Comet Phase 3: Plan and Build (Build)

Prerequisites

• Design Doc has been created (Phase 2 complete)
• Active change exists

Steps

0. Entry State Verification (Entry Check)

Use the stable comet CLI described in comet/reference/scripts.md, then run entry verification. When resuming from any entry point, first run the recovery check in comet/reference/context-recovery.md:

comet state select <change-name>
comet state check <name> build

Proceed to Step 1 after verification passes. The script outputs specific failure reasons when verification fails.

If the select / check output is BLOCKED because bound_branch does not match the current branch, immediately pause under comet/reference/decision-point.md and let the user choose one option: switch back to the bound branch and rerun entry verification, or run comet state rebind <change-name> after the user explicitly confirms the current branch should take over this change, then rerun entry verification. Do not switch branches or rebind on your own.

Idempotency: All build phase operations can be safely re-executed. Read .comet.yaml phase to confirm build, read the plan header base-ref, then parse tasks.md checkboxes in document order and resume from the first unchecked task. Already-committed tasks must not be re-committed.

1. Create Plan (Subagent Offload)

Create the implementation plan through a subagent, avoiding planning skill occupying main session context. Plan files and execution feedback must use the configured Comet artifact language from comet state get <name> language.

Subagent instructions:

You are an implementation planning expert. Create an implementation plan based on the following inputs:

1. Immediately execute: Use the Skill tool to load the Superpowers writing-plans skill. Skipping this step is prohibited. After the skill loads, ARGUMENTS must include: Language: Use the configured Comet artifact language from comet state get <name> language
2. Read the Design Doc (technical design document under docs/superpowers/specs/)
3. Read openspec/changes/<name>/tasks.md (task boundaries)
4. Follow the skill's guidance to create the plan

Plan requirements:

• Save to docs/superpowers/plans/YYYY-MM-DD-<feature>.md
• Reference design document, break down into executable tasks
• Plan file header must contain associated metadata:

---
change: <openspec-change-name>
design-doc: docs/superpowers/specs/YYYY-MM-DD-<topic>-design.md
base-ref: <git rev-parse HEAD before implementation>
---

base-ref is used during verification to measure committed changes across the full implementation range. Record the current commit when creating the plan:

git rev-parse HEAD

Write the plan to file, then return the file path.

Execute subagent: Use the current platform's subagent dispatch mechanism to send the above task.

After the subagent completes:

• If a valid file path is returned and the file exists, record it as the plan
• If the subagent fails or returns an invalid path, fall back to loading the Superpowers writing-plans skill inline in the main session (degraded fallback)

2. Update Plan Status and Jointly Confirm Workflow Configuration

Record plan path:

comet state set <name> plan docs/superpowers/plans/YYYY-MM-DD-feature.md

No manual phase update needed — guard auto-transitions when exit conditions are met.

Check current platform capabilities before presenting the joint decision: verify whether using-git-worktrees is available, whether a real background subagent/Task/multi-agent dispatcher exists, and whether the repository can safely create a branch. Show only isolation and execution options that are currently executable. If a field has only one valid value, explain why and apply it without manufacturing another pause.

After recording the plan, provide exactly one joint decision point that collects whether to continue now, available workspace isolation, available execution method, TDD mode, and code review mode. The branch name must be confirmed in the same Step 2 joint decision when branch is selected. Do not ask continue/pause first and then create another configuration or naming blocker.

This is a user decision point. Follow comet/reference/decision-point.md once and show the plan summary, pause option, and every executable Step 3 setting together. Continuing requires all settings and any conditional branch name in the same response. Do not auto-select or write the pause into build_mode.

When the user chooses to continue and supplies complete configuration:

comet state set <name> build_pause null

When the user chooses to pause:

comet state set <name> build_pause plan-ready

After setting build_pause: plan-ready, stop the current invocation. Do not choose isolation or build_mode, and do not load an execution skill.

3. Apply the Confirmed Workflow Configuration

If resuming with build_pause: plan-ready and the plan file exists, do not rerun writing-plans. Reissue the same joint Step 2 decision and clear the pause only after the user supplies complete configuration:

comet state set <name> build_pause null

Then apply the workspace isolation, execution method, TDD mode, and code review mode below.

