feat(ce-dispatch): single-unit sync MVP rewrite

[devin-ai-integration]

Summary

Rewrites ce-dispatch from multi-unit fan-out to a single-unit sync MVP. Each invocation creates one GitHub issue for one implementation unit; orchestrator and the in-workspace agent coordinate via issue comments and the resulting PR; the user pings each side manually.

The skill stays at plugins/compound-engineering/skills/ce-dispatch/ (no rename) and keeps the beta-skills triplet (name: ce-dispatch, [BETA] description prefix, disable-model-invocation: true) — the directory was never -beta; the markers ride in the frontmatter, per the beta-skills framework.

Behavior changes

Phase 0 — drop dispatch_mode and dispatch_auto_review; keep dispatch_branch_prefix, dispatch_base_branch, dispatch_labels. New Phase 0.3 asks the user for the worktree absolute path (the user creates the Conductor workspace before invoking the skill, which closes the chicken-and-egg of needing the worktree before the issue exists). Worktree dirname becomes the agent name used in comments.

Phase 1 — pick exactly one implementation unit. Drop dependency-graph construction, parallel-safety overlap checks, unit coalescing.

Phase 2 — render a single self-contained dispatch prompt. Populate three new template sections (<orientation>, <agent-identity>, <comment-protocol>) alongside existing ones. Drop multi-unit coalescing branches.

Phase 3 — create exactly one issue via gh.

Phase 4 — collapse the six-option monitor loop to a four-option respond menu (Reply to agent comment / Review the PR / Mark unit complete / Done for now). Drop dependency-aware merge gating, dependency-graph rendering, auto re-dispatch, auto-review. Compose existing CE skills (ce-code-review, ce-resolve-pr-feedback) for the PR-feedback round-trip.

Prompt template changes

• Add <orientation> for progressive context exposure (path list, not inlined content) — README, AGENTS.md, plan, architecture doc, unit's pattern files, unit's Files: paths.
• Add <agent-identity> carrying agent-name and worktree-path.
• Add <comment-protocol> with the [<agent-name> -> orchestrator] timestamped format and explicit STOP-after-asking directive — closes the airgap concern (in-workspace agent must not make architectural decisions without orchestrator input).
• Rework <ce-plugin> as an explicit nine-step compound-engineering loop: read orientation -> /ce-work -> implement and verify -> /ce-code-review -> /ce-compound (optional) -> /ce-commit-push-pr -> comment with PR URL -> stop and wait -> on ping run /ce-resolve-pr-feedback. Loop.
• Drop dependencies: from metadata footer; rename unit_ids: -> unit_id:; add agent_name: and worktree_path:.

Configs / ce-plan routing / conductor-notes

• Mirror dropped/retained keys across .compound-engineering/config.local.example.yaml and ce-setup/references/config-template.yaml.
• Update the inline routing description for the "Dispatch to external agents" menu option in ce-plan SKILL.md and plan-handoff.md to reflect single-unit shape; drop fan-out / parallel-execution language. Menu label and position (option 4 of 5) are unchanged; platform skill-invocation primitive guidance is preserved.
• Light revisions to conductor-notes.md metadata-comment description so it lists the single-unit-MVP keys and states dependency-graph state is intentionally absent.

Regression guards preserved

Verbatim from EveryInc#762:

• gh pr list invocations in Phase 4 require --state all (otherwise merged PRs are invisible to the respond loop).
• git symbolic-ref invocations for origin/HEAD require --short (otherwise the full ref path leaks into dispatch metadata).

Contract tests

Rewrote tests/skills/ce-dispatch-contract.test.ts end-to-end: dropped tests for removed behavior, updated tests for changed behavior, added tests for new behavior (orientation / agent-identity / comment-protocol sections; nine-step ce-plugin sequence; STOP-and-wait directive; comment-protocol prefix shape; ce-resolve-pr-feedback routing in Phase 4; ce-plan routing wording without fan-out language). 63/63 pass.

