@lifeaitools/rdc-skills 0.37.11 → 0.38.1

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "rdc",
-  "version": "0.37.11",
+  "version": "0.38.1",
   "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",
@@ -2121,6 +2121,46 @@
       "codeflow_required": false,
       "purpose": "orient before working: position, target from the registry, and the harness shape"
     },
+    "close": {
+      "name": "close",
+      "slash": "rdc:close",
+      "category": "dev-loop",
+      "usage": "rdc:close",
+      "args": {
+        "positional": [],
+        "flags": []
+      },
+      "requires": [
+        "git",
+        "work-items-rpc"
+      ],
+      "produces": [
+        "git_commits",
+        "git_pushes",
+        "deploys"
+      ],
+      "default_model": "sonnet",
+      "triggers": [
+        "can I exit",
+        "close the session",
+        "are we done",
+        "is everything landed",
+        "rdc:close",
+        "wrap up the session",
+        "end the session"
+      ],
+      "follows": [
+        "build",
+        "fixit",
+        "overnight"
+      ],
+      "leads_to": [],
+      "sandbox_aware": false,
+      "output_contract": "guides/output-contract.md",
+      "enabled_default": true,
+      "codeflow_required": false,
+      "purpose": "decide whether the session can exit (commits, land, deploy, stash), fix what is unfinished, and record it as closed via rdc-close"
+    },
     "new": {
       "name": "new",
       "slash": "rdc:new",

package/README.md

@@ -318,6 +318,7 @@ skills/
   build/SKILL.md       (dispatch agents in waves — mandatory validator gate)
   channel-formatter/SKILL.md (channel-native formatting and content packs)
   clean-code-analyzer/SKILL.md (naming, dead code and complexity smells)
+  close/SKILL.md       (can this session exit: verify, fix, record closed)
   co-develop/SKILL.md  (peer-aware Claude/Codex co-development)
   collab/SKILL.md      (bidirectional claude.ai ↔ Claude Code relay)
   convert/SKILL.md     (Office/Markdown conversion)

package/clauth-plugin.json

@@ -1,7 +1,7 @@
 {
   "schema": "lifeai.plugin.v1",
   "id": "rdc-skills",
-  "version": "0.37.11",
+  "version": "0.38.1",
   "publisher": "LIFEAI",
   "core": false,
   "documentation": {

package/git-sha.json

@@ -1,3 +1,3 @@
 {
-  "sha": "a26cb2cc3a96f0b96f52c7b28c5dfef0264856e0"
+  "sha": "597fd36853cf0e514b5b8729ac1de893ea026b00"
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lifeaitools/rdc-skills",
-  "version": "0.37.11",
+  "version": "0.38.1",
   "description": "RDC typed-agent dispatch skill suite for Claude Code - plan, build, review, overnight builds",
   "keywords": [
     "claude-code",
@@ -19,7 +19,7 @@
     "type": "plugin",
     "skills": "skills/",
     "guides": "guides/",
-    "version": "0.37.11",
+    "version": "0.38.1",
     "commands": "commands/"
   },
   "bin": {

