Execute Code Plan

Purpose

Drive an approved code plan to completion by orchestrating the agent pipeline from the main conversation. This skill conducts; it does not implement. It dispatches task-implementer per plan task in dependency order, gates each checkpoint with completion-verifier, routes failures to root-cause-investigator, runs code-reviewer and security-reviewer on the finished change, and live-validates runnable surfaces through the manual-tester agent once the change is green. It stays thin because the agents own the methodology and already speak compatible output contracts; the skill's whole job is sequencing, gating, and honest failure handling. Do not begin until an approved plan exists, and do not let the orchestrator drift into doing the work itself.

This must run in the main conversation

Subagents cannot spawn subagents. The orchestration that dispatches task-implementer and completion-verifier can only run where the Agent tool reaches them: the main thread. This skill therefore cannot be wrapped as an agent, and the main loop must not hand the orchestration to one. If you cannot spawn the pipeline agents, you are inside a subagent and must hand back to the main conversation.

Workflow

Copy this checklist and track progress:

Execute Plan:
- [ ] 1. Load, validate, and (if large or complex) adversarially review the plan
- [ ] 2. Establish a green baseline
- [ ] 3. Execute tasks in dependency order (parallel where marked)
- [ ] 4. Gate every checkpoint with completion-verifier
- [ ] 5. Final review pass (code + security)
- [ ] 6. Validate live behavior on runnable surfaces (manual-tester)
- [ ] 7. Report; the user owns commit and merge

1. Load and validate the plan

Read the approved plan (from create-code-plan or the implementation-planner agent). Extract the task table: each task's description, files, verify command, dependencies, and any parallel marker. If no approved plan exists, stop and route to create-code-plan; executing an unwritten plan is improvising, not orchestrating. If the plan is stale against the current code (files moved, interfaces changed), surface that before starting rather than dispatching tasks against a map that no longer matches the territory. When the plan was written from a spec that is still available, reconcile the plan against it too; a plan that drifted from the spec gets faithfully executed into the wrong thing, and execution is the most expensive place to discover that divergence.

Pre-flight review for large or complex plans. Before establishing the baseline, dispatch a plan-reviewer for any plan that is large or complex: more than roughly six tasks, spanning multiple subsystems, touching production data or infrastructure, carrying a High-impact risk, or introducing a new architectural pattern or data model. Pass it the plan and the source spec or request. It returns severity-ranked findings and a verdict (approve, revise-then-approve, rework, or reject) against the goal, and never edits the plan. Treat a blocker finding or a rework/reject verdict as a stop: surface it and route the fix back through create-code-plan or to the user before any task runs, because a flaw caught in the plan is one cheap edit while the same flaw caught mid-execution is unwound work across every task built on it. Skip the pre-flight for small, single-subsystem plans; their create-code-plan step 6 self-review already covers them, and a fresh reviewer on a three-task plan is latency for no gain.

2. Establish a green baseline

Run the build and test suite once before touching anything, and record the result. A green baseline is what lets you attribute any later failure to the work. If the baseline is red, stop and route the existing failure to fix first, because you cannot tell your breakage from the pre-existing one on a red suite.

3. Execute tasks in dependency order

Walk the task graph. For each task, dispatch a task-implementer subagent with the full task text (description, files, verify command) plus the plan context it needs.

• Parallelize only what the plan marks independent. Spawn independent tasks in one message so they run concurrently. Tasks that share a contract wait for the task that defines it.
• Isolate parallel tasks that touch the same files with isolation: worktree, or they corrupt each other's edits.
• Collect each report. A task-implementer returns Complete with quoted evidence, or Blocked with what is missing. A Blocked report halts that branch; do not invent the missing piece in the main thread.

4. Gate every checkpoint with completion-verifier

After each task, or each phase checkpoint the plan defines, dispatch completion-verifier with the task's requirements and the implementer's report. It returns Verified or Rejected with per-requirement evidence it gathered itself.

• Never advance past a Rejected. A rejection means a requirement is unmet or the green is decorative; feed the verifier's findings to a fresh task-implementer and re-run. Bound this to two attempts, then stop and surface to the user. A verify-implement loop that will not converge is a sign the task or the plan is wrong, not something to grind on.
• When a task fails for an unclear reason rather than a named unmet requirement, dispatch root-cause-investigator to diagnose before re-implementing, so the retry targets the cause instead of guessing.

