@lifeaitools/rdc-skills 0.35.7 → 0.35.9

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "rdc",
-  "version": "0.35.7",
+  "version": "0.35.9",
   "description": "RDC typed-agent dispatch skill suite for Claude Code — plan, build, review, overnight unattended builds with work-item tracking and TDD enforcement.",
   "author": {
     "name": "LIFEAI",

package/clauth-plugin.json

@@ -1,7 +1,7 @@
 {
   "schema": "lifeai.plugin.v1",
   "id": "rdc-skills",
-  "version": "0.35.7",
+  "version": "0.35.9",
   "publisher": "LIFEAI",
   "core": false,
   "documentation": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lifeaitools/rdc-skills",
-  "version": "0.35.7",
+  "version": "0.35.9",
   "description": "RDC typed-agent dispatch skill suite for Claude Code - plan, build, review, overnight builds",
   "keywords": [
     "claude-code",

package/skills/collab/SKILL.md

@@ -505,6 +505,41 @@ chitchat MCP + SSE; you are the build half of a live session.
 - `type: stop` ends the session; send a final summary, then `chitchat_stop`.
 - Stream progress mid-work with `chitchat_reply` on long tasks.
 
+### File-relay transport (the second way `listen` arrives)
+
+Merged from `commands/collab.md` on 2026-08-29, which is now removed. This
+skill documented only the chitchat/SSE transport; the command documented a
+FILE relay, and neither mentioned the other. Two transports for one mode, each
+written down in a place the other's reader would not look.
+
+Invoked as `/rdc:collab --session <session_id>`. claude.ai writes tasks into an
+inbox; you read, act, commit, write the response to an outbox, and loop.
+
+```
+sessionDir = .rdc/relay/sessions/<session_id>/
+inbox      = sessionDir/inbox/
+outbox     = sessionDir/outbox/
+```
+
+1. **Parse** `--session <uuid>`. With no `--session`, list
+   `.rdc/relay/sessions/` and show what is available. If the directories do not
+   exist, say so — `chitchat_start` has to run from claude.ai first. Do not
+   create them.
+2. **Announce.** Set `sessionDir/status.json` to
+   `{ "status": "active", "cli_connected_at": "<iso>", "session_id": "<id>" }`
+   and write a ready signal to the outbox, so claude.ai knows the CLI attached.
+3. **Poll** the inbox for `.md` files not ending in `.processed`, oldest first
+   by name. Nothing found → wait 5s and poll again; after 10 minutes idle print
+   a `Still listening...` heartbeat and keep waiting.
+4. **Process one message.** Read its frontmatter `type`. `stop` ends the
+   session. Anything else is a task: rename the file to `<name>.processed`
+   FIRST so a crash cannot replay it, then act with full capabilities — edits,
+   commits to the lane, `npx tsc --noEmit` (never `pnpm build`), other skills —
+   and write the response to the outbox when done.
+
+`agent-bootstrap.md` rules apply throughout, and Dave typing in the terminal is
+a high-priority override in this transport exactly as in the other.
+
 ---
 
 ## Dave interjections

package/skills/deploy/SKILL.md

@@ -38,8 +38,16 @@ No raw MCP dumps. No UUIDs unless asked.
 - `rdc:deploy audit` — fleet-wide scan for missed failures
 - `rdc:deploy audit --fix` — fleet scan + auto-remediate safe issues
 - `rdc:deploy maintenance <service>` — create, update, or verify one allowlisted private infrastructure service (Mode 7)
+- `rdc:deploy dev <slug>` — explicit alias for the first form above. Plain
+  `rdc:deploy <slug>` already targets PM2 dev, so this only makes the intent
+  unmissable when a reader expects Coolify.
 - `rdc:deploy` (no args) — print mode menu, ask which
 
+> `<ref>` is what `commands/deploy.md` called `<build-id>` before that file was
+> removed (2026-08-29). Same thing — a registered manifest ref, a commit, or a
+> tag. With no ref, the deployment is the latest commit on the app's WATCHED
+> branch, not on whatever branch happens to be checked out locally.
+
 ## Modes
 