The plan is on the current branch. These settings are all part of the single Step 2 decision:

Workspace Isolation:

Recommendation rules:

• User explicitly wants to keep the current branch, or the current branch is already the target branch for this change → Recommend A
• Change involves ≤ 3 files and the current branch is clean → Recommend B
• Need parallel development, current branch has uncommitted work → Recommend C

Execution Method:

Execution method recommendation rules:

• Task count ≥ 3 → Recommend A
• Task count ≤ 2 and no cross-module dependencies → Recommend B
• From hotfix path → Recommend B

These tables are part of the Step 2 joint decision and do not create another pause. First remove options that capability preflight found unavailable. When multiple valid options remain, do not choose current, branch, or worktree, execution method, TDD mode, or review mode from recommendations. Recommendations explain a preference; they never replace user confirmation.

After user selection, update isolation, execution method, TDD mode, and code review mode fields:

comet state set <name> isolation <current|branch|worktree>

• If the user chooses executing-plans: run comet state set <name> subagent_dispatch null, then run comet state set <name> build_mode executing-plans
• If the user chooses subagent-driven-development: first confirm the current platform has real background subagent / Task / multi-agent dispatch capability; after confirming, run comet state set <name> subagent_dispatch confirmed, then run comet state set <name> build_mode subagent-driven-development
• If real background dispatch capability cannot be confirmed, do not show or write build_mode: subagent-driven-development. If recovered state already records that mode but capability is unavailable, return to the same Step 2 joint decision with only executable modes; do not create a separate "switch to executing-plans" pause

TDD Mode:

Run comet state set <name> tdd_mode <tdd|direct>

Code Review Mode:

Run comet state set <name> review_mode <off|standard|thorough>

isolation is a script-enforced hard constraint. Full workflow init may temporarily leave it as null, but only before this step. If it remains null, both the build → verify guard and comet state transition build-complete will fail. Full workflow allows current, branch, or worktree, but current must be written only after the user explicitly selects it in Step 2; never make it a silent default.

subagent_dispatch is a script-enforced hard constraint. build_mode: subagent-driven-development requires subagent_dispatch: confirmed before leaving the build phase, otherwise both comet guard build --apply and comet state transition build-complete will fail.

tdd_mode is a script-enforced hard constraint. Full workflow must have tdd_mode selected as tdd or direct before leaving the build phase, otherwise both comet guard build --apply and comet state transition build-complete will fail.

review_mode is a script-enforced hard constraint. Full workflow must have review_mode selected as off, standard, or thorough before leaving the build phase, otherwise both comet guard build --apply and comet state transition build-complete will fail. Legacy state files without this field follow a compat path, but should be backfilled on recovery.

build_mode defaults to direct only for hotfix/tweak presets. Full workflow must not default to direct. Use it only when the user explicitly asks to bypass the plan execution skills and you record an explicit override:

comet state set <name> direct_override true
comet state set <name> build_mode direct

Without direct_override: true, build_mode=direct in full workflow is blocked by both guard and state transition.

Execute isolation:

• current: Do not create a new branch or worktree; execute directly on the current Git branch. Run comet state set <name> isolation current immediately; the command writes the current branch to bound_branch. If HEAD is detached, stop and ask the user to check out a real branch first, because there is no auditable branch binding.

• branch: Use the branch name already confirmed in Step 2; do not pause again. If legacy recovery no longer has the branch name from that joint decision, re-enter the same Step 2 decision instead of creating a separate branch-naming decision.

  Branch naming convention:

  • Read the workflow field from .comet.yaml to determine the prefix
  • workflow: full → recommend feature/YYYYMMDD/<change-name>
  • workflow: hotfix → recommend hotfix/YYYYMMDD/<change-name>
  • workflow: tweak → recommend tweak/YYYYMMDD/<change-name>
  • Format the current runtime date as YYYYMMDD; do not depend on one shell's date command

  Example: if change name is fix-login-bug and today is 2026-06-09, recommend feature/20260609/fix-login-bug

  Immediately after Step 2 confirms the branch name, run git checkout -b <branch-name>, then run comet state set <name> isolation branch to write the new branch to bound_branch. Continue on the new branch.

• worktree: Must use the Skill tool to load the Superpowers using-git-worktrees skill to create isolated workspace. Do not bypass this skill with plain shell commands or native tools; if the skill is unavailable, stop the process and prompt to install or enable Superpowers skills.

