feat: Phase-2 canonical history — native tool_use/functionCall replay across backends

[saucam]

What

Phase 2 of the canonical-history plan (docs/multi-provider-meta-harness.md). Capture was already structured — CanonicalTurn carries content, thinking, and CanonicalToolCall[] with ids/inputs/outputs — but every to<Provider>Messages converter flattened tool calls into [Tool: …] prose, so a switched session got a narrated summary instead of a native continuation. This PR rewrites the rendering side only; no new capture, no switch-loop changes.

Converters

Converter	Native shape emitted
toAnthropicMessages	assistant tool_use blocks + a paired user message of tool_result blocks. tool_use_id round-trips from CanonicalToolCall.id; failures carry is_error: true. Empty text blocks are omitted (API rejects them); consecutive same-role messages are fine (API merges).
toGeminiContent	functionCall parts on the model turn + a role "function" turn of functionResponse parts. Verified against @google/generative-ai's validateChatHistory (VALID_PARTS_PER_ROLE) — the design doc sketched role "user" for responses, which the SDK rejects; doc corrected. Gemini pairs by name (its wire format has no call id). functionResponse.response must be an object, so output rides { output, success }.
toOpenAIMessages	assistant tool_calls[] + role:"tool" messages paired by tool_call_id. Failures serialize as Error: <output> per design doc §11.3 (OpenAI has no is_error); tool-calls-only turns carry content: null.

Stateless backends (gemini, openai) benefit immediately — they run the converter every turn, so prior tool activity (from any backend) now arrives as real function-call turns.

Warm-backend seed

renderHistorySeed (claude, pi seedFromHistory) upgrades from the one-line flattening to structured blocks — tool name + status, full input JSON, fenced output (2k per-tool cap inside the existing 24k oldest-dropped budget). Investigated the higher-fidelity option: neither the Claude Agent SDK nor pi RPC accepts synthesized native-history injection (the SDK resumes only from its own persisted session files, an internal format), so the string seed remains the warm-backend ceiling. That contract is now stated in docs/providers-pi.md and the session.set_provider protocol comment.

Honest limits (documented in canonical.ts)

• thinking is captured but not replayed into any provider payload: Anthropic requires a cryptographic signature on replayed thinking blocks (synthesized ones are rejected — and the API ignores prior-turn thinking anyway); Gemini/OpenAI don't accept imported reasoning. Display-only by design.
• A CanonicalTurn flattens an agent loop (text → tool → text …), so converters emit the parallel-tool-call shape — valid everywhere, but intra-turn interleaving isn't reconstructed.

Tests

• Converter units: tool_use/tool_result id round-trip + is_error, Gemini functionCall/functionResponse on role function, OpenAI tool_calls + tool-role pairing + Error: prefix, tool-calls-only edge cases (no empty text block / content: null), thinking absent from Anthropic output.
• Seed: structured-block form (### Tool call: run_shell → ok, input JSON, fenced output), failure marking, per-tool truncation, old [Tool: form asserted gone.
• Round-trip: claude→pi switch over fake-pi asserts the first post-switch prompt carries the structured tool history.

bun run test (1198 pass), typecheck, lint all green. Wire-neutral — protocol change is comment-only.

🤖 Generated with Claude Code

[github-actions]

🤖 Gemini code review

This PR implements Phase 2 of the canonical history plan, rendering history in each provider's native structure (Anthropic tool_use/tool_result, Gemini functionCall/functionResponse, OpenAI tool_calls/tool-role messages) instead of flattened prose. However, there is a critical compatibility issue with the Anthropic converter where consecutive same-role messages are emitted, which will be rejected by the Anthropic Messages API.

Findings: 🔴 0 · 🟠 1 · 🟡 1 · 🟢 0

---

_{Tokens spent · ⬆️ Input: 10,152 · ⬇️ Output: 965 · Σ Total: 23,167
Total may be higher due to thinking token counts.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 81.61%. Comparing base (8cf7388) to head (140f133).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #138      +/-   ##
==========================================
+ Coverage   81.57%   81.61%   +0.03%     
==========================================
  Files          95       95              
  Lines       16406    16439      +33     
==========================================
+ Hits        13384    13417      +33     
  Misses       3022     3022              

Flag	Coverage Δ
daemon	81.61% <100.00%> (+0.03%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
packages/protocol/src/types.ts	100.00% <ø> (ø)
src/daemon/providers/canonical.ts	100.00% <100.00%> (ø)
src/daemon/providers/gemini/index.ts	100.00% <ø> (ø)
src/daemon/providers/openai/index.ts	100.00% <ø> (ø)

🚀 New features to boost your workflow:

• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.