feat: trace reasoning content and offered tools on generations

[Ali-Aleph-Alpha]

Change

Traces recorded the model's answer text but not its reasoning: thinking blocks were extracted for the history input but dropped from the generation output, so a Langfuse trace showed token counts for reasoning (output_reasoning_tokens) without the reasoning itself.

Generation output now carries the reasoning as Langfuse thinking blocks ({ type: "thinking", content }, the same ChatML shape the history input already uses, which Langfuse renders as collapsible reasoning blocks):

• Structured thinking blocks from pi's message content are included in the generation output.
• Reasoning streamed inline as <think> tags — the vLLM default and common on OpenAI-compatible servers that do not send a structured reasoning field — is split out of the answer text. An unclosed tag (aborted stream) is treated as reasoning until the end. Structured blocks win over inline tags so a model emitting both is not traced twice.
• Thinking content goes through the secret redactor: models regularly quote file contents (including .env) into their reasoning.
• thinkingSignature stays untraced (provider replay data; already covered by an existing test).

Time to first token now counts thinking tokens, which reasoning models stream before any text — previously TTFT was inflated by the full reasoning phase.

Each generation also records the tool definitions the model was called with (name, description, parameter schema) as available_tools metadata, and the turn root lists the active tool names, so a trace can distinguish "tool not offered" from "tool not chosen".

Notes

• The available_tools part overlaps with feat: attach the offered tool definitions to every generation #34 (which attaches tool definitions to the generation input for playground replay, closing Attach tool definitions to generations #23). Happy to drop it in favor of feat: attach the offered tool definitions to every generation #34 and keep only the reasoning changes.
• Verified end-to-end against a self-hosted Langfuse with a reasoning model served through an OpenAI-compatible proxy that streams <think> tags inline.

Test plan

• pnpm typecheck, pnpm test (111 tests: unit + integration against the real pi CLI with a mock provider and a fake Langfuse ingest)
• New coverage: think-tag splitting (closed/unclosed/multiple/redaction), structured-over-inline precedence, redaction of secrets in reasoning, readAvailableTools filtering and older-pi fallback, integration assertions for thinking in generation output and available_tools metadata

[CLAassistant]

All committers have signed the CLA.