package/skills/close/SKILL.md

@@ -0,0 +1,85 @@
+---
+name: close
+description: rdc:close () — decide whether this session can exit, fix what is unfinished, and record it as closed
+---
+
+> **⚠️ 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.
+
+> If dispatching subagents or running as a subagent: read `{PROJECT_ROOT}/.rdc/guides/agent-bootstrap.md` first (fallback: `{PROJECT_ROOT}/.rdc/guides/agent-bootstrap.md`).
+
+
+# rdc:close — Can this session exit?
+
+## When to Use
+- Before ending any session that changed something.
+- Dave asks "are we done", "can I exit", "is everything landed".
+- As the last step of `rdc:build`, `rdc:fixit` and `rdc:overnight`.
+
+## What it is
+
+`rdc-close` (`$LIFEAI_ENV/bin/rdc-close.mjs`) holds every mechanical check and the session marker. This skill is the
+judgment layer on top: it fixes what the engine reports. The engine never commits, edits or deploys; those need a
+decision, and the decision is yours.
+
+If a session ends without a close, a SessionEnd hook (`rdc-close --hook --detach`, used by the harness and not by
+you) starts a detached worker that lands what is **already committed** and writes an unclean record for the rest.
+The next session in that lane is told what was left, and is also told when a lane starts with uncommitted paths even
+though no exit ran (a crash fires no SessionEnd). So close properly; the hook is a safety net, not the plan.
+
+## Procedure
+
+1. **Get the verdict.** `node "$LIFEAI_ENV/bin/rdc-close.mjs" check` (add `--json` to parse). Exit 0 is clear.
+   It reads root, integration branch and ship route from the work contract, so the same call serves every repository.
+
+2. **Fix each blocking check, then re-run `check`.**
+
+   | Check | Blocks when | Fix |
+   |---|---|---|
+   | `contract` | a proof row is unproved | `rdc-work verify --all`; then prove it, or `drop` / `ship-row` / `hold` with a reason. Never edit code to turn a gate green. |
+   | `clean` | files are uncommitted or untracked | Commit under the declared contract with its `Work-Item:` trailer, or remove scratch with `rm`. Stage explicit paths, never `git add -A`. Stash only work that is truly deferred, tagged `rdc-<first 8 of the session id>`. |
+   | `landed` | a commit is not on `origin/<integration>` | Run the contract's ship route (`rdc-work status` prints it: `node scripts/land.mjs`, `node machines/land.mjs`, ...). A pushed lane branch is not landed. |
+   | `other_repos` | a repository this session touched is not clean | Repeat the two rows above in that repository. |
+   | `deploy` | `fail`: a deployable changed and no deploy followed | `/rdc:deploy <slug>`. `unknown`: the deploy manager's job window does not reach the change, so deploy if this session changed that app and say so. `na`: nothing to do. |
+   | `stash` | a stash entry carries this session's tag | Pop, commit or drop it with a reason. Entries tagged for other sessions are not yours; leave them. |
+
+3. **Housekeeping the engine cannot decide.**
+   - An `apps/*` or `sites/*` change bumps that app's displayed version in the same commit (`version-numbering.md`).
+     A published package bumps only inside `/rdc:release`.
+   - A non-obvious lesson whose fix is not encoded in a commit: `rdc-lesson submit`.
+   - `/rdc:report` when the work contract's tier asks for a session report.
+   - Finalize the CodeFlow session (`codeflow_finalize`).
+
+4. **Close.** `node "$LIFEAI_ENV/bin/rdc-close.mjs" run`:
+   - runs every check, even when a marker already exists (the marker covers HEAD and the tree, not stash, other
+     repositories, deploy or the contract);
+   - if the lane is clean and a commit is not yet on `origin/<integration>`, lands it through the contract's ship
+     route. In lifeai-env that route publishes the package by default, so `run` uses it land-only
+     (`--existing --no-publish`): no publish and no global install. Releasing stays a deliberate `/rdc:release`;
+   - runs `clauth scrub`;
+   - writes the session marker only when every check is clear. Otherwise it writes an unclean record and retires
+     any marker.
+
+   Exit 0 means closed. Exit 1 prints the failing checks: go back to step 2. `rdc-close status` says whether this
+   session's marker is current (exit 0) or not (exit 1). Exit 2 is a usage error.
+
+5. **Report one line.** Closed (marker current, with the commit it names), or the blockers that remain and the one
+   decision that clears them.
+
+## Rules
+
+- Never say "closed" without `rdc-close run` exiting 0 **in this turn**. A marker goes stale on any later commit or
+  edit, so a close claimed earlier does not carry. A stash or a dirty second repository does not change the marker,
+  which is why `run` checks them every time.
+- Never commit, reset or delete another session's work. A dirty sibling worktree is preserved, not cleaned.
+- Deploy only when the engine says a deployable changed. `deploy: na` is the answer for a docs-only or tooling-only
+  session; do not run `/rdc:deploy` for the sake of it.
+- A blocker needs an independent look before you report it as one: probe it, and have another agent confirm.
+- A question for Dave goes through `AskUserQuestion`, never as a line in the report.
+
+## Capture lessons (exit step)
+
+If this close taught something non-obvious, follow `.rdc/guides/lessons-learned-spec.md` § Capture procedure: a fix
+already encoded goes in its commit message; a fix not encoded is `node "$LIFEAI_ENV/bin/rdc-lesson.mjs" submit`.
+Note "N lessons captured" in the verdict. A close that taught nothing records nothing.