[Pepps233]

Summary

Build a deterministic, self-contained evidence document from each complete or partially recoverable invocation trace.

The materializer validates trace structure, resolves referenced session segments, and preserves source provenance without duplicating transcript content into trace records.

Reasoning

Invocation traces contain execution facts while session JSONL remains authoritative for messages, tool calls, tool results, usage, and compactions.

Keeping those ownership boundaries explicit avoids competing copies and lets missing or changed session data surface as diagnostics instead of silently producing unsupported conclusions.

The normalized model is runtime-neutral, so coding-agent v3 entry cursors and AgentHarness durable sequence cursors produce the same evidence shape without coupling either runtime to the other's internal types.

Developers gain a bounded, LLM-ready diagnostic document that remains useful when an invocation or one of its referenced sessions is incomplete.

Changes

• Add the versioned MaterializedInvocation model and stable JSON serialization.
• Consolidate trace span starts, events, and settlements while preserving every contributing trace record ID and span ID.
• Stream coding-agent v3 and AgentHarness session files, verify cursor bounds and digests, and support relocated session paths.
• Preserve causal identifiers, session binding IDs, entry IDs, sequence numbers, timestamps, parent span IDs, and explicit ordering ambiguity.
• Report missing, changed, truncated, incompatible, writer-degraded, and crashed evidence as partial diagnostics.
• Add fixtures covering coding-agent transcript content and AgentHarness durable records.
• Document evidence ownership and exclude claim extraction, proposal aggregation, approval, and behavior mutation work.

Testing

• node ../../node_modules/vitest/dist/cli.js --run test/harness-trace.test.ts test/materializer.test.ts
• npm run check
• git diff --check feat/invocation-trace-lifecycle...HEAD