comet-build

Comet Phase 3: Plan and Build (Build)

Before starting or recovering, read and follow comet-classic/reference/classic-layout.md. Every OpenSpec CLI call in this file must use the adapter, and every file path must use the <classic-*> logical roots bound by that protocol.

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-classic/reference/scripts.md, then run entry verification. When resuming from any entry point, first run the recovery check in comet-classic/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-classic/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.

Before dispatching, the main session fixes the full plan path docs/superpowers/plans/<YYYY-MM-DD>-<change-name>.md (e.g. docs/superpowers/plans/2026-08-21-rename-alert.md) and passes it into the instructions below verbatim.

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. If the Skill tool is unavailable or the skill cannot be found, end immediately with a single-line final reply SKILL_UNAVAILABLE; do not retry and do not write the plan yourself without the skill
2. Read the Design Doc (technical design document under docs/superpowers/specs/)
3. Read <classic-change-dir>/tasks.md (task boundaries)
4. Follow the skill's guidance to create the plan; no user will answer you during the run, so skip steps that ask the user a question (such as the writing-plans Execution Handoff at the end) and do not ask the user anything

Plan requirements:

• Save to the plan path given in the instructions; do not change the file name
• Cover only the tasks listed in tasks.md; do not expand scope
• 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

After writing the plan to file, the last line of your final reply must be exactly:

PLAN_PATH: <relative path of the plan file>

Execute subagent: Dispatch the above task to a subagent.

After the subagent completes:

• Read the PLAN_PATH: line at the end of the reply; if the file exists, record it as the plan; if the line is missing, look for an existing valid path in the reply text
• If the subagent reports SKILL_UNAVAILABLE, the dispatch fails, or no valid path can be obtained, fall back to loading the Superpowers writing-plans skill inline in the main session (degraded fallback), remember that dispatch is unavailable for this session, and go inline directly on later Step 1 entries

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.

When presenting the joint decision, provide only the execution method, TDD mode, and code-review mode supported by this workflow. The workspace was prepared and bound during Open; if the change has no valid isolation, return to /comet-open instead of creating or switching a workspace in Build.

After recording the plan, provide exactly one joint decision point that collects whether to continue now, execution method, TDD mode, and code review mode. Do not ask continue/pause first and then create another configuration blocker.

This is a user decision point. Follow comet-classic/reference/decision-point.md once and show the plan summary, pause option, and every executable Step 3 setting together. 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 execution method, TDD mode, and code review mode below.

The plan is in the workspace prepared during Open. First verify the existing binding:

comet state get <name> isolation

If the result is empty, stop Build and return to /comet-open for workspace resolve/prepare; do not make the first current/branch/worktree choice or create a workspace here.

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

The execution, TDD, and review tables are part of the Step 2 joint decision and do not create another pause. Recommendations never replace user confirmation.

After user selection, update only the execution method, TDD mode, and code-review mode fields; preserve the isolation and bound_branch established during Open.

• 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: run comet state set <name> subagent_dispatch confirmed to record the selected subagent execution, then run comet state set <name> build_mode subagent-driven-development

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 must write current, branch, or worktree during Open and complete the corresponding workspace preparation and bound_branch binding before entering Build; if it is missing, Build may only stop and return to Open for repair.

subagent_dispatch is a script-enforced hard constraint that records the user's selected subagent execution. 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; it is not a capability check.

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.

Execution location:

Open has already prepared the current directory, branch, or Worktree according to isolation and returned the actual projectRoot. On resume, run comet classic workspace resolve <name> --json, enter the returned directory, then run comet state select <change-name>; do not create a Worktree, switch branches, commit a plan to transfer it across Worktrees, or rebind isolation in Build.

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 loading fails, stop and report the error; 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-classic/reference/subagent-dispatch.md for Comet-specific extensions (subagent 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 subagent dispatch fails, follow comet-classic/reference/subagent-dispatch.md to record the current task as BLOCKED with the failure reason; the main session must not take over implementation.

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-classic/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 loading requesting-code-review fails under standard or thorough, stop and ask the user to resolve the error 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-classic/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-classic/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

Handoff synchronization: adding, modifying, or removing a delta spec expires the design handoff pack (handoff_hash). During Build you can regenerate it directly without reverting the current phase or step:

comet handoff <change-name> design --write

Regeneration rebuilds the handoff from the current OpenSpec artifacts and updates handoff_hash; it does not change the phase field or the Runtime currentStep, so you can continue in the Build phase after refreshing.

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-classic/reference/context-recovery.md with phase set to build.
• User manual-change resume: handle uncommitted changes through comet-classic/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-classic/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