Adopt pi 1.0, improve the loop and compaction, and go English-only

[Justar96]

Summary

• pi 1.0 adoption. DeepSeek runs through llm-pi-ai on @earendil-works/pi-ai 1.0.0, using the shipped deepseek-official route. This removes llm-deepseek and the three packages only it used: /usage, Files-API upload, request extensions, and DeepSeek image pricing are gone. MCP uses @earendil-works/pi-mcp. run_code runs in a pi-codemode QuickJS sandbox, while workflow scripts keep the Node runtime.
• Agent loop. When a reply is cut off at the output limit, the model is told which tool calls were dropped and the turn continues once. Hooks and tool policies can halt a turn through ToolHalt, including from inside run_code. The TUI shows the reason for a halt. A goal keeps going after a recovered turn.
• Live tool progress. bash and pwsh stream their newest output to the TUI through the process-local events agent/tool-progress and agent/tool-executed. Nothing new is persisted or shown to the model.
• Compaction. There is a live compaction-basic section in settings.yaml, with provider-wide modelPolicies and thresholdTokens, so a cliproxyapi route can set its own compaction point. Checkpoints keep <read-files>, <modified-files>, and <todo-list>. When the summary request itself overflows, it falls back to a bounded transcript. Transient summary errors are retried.
• English-only. Removed all .zh.md and .i18n.yaml files, the zh UI locale and locale setting, and the translation-pairing tooling. Frozen history in snapshots/, archived notes, and historical formats is untouched.

Session format

There is no version bump and no persistence digest moved. session-log-deepseek/delivery-accepted was added to the retired vocabulary. bun scripts/persistence-changes.ts --check lists request/tool-update and subagent/routing-decision as unacknowledged same-version root additions. These need acknowledging before release.

Eval

evals/agent-loop/versions/unreleased/2026-10-02-pi-migration-compaction has a paired record against 7428f19adb: 7 models, the standard suite, 3 trials each.

• Total tokens and request counts are within noise for every model. No tool errors were added.
• Flagged regression: gpt-6.1-sol, failures 1 → 3. I accepted it. All three candidate failures hit the 180-second wall-clock limit while waiting on a model response with no tool running. They happened in the same trial at the same time, and the base arm and gpt-6-astra show the same gateway stall.
• DeepSeek uncached input rose 9.2% (flash) and 13.6% (v4-pro) in that run. Most of it was a one-time cold cache for the new prefix. The rest was pi-ai's wire: the deferred-tool placeholder, cache_control markers, and the prompt and runtime context sent as two separate user messages. The new llm-pi-ai messagesWire option on deepseek-official restores the old adapter's shape. A follow-up paired record, 2026-10-02-deepseek-messages-wire, shows uncached input at −2.6% (flash) and −8.3% (v4-pro), within noise, with 108/108 successes and no regressions.

Checks

• bun run preflight passes: 26 passed and 1 skipped, because actionlint isn't installed. This includes runtime and e2e PTY, with the new live-output scenario.
• Not run: Windows/macOS, a packed install, and key-gated e2e outside the eval.

Known follow-ups

• llm-pi-ai/src/context.ts joins multi-block text without a separator. Fixing it changes what the model sees, so it needs an eval.
• agent-presets/src/display.ts and its ./display subpath are dead code.
• pi-codemode is pinned exactly, because the meter uses its worker internals.