 ### Mode 1 — deploy <slug> [ref]

package/skills/handoff/SKILL.md

@@ -159,6 +159,9 @@ Est: <hours>',
   p_labels      := ARRAY['<label>'],
   p_estimated_hours := 2,
   p_source      := 'planning'
+  -- If the epic has architecture_ref set, also add a required
+  -- architecture-fidelity-<slug> checklist row via p_checklist here — the exit gate
+  -- hard-rejects `done` on any task under an architecture_ref epic that lacks one.
 );
 ```
 

package/skills/plan/SKILL.md

@@ -33,6 +33,30 @@ description: rdc:plan (topic) — produce architecture, decisions and an epic wi
    - Relevant CLAUDE.md files from affected packages
    - Existing Supabase epics: `SELECT get_open_epics()`
 
+1b. **Identify the affected domains, then load the matching architecture doc.**
+
+   Merged from `commands/plan.md` on 2026-08-29, which is now removed. This file
+   referenced architecture docs 9 times and the command 28 — but only the
+   command carried the ROUTING, so the instruction to load the right doc lived
+   in the copy a reader might never open.
+
+   | Domain keywords in the topic | Architecture doc to read |
+   |---|---|
+   | PRT, trust, capital, NAV, investor, land, DST | `docs/systems/prt/ARCHITECTURE.md` |
+   | CS 2.0, HAIL, PAL, virtue, quad-pixel, ontology, BPMN, cognitive | `docs/systems/cs2/ARCHITECTURE.md` |
+   | marketing, CRM, campaign, contact, outreach, RDC app | `docs/systems/rdc/ARCHITECTURE.md` |
+   | Claude workflow, skills, agents, dispatch, rdc:build | `docs/systems/claude-workflow/ARCHITECTURE.md` |
+   | Life AI, LIFEAI platform, life.ai | `docs/systems/lifeai/ARCHITECTURE.md` |
+   | media, R2, images, regen-media, MCP image | `docs/systems/media/ARCHITECTURE.md` |
+   | UI, component, brand, design token, shared, OG image | `docs/systems/shared/ARCHITECTURE.md` |
+
+   `.claude/rules/system-quick-links.md` is the routing map to all of them. A
+   topic spanning several domains reads ALL the matching docs before proceeding.
+
+   **A plan that contradicts an existing architecture doc is invalid.** Load them
+   first, and when a decision conflicts with one, flag the conflict rather than
+   quietly planning around it.
+
 2. **Read the codebase** — understand current state:
    - What packages are affected?
    - What types/interfaces already exist?

package/skills/release/SKILL.md

@@ -74,6 +74,31 @@ If blocked, abort immediately with the message above. Do NOT proceed to the vers
 
 If PUBLISH.md is absent and app has no `app_deployments` row (library/package), skip this gate.
 
+## Rules
+
+Merged from `commands/release.md` on 2026-08-29, which is now removed. Each of
+these was in the command and NOT in this file, which is precisely the drift that
+shipping one verb as two documents produces.
+
+- Do not release without explicit user authorization.
+- Prefer repo-local release instructions in `.rdc/release.json`, README, package
+  scripts, or CI workflows.
+- **Never force push or bypass hooks.**
+- **Never declare success without verifying the installed or deployed version.**
+  A publish that exits 0 is not a published package — the registry can lag
+  minutes behind, so check the version endpoint, not the success line.
+
+For a `package`-class target that already resolves through `rdc-harness` (a real
+monorepo subtree, not a standalone repo like this one),
+`packages/deploy/src/runners/registry-release.mjs` already proves the
+"Tests/self-test passed" through "Local install/update executed" rows safely:
+real `npm pack`, isolated-prefix install (never the real global store), real
+verify, and `--live` explicitly gating the actual publish. Where applicable,
+`node C:/Dev/rdc-harness/bin/rdc-harness.mjs deploy <slug> [--live]` can supply
+that evidence directly instead of hand-rolling the same pack/install/verify
+cycle. It does **not** replace version bump, tag or push — the harness CLI does
+none of those.
+
 ## Resolution Order
 
 1. Current repo if `<repo>` is `.` or omitted and the user clearly refers to the current workspace.

package/tests/help-surface.test.mjs

@@ -7,10 +7,22 @@ import { fileURLToPath } from 'node:url';
