Use skill descriptions as Claude adapter preview text

.claude/commands/qa.md

@@ -1,4 +1,4 @@
-You are executing the `/qa` skill. Your argument is: $ARGUMENTS
+View the QA queue or review a specific issue
 
 ---
 
@@ -158,4 +158,4 @@ After applying labels, tell the user what was done:
 - Never post the comment without showing it to the user first.
 - Never apply labels without user confirmation (the confirmation in Step 3 is the single gate for both posting the comment and applying labels).
 - Never merge, deploy, or take any action beyond labeling and commenting.
-<!-- generated by CodeCanon/sync.sh | skill: qa | adapter: claude | hash: 447216d6 | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->
+<!-- generated by CodeCanon/sync.sh | skill: qa | adapter: claude | hash: 66d58a4f | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->

.claude/commands/release.md

@@ -1,4 +1,4 @@
-You are executing the `/release` skill. Your argument is: $ARGUMENTS
+Create a GitHub Release and promote the pre-production branch to production
 
 ---
 
@@ -357,4 +357,4 @@ After the command runs, note the release URL from the output.
 Tell the user:
 
 > "Released vX.Y.Z. Issues #N, #M closed automatically. GitHub Release vX.Y.Z created at `<url>`. Run `make deploy-prod` to ship to production."
-<!-- generated by CodeCanon/sync.sh | skill: release | adapter: claude | hash: b470d522 | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->
+<!-- generated by CodeCanon/sync.sh | skill: release | adapter: claude | hash: 26860d5c | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->

.claude/commands/review.md

@@ -1,4 +1,4 @@
-You are executing the `/review` skill. Your argument is: $ARGUMENTS
+Run a standalone code review on a pull request
 
 ---
 
@@ -54,4 +54,4 @@ After the review agent completes, relay its verdict to the user:
 - This command does not commit, push, or merge anything. It only reviews.
 - It may be called manually at any time, not just from `/ship`.
 - The review comment is posted to GitHub for the audit trail.
-<!-- generated by CodeCanon/sync.sh | skill: review | adapter: claude | hash: 06aee22d | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->
+<!-- generated by CodeCanon/sync.sh | skill: review | adapter: claude | hash: 952b7baa | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->

.claude/commands/setup.md

@@ -1,4 +1,4 @@
-You are executing the `/setup` skill. Your argument is: $ARGUMENTS
+Detect setup state, guide first-time configuration, populate labels, and walk through optional config values
 
 ---
 
@@ -469,4 +469,4 @@ Add a note: `/start` can be used to create well-formed GitHub issues without wri
 - Never fetch more than 100 labels in a single command. `gh label list --limit 100` is the ceiling.
 - Do not skip any human gate in Phase 2 or Phase 3 — each write requires confirmation.
 - If the user skips a config value, do not ask again. Move on.
-<!-- generated by CodeCanon/sync.sh | skill: setup | adapter: claude | hash: 5e26ca1c | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->
+<!-- generated by CodeCanon/sync.sh | skill: setup | adapter: claude | hash: 438c2794 | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->

.claude/commands/ship.md

@@ -1,4 +1,4 @@
-You are executing the `/ship` skill. Your argument is: $ARGUMENTS
+Type-check, commit, open PR, review, and merge to the integration branch
 
 ---
 
@@ -191,4 +191,4 @@ Report success based on mode:
 - When `ai` is `"off"`, skip the review agent entirely — merge immediately after checks pass.
 - In trunk mode, merges target `main`. Otherwise, `/ship` merges only to `dev` — never directly to `` or `main`.
 - If `make merge` fails for any reason, report it and stop — do not attempt workarounds.
-<!-- generated by CodeCanon/sync.sh | skill: ship | adapter: claude | hash: d98aef47 | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->
+<!-- generated by CodeCanon/sync.sh | skill: ship | adapter: claude | hash: ac83f87e | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->

.claude/commands/start.md

@@ -1,4 +1,4 @@
-You are executing the `/start` skill. Your argument is: $ARGUMENTS
+Start a new feature or bugfix
 
