name	tp-spike-implement
description	Execute a spike from plan.md — serial exploration with human review gates at phase boundaries. Speed over ceremony.
argument-hint	{spike-name} [phase-number] [--auto] [--force-takeover]

Spike Implement

Execute phases from a spike plan serially, with human review gates between phases.

Arguments:

{spike-name} (required), optionally a phase number (e.g., my-spike 2). Without a phase number, executes the next incomplete phase.
--auto (optional) — autonomous mode. Skips phase confirmation, auto-continues at phase boundaries, retries blocked tasks with simplification (max 3 attempts), logs all decisions to decisions.md. See skills/_shared/auto-mode.md for convention.

Prerequisites

three-pillars-docs/tp-designs/{spike-name}/plan.md must exist. If not, tell the user to run /tp-spike-plan {spike-name} first and stop.

Steps

Run first-run preflight per skills/_shared/first-run.md.

0a. Run cwd preflight per skills/_shared/cwd-preflight.md: python3 "$TP_ROOT"/skills/_shared/cwd_preflight.py {spike-name}. Exit 3 → stop and show the cd fix. Exit 0 → continue.

Run collaboration preflight per skills/_shared/collaboration.md with phase: "spike-implement". This verifies the branch and lock before writing code. Honor --force-takeover if passed. In --auto mode, do not prompt — if the lock is held by another developer, log the conflict to decisions.md and stop.
Read plan.md from the design directory.
Determine which phase to execute:
- If a phase number was given, use that.
- Otherwise, find the first phase with unfinished tasks (no **Status**: line or status is not Done/Skipped).
Phase confirmation:
- Normal mode: Show the user which phase you are about to execute and how many tasks it contains. Ask for confirmation.
- --auto mode: Log the phase start to decisions.md and proceed without confirmation.
For each task in the phase, execute serially (no worktrees, no parallel agents):

For each task:
- Hypothesis: Read what the task expects to learn or build.
- Try: Implement the task directly. Prioritize speed — no mandatory test-first, no strict red-green-refactor. Write tests when they add value, skip ceremony when they do not.
- Evaluate: Check if the result works. Run any existing tests for affected files. Judge whether the task outcome matches the Hypothesis.
- Update status: Add **Status**: Done to the task in plan.md. If the task fails after reasonable effort, mark it **Status**: Blocked with a short reason and continue to the next task.
- Commit: Per skills/_shared/commit-after-work.md, commit the experiment. Artifact paths to stage: the tracked code files the experiment touched, three-pillars-docs/tp-designs/<spike-name>/plan.md (updated status), any new files under three-pillars-docs/tp-designs/<spike-name>/demos/ produced by this experiment (typically demos/<NN>-* where <NN> is the experiment number), and three-pillars-docs/tp-designs/<spike-name>/decisions.md if --auto mode appended an entry. Demos and decisions are tracked — they are the spike's evidence record. Stage only files this experiment touched, not the entire demos/ directory. Commit message: Spike: <spike-name> <phase>.<experiment>.
- --auto mode — retry on failure: If a task fails, don't immediately mark it Blocked. Instead:
  1. Log a simplification entry to decisions.md describing what failed and why.
  2. Simplify the approach (reduce scope, use mocks, skip edge cases, use simpler implementation).
  3. Retry the task with the simplified approach.
  4. Repeat up to 3 attempts total. If all 3 fail, mark as **Status**: Abandoned (3 attempts exhausted) and log the full history to decisions.md.
Phase complete:
- Normal mode — stop and present the Deliverable:
  - Summarize what was built or learned in this phase.
  - List any blocked tasks with reasons.
  - Present the phase Deliverable from plan.md.
  - Wait for the user to decide:
    - Continue: proceed to the next phase.
    - Pivot: help the user update plan.md (add/remove/reorder tasks or phases) before resuming.
    - Stop: end the spike. Save progress in plan.md.
  - Do NOT auto-advance to the next phase.
- --auto mode — assess and continue:
  - Log a phase boundary entry to decisions.md with: tasks completed/blocked/abandoned, summary of findings, decision to continue or stop.
  - If there are remaining phases, auto-advance to the next phase.
  - If all tasks in the phase were blocked/abandoned, still continue to the next phase (later phases may not depend on earlier ones). Only stop the entire spike if ALL remaining phases have been attempted.
Update plan.md — statuses should already be updated per-task in step 5. Verify all tasks in the completed phase have a status line.

