pi-context-diet

A Pi extension that keeps long sessions workable.

Large discovery tool outputs — search hits, logs, MCP results — are archived to disk and replaced in the transcript by a short fact list written by a cheaper model, with a pointer to the full text. Exact reads, small outputs, failures, and anything that looks like a credential pass through untouched.

[context-diet: ~8205 -> ~229 tokens via openrouter/~deepseek/deepseek-flash-latest;
 full output: ~/.cache/pi-context-diet/raw/373a827986f8e3ae.txt]
- 412 .d.ts files under dist/themes; no matches under dist/core/plugins
- ToolResultEvent exposes toolName, input, content, isError, parentToolCallId

This is a Pi port of the PostToolUse technique in "Claude Context Diet" (archived copy, in case the artifact expires). Same policy and guardrails, expressed as Pi's tool_result hook instead of a settings.json hook chain.

Install

pi install git:github.com/lscoughlin/pi-context-diet

Pin a release if you prefer: pi install git:github.com/lscoughlin/pi-context-diet@v0.1.0. To try it for one invocation without touching settings:

pi -e git:github.com/lscoughlin/pi-context-diet

Restart Pi afterwards. The extension loads and registers /context-diet at startup.

Use

/context-diet	status: active helper, model chain, thresholds, this session's savings
/context-diet config	interactive menu for enabled, helperModels, minTokens, maxSummaryTokens, minSaving

/context-diet config needs an interactive session — Pi's ui.select is a no-op under pi -p, so scripted runs silently skip the menu. status works everywhere.

Configure

The extension works with no configuration. The first run reports built-in defaults; the first change through /context-diet config writes $PI_CODING_AGENT_DIR/extensions/context-diet/config.json (default ~/.pi/agent/extensions/context-diet/config.json). To edit it by hand, copy src/context-diet/config.example.json there first — the //-prefixed keys document each field.

The file is read once at startup, so hand edits need a restart. Menu edits apply immediately.

key	default	meaning
enabled	true	master switch
minTokens	3500	outputs at or below this stay raw
maxSourceChars	200000	larger outputs stay raw, capping helper cost
maxSummaryTokens	1800	a longer replacement is rejected — a bad summary is worse than the raw text
minSaving	0.3	reject a replacement that isn't at least this much smaller
helperTimeoutMs	90000	wall-clock budget for the helper call
helperModels	four cheap DeepSeek/GLM/Luna ids	[provider, modelId] pairs, cheapest-first; first one with credentials wins
exemptTools	read, edit, write, powershell	never compressed
bashAllow	rg/grep/find/test/log commands	bash is eligible only if it matches, exits 0, and prints nothing on stderr
secret	api-key/password/bearer/private-key	matches in the output or the arguments leave the result raw
archiveDir	~/.cache/pi-context-diet/raw	where full outputs go; ~ is expanded

MCP tools are eligible by default — in the original's measurements they were the large majority of compressed tokens. Add a tool name to exemptTools to opt it out.

Choosing a helper model

/context-diet status prints every configured entry with its resolve status, so a typo is visible rather than silently skipped:

helper in use: openrouter/~deepseek/deepseek-flash-latest
model chain (first with credentials wins):
  1. openrouter/~deepseek/deepseek-flash-latest  ok ($0.0243/Mtok in)
  2. openrouter/~deepseek/does-not-exist         MISSING (no such model id)
  3. openrouter/~z-ai/glm-flash-latest           ok ($0.02/Mtok in)

Model ids must match ~/.pi/agent/models-store.json exactly, including any ~ prefix. A fast, cheap model is the right choice: the helper only summarises text it is handed. It must be accurate above all — a model that invents file names is worse than no compression, because a plausible summary is not reviewable the way raw output is. The default chain was picked with that in mind; a strong model costs roughly 80× more per input token and is not usually worth it here.

Uninstall

pi remove git:github.com/lscoughlin/pi-context-diet

Archives under ~/.cache/pi-context-diet/ are left behind; delete them manually. To just switch it off, set enabled to false via /context-diet config.

How it behaves

• Fails open. Any error, timeout, unparseable answer, or oversize summary leaves the original result in place. The extension never blocks a tool.
• Never touches what you'd edit. read, edit, and write results stay byte-exact, since the model needs them to make changes.
• Bash is narrow by default. Only successful, stderr-free discovery commands qualify. A failing test run is a signal, not noise, so it stays raw.
• Credentials are checked twice — against the output and against the tool arguments, since a secret can arrive on a command line without appearing in the result.
• Nested calls are skipped. Sub-results from codemode scripts are left alone.
• Summaries are adversarial input. The helper prompt treats tool output as untrusted data and bounds itself to 1–6 facts and 0–3 unknowns.

Develop

npm install
npm test          # 20 offline behavioural checks, no model calls
npm run typecheck

test/selftest.mjs loads the extension through jiti with the same module aliases Pi uses, drives the command and the tool_result hook against a fake model registry, and asserts the guardrails — comment keys surviving a config save, derived values not leaking to disk, the helper reorder keeping fallbacks, and credential-bearing argv being left raw.

Cost, honestly

The technique trades helper tokens for context occupancy. A compressed tool result is re-read on every subsequent turn, so the saving compounds — a 12,000-token blob becomes a ~380-token stub and stays that way. The original's 7-day ledger measured −89.9% of tool-output tokens in the main context.

Dollar savings, though, depend on the gap between your main model and the helper. If you run an expensive main model, this is close to free money. If you already run a cheap model, the win is working room and longevity, not the bill: less compaction churn, fewer truncated contexts, more turns before a session degrades.

Credits

Technique and policy from "Claude Context Diet". This port differs from the original in one structural way: Pi exposes tool results through an extension tool_result hook, which can replace what the model reads, whereas Pi's YAML hook system (tool.after.*) explicitly cannot modify a completed tool's output.

License

MIT