 ---
 
@@ -226,4 +226,4 @@ When done, say: **"The code is ready for review. Please run `make dev` and test
 - `gh issue create` must use `--title` and `--body` flags. Never open an interactive editor.
 - The issue is assigned to `@me` at creation. If you are creating a ticket on someone else's behalf, remove the assignee after creation with `gh issue edit <number> --remove-assignee @me`.
 - Apply resolved labels and milestone to every new issue. Label resolution order: per-invocation flag → pool selection from `` → omit (or create if `false` is `true`). Never apply a label not in `` unless `false` is `true`. Milestone resolution order: per-invocation flag → `` config → auto-detected from GitHub open milestones. Never prompt for a milestone more than once per invocation.
-<!-- generated by CodeCanon/sync.sh | skill: start | adapter: claude | hash: 0bb0f660 | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->
+<!-- generated by CodeCanon/sync.sh | skill: start | adapter: claude | hash: 857d7656 | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->

.claude/commands/status.md

@@ -1,4 +1,4 @@
-You are executing the `/status` skill. Your argument is: $ARGUMENTS
+Summarize in-progress and recently completed work from GitHub and git
 
 ---
 
@@ -175,4 +175,4 @@ Do not post, comment, write files, or take any action. Output only.
 - If `gh` is unauthenticated or any fetch fails, report the error and stop immediately.
 - Do not retry failed commands.
 - Strip the leading `@` from the subject when passing to `gh` flags that do not accept it.
-<!-- generated by CodeCanon/sync.sh | skill: status | adapter: claude | hash: 3e7a06c3 | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->
+<!-- generated by CodeCanon/sync.sh | skill: status | adapter: claude | hash: 79583183 | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->

.claude/commands/version.md

@@ -1,4 +1,4 @@
-You are executing the `/version` skill. Your argument is: $ARGUMENTS
+Bump the project version, tag, and push — run before deploying to preview
 
 ---
 
@@ -79,4 +79,4 @@ Report based on mode:
 - **Trunk mode** (`dev` empty): "Tagged vX.Y.Z. Run `/release` to create the GitHub Release."
 - **Two-branch mode** (`dev` set, `` empty): "Tagged vX.Y.Z. Run `make deploy-preview` to deploy to preview for testing. When testing is complete, run `/release`."
 - **Three-branch mode** (both set): "Tagged vX.Y.Z. Run `make deploy-preview` to deploy to staging. When testing is complete, run `/release`."
-<!-- generated by CodeCanon/sync.sh | skill: version | adapter: claude | hash: ab8011ef | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->
+<!-- generated by CodeCanon/sync.sh | skill: version | adapter: claude | hash: 2dcb66e2 | DO NOT EDIT — run CodeCanon/sync.sh to regenerate -->

ROADMAP.md

@@ -0,0 +1,21 @@
+# Roadmap
+
+Ideas and future work. Not prioritized — just captured so they don't get lost.
+
+## Swarm mode / multi-agent workflows
+
+The current skill set (start/ship/review/version/release) assumes **deep mode**: one issue, one branch, one coherent PR. This works well for disciplined, sequential work.
+
+In practice, developers often run multiple agents simultaneously on unrelated tasks — "swarm mode." This resists the ticket-per-change structure because the agents share a working directory and branch. Worktrees could isolate agents, but only Claude Code supports them natively; Cursor, Codex, and Gemini don't.
+
+Possible additions:
+
+- **`/checkpoint` skill** — commit and push WIP without the full ship ceremony. Gives save points during swarm mode without pretending each save is a reviewable unit.
+- **Worktree launcher** — a `make agent name=<task>` target that creates a worktree, launches an agent with `/start`, and registers cleanup after `/ship`. Orchestration layer outside the skills themselves.
+- **Conflict detection** — warn when multiple agents are modifying overlapping files on the same branch.
+
+Decision for now: better discipline (one agent per issue, sequential) is the right path. Revisit when the pain of sequential work outweighs the cost of coordination tooling.
+
+## Three-branch Makefile targets
+
+GitHub issue #6. `Makefile.agents.mk` lacks `STAGING_BRANCH` support — no `promote` target, no staging guard rails. The skills handle three-branch mode but the Makefile doesn't.

adapters/claude/header.md

@@ -1,4 +1,4 @@
-You are executing the `/{skill}` skill. Your argument is: $ARGUMENTS
+{description}
 
 ---
 

skills/release.md

@@ -1,7 +1,7 @@
 ---
 skill: release
 type: skill
-description: "Create a GitHub Release; in two-branch and three-branch mode, also promotes the pre-production branch to `{{BRANCH_PROD}}`"
+description: Create a GitHub Release and promote the pre-production branch to production
 args: none
 ---