After creating isolation, confirm plan file is accessible (naturally accessible with branch method; for worktree method, confirm plan has been committed). If the plan file has not been committed under worktree mode, commit it first before creating the worktree:

git add docs/superpowers/plans/YYYY-MM-DD-feature.md
git commit -m "chore: add implementation plan"

After entering the final execution branch or worktree, bind the current change again inside that actual workspace. Branch mode is bound after checkout with isolation branch; worktree mode must run comet state set <name> isolation worktree inside the new workspace to write that worktree's current branch to bound_branch. A new worktree does not inherit the original workspace-local selection file, so select the current change too:

comet state select <change-name>

Do not begin source writes until this binding succeeds.

Execute plan: Must handle execution according to the actual runtime of build_mode.

• build_mode: executing-plans: Immediately execute: Use the Skill tool to load the Superpowers executing-plans skill. Skipping this step is prohibited. If the skill is unavailable, stop the process and prompt to install or enable the corresponding skill; do not substitute with normal conversation. After the skill loads, ARGUMENTS must include the same Language constraint as Step 1: Language: Use the configured Comet artifact language from comet state get <name> language. Execute according to plan.
• build_mode: subagent-driven-development: The main session only coordinates and must not write implementation code directly. Immediately execute: Use the Skill tool to load the Superpowers subagent-driven-development skill. After the skill loads, read comet/reference/subagent-dispatch.md for Comet-specific extensions (real background dispatch, task isolation, checkoff verification, TDD constraints, continuous execution, context recovery) and apply them alongside the skill's workflow. If they conflict, the more specific Comet extensions take precedence.
• If the execution preflight finds that background dispatch capability has disappeared, do not execute directly in the main window and do not create a new second decision. Return to the same Step 2 joint decision with the unavailable mode removed. After the user selects main-window execution there, run comet state set <name> build_mode executing-plans, then continue through that branch.

TDD Mode Execution Constraints:

If tdd_mode: tdd:

• build_mode: executing-plans: After loading the execution skill and before executing the first task, Immediately execute: Use the Skill tool to load the Superpowers test-driven-development skill once. Skipping this step is prohibited. After the skill loads, start from the first unchecked task and follow the loaded TDD Red-Green-Refactor cycle for each task. Must not skip the failing test verification phase. Do not reload this skill for subsequent tasks; follow the already-loaded flow. If resuming after context compaction, re-run this step to load the TDD skill once, then continue from the first unchecked task.
• build_mode: subagent-driven-development: The main session does not load the TDD skill. TDD constraints and evidence thresholds are defined in comet/reference/subagent-dispatch.md; every background implementer and fix agent must use the Skill tool to load the Superpowers test-driven-development skill and follow the Comet-injected TDD hard constraint.

If tdd_mode: direct: Follow normal flow, no enforced TDD.

executing-plans review gate:

Under executing-plans, the main session executes tasks directly (no isolated implementer subagent), so there is no per-task reviewer as in subagent-driven-development. Code review is done against completed diffs and scales with review_mode:

• review_mode: off: No automatic code review. Do not load requesting-code-review. Record the skip reason in the verification report draft or tasks.md.
• review_mode: standard: After all planned tasks are complete and before the build → verify phase guard, use the Skill tool to load the Superpowers requesting-code-review skill once and request one lightweight code review (correctness, security, edge cases) scoped to the whole change.
• review_mode: thorough: In addition to the single final review, request one segmented code review per task segment (every 3 tasks, scoped to that segment's diff). If total tasks ≤ 3, skip the mid-execution segments and only do the final review. Each segment review uses requesting-code-review against the segment's commit range. This is the closest equivalent to subagent-driven-development's per-task review that executing-plans can offer, since it has no isolated implementer to review per task.

Requirements (apply to standard and thorough):

• the requesting-code-review skill must be loaded before comet guard <change-name> build --apply
• if requesting-code-review is unavailable under standard or thorough, stop and ask the user to install/enable it and retry, or explicitly switch to review_mode: off with a recorded reason; never skip the gate or continue guard before that explicit switch
• CRITICAL review findings (security vulnerabilities, data loss risk, build/test failures) must be fixed first and must not be carried into verify
• if non-CRITICAL review findings are accepted, record the acceptance reason and impact scope in tasks.md, the commit body, a verification report draft, or another durable artifact

3b. In-Execution Debugging (Debug Gate)

During task execution, whenever a crash, unexpected behavior, test failure, or build failure appears while running the program, tests, build, or manual verification, must use the Skill tool to load the Superpowers systematic-debugging skill. Before root-cause investigation is complete, must not propose or implement source-code fixes.