 const __dirname = dirname(fileURLToPath(import.meta.url));
 const root = resolve(__dirname, '..');
 
+/**
+ * `help` has ONE surface now.
+ *
+ * It used to ship as both commands/help.md and skills/help/SKILL.md, which is
+ * half of why a single verb appeared four times in the command list. The
+ * command was removed on 2026-08-29 after checking substance rather than
+ * counting lines: the skill already carried the manifest resolution order, the
+ * plugin.json path and the slash forms.
+ *
+ * Kept as a map rather than collapsed to one constant so the loop below still
+ * names which document failed — and so restoring a second surface, if that ever
+ * becomes right, is one line.
+ */
 const files = {
   readme: join(root, 'README.md'),
   skillHelp: join(root, 'skills', 'help', 'SKILL.md'),
-  commandHelp: join(root, 'commands', 'help.md'),
 };
 
 const docs = Object.fromEntries(
@@ -40,12 +52,14 @@ assert.match(docs.readme, /Nineteen[\s\S]*\/rdc:\*` command shorthands/i, 'READM
 assert.match(docs.readme, /Use `rdc_skill_list` for the authoritative live catalog/, 'README should point callers to live MCP catalog');
 assert.doesNotMatch(docs.readme, /All user-invocable skills become available as slash commands/, 'README must not imply all MCP skills are slash commands');
 assert.doesNotMatch(docs.readme, /29 skills organized into 6 categories/, 'README must not carry stale category count');
-assert.match(docs.commandHelp, /all MCP skills/, 'command help should refer to MCP skill catalog');
+// One surface, so these assert once. They were duplicated across commandHelp
+// and skillHelp; the negative pair moves to the surviving document rather than
+// being dropped — a stale-wording check is worth keeping regardless of which
+// file carries the text.
 assert.match(docs.skillHelp, /all MCP skills/, 'skill help should refer to MCP skill catalog');
 assert.match(docs.skillHelp, /manifest-driven/i, 'skill help should be manifest-driven');
-assert.match(docs.commandHelp, /manifest-driven/i, 'command help should be manifest-driven');
-assert.doesNotMatch(docs.commandHelp, /Print the full usage menu below verbatim/, 'command help must not use stale static menu wording');
-assert.doesNotMatch(docs.commandHelp, /get\/<service>/, 'command help must use current clauth /v/<service> wording');
+assert.doesNotMatch(docs.skillHelp, /Print the full usage menu below verbatim/, 'help must not use stale static menu wording');
+assert.doesNotMatch(docs.skillHelp, /get\/<service>/, 'help must use current clauth /v/<service> wording');
 
 const skillDirs = readdirSync(join(root, 'skills'))
   .filter((name) => {

package/commands/build.md

@@ -1,223 +0,0 @@
----
-name: build
-description: rdc:build (epic-id) - [--no-review] — execute a planned epic, then gate and ship to dev
----
-
-> **⚠️ OUTPUT CONTRACT (READ FIRST):** `guides/output-contract.md`
-> Checklist-only output. No tool-call narration. No raw MCP/JSON/log dumps.
-> One checklist upfront, updated in place, shown again at end with a 1-line verdict.
-
-> **Sandbox contract:** This skill honors `RDC_TEST=1` per `guides/agent-bootstrap.md` § RDC_TEST Sandbox Contract. Destructive external calls short-circuit under the flag.
-
-
-# rdc:build — Typed Agent Dispatch Engine
-
-## When to Use
-- Plan is approved and ready to execute
-- Project lead says "build it", "go", "execute", "do not stop"
-- An epic exists with child tasks ready for implementation
-- Called by `rdc:overnight` as part of the automated build loop
-
-## Arguments
-- `rdc:build <epic-id>` — build from a specific Supabase epic
-- `rdc:build <topic>` — find the epic by label/title match
-- `rdc:build` (no args) — show open epics and ask which to build (interactive only)
-- `rdc:build <epic-id> --unattended` — silent mode for overnight builds
-
-## Agent Types & Guide Files
-
-Every dispatched agent MUST read two files before starting — in this order:
-1. `{PROJECT_ROOT}/.rdc/guides/agent-bootstrap.md` — credentials, git rules, completion report format
-   (fallback: `{PROJECT_ROOT}/.rdc/guides/agent-bootstrap.md` if `.rdc/` does not exist)
