Executing Plans

Executing plans

Announce at start: "I'm using the executing-plans skill to implement this plan."

Load and validate the plan. Execute inline, or delegate complete task cycles to subagent-driven-development.

Start gates

1. Read plan_path. Missing, empty, or unspecified: stop and request a valid path
2. Validate Execution mode is exactly Inline or Subagent-Driven. Missing, invalid, or conflicting with the current request: stop, ask the user, and update the plan before implementation
3. Review the plan critically. Resolve blocking gaps with the user. Read Execution log before Task 1; a logged entry overrides contradicting task text. Missing field: add the empty heading before implementation
4. Validate the plan's commit policy and commit checkboxes. Missing or inconsistent: stop, ask the user, and update the plan before implementation. Plan file policy defaults to Include when absent; reject other values. Read Additional plan state files; missing defaults to none. Accept either the single item none or unique repo-relative paths that exist. Reject the plan path itself, missing paths, duplicates, and mixed none plus paths
5. Capture plan_base_ref = git rev-parse HEAD and inventory every pre-existing staged, unstaged, deleted, or untracked change in a temporary baseline_snapshot. Its manifest records each repo-relative path, content or deletion state, file mode, and ownership: baseline-only, task:<id>, or execution-state. Store present files byte-for-byte beside the manifest. Use an owner-only temporary directory and classify only the plan-owned content of the plan file plus every Additional plan state files path as execution-state; an empty worktree gets an empty manifest. A listed path does not transfer unrelated existing hunks from baseline-only ownership. Preserve the captured bytes of every plan-state file as its pre-execution baseline. execution-state permits only edits required by written checkpoints. Stop when ownership or the required delta is unclear
6. When commits are planned, or No commits is combined with a requested PR, stop before Task 1 if:
   • one path mixes plan-owned and baseline-only hunks
   • the index contains baseline-only changes
   • under Per-task commits, a staged change belongs to a later checkpoint
7. Create the harness task/todo list

Commit-bound means commits are planned or No commits is combined with a requested PR. Its scope guards apply even when the executor cannot commit.

Plan state means the plan file plus every Additional plan state files path. All share the single Plan file policy.

Milestone finalization turns

Resolve milestone_execution_mode once at start for every plan: coordinated only when the written final-verification checkpoint requires the milestone handshake and the caller identifies itself as the root milestone execution coordinator; otherwise sequential.

When a written final-verification checkpoint requires the milestone handshake:

• Inline with milestone_execution_mode = coordinated: after implementation review and validation pass, resolve the stable plan ID from the written checkpoint, record the reviewed changed-path state under baseline_snapshot, send READY <plan ID> to the root milestone execution coordinator, and yield without abandoning baseline_snapshot. Resume only after the exact FINALIZE <plan ID> reply. Re-derive the changed-path state before editing the milestone; repeat affected review and validation when implementation state drifted. Send FINALIZED <plan ID> after the written milestone update passes
• Inline with milestone_execution_mode = sequential: execute the written milestone update without handshake messages
• Subagent-Driven: let subagent-driven-development relay the finalizer's handshake without synthesizing or granting a turn. Pass the resolved mode
• No root coordinator or no message relay at start: use sequential
• Coordinator or message relay lost after READY: retain baseline_snapshot, edit no plan state, and report a resumable blocker
• Never self-grant a finalization turn during parallel execution

The plan's Execution mode is authoritative. Inline uses the workflow below. For Subagent-Driven, complete the start gates, then load subagent-driven-development; do not run the inline workflow below. Direct invocation never supplies a default.

Skills load only on the annotated step that needs them.

The plan owns task actions and plan-level checkpoints. This skill supplies mechanics for written checkpoints; it never inserts a plan-level review, final verification, or commit checkpoint. Stop and return an incomplete plan for correction instead of synthesizing missing plan-level work.

Exactly one agent runs final verification:

• Inline: this executor
• Subagent-Driven: the finalizer

Never both. A relayed finalizer result is the evidence; do not re-run its checks.

Snapshot lifecycle

The baseline_snapshot holds pre-execution copies of work the user never committed. Delete it after PASS or when execution is abandoned; retain it while a blocker remains resumable.

