Upgrade Pi to 0.81.1 and reconcile proactive-compaction retry ownership

[ArneDeutsch]

Problem

PiCC currently pins the coordinated Pi packages at 0.80.10. A transient Codex WebSocket failure during proactive compaction can make each physical ctx.compact() call fail because that Pi version does not apply its configured provider retry policy to summarization.

In an observed session, PiCC exhausted its three bounded outer attempts and paused safely. Manual /compact succeeded shortly afterward. The same session had compacted successfully earlier, and its persisted provider diagnostic identified a transport failure after the message stream started rather than malformed state, credentials, hooks, or restoration logic.

Pi 0.81.1 includes the upstream summarization retry change from pi-mono#6901, addressing pi-mono#6647.

Expected behavior

A transient summarization transport drop should retry under one coherent bounded policy and normally complete without requiring manual /compact. Persistent failure should continue to pause safely with actionable recovery and retained transcript state.

Proposed scope

• Upgrade @earendil-works/pi-coding-agent, @earendil-works/pi-agent-core, @earendil-works/pi-ai, and @earendil-works/pi-tui to at least 0.81.1.
• Adapt any changed APIs or events.
• Reconcile Pi's new inner summarization retries with PiCC's three outer compaction attempts so the effective retry budget and delay do not multiply unexpectedly.
• Add regression coverage for a transient summary transport failure followed by success.
• Preserve PiCC's fail-closed checkpoint, generation-bound recovery token, manual recovery, and transcript-retention semantics.

Non-goals

• Implementing a PiCC WebSocket client.
• Retrying quota, authentication, abort, or deterministic provider failures.
• Changing proactive compaction thresholds.
• Exposing private provider diagnostics in transcripts.

🤖 Generated with the implement-feature skill — agent-authored, posted under the maintainer's authenticated gh account, not hand-written.

[ArneDeutsch]

Automatic compaction diagnostics — three consecutive failures

A second persisted coordinator session on Pi 0.80.10 exhausted all three proactive-compaction attempts in one checkpoint generation. Attempts 1 and 2 necessarily followed the operational retry path, and attempt 3 persisted checkpoint-exhausted with failureCategory: operational and manual-recovery guidance.

The preceding zero-usage Operation aborted record was the expected stop of the ordinary run before compaction, not the compaction failure itself. No summary committed during the failed attempts. The same session had compacted successfully earlier, and a later manual compaction committed successfully.

The transcript does not retain the individual provider exception text: the current main-session adapter reduces each onError value to a lifecycle category. Therefore this recurrence is consistent with the transient summarization transport failure already described here, but does not independently prove the same WebSocket root cause.

🤖 Generated with the implement-feature skill — agent-authored, posted under the maintainer's authenticated gh account, not hand-written.

[ArneDeutsch]

What was built for #156

PiCC now uses Pi 0.81.1 summarization retries inside one proactive compaction transaction for both main sessions and PiCC-created children. Retry eligibility, budgets, and backoff remain Pi-owned while PiCC preserves checkpoint admission, retained state, generation authority, and safe recovery.

How behaviour changes

An eligible transient summary transport failure can recover automatically and resume the paused work once. Persistent or non-retryable failures still pause or terminate by category, with pre-commit manual recovery retained and post-commit failures requiring replacement rather than recompaction.

Differences & extensions vs. the original ask

Main-session cancellation is logical because Pi's extension API cannot physically abort compaction; PiCC waits for the bounded configured transaction to settle and prevents continuation. Child sessions physically abort compaction through the public SDK. The implementation also adds sanitized child retry progress and real-stack privacy, lifecycle, request-budget, and transcript-retention proofs.

Not included this pass

A public main-extension compaction-abort API, uniform retry-wait visibility across every mode, redaction of Pi-owned native retry diagnostics, and an upstream Windows RPC shutdown fix are not included.

🤖 Generated with the implement-feature skill — agent-authored, posted under the maintainer's authenticated gh account, not hand-written.