-2. `{PROJECT_ROOT}/.rdc/guides/<type>.md` — role-specific guide
-   (fallback: `{PROJECT_ROOT}/.rdc/guides/<type>.md`)
-
-Include both lines in every agent prompt:
-```
-"Read {PROJECT_ROOT}/.rdc/guides/agent-bootstrap.md first (fallback: .rdc/guides/agent-bootstrap.md), then {PROJECT_ROOT}/.rdc/guides/<type>.md (fallback: .rdc/guides/<type>.md) before starting."
-```
-
-| Agent Type | Guide File | When to dispatch |
-|-----------|-----------|-----------------|
-| `frontend` | `.rdc/guides/frontend.md` | React components, pages, UI, Tailwind, animation |
-| `backend` | `.rdc/guides/backend.md` | API routes, server components, database queries, auth |
-| `data` | `.rdc/guides/data.md` | Migrations, schema changes, RPC functions |
-| `design` | `.rdc/guides/design.md` | Visual design, brand palettes, OG images, token work |
-| `infra` | `.rdc/guides/infrastructure.md` | CI/CD, deployment, DNS, SSL |
-| `content` | `.rdc/guides/content.md` | Marketing copy, messaging, tone |
-| `cs2` | `.rdc/guides/cs2.md` | CS 2.0 paradigm work (generic) |
-| `hail` | `.rdc/guides/cs2.md` + `packages/hail/CLAUDE.md` | Grammar, DSL compiler, evolution |
-| `pal` | `.rdc/guides/cs2.md` + `packages/pal/CLAUDE.md` | Sessions, moment windows, graph memory |
-| `bpmn` | `.rdc/guides/cs2.md` + `docs/systems/<domain>/flowable-bpmn-architecture.md` | BPMN flows, governance |
-| `virtue` | `.rdc/guides/cs2.md` + `packages/virtue-engine/CLAUDE.md` | Virtue weights, coherence, certification |
-| `viz` | `.rdc/guides/frontend.md` + `.rdc/guides/design.md` | Custom viz components, charts, diagrams |
-
-### How to classify a task → agent type
-
-Read the task title and description, then:
-- Mentions React, component, page, UI, Tailwind → `frontend`
-- Mentions API route, server, database query, auth → `backend`
-- Mentions migration, schema, table, RPC → `data`
-- Mentions brand, palette, typography, OG image → `design`
-- Mentions deploy, infrastructure, CI, DNS → `infra`
-- Mentions copy, messaging, email template → `content`
-- Mentions grammar, DSL, compiler → `hail`
-- Mentions session, moment, memory graph → `pal`
-- Mentions BPMN, flow, governance → `bpmn`
-- Mentions virtue, coherence, certification → `virtue`
-- Mentions visualization, chart, diagram, SVG → `viz`
-- Multiple types? Dispatch multiple agents, each with its guide.
-
-### Execution primitive for create/open/build/deploy checklist rows
-
-When a dispatched agent's checklist row is to materialize a product shape,
-open a signed edit session, run a target's declared build gates, or deploy
-to dev-PM2/npm-registry, it uses the real, tested `rdc-harness` CLI instead
-of hand-rolled bash/curl:
-
-```bash
-node C:/Dev/rdc-harness/bin/rdc-harness.mjs <create|open|edit|build|deploy> <slug> --monorepo-root <the dispatched agent's own worktree>
-```
-
-One JSON receipt per call, exit 0/1 — tick the checklist row with the parsed
-receipt as evidence, not the raw dump. No Coolify awareness (production
-deploy stays `/rdc:deploy`'s own path) and no live co-editing surface
-outside `site-html`/`site-ts` (other classes get file-boundary save only —
-real, currently-unbuilt gap for other product classes, not something to
-paper over here). `open`/`edit` require `RDC_HARNESS_ISSUER_SECRET` set
-explicitly per-session — never a default.
-
-## Procedure
-
-1. **Load the epic and its durable admission decisions:**
-   ```sql
-   SELECT get_work_items_by_epic('<epic-id>');
-   ```
-   - Read `design_review_state` and `status` for every executable child.
-   - Only `automatic_approved`, `human_approved`, or legacy `not_required` rows may be considered for dispatch.
-   - For `pending`, `needs_human`, or `rejected`, write an `admission_refocus` receipt, keep the child blocked, and route it to the reviewer/planner. **Do not dispatch it, retry it, or call the epic complete.**
-   - Interactive (no args): show open epics, ask which to build
-   - Unattended (no tasks found): escalate via advisor tool