Validation

• Targeted contract test: 63/63 pass.
• bun test: 1307/1308 pass. The single failing test (resolve-base.sh > resolves against origin/HEAD in a detached shallow checkout) is a pre-existing environmental issue unrelated to this change — already failing on baseline main (verified Loop 1 baseline before any edits).
• bun run release:validate: clean.

Review & Testing Checklist for Human

This is yellow risk: structural rewrite of an in-flight beta skill. The skill is disable-model-invocation: true, so accidental invocation by the planner is gated, but human-driven /ce-dispatch <plan> flows go through this code.

• Skim the new <ref_file file="/home/ubuntu/repos/compound-engineering-plugin/plugins/compound-engineering/skills/ce-dispatch/SKILL.md" /> top-to-bottom and confirm the four phases match your mental model of the sync MVP. Phase 0.3 (worktree confirmation), Phase 1 (single unit), Phase 4 (four-option respond menu) are the meaningful behavioral changes.
• Skim the new <ref_file file="/home/ubuntu/repos/compound-engineering-plugin/plugins/compound-engineering/skills/ce-dispatch/references/dispatch-prompt-template.md" /> and confirm <orientation>, <agent-identity>, <comment-protocol> carry the shape you want, and that the nine-step <ce-plugin> block prescribes the compound-engineering loop the way you want the in-workspace agent to walk it.
• Spot-check the contract test rewrite: <ref_file file="/home/ubuntu/repos/compound-engineering-plugin/tests/skills/ce-dispatch-contract.test.ts" />. Confirm the not.toMatch guards for removed behavior (Parallel Safety Check, six options, dependencies key) and the new expect(...).toContain assertions for added behavior reflect the right shape.
• Confirm the ce-plan routing wording change reads correctly: <ref_snippet file="/home/ubuntu/repos/compound-engineering-plugin/plugins/compound-engineering/skills/ce-plan/SKILL.md" lines="907-907" /> and <ref_snippet file="/home/ubuntu/repos/compound-engineering-plugin/plugins/compound-engineering/skills/ce-plan/references/plan-handoff.md" lines="67-67" />.

End-to-end test plan (manual, after merge)

1. Create a Conductor workspace from mvp/ce-dispatch-beta-rewrite for a small repo with a plan that has a single implementation unit. Note the worktree path.
2. From the orchestrator session: /ce-dispatch <plan-path>. Confirm Phase 0.3 asks for the worktree path; supply it.
3. Confirm Phase 1 surfaces the unit list and lets you pick one.
4. Confirm Phase 3 creates one GitHub issue with the rendered prompt as the body and the four sections (<orientation>, <agent-identity>, <comment-protocol>, plus the legacy XML).
5. From the Conductor workspace: tell the agent to read the issue. Walk the nine-step loop end-to-end (/ce-work -> /ce-code-review -> /ce-compound -> /ce-commit-push-pr).
6. Back in the orchestrator: re-invoke /ce-dispatch to enter the respond loop. Confirm "Review the PR" pulls the new PR; confirm "Reply to agent comment" surfaces the latest agent comment if you posted one; confirm "Mark unit complete" closes the issue once the PR is merged.

Notes

• This PR does not delete or rename any directories. The skill remains at ce-dispatch/. No additions to STALE_SKILL_DIRS, STALE_AGENT_NAMES, STALE_PROMPT_FILES, or EXTRA_LEGACY_ARTIFACTS_BY_PLUGIN are needed.
• The four architectural decisions you approved are all reflected: (1) single skill with Phase 4 respond menu; (2) dependencies: dropped from metadata footer; (3) branch mvp/ce-dispatch-beta-rewrite off main; (4) everything marked for removal in the plan was removed.
• Drafted intentionally — pending your review before flipping to ready.
• The pre-existing resolve-base.sh failure is environment-related (detached shallow checkout). I baselined it before any edits; not introduced by this PR.

Link to Devin session: https://app.devin.ai/sessions/56b17768c2ef4657ba155e2435bf1548
Requested by: @shubness

---

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring

[devin-ai-integration]

Devin Review found 1 potential issue.

View 4 additional findings in Devin Review.

