🏋️ pi-write-coach

A coach for large files — game plan before, whistle after.

When a model tries to write a large file in one shot, the output can hit max_tokens mid-stream and land a truncated, broken file on disk. pi-write-coach prevents this by teaching the model to build large files incrementally: skeleton first, then small edits — each step too small to truncate into garbage.

pi install npm:@nukcole-xinluo9510/pi-write-coach

The problem

model wants to write 800-line file
  → emits 15,000+ output tokens in one shot
  → hits max_tokens limit mid-stream
  → auth.ts lands on disk: broken, half-written
  → you need to ask it to rewrite anyway

The solution

STEER injects the game plan:
  → write a skeleton first (imports, signatures, // TODO markers)
  → then edit each section one at a time

GUARD blows the whistle if a write is still too big:
  → blocks single writes/edits over 200 lines / 8000 chars
  → tells the model exactly how to fix it
  → releases path after 2 blocks to prevent retry loops

Two honest layers

STEER — before_agent_start

Injects a large-file policy into the system prompt before the model generates anything. This is the part that actually saves output tokens — it changes model behaviour ahead of generation.

GUARD — tool_call (write + edit)

Blocks oversized single writes and edits. Its job is file safety: a giant single write can be truncated mid-stream and land a broken file. The guard forces the work into small, untruncatable pieces.

Honest note: tool_call fires after the model already emitted the call as output tokens — the GUARD does not recover those tokens. It protects the file on disk. Token savings come from STEER only.

What it guards

| Operation | What is measured | When blocked | |-----------|-----------------|--------------| | write | content length | > 200 lines OR > 8000 chars | | edit | sum of all edits[].newText | > 200 lines OR > 8000 chars |

The "3-line skeleton + 1 giant edit" bypass is caught: the guard sums the total replacement output of the call, not per-edit.

Never blocked: package-lock.json, *.min.js, *.lock, *.generated.*, dist/** (add custom paths by extending ALLOWLIST in core.ts).

Loop-breaker

After a path is blocked RELEASE_AFTER_BLOCKS times (default: 2), it is permanently released for the session. A model that genuinely must write a large file is never trapped in an infinite retry loop.

Guarded models

Only applies to big-context models where chunking matters most: claude and deepseek (substring match on provider/id). Other models (gpt, gemini, local) are never affected.

Install

pi install npm:@nukcole-xinluo9510/pi-write-coach

Requirements

• pi 0.79+

Testing

cd packages/pi-write-coach
npm test   # 49 unit tests, no pi runtime needed

Covers: countLines, isGuardedModelKey, isAllowlistedPath, measureEdits, createSizeGuard state machine (block → loop-breaker → release → permanent pass), guidance message content.

License

MIT