean-room-skill run) work on every supported runtime.

Installed project-local. The workflow script ships to project-local .claude/workflows/ (not global). Workflow({ name }) discovers it from the current project's .claude/workflows/ (or ~/.claude/workflows/ if a personal copy exists). If the current project does not have it, initialize a project-local install before launching - see Step 3.

What this is (and is NOT)

• It drives the six clean-room roles with the workflow's OWN agent() subagents (in-session,

subscription, no claude -p), gating every wall crossing with the real clean-room-skill artifact validate --role leakage + schema hooks.

• It is a cost-free path with context-level separation, NOT the OS-enforced wall. A workflow

cannot set CLEAN_ROOM_* env or install hooks, so nothing stops a clean subagent from reading source off disk except the neutral-artifact discipline + the leakage gate. If the user needs the enforced boundary, use clean-room-skill run --agent-runtime claude instead (that path costs API tokens by design).

• The workflow READS the authorized source and WRITES clean specs, plans, code, and reports under the

external artifact roots. Confirm authorization and paths before launching.

Steps

1. Get the brief. Take the end goal from the invocation if present. If missing, ask what they are

reimplementing and why they are authorized to.

1. Run the discussion. Collect the fields below with AskUserQuestion (batch - max 4 per call,

~3 calls); ask free-text ones plainly. Offer defaults so a terse brief is still runnable. Do NOT infer the end goal or target stack from source - clean-room forbids it; if unknown, ask.

Batch A (goal + stack):

• endGoal (free-text): what the clean implementation must do; its success definition.
• targetStack: language / runtime / framework / packageManager / testFramework. Pass as a nested object.
• compatibilityPolicy: public-behavior-and-API-names only (default) or public-behavior-only.

Private structure/comments/internal names are NEVER mirrored.

• featurePolicy (skippable): features to preserve / remove / add / non-goals.

Batch B (policy):

• licensePolicy: destination license + any blocked dep licenses.
• dependencyPolicy: allow new deps? prefer stdlib? require approval for native deps.
• codeHygienePolicy (skippable): max lines per code/test file, max files per iteration.
• schemaProfile: speckit-feature-folder (default) / openspec-delta / gsd-planning-package / kiro-spec-folder.

Batch C (roots + bounds - all safety-relevant):

• sourceRoots (REQUIRED): absolute path(s) to the authorized source. No source = cannot run.
• artifactBase: where run artifacts live (default ~/Documents/CleanRoom). Must be OUTSIDE the

source tree and neutral-named.

• project / taskId (skippable): neutral names; the CLI generates neutral ones if omitted.
• maxIterations: finite inner-loop cap (default 3).

1. Ensure the workflow is installed project-local, preview, STOP for confirmation, then launch.

First confirm this project has the workflow. Dynamic workflows load from project-local .claude/workflows/. If .claude/workflows/clean-room-loop.js is absent in the current project (and no ~/.claude/workflows/clean-room-loop.js personal copy exists), initialize a project-local install before launching:

bash clean-room-skill --claude --local --yes # or, without a global CLI: npx clean-room-skill@latest --claude --local --yes

That writes clean-room-loop.js into the current project's .claude/workflows/ so Workflow({ name }) can discover it. Then show a compact preview of the args (at least endGoal, sourceRoots, artifactBase, targetStack, maxIterations) so a wrong source path or output root is caught BEFORE a filesystem-writing, source-reading run starts. STOP and wait for an explicit "yes". Do NOT call Workflow in the same turn as the preview. Only after the user confirms:

Workflow({ name: "clean-room-loop", args: { endGoal, targetStack, // { language, runtime, framework, packageManager, testFramework } compatibilityPolicy, featurePolicy, licensePolicy, dependencyPolicy, codeHygienePolicy, sourceRoots, // array of absolute paths (REQUIRED) artifactBase, project, taskId, schemaProfile, maxIterations, specSliceRef, // optional }})

Pass only what was gathered; omit the rest (the workflow defaults them).

1. Fallback. If Workflow() errors or is unavailable (non-Claude host, or dynamic workflows

disabled), do NOT hand-run the roles here. Route the user to /clean-room:unattended, which prefers fresh-context in-harness roles on that harness and drops to the durable runner only as a last resort: clean-room-skill run --agent-commands <adapter> on Codex/Pi/other runtimes (spawns the harness CLI, shell: false), or --agent-runtime claude (spawns claude -p, Claude only, per-token) last. Use the runner only once a runner-ready manifest with loop_context exists.

1. Hand back the workflow's result (the terminal clean-room-result.json result string, task

root, and clean/implementation roots).