-   - **Read the epic's `plan_ref`, `spec_ref`, `architecture_ref`, and `scoping_statement` fields** (returned on the epic row itself). `scoping_statement` bounds what this build may touch — do not silently expand past it. If `architecture_ref` is set, this epic crosses an architectural boundary: read that doc now, before classifying or dispatching any task, and carry it into every agent prompt in step 7.
-
-1a. **Run the durable CodeFlow supervisor before each wave and after every gate-changing action.**
-   - Invoke `runOrchestrator()` with the project manifest, `SupabaseStateStore`, and the real phase dispatcher. It is the sole authority for resuming/refocusing a phase DAG; do not reconstruct waves by hand from task prose.
-   - A returned `admission_refocus` or `pipeline_blocked` is a durable hold, not a failed attempt to work around. Preserve its task state and route the required Design Review or validator closure.
-   - Only a returned `pipeline_complete` whose phase tasks are all design-review admitted **and** durably `done` permits an epic completion claim. If the project lacks a real dispatcher/manifest, report `BLOCKED: CodeFlow supervisor entrypoint unavailable` rather than emulating completion.
-
-2. **CHECK FOR EXISTING WORK (mandatory — never skip):**
-   ```sql
-   -- Check if prototypes exist from earlier sessions
-   SELECT name, component, source_path, status, notes
-   FROM prototype_registry
-   WHERE status IN ('prototype', 'converting')
-   ORDER BY created_at DESC;
-
-   -- Check for design decisions on this topic
-   SELECT topic, context_type, summary, source
-   FROM design_context
-   WHERE topic ILIKE '%<epic-topic>%'
-   ORDER BY created_at DESC;
-   ```
-   **If a prototype exists: ADAPT IT. Do not build from scratch.**
-   Tell the agent: "Read <source_path> first and convert it to the production contract."
-   
-   **If design decisions exist: follow them.** Include the summary in the agent prompt.
-
-3. **Load the plan** (if exists): check `.rdc/plans/` for matching topic (fallback: `.rdc/plans/`).
-
-4. **Read CLAUDE.md files** for all affected packages, plus `docs/CODING-STANDARDS.md`
-   (SOLID/Clean-Architecture standard — regen-root; skip if absent) — carry it into every
-   dispatched agent prompt.
-
-5. **Classify each task** → assign agent type from the table above.
-
-6. **Use the supervisor-resolved waves** — parallelize only phases returned by `runOrchestrator()` after its durable admission check:
-   - Wave 1: independent tasks (different packages/files)
-   - Wave 2: tasks that depend on Wave 1 outputs
-   - Wave 3: integration tasks
-
-7. **For each wave — dispatch typed agents in parallel:**
-   - Set work item to `in_progress` before dispatching
-   - Each agent prompt MUST include:
-     - `"Read {PROJECT_ROOT}/.rdc/guides/agent-bootstrap.md first (fallback: .rdc/guides/agent-bootstrap.md), then {PROJECT_ROOT}/.rdc/guides/<type>.md (fallback: .rdc/guides/<type>.md) before starting."`
-     - Specific files to create/modify
-     - Exact deliverables and commit message
-     - The epic's `scoping_statement` — explicit boundary on what this task may and may not touch
-     - `"NEVER run pnpm build/test. NEVER modify files outside your scope."`
-   - **If the epic's `architecture_ref` is set:** include `"Read <architecture_ref> before implementing. Your task's checklist requires a checked architecture-fidelity-<slug> row before this item can close — when you tick it, its evidence must cite the specific section/boundary of <architecture_ref> your implementation conforms to, not just 'done'."` A task under an `architecture_ref` epic will hard-fail at the exit gate (step 9) without this row checked with real evidence.
-   - Use `run_in_background: true` for parallel execution
-   - NEVER let agents overlap on the same files
-
-8. **Post-wave test gate (mandatory):**
-   After all agents in a wave complete, before marking tasks done:
-   ```bash
-   # For each package modified in this wave:
-   cd packages/<name> && npx vitest run 2>&1 | tail -20
-   ```
-   - All tests must pass before proceeding to next wave
-   - If tests fail: fix before marking the wave done