Rules

Validate {spike-name} per skills/_shared/validate-name.md.
Respect the lock per skills/_shared/collaboration.md — the preflight step can refuse to proceed if another developer holds this spike.
Serial execution only — no worktrees, no parallel subagents. One task at a time in the main agent.
Speed over ceremony — no strict red-green-refactor. Write the code, check if it works, move on. Tests are encouraged but not required for every task.
Demo / artifact convention — every artifact a future reader of spike-results.md would want to inspect goes in three-pillars-docs/tp-designs/<spike-name>/demos/. This includes: tool inputs (the actual repo maps, config files, wrapper scripts, prompts you fed to the tool), tool outputs (generated files, logs, transcripts), measurements (CSV/markdown tables), and rendered media (MP4s, screenshots). Pure scratch state (test-project clones, container images, build caches) can live in /tmp/<spike-name>/ — it is reproducible and bulky. Rule of thumb: if /tmp/ were wiped right now, would spike-results.md still be reproducible from demos/ plus the project source it points to? If not, the missing piece belongs in demos/. Never treat /tmp/ as durable storage for evidence — it gets cleaned and the finding becomes unverifiable. The directory is tracked (not gitignored) — demos are the spike's permanent evidence record. Each experiment commits its own demo files alongside the plan.md status update. Create the directory on first use.
Human review gates — in normal mode, always stop at phase boundaries. Never auto-advance. In --auto mode, auto-advance after logging a phase boundary entry to decisions.md.
Pivot support — in normal mode, if the user wants to change direction, help them edit plan.md before continuing. In --auto mode, no pivots — continue with the plan as-is.
If a task is already done (artifact exists, test passes), mark it **Status**: Skipped (pre-exists) and move on.
--auto mode: Follow the auto-mode convention in skills/_shared/auto-mode.md. Log all phase boundaries, task failures, simplification attempts, and retry outcomes to decisions.md. Never prompt the user. Execute ALL phases in sequence without stopping.

Más de este repositorio

mismo repositorio

council

CurtisThe/three-pillars-plugin

Convene the Council of High Intelligence — multi-persona deliberation with historical thinkers for deeper analysis of complex problems.

2026-07-134

tp-design-audit

CurtisThe/three-pillars-plugin

Design Audit — multi-angle review of a detailed design against the actual codebase before implementation.

2026-07-134

tp-design-complete

CurtisThe/three-pillars-plugin

Mark a TDD design as complete — stamp completion, archive handoff.md, archive to three-pillars-docs/completed-tp-designs/, and optionally commit + open a PR merging the design branch back to the base branch.

2026-07-134

tp-docs-init

CurtisThe/three-pillars-plugin

Scaffold architecture.md, product_roadmap.md, and known_issues.md in three-pillars-docs/ from codebase analysis. Creates the three project docs that the TDD pipeline reads for context. Assumes three-pillars-docs/vision.md already exists — if missing, recommends /tp-setup first.

2026-07-134

tp-guide

CurtisThe/three-pillars-plugin

Read project docs and propose what to do next. Surfaces highest-impact work from roadmap, known issues, and in-flight designs.

2026-07-134

tp-implementation-audit

CurtisThe/three-pillars-plugin

Final audit of a completed plan — verify the full implementation against both design.md and detailed-design.md.

2026-07-134

Spike Implement

Execute phases from a spike plan serially, with human review gates between phases.

Arguments:

{spike-name} (required), optionally a phase number (e.g., my-spike 2). Without a phase number, executes the next incomplete phase.

--auto (optional) — autonomous mode. Skips phase confirmation, auto-continues at phase boundaries, retries blocked tasks with simplification (max 3 attempts), logs all decisions to decisions.md. See skills/_shared/auto-mode.md for convention.

Prerequisites

three-pillars-docs/tp-designs/{spike-name}/plan.md must exist. If not, tell the user to run /tp-spike-plan {spike-name} first and stop.

Steps

Run first-run preflight per skills/_shared/first-run.md.

0a. Run cwd preflight per skills/_shared/cwd-preflight.md: python3 "$TP_ROOT"/skills/_shared/cwd_preflight.py {spike-name}. Exit 3 → stop and show the cd fix. Exit 0 → continue.

Run collaboration preflight per skills/_shared/collaboration.md with phase: "spike-implement". This verifies the branch and lock before writing code. Honor --force-takeover if passed. In --auto mode, do not prompt — if the lock is held by another developer, log the conflict to decisions.md and stop.

Read plan.md from the design directory.

Determine which phase to execute:

If a phase number was given, use that.
Otherwise, find the first phase with unfinished tasks (no **Status**: line or status is not Done/Skipped).

Phase confirmation:

Normal mode: Show the user which phase you are about to execute and how many tasks it contains. Ask for confirmation.
--auto mode: Log the phase start to decisions.md and proceed without confirmation.