[devin-ai-integration]

🟡 Menu label still describes multi-unit fan-out but routing text was updated to single-unit

The PR updated the routing text for the "Dispatch to external agents" option to single-unit language (line 907: "hands one implementation unit at a time off...via a single GitHub issue"), but the corresponding menu label at plugins/compound-engineering/skills/ce-plan/SKILL.md:899 still says "Create GitHub issues for each implementation unit" (plural, implying fan-out of all units). The same stale label exists in plugins/compound-engineering/skills/ce-plan/references/plan-handoff.md:44. This inconsistency means the ce-plan agent presents users with a menu option whose description promises multi-unit dispatch while the actual behavior (routed through ce-dispatch) only creates one issue for one unit. The new test at tests/skills/ce-dispatch-contract.test.ts:528-546 only checks the routing bullet for fan-out language, not the menu label, so this drift isn't caught.

Prompt for agents

The menu option label at line 899 of ce-plan/SKILL.md says "Create GitHub issues for each implementation unit, ready for pickup by Conductor workspaces or other issue-driven agent workflows" which implies multi-unit fan-out. The same stale label exists at line 44 of plugins/compound-engineering/skills/ce-plan/references/plan-handoff.md. Both need updating to reflect single-unit MVP behavior (e.g., "Dispatch one implementation unit to a Conductor workspace via a single GitHub issue" or similar). The corresponding test at tests/skills/ce-dispatch-contract.test.ts:528-546 should also be extended to check the menu label text (match pattern like /4\.\s+\*\*Dispatch to external agents\*\*[^\n]+/) for plural/fan-out wording.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

Battle-tested this rewrite via Anthropic's skill-creator eval protocol before flipping out of draft. Evidence on the stacked PR #5 (mvp/ce-dispatch-evals → this branch).

Headline (Opus 4.7, ~$1.50 in API costs):

Configuration	Pass rate (24 expectations across 4 evals)
with_skill	100% (24/24) after iteration-2 assertion fix; 95% (23/24) raw iteration-1
without_skill baseline	51% (12/24) iter-1 / 56% (13/24) iter-2
delta	+44 to +49 percentage points

4 evals run, each with quantitative expectations (5–9 each):

• 1-happy-path-single-unit-dispatch — Phase 0–3 issue rendering (orientation, agent-identity, comment-protocol, ce-plugin block, metadata footer, single gh issue create): 9/9 vs. 4/9 baseline
• 2-phase-4-respond-review-pr — four-option menu + PR-review routing through /ce-code-review: 4/5 → 5/5 iter-2 vs. 3/5 → 4/5 baseline
• 3-phase-4-respond-reply-to-agent-comment — [orchestrator -> agent] <ISO 8601> reply prefix: 5/5 vs. 3/5 baseline
• 4-phase-4-respond-mark-unit-complete — gh issue close + worktree archival prompt + PR-merged verification gate: 5/5 vs. 2/5 baseline

Findings: No skill bugs. The skill correctly renders all six required prompt-template sections, scopes content from the right unit on a multi-unit plan, surfaces exactly four Phase 4 options (no six-option monitor menu, no dependency graph, no auto-review), routes review through /ce-code-review, uses --state all consistently, and verifies PR is MERGED before closing the issue. The one initial assertion failure was an over-prescriptive eval (caught by the grader's own eval_feedback); refining the assertion in iteration 2 gave a clean 100% pass.

See:

• PR test(ce-dispatch): add skill-creator-style eval pack + battle-test results #5 description for full breakdown and reproducibility commands
• evals/ce-dispatch-workspace/REPORT.md for the human-readable battle-test report
• evals/ce-dispatch-workspace/iteration-1/benchmark.md for the iteration-1 aggregate
• Spot-check any iteration-1/<eval-name>/with_skill/grading.json for per-expectation evidence

Caveat: the runner is single-shot Chat Completions, so the agent describes commands rather than executing them. Real gh issue create/gh issue close/multi-turn comment-protocol roundtrips remain the manual end-to-end test plan. The contract test in this PR (63/63 passing) covers the loader/template-render path.

Ready for review whenever you are.