mikeparcewski/propose-process

Propose Process (Delegation Shim)

Thin shim. Full rubric lives in agents/crew/process-facilitator.md (extracted in #652 item 3 so callers no longer pay the ~200-line rubric context cost on every invocation).

Inputs

• description — project description, issue text, or the task that just completed.
• priors (optional) — wicked-brain:search output on related projects/gotchas.
• constraints (optional) — hard constraints (deadline, stack, compliance envelope).
• mode — propose | re-evaluate | yolo; default propose.
• project_slug OR project — short snake_case identifier (required for real project calls; omit only on output=json measurement / no-project calls).
• project_dir — absolute path; agent writes the draft plan here. If absent, the shim resolves it via the resolve_path.py subpath form: sh "${CLAUDE_PLUGIN_ROOT}/scripts/_python.sh" "${CLAUDE_PLUGIN_ROOT}/scripts/resolve_path.py" wicked-crew projects {project_slug}. On output=json with no project context, an ephemeral tmp dir is used (see Step 0).
• bookend (phase-start | phase-end) and phase — required for re-evaluate from phase-executor.
• current_chain — required for re-evaluate: tasks created so far, status, evidence.
• auto_proceed — bool; set by yolo mode; suppresses user confirmation.
• output — json | unset. json returns the JSON content directly to the caller without creating tasks (used by crew:start Step 5 + measure_facilitator.py).

Outputs (file-handoff contract)

The dispatched agent writes its plan to ${project_dir}/process-plan.draft.json matching skills/propose-process/refs/output-schema.md. The shim reads it back after Task() returns and forwards the JSON to the caller.

Two call shapes are supported:

• Real project call (mode=propose|re-evaluate, project_slug or project provided): ${project_dir} resolves under ~/.something-wicked/wicked-garden/projects/{session-slug}/wicked-crew/projects/{project_slug}/ (the {session-slug} segment is the per-cwd slug from _paths._get_project_slug()). The draft persists at ${project_dir}/process-plan.draft.json. The caller (commands/crew/start.md Step 7, etc.) is responsible for any subsequent persistence (process-plan.json + process-plan.md) and for cleaning up the draft if desired.
• Measurement / no-project call (output=json AND none of project_dir, project_slug, project provided — used by scripts/ci/measure_facilitator.py capture flows and ad-hoc rubric probes): the shim allocates an ephemeral handoff dir under ${TMPDIR:-/tmp}/wicked-garden-measure-{random_id}/, writes the draft there, reads it back, returns the JSON content, and removes the directory before returning. Nothing persists past the call.

Delegation

When invoked:

0. Resolve project_dir (in order):

   • If project_dir is set, use it as-is and mkdir -p it.
   • Else if project_slug or project is set, resolve via sh "${CLAUDE_PLUGIN_ROOT}/scripts/_python.sh" "${CLAUDE_PLUGIN_ROOT}/scripts/resolve_path.py" wicked-crew projects {project_slug} (the resolve_path.py subpath form auto-creates intermediate dirs).
   • Else if output=json AND no project context was provided (measurement / no-project call), allocate an ephemeral dir ${TMPDIR:-/tmp}/wicked-garden-measure-{random_id}/, mkdir -p it, and remember to remove it after Step 3 returns.
   • Otherwise STOP and surface "missing project_dir/project_slug/project (and output != json)" — refer the user to /wicked-garden:report-issue.

1. Dispatch the facilitator agent. Forward EVERY documented input as its own labeled field — no parenthetical "when set" hints, no bundled lists. Substitute each {token} with the actual caller-provided value (e.g. mode=propose, project_slug=auth_rewrite). For optional inputs the caller did not provide, substitute the literal string none so the agent sees an explicit absence:

Task(
  subagent_type="wicked-garden:crew:process-facilitator",
  prompt="""
    Run the propose-process rubric. Documented inputs (one field per line):

    description: {description}
    priors: {priors}
    constraints: {constraints}
    mode: {mode}
    current_chain: {current_chain}
    auto_proceed: {auto_proceed}
    project_dir: {project_dir}
    project_slug: {project_slug}
    bookend: {bookend}
    phase: {phase}

    For any field whose value is `none`, treat it as absent (use the input's
    documented default).

    Write the resulting JSON to {project_dir}/process-plan.draft.json before
    returning. Do NOT issue TaskCreate calls — the caller emits the chain.
  """
)

output is intentionally NOT forwarded — the agent always writes the draft file regardless, and the shim (Steps 3 / 4 below) decides whether to render process-plan.md based on the caller's output value.

See refs/dispatch-example.md for a concrete example with documented inputs filled in for the crew:start Step 5 call.

2. After the agent returns, read ${project_dir}/process-plan.draft.json from disk.

3. If output=json: return the JSON content directly (do NOT create tasks). This is the path used by crew:start Step 5, all re-evaluate callers (crew:execute, crew:just-finish, phase-executor ×2), and measure_facilitator.py. If Step 0 allocated an ephemeral ${TMPDIR}/wicked-garden-measure-{random_id}/ (no-project call), rm -rf that directory after the JSON has been read.

4. Otherwise: render process-plan.md from the JSON, persist ${project_dir}/process-plan.json, then emit the task chain.

Failure modes

Draft file missing or invalid JSON → STOP, surface the first 200 chars of agent output (or the JSON parse error + path). No legacy fallback in v6. Refer the user to /wicked-garden:report-issue.

See also

agents/crew/process-facilitator.md (full rubric) — refs/output-schema.md (JSON contract) — refs/ (factor-definitions, specialist-selection, phase-catalog, evidence-framing, ambiguity, plan-template, interaction-mode, gate-policy, re-eval-addendum-schema, spec-quality-rubric).