5. Final review pass

When every task is Complete and Verified, dispatch code-reviewer on the cumulative diff, and security-reviewer too when the change touched any sensitive surface (auth, input handling, secrets, external integrations). Route each confirmed finding back through a task-implementer the same way, then re-verify.

6. Validate live (runnable surfaces)

Green and reviewed is not the same as working: the suite and the reviewers exercise code paths and the diff, and neither renders the UI, hits the live endpoint, or runs the built artifact. When the change has a runnable surface, validate it after the pipeline is green. Conduct here too: dispatch the manual-tester agent with the change summary and the running app URL; it drives the surface and returns a pass/fail-with-evidence report. Fall back to running validate-web or validate-api directly in the main thread only when the manual-tester agent is unavailable.

Validation needs the app actually running, so start it (or have the user start it) first. If no target can be brought up, report that validation was not possible rather than skipping it silently. Fold the evidence (pass/fail per behavior, screenshots, the saved Bruno collection path) into the report. A change with no runnable surface (a pure library, an internal refactor with no behavioral change) skips this step; say so.

7. Report

Summarize for the user: tasks completed with their evidence, what each checkpoint verified, the review findings and their resolution, and the final suite state. Surface any task left Blocked and why. Do not commit or merge unless the plan or the user explicitly says to; the user owns git.

Orchestration rules

• Conduct, do not perform. The agents do the work; the main thread sequences and gates. If you are editing files or writing tests in the main conversation, you have stopped orchestrating and lost the clean per-task context the pipeline exists to provide.
• One task, one fresh implementer. Each task-implementer gets isolated context by design; do not try to carry one agent across tasks.
• A green checkpoint is the only gate forward, meaning the verifier's verdict, never the implementer's own word.
• Stop and surface to the user on: no plan, a red baseline, a blocker the pipeline cannot resolve, a non-converging verify loop, or any scope decision the plan did not settle. The orchestrator escalates honestly; it does not paper over.

Gotchas

• This skill cannot be an agent. The whole pipeline rests on spawning subagents, which only the main thread can do; wrapping the orchestrator as an agent breaks it silently at the first dispatch.
• A red baseline poisons attribution. Without step 2, the first test failure is unattributable and you burn a diagnosis cycle discovering it predated your work.
• Parallel tasks on shared files corrupt silently. Two implementers editing one file without worktree isolation overwrite each other; the plan's parallel marker is a claim of independence you verify, not assume.
• The retry loop has a floor. Re-dispatching forever on a Rejected is the orchestrator refusing to admit the task or plan is wrong. Two rounds, then escalate.
• Review findings do not reopen the plan. A code-reviewer finding becomes a new task routed through the pipeline, not a license to re-architect mid-execution; if a finding reveals the plan itself was wrong, stop and say so instead of improvising a rewrite.
• Green is not validated. A verified, reviewed, green change can still render a broken UI or return 500 on a real request, because the suite proves the logic, not the running system. Live validation (the manual-tester agent wrapping validate-web / validate-api, or driving the built artifact directly) is the post-pipeline step that closes that gap; skip it only when there is no runnable surface, and say so.

Example

A four-task plan (T1 schema, T2 endpoint depends on T1, T3 client depends on T1, T4 docs independent):

1. Load plan; baseline `npm test` green (42 passing).
2. Dispatch T1 (schema) and T4 (docs) in parallel; both independent.
   T1 -> Complete (RED/GREEN quoted). T4 -> Complete.
3. Verify T1 and T4 -> both Verified.
4. T1 done, so dispatch T2 (endpoint) and T3 (client) in parallel,
   isolation: worktree (both import the new schema file).
   T2 -> Complete. T3 -> Blocked: "client SDK version in plan does not exist".
5. T3 blocked: stop that branch, surface the report to the user; do not guess a version.
6. Verify T2 -> Verified. Dispatch code-reviewer + security-reviewer on the T1+T2+T4 diff.
   One finding (missing authz check on the endpoint) -> new task -> implement -> re-verify.
7. Validate live: T2 is a new endpoint, so dispatch manual-tester against the running service
   (it runs validate-api); the saved request returns 201 and the asserted body shape holds.
   Collection saved to bruno/.
8. Report: T1/T2/T4 complete, verified, and live-validated, review clean after the fix, T3
   blocked pending the SDK question. Suite green (51 passing). No commit; the user owns git.