-   - NEVER use `pnpm build` or `pnpm turbo test` — vitest only per package
-   - New code must have tests: if a modified package shows 0 new test files, flag it
-
-9. **As agents complete:**
-   - Verify commit landed on the development branch
-   - Push to origin *(skip if `$RDC_TEST=1` — echo `[RDC_TEST] skipping git push` instead)*
-   - Ensure the agent submitted `implementation_report.codeflow_post`, then set the work item to `review`; the validator closes `done`
-   - Re-invoke `runOrchestrator()` after the durable status/gate update. A task in `review` remains incomplete even when its phase gate passed.
-   - **If the epic's `architecture_ref` is set:** before the validator attempts `done`, confirm the task's checklist has a checked `architecture-fidelity-*` row with real evidence (a cited doc section, not a bare "matches"). `update_work_item_status(..., 'done')` will hard-reject otherwise — catching this here avoids a wasted validator round-trip.
-   - Continue to next wave
-
-   **If an agent fails:**
-   - Interactive: diagnose before retrying
-   - Unattended: retry once; on second failure escalate via advisor
-     ```
-     BUILD_STATUS: { wave, tasks_done, tasks_failed, commits, escalated: true }
-     ```
-
-10. **Final verification gate (mandatory — before marking work or epic done):**
-    Dispatch the verify agent (see `guides/agents/verify.md`) across every package/app touched in this build.
-    The Iron Law: **NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE.**
-    - Run `npx vitest run --dir <pkg>` fresh for each touched package
-    - Run `npx tsc --noEmit --project <pkg>/tsconfig.json` for each
-    - Read the full output — zero failures, zero type errors
-    - If any step fails: fix and re-run the entire gate. Do not skip.
-    - NEVER `pnpm build` / `pnpm test` / `pnpm -r` (crashes machine)
-    - **ATF Test-Ladder / rdc-harness (WIP — best-effort, not a hard gate yet):** if a
-      touched package/repo ships an ATF `STP-001.md` or an `rdc-harness`-style
-      `tools/mutate-check.mjs`/`tools/proof-ledger.mjs` pair, run it and quote the result
-      alongside vitest/tsc. A red mutation gate is a real finding — report it, do not
-      silently drop it because it isn't wired into this checklist as required yet.
-      Absence of either system in the target is not a failure; do not install one ad hoc.
-
-11. **After verification passes:**
-    - Push all commits:
-      ```bash
-      if [ "$RDC_TEST" != "1" ]; then
-        git push origin {development-branch}
-      else
-        echo "[RDC_TEST] skipping git push origin {development-branch}"
-      fi
-      ```
-    - Re-invoke `runOrchestrator()` and require its `pipeline_complete` receipt before `bump_epic_version()` or any epic completion claim. A clean code review or green test suite is not a substitute for admitted, validator-closed work items.
-    - Report summary with verification evidence quoted
-
-## Agent TDD Requirements
-
-When dispatching agents, include in every prompt:
-```
-TDD REQUIREMENT: Write tests FIRST for new functions/modules.
-Run: npx vitest run packages/<name> to verify red → implement → verify green.
-NEVER run pnpm build or pnpm turbo. Use npx vitest run only.
-```
-
-## Rules
-- Branch: development branch only (auto-commit, no confirmation needed)
-- NEVER let two agents edit the same file
-- NEVER run `pnpm build` (crashes system) — code only
-- Every agent reads its guide file — no exceptions
-- Update Supabase work items IN REAL TIME — not batch at end
-- **Never dispatch, resume, or complete around `design_review_state`; the durable database result and `runOrchestrator()` receipt win over an agent's narrative**
-- Push after each wave, not just at the end
-- Unattended: NEVER pause — continue automatically
-- Unattended: max 2 retries per task before escalating to advisor

package/commands/collab.md

@@ -1,179 +0,0 @@
----
-name: collab
-description: rdc:collab () - [--session] — run a structured agent-to-agent conversation
----
-
-> **⚠️ OUTPUT CONTRACT (READ FIRST):** `guides/output-contract.md`