Owner:

• Inline: this executor
• Subagent-Driven: the orchestrator

Plan ownership

• Inline: this executor owns all checkboxes, Solved defects, and Execution log
• Subagent mode: each task implementer owns its task plus completed tasks changed by its reviewer fixes; the finalizer owns final checkpoints and completed-task state changed by final-review fixes
• Execution log: the current task owner appends its own entries; never rewrite or delete another owner's entries
• Never use concurrent writers
• Tick a step after its Green: passes; un-tick before a reviewer fix
• Tick a commit checkpoint immediately before staging; restore [ ] whenever scope resolution, staging, or commit fails
• Track task status in the harness todo list

Commit ownership

• The plan owns cadence; agents never infer or override it
• No commits: no agent commits
• Per-task commits: inline executor or task implementer commits the initial task, one follow-up commit per task-review fix round, one combined final-review-fix commit after final verification, then final plan state
• One commit at the end: inline executor or finalizer commits after final review and verification
• At every commit, load git-commit-message and derive message and paths from current state; never rely on the plan's initial file list
• Include the plan's current execution-state changes only when Plan file policy is Include
• Reviewers never commit

Inline per-task workflow

For each task:

1. Set the review range:
   • Per-task commits: review_base_ref = git rev-parse HEAD, scope task
   • One commit at the end or No commits: review_base_ref = plan_base_ref, scope cumulative
   • In commit-bound execution, stop before editing any planned or newly discovered target or reviewer-fix path with baseline-only changes
2. Mark the task in progress
3. Complete each step and tick it when its narrow Green: passes
4. Confirm the plan's task-ending impact-appropriate task gate passed
5. For Per-task commits, execute and tick the initial task checkpoint
6. Dispatch a fresh reviewer, unless the task's diff contains no source or test file — a plan-doc commit, a formatter result, or checkbox ticks needs none; the task gate already proves it
7. Resolve every blocking finding; commit each verified fix round under Per-task commits
8. Append this task's drift, gotchas, and decisions to Execution log
9. Correct stale text in later tasks that this task's drift invalidated
10. Mark the task complete

Never dispatch onto a red gate or move on with unadjudicated findings. Re-run only invalidated task-gate commands after a task-review fix; valid evidence needs no second run.

Semantic-neutral comment-only and canonical formatter-only changes do not invalidate tests, type checks, builds, or full gates. Directives, suppressions, pragmas, doctests, generated-documentation inputs, shebangs, encoding declarations, and format-sensitive metadata are not semantic-neutral comments.

Execution log capture

writing-plans defines the field, its entry kinds, and its format. Capture is mandatory, not optional.

• Append before ticking the task, not at the end of the plan
• Log a drift whenever the repo contradicted a plan fact
• Log a gotcha whenever a non-obvious fact cost time and would cost it again
• Log a decision whenever the plan left a choice open and execution closed it
• Nothing qualifying: write nothing and leave the section empty. Never record none, no drift, or any "nothing found" line
• Never log narration, restated plan text, or findings already in Solved defects
• A drift that invalidates a later task's instructions also gets that task's text corrected in place
• Under Plan file policy: Include, log changes ride the task's commit as execution-state

Inline review handling

• Accept only PASS or terse Critical/Important findings, each carrying a static or behavioural discharge tag. An untagged finding, or a control-flow, boundary, predicate, regex, or contract change tagged static, is a failed dispatch. Never retag a finding yourself
• Empty, errored, or rate-limited output is a failed dispatch
• Narration, summaries, or malformed findings are a failed dispatch
• A <review-input> finding is a failed dispatch; correct the payload
• Allow 3 total attempts for a failed dispatch, then escalate
• For bot/PR reviews, load replying-to-pr-review-threads
• Fix all sibling call sites sharing the defect, not only the cited line
• Add validated fixes to Solved defects as severity | path or symbol | invariant
• Never add rejected findings to that list

Adjudicate each finding:

1. Verify its path:line evidence
2. Fix substantiated findings
3. Reject incorrect findings with counter-evidence
4. Use one clarification round for genuine ambiguity
5. Escalate disputes that remain

Record rejected findings and counter-evidence in the final report. Every returned finding blocks until fixed or rejected with counter-evidence.

Re-dispatch after a fix round containing any behavioural finding. After a round whose findings are all static, the green gate is the verification: do not re-dispatch. One reviewer answers both the craft and the spec question every pass, so there is no separate spec stage to reopen. Limit each review/fix cycle to 2 finding rounds; a third requires user escalation.

