pi-checkpoint-compaction

Zero-LLM compaction for pi coding agent.

Instead of re-summarizing your whole history with an LLM every time the context window fills up, this extension keeps a model-maintained checkpoint file (~/.pi/agent/checkpoints/<session>.md) and folds it into context at compaction time — no LLM call. If the checkpoint is stale or missing, it degrades to a mechanical span digest (still no LLM), and only as a dead-last fallback does it hand off to pi's built-in LLM summarization.

This is a port of the dsh-checkpoint engine (v0.2.4) used by deepseek-harness.

Install

pi install npm:pi-checkpoint-compaction

or try it for one run without installing:

pi -e npm:pi-checkpoint-compaction

How it works

The extension registers one tool and one command:

Resource	Purpose
checkpoint_update tool	Overwrites this session's checkpoint file with the model's current state (goal / done / next).
/checkpoint command	Prints the current checkpoint.

At compaction (automatic or manual /compact), the session_before_compact hook applies a 3-tier policy:

Tier	Condition	What happens	LLM?
1	checkpoint newer than the span (60 s epsilon)	checkpoint text folded in as the new summary	no
2	checkpoint exists but older than the span (max 12 h)	mechanical digest of the span (user asks, assistant decisions, tool calls, [TOOL-ERROR] lines) with the checkpoint as goal anchor	no
3	no usable checkpoint and nothing digestible in the span, or custom compaction instructions	defer to pi's default LLM compaction	yes

Split turns (pi cutting mid-turn, keeping a turn prefix) are handled mechanically too: the span digest/checkpoint fold is merged with a mechanical digest of the turn prefix, in pi's own **Turn Context (split turn):** format. That path is where pi's default summarizer spends two LLM calls — here it spends zero.

The checkpoint file is per-session (<sessionId>.md); a shared active.md is only read as a fallback anchor and is never overwritten when a session ID is known. Checkpoints older than 12 h are ignored.

Typical usage

Tell the agent once:

After every meaningful milestone, call checkpoint_update with the current goal, what is done, and what is next.

You don't even need to trust the agent

A turn_end hook rewrites the checkpoint file after every turn: it preserves the model's last semantic checkpoint verbatim and appends a fresh "Recent activity" section (last user prompt, assistant text, tool calls, tool errors of the turn that just ended). If the model called checkpoint_update during the turn, the harness skips — the model's word wins.

So the file is fresh at every compaction boundary even if the model never calls the tool at all; the tool then exists to add semantic state (goals, decisions, plans), not to keep the file fresh. Compaction costs zero LLM tokens in either case.

Per-tool-call updates are deliberately not done: pi's compaction always keeps the recent turns, and mid-turn compaction digests the turn prefix mechanically, so turn granularity is exactly as fresh as compaction ever needs — at a fraction of the file churn.

Finally, the extension appends a one-paragraph standing instruction to the system prompt each run (via before_agent_start) reminding the model to call checkpoint_update at milestones — so the semantic quality of the file tracks what the model is actually thinking, even though freshness no longer depends on it.

Development

The extension is a single file (extensions/checkpoint-compaction.ts) with no runtime dependencies beyond what pi bundles (typebox, the pi core).

Tests

test-checkpoint.mjs at the repository root is a self-contained suite — no test framework, no network, runs against a temp directory. 46 checks covering:

• registration of the checkpoint_update tool, /checkpoint command, and the compaction / turn_end / before_agent_start hooks
• checkpoint file lifecycle: per-session writes, 12 h staleness cap
• Tier 1: fresh checkpoint folded in verbatim, zero LLM
• Tier 2: stale checkpoint → mechanical span digest (user asks, assistant steps, [TOOL-ERROR] lines) with the checkpoint as goal anchor
• Tier 3: no usable checkpoint and nothing digestible → defers to pi's default LLM compaction; custom compaction instructions always defer
• split turns: turn prefix digested in pi's **Turn Context (split turn):** format and merged with the span digest
• turn_end auto-maintenance: "Recent activity" appended every turn, the model's last semantic block preserved verbatim, 60 s freshness epsilon
• system-prompt standing-instruction injection via before_agent_start

npm test            # or: node test-checkpoint.mjs

License

MIT