For specific investigation, minimal failing test, fix verification, and keeping the current change verification loop, follow comet/reference/debug-gate.md.

4. Spec Incremental Updates

When the initial spec is found incomplete during implementation, handle by scale:

50% Threshold Determination: Using initial task count in tasks.md as baseline, if new tasks exceed half of that total, it's considered outside original plan scope, must follow the comet/reference/decision-point.md protocol to pause and wait for the user to decide whether to split into a new change.

When creating an independent change, must invoke /comet-open, not /opsx:new directly. /comet-open creates both OpenSpec artifacts and .comet.yaml, preventing the new change from leaving the Comet state machine.

User choices must include:

• "Split into new change" — create independent change via /comet-open
• "Continue in current change" — record scope-expansion decision, update tasks.md and delta spec, then continue

Principles:

• Delta spec is a living document, can be modified at any time during this phase
• Each update should be committed with commit message explaining the change reason
• Do not sync to main spec in advance, sync uniformly during archiving
• For small-scale incremental direct delta spec edits, note in commit message to facilitate design doc drift assessment during archiving

5. Context Management

Build is the longest phase and may span many tasks. To support resume after context compaction:

• After each task: complete acceptance per the current execution branch and review_mode before checking off and committing. subagent-driven-development dispatches no per-task reviewer under off; under standard, a per-task reviewer fires only when the task hits a risk signal; under thorough, every task gets a per-task reviewer. All modes must perform targeted verification by unique task text. Parse tasks.md checkboxes to count remaining work without rereading unrelated task bodies
• Context compression recovery: Follow comet/reference/context-recovery.md with phase set to build.
• User manual-change resume: handle uncommitted changes through comet/reference/dirty-worktree.md. That protocol defines checks, attribution, and prohibitions. Build-specific handling:
  1. After attribution, if the diff implies plan or spec changes, handle it through Step 4 "Spec Incremental Updates"
• Long task split: if a single task exceeds 200 lines of code changes, consider splitting it into multiple subtasks and commits

Exit Conditions

• All tasks.md checked
• Code committed
• Project-specific build/tests explicitly run and pass; do not rely only on guard auto-detection
• isolation has been written as current, branch, or worktree
• build_mode has been written as subagent-driven-development, executing-plans, or direct with explicit override; if subagent-driven-development, subagent_dispatch must be confirmed
• tdd_mode has been written as tdd or direct
• review_mode has been written as off, standard, or thorough
• Code review has been completed per the executing-plans review gate (Section "Execute plan") for the chosen review_mode: under standard or thorough, code review has been requested and CRITICAL review findings fixed or non-CRITICAL acceptance rationale recorded; under review_mode: off, the reason for skipping automatic code review has been recorded in a persistent artifact
• Phase guard: Run comet guard <change-name> build --apply; after all PASS, state advances to phase: verify

Guard runs the inferred project build check (npm run build, Maven, or Cargo when detected). When the inferred command fails, guard prints the command output as evidence for debugging.

If the project has no automatically inferred build command, the user or Agent must run the real build command first, then record its evidence separately:

comet state record-check <change-name> build --command "<actual build command>" --exit-code 0

--command records command text only; Comet never executes it. Build and verify evidence are separate and cannot substitute for each other. COMET_SKIP_BUILD=1 is only a compatibility bypass for legacy workflows, not auditable build evidence.

Before exit, run guard to auto-transition:

comet guard <change-name> build --apply

State file is automatically updated to phase: verify, verify_result: pending.

Automatic Handoff to Next Phase

Follow comet/reference/auto-transition.md. Key command:

comet state next <change-name>

• NEXT: auto → invoke the skill pointed to by SKILL to enter the next phase
• NEXT: manual → do not invoke the next skill; return control with HINT, end the invocation, and do not create another confirmation point
• NEXT: done → workflow is complete, no further action needed