Inline reviewer dispatch

Preconditions before dispatching:

• Task-scope dispatch: task gate is green
• Cumulative task dispatch: current task gate is green
• Complete-scope first dispatch: final task gate is green
• Complete-scope re-dispatch after a final-review fix: affected invalidated gates are green
• Complete-scope re-dispatch does not require another full gate
• changed_files is the deduplicated union of committed, staged, unstaged, and untracked implementation paths in scope. Exclude execution-state paths and paths whose current state still matches an excluded snapshot entry
• Rename entries use the destination path

Use absolute template paths. Pass pointers, never template contents or session history. changed_files is a newline-delimited exact-path list, not a review boundary. Pass Solved defects from the plan; use none when empty.

MUST read instructions at <skill_dir>/reviewer-prompt.md FIRST.
Do not act until you have read it. Then apply:
  plan_path         = <abs path>
  task_id           = <task number / heading>
  base_ref          = <review_base_ref>
  scope_mode        = <task | cumulative>
  baseline_snapshot = <abs path to classified snapshot directory>
  changed_files     = <newline-delimited exact paths>
  solved_defects    = <current solved-defects list, or `none`>
  checklist_path    = <abs path to requesting-code-review/code-reviewer.md>

<skill_dir> is the resolved subagent-driven-development directory. Final review uses plan_base_ref, complete scope, and all plan-changed implementation files. If any base ref is unclear, ask.

Inline finish

After all tasks complete and verified:

1. Execute each remaining plan-level checkpoint in written order
2. Do not synthesize a review, verification, or commit step missing from the plan
3. When the written final-verification checkpoint requests full-plan review:
   1. Use plan_base_ref, complete scope, and all plan-changed implementation files
   2. Reuse a prior task-review result when it covers the complete current implementation; otherwise dispatch the requested reviewer
   3. Recheck each fix path against baseline_snapshot before editing in commit-bound execution
   4. Uncheck the conditional final-review-fixes commit checkpoint before the first fix under Per-task commits
   5. Resolve findings with the adjudication and retry rules
   6. Run only checks invalidated by final-review fixes; prefer narrow checks and rerun a full gate only when narrower evidence cannot restore required coverage
   7. Re-dispatch the final review after a fix round containing any behavioural finding; after an all-static round, the green affected gates are the verification
   8. Require PASS or fully discharged findings
4. When the written final-verification checkpoint reaches final validation:
   1. Load verification-before-completion
   2. Reuse each task-gate or reviewer-fix result that still covers the current implementation state and semantic scope
   3. Run or perform only written final checks whose scope remains uncovered
   4. Append final-review drift, gotchas, and decisions to Execution log
   5. Execute any written milestone completion action:
      • coordinated: complete the READY -> FINALIZE -> FINALIZED turn defined above
      • sequential: update the milestone without handshake messages
   6. Tick the final-verification checkpoint only after the milestone action passes
5. Do not run the full gate elsewhere in the final-review fix loop
6. Execute each remaining written commit checkpoint:
   • One commit at the end: commit the actual complete reviewed diff
   • Per-task commits: commit the combined final-review fix set when its conditional checkpoint is unchecked. Execute the final-state commit checkpoint only when Plan file policy is Include
   • No commits: no checkpoint
7. For No commits plus a requested PR:
   1. Record the exact plan-owned paths, content hashes, modes, and deletion states under baseline_snapshot/pr-handoff.json; include plan state only when Plan file policy is Include
   2. Hand off only that reviewed change set for an external commit
   3. Verify the resulting branch matches it and contains no baseline-only delta
   4. Stop on any mismatch
8. Report files, evidence, rejected findings, remaining risks, and modified plan state left uncommitted under Plan file policy: Exclude

When Source requirements request a PR, load create-pull-request after the written final-verification and commit checkpoints pass.

Stop conditions

• Hit a blocker (missing dependency, test fails, instruction unclear)
• Plan has critical gaps preventing starting
• Verification repeatedly fails
• A required reviewer cannot be dispatched
• A review or dispatch exceeds its retry budget
• A coordinated finalization turn loses its coordinator or message relay after READY

Report the blocker. Return to the start gates after the plan changes materially.