For each task in the phase, execute serially (no worktrees, no parallel agents):

For each task:

Hypothesis: Read what the task expects to learn or build.
Try: Implement the task directly. Prioritize speed — no mandatory test-first, no strict red-green-refactor. Write tests when they add value, skip ceremony when they do not.
Evaluate: Check if the result works. Run any existing tests for affected files. Judge whether the task outcome matches the Hypothesis.
Update status: Add **Status**: Done to the task in plan.md. If the task fails after reasonable effort, mark it **Status**: Blocked with a short reason and continue to the next task.
Commit: Per skills/_shared/commit-after-work.md, commit the experiment. Artifact paths to stage: the tracked code files the experiment touched, three-pillars-docs/tp-designs/<spike-name>/plan.md (updated status), any new files under three-pillars-docs/tp-designs/<spike-name>/demos/ produced by this experiment (typically demos/<NN>-* where <NN> is the experiment number), and three-pillars-docs/tp-designs/<spike-name>/decisions.md if --auto mode appended an entry. Demos and decisions are tracked — they are the spike's evidence record. Stage only files this experiment touched, not the entire demos/ directory. Commit message: Spike: <spike-name> <phase>.<experiment>.
--auto mode — retry on failure: If a task fails, don't immediately mark it Blocked. Instead:
1. Log a simplification entry to decisions.md describing what failed and why.
2. Simplify the approach (reduce scope, use mocks, skip edge cases, use simpler implementation).
3. Retry the task with the simplified approach.
4. Repeat up to 3 attempts total. If all 3 fail, mark as **Status**: Abandoned (3 attempts exhausted) and log the full history to decisions.md.

Phase complete:

Normal mode — stop and present the Deliverable:
- Summarize what was built or learned in this phase.
- List any blocked tasks with reasons.
- Present the phase Deliverable from plan.md.
- Wait for the user to decide:
  - Continue: proceed to the next phase.
  - Pivot: help the user update plan.md (add/remove/reorder tasks or phases) before resuming.
  - Stop: end the spike. Save progress in plan.md.
- Do NOT auto-advance to the next phase.
--auto mode — assess and continue:
- Log a phase boundary entry to decisions.md with: tasks completed/blocked/abandoned, summary of findings, decision to continue or stop.
- If there are remaining phases, auto-advance to the next phase.
- If all tasks in the phase were blocked/abandoned, still continue to the next phase (later phases may not depend on earlier ones). Only stop the entire spike if ALL remaining phases have been attempted.

Update plan.md — statuses should already be updated per-task in step 5. Verify all tasks in the completed phase have a status line.

Rules

Validate {spike-name} per skills/_shared/validate-name.md.

Respect the lock per skills/_shared/collaboration.md — the preflight step can refuse to proceed if another developer holds this spike.

Serial execution only — no worktrees, no parallel subagents. One task at a time in the main agent.

Speed over ceremony — no strict red-green-refactor. Write the code, check if it works, move on. Tests are encouraged but not required for every task.

Demo / artifact convention — every artifact a future reader of spike-results.md would want to inspect goes in three-pillars-docs/tp-designs/<spike-name>/demos/. This includes: tool inputs (the actual repo maps, config files, wrapper scripts, prompts you fed to the tool), tool outputs (generated files, logs, transcripts), measurements (CSV/markdown tables), and rendered media (MP4s, screenshots). Pure scratch state (test-project clones, container images, build caches) can live in /tmp/<spike-name>/ — it is reproducible and bulky. Rule of thumb: if /tmp/ were wiped right now, would spike-results.md still be reproducible from demos/ plus the project source it points to? If not, the missing piece belongs in demos/. Never treat /tmp/ as durable storage for evidence — it gets cleaned and the finding becomes unverifiable. The directory is tracked (not gitignored) — demos are the spike's permanent evidence record. Each experiment commits its own demo files alongside the plan.md status update. Create the directory on first use.

Human review gates — in normal mode, always stop at phase boundaries. Never auto-advance. In --auto mode, auto-advance after logging a phase boundary entry to decisions.md.

Pivot support — in normal mode, if the user wants to change direction, help them edit plan.md before continuing. In --auto mode, no pivots — continue with the plan as-is.

If a task is already done (artifact exists, test passes), mark it **Status**: Skipped (pre-exists) and move on.

--auto mode: Follow the auto-mode convention in skills/_shared/auto-mode.md. Log all phase boundaries, task failures, simplification attempts, and retry outcomes to decisions.md. Never prompt the user. Execute ALL phases in sequence without stopping.