-> Checklist-only output. No tool-call narration. No raw MCP/JSON/log dumps.
-> One checklist upfront, updated in place, shown again at end with a 1-line verdict.
-
-
-# /rdc:collab — Claude Code Collab Session Listener
-> Invoked as: `/rdc:collab --session <session_id>`
-> You are the build/execute half of a live collab session with claude.ai.
-> Transport: file relay via `.rdc/relay/sessions/<id>/inbox/` + `outbox/`
-> Dave is watching this terminal and can interject at any time.
-
----
-
-## What This Is
-
-claude.ai writes tasks into your inbox. You read, act, commit, write the response
-to outbox, and loop. clauth relays messages between claude.ai and this session via
-the `chitchat_*` MCP tools. Dave can watch everything in this terminal and interject
-by typing — treat anything Dave types as a high-priority override.
-
----
-
-## Step 1 — Parse session ID
-
-Extract `--session <uuid>` from args.
-
-If no `--session`, list `.rdc/relay/sessions/` and show available sessions.
-
----
-
-## Step 2 — Initialize
-
-```
-sessionDir = .rdc/relay/sessions/<session_id>/
-inbox      = sessionDir/inbox/
-outbox     = sessionDir/outbox/
-```
-
-Verify both dirs exist. If not:
-```
-Session directory not found. Run chitchat_start from claude.ai first.
-```
-
-Update `sessionDir/status.json`:
-```json
-{ "status": "active", "cli_connected_at": "<iso>", "session_id": "<id>" }
-```
-
-Write a ready signal to outbox so claude.ai knows the CLI is connected:
-
-```
-outbox/<iso-ts>-ready.md
----
-from: claude-code
-to: claude-ai
-session_id: <id>
-type: ready
-responded_at: <iso>
----
-Claude Code connected. Ready to receive tasks.
-cwd: <rootPath>
-```
-
-Print to terminal:
-```
-[rdc:collab] Session <id> active.
-Inbox:  .rdc/relay/sessions/<id>/inbox/
-Outbox: .rdc/relay/sessions/<id>/outbox/
-Waiting for messages from claude.ai... (Ctrl+C to end)
-```
-
----
-
-## Step 3 — Poll inbox
-
-Scan `inbox/` for `.md` files that do NOT end in `.processed`. Sort ascending by name.
-
-- **No files:** wait 5 seconds, poll again.
-- **After 10 min idle:** print `[rdc:collab] Still listening...` heartbeat, keep waiting.
-- **Files found:** process the oldest one first (Step 4).
-
----
-
-## Step 4 — Process message
-
-Read the file. Parse frontmatter `type` field.
-
-**`type: stop`** → go to Step 6.
-
-**Anything else (default: task/message):**
-
-Print to terminal:
-```
-[rdc:collab] Turn <N> from claude.ai:
-──────────────────────────────────────
-<message body>
-──────────────────────────────────────
-```
-
-Rename the inbox file to `<filename>.processed`.
-
----
-
-## Step 5 — Do the work
-
-Act on the message. Full Claude Code capabilities:
-- File edits, git commits to `develop`
-- Supabase RPC queries
-- Type-checks: `npx tsc --noEmit` (never `pnpm build`)
-- Run skills: `/rdc:plan`, `/rdc:fixit`, etc.
-- Answer questions directly
-
-Follow `.rdc/guides/agent-bootstrap.md` rules throughout.
-
-When done, write response to outbox:
-```
-outbox/<iso-ts>-turn<N>.md
----
-from: claude-code
-to: claude-ai
-session_id: <session_id>
-turn: <N>
-responded_at: <iso>
-commits: <sha1, sha2 or none>
----
-
-<what you did, what you found, any questions or decisions needed from claude.ai>
-```
-
-Print to terminal:
-```
-[rdc:collab] Turn <N> done. Response written to outbox.
-Waiting for next message...
-```
-
-Return to Step 3.
-
----
-
-## Step 6 — End session
-
-Received `type: stop` in inbox, or Dave pressed Ctrl+C.
-
-Write final summary to outbox:
-```
-outbox/<iso-ts>-final.md
----
-from: claude-code
-to: claude-ai
-session_id: <session_id>
-type: final
-responded_at: <iso>
----
-Session complete.
-Turns: <N>
-Commits: <list or none>
-Open items: <anything unresolved>
-```
-
-Update `status.json` → `{ "status": "done", "ended_at": "<iso>" }`
-
-Print:
-```
-[rdc:collab] Session ended.
-```
-
----
-
-## Dave Interjections
-
-If Dave types in this terminal during a turn:
-- Treat it as an override injected into the current task
-- Acknowledge it in your outbox response
-- If it changes direction mid-task, note what you stopped and why