Skip to content

v0.5.3 — Independent extension preview for Pi 1.3.0, 1.2.0 and 1.1.0

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 21 Oct 01:22
5cde2de

Pi Context Compaction v0.5.3 supplies an ordinary extension for unmodified official Pi 1.3.0, 1.2.0 and 1.1.0.

Install this release's extension package (a fixed package tag) from your terminal, then restart Pi:

pi install git:github.com/youxi-huang/pi-context-compaction@extension-v0.5.3

Start a new persistent text session and run /compaction-doctor to check it without calling a model. Source tag v0.5.3 selects the development monorepo, not the package. To follow later previews instead, install the moving extension-preview branch in place of the tag and update only this package with pi update --extension git:github.com/youxi-huang/pi-context-compaction@extension-preview. A plain pi update updates Pi itself.

This pre-release supports manual /compact and Pi's threshold compaction after completed text turns, source-linked history retrieval and explicit disk-confirmed failure recovery. In the sessions it handles, overflow recovery still requires explicit /compact, and background cache warming is paused. Sessions it never protected and cannot check, such as ones with native checkpoints, images or another Pi version, are left to Pi without a failure marker. Other checkpoint formats, parent-history grants, attachments, concurrent writers and Windows are outside its verified scope.

Ordinary Agent retries after transient provider failures are supported, including partial-response connection drops. Existing retry-related blocks and text-only context-overflow failures can use explicit /compact recovery when sufficient completed source history exists. Arbitrary context edits and unfinished tool-call omissions remain unsupported.

Optional preparation has two arms, both disabled by default: A prepares with the configured writer; B merges data-only work notes from the session model. A validated candidate can commit its covered prefix without a writer call while every later original stays verbatim. B carries covered user messages in full and falls back to the configured writer when they do not fit. Every writer call is journaled, including cancelled or discarded preparation and unknown usage. These are experimental paths; real-provider pause, total cost, cache behavior and recovery benefits remain unproven.

Validation: this publication run passed 88 stock-Pi package and lifecycle checks on official Pi 1.3.0, 1.2.0 and 1.1.0 and the full default-settings scenario suite on Pi 1.3.0 with strict acceptance and no performance exceptions or known gaps. Scenario evidence is bound to this exact package and source commit; the 50 ms request-overhead target and checkpoint-growth bound are required checks. Focused, security and frozen evaluator checks also remain required on final main. Services are synthetic and loopback only; no real-provider quality, latency or interactive TUI claim is made.

Source commit: 5cde2de. Preview runtime: stock-preview+src.cb5dcce5a76c043e. Resident runtime keeps its v0.3.0 fingerprint.

Usage and recovery · Validation

Changes

v0.5.3 adapts the independent extension's note writer to GPT-6 Luna. The session writer sends the session's ID with its request, so Pi's Codex provider can serve it from the session's prompt cache. Writers are asked for the note itself, and the two schema envelopes Luna returned around complete notes are removed before every commit check runs. A fresh session without arm B no longer offers context_note, saving about 1,209 Pi-estimated tokens per request. In a writer bench beside the v0.5.2 release build on the same day, DeepSeek V4.1 Flash committed 64 of 64 notes against 62 of 64; GPT-6 Luna at max, through a third-party relay, committed 47 of 64 against 37 of 64 before, and the notes it still loses carry next steps without sources. The changes are in the independent preview only: the resident runtime, its build fingerprint, the checkpoint schema and the evaluation pin are unchanged.

Changed

  • The independent preview's session writer sends the session's ID with its request, as Pi does with each of the session's requests. Pi's Codex provider keys its prompt cache on that ID: without it, GPT-6 Luna's writer read 7.1% of its input from cache, against 93.6% on DeepSeek. Providers that need no key ignore it.
  • The independent preview asks its note writers for the note itself rather than for an object matching the note's JSON Schema. Before validation it removes the two schema envelopes GPT-6 Luna returned around complete notes: a string type beside exactly the seven sections, and type, properties and required with exactly the seven sections under properties. Nothing else is rewritten, every commit check still applies, and /compaction-status names a removed envelope under envelope in writerCallDetails. Answers that are not JSON are refused as before.
  • With arm B off, a fresh independent-preview session no longer offers context_note, saving about 1,209 Pi-estimated tokens per request. A branch with an earlier call, saved or refused, activates it on reopen or branch navigation; explicit activations and saved notes remain usable.

Maintenance

  • The preview bench measures openai/gpt-6-luna through an OpenAI-compatible relay with --provider relay --base-url, keeping Pi's catalog settings and the Responses API, so the session writer's cache key reaches the relay. A relay's own usage counts make its cache share and cost incomparable with other runs; refusals, notes and probe scores compare. Writer rows record each relay response's shape as they do for Codex, and both now name the keys and kinds of each section's items, which is how the relay run traced its refused notes to next steps without sources.
  • The preview bench stops a unit still running after --unit-minutes, 10 by default, so one request that never returns cannot hold a run; the unit settles as crashed and a probe it held is scored as blocked. Writer rows and the summary also count the schema envelopes removed from notes.
  • The preview bench measures openai-codex/gpt-6-luna on the ChatGPT subscription with --provider codex, in writer, probes and smoke modes, under the same request, time and cost limits. In writer mode it records the shape of every Codex response, never its text, which is how the first run traced its refused notes. In Claude Code cloud sessions the access token stays in the environment's network proxy and Pi holds an unsigned placeholder; docs/context-memory/codex-subscription.md covers the setup and the renewal every ten days or so.