traces-v2 / data

Commit History

  1. v2.19: ensure every tool def has a 'parameters' field (Lena follow-up). Cross-session lookup now prefers variants WITH parameters; final pass adds empty `{type:object, properties:{}}` for tools that genuinely have no args (e.g., 'finish') where no sibling captured a real schema. ~1.99M tool defs upgraded. Regression suite now checks this.

    Verified signature
    3968af0

  1. v2.18: normalize all '[]' input/output.messages → null (matches v1 convention). Covers (a) the 45 output.messages '[]' from v2.14's empty-text-part drop and (b) 201 pre-existing input.messages '[]' from v2 ingest. Consumers doing `out[0]` / `in[0]` access stay consistent with v1's null convention. Regression suite checks both fields.

    Verified signature
    cdec6cd
  2. v2.17: normalize '[]' output.messages → null on 45 spans. v2.14's empty-text-part drop left some output.messages as the JSON string '[]' when all messages were dropped; v1 + the rest of v2 use null for 'no output' (consumers' `out[0]` access crashes on []). Regression suite now checks for this so it can't recur.

    Verified signature
    9437cd5
  3. v2.16: backfill empty start_time/end_time on 2,863 chat spans (all gpt-5.2). The v2.8 'recover zero-call sessions from agent.log' commit reconstructed spans whose smolagents-style logs don't always emit per-step timestamps. Inherited the parent invoke_agent's start_time + monotonic millisecond offsets per child (preserves ordering). Closes Lena's build failures on 242 sessions. Regression suite now checks empty timestamps too — gap closed.

    Verified signature
    d22359f
  4. v2.15: cross-session tool-definitions recovery. Pre-pass builds (benchmark, harness, tool_name) → real-definition lookup over the corpus; sessions that had stub-only or no defs now inherit real schemas from sibling sessions in the same (bench, harness) combination. 100% of called-tool names in previously-stub sessions matched a real def in source. 1,245 stub sessions upgraded to real definitions (with parameter schemas). Lena 0 errors, regression bit-for-bit clean, smolagents untouched (synthesis path doesn't trigger for it).

    Verified signature
    6636f52

  1. v2.14: tool-message shape normalization + tool-defs propagation/synthesis + empty-part drop. (1) Convert {role:tool, parts:[{type:text}]} → tool_call_response: real id from preceding assistant.tool_call when available, positional fallback `call_<msg>_<part>` otherwise (drops to 0 orphan ids / 2.47M tcr parts). (2) Propagate gen_ai.tool.definitions across sibling chat spans (+94k); synthesize from real called-tool names for sessions where defs never existed. (3) Drop empty-content text parts. Lena 0/241k, tracelab 1,378 (all Issue #13 stragglers, same as v2.13), Yoav 0 model mismatches. Bit-for-bit regression-clean.

    Verified signature
    18121ab
  2. v2.13: backfill gen_ai.provider.name from canonical model (+58k spans) per OTel semconv requirement. Keeps v2.12 typing + canonical models + recomputed top-level. Verified bit-for-bit against v3-struct-recovered (span/row/id/token sums identical, only delta is the intended provider backfill).

    Verified signature
    6b92784

  1. v2.12: native typing (int tokens, list finish_reasons, float temperature) + canonical model names (no routing-prefix variation) + recomputed top-level models from chat spans. Addresses Yoav audit.

    Verified signature
    050c884

  1. Recover 500 zero-call sessions / 5147 turns from agent.log via Phase 2-4 parsers

    Verified signature
    16922fa

  1. Revert experimental retag — preserves all agent data; consumer-side fix expected (check op=='chat' before checking input.messages key)

    Verified signature
    8b41405

  1. Conservative retag: only tag chats with null input.messages as missed sim/judge (preserves legitimate alt-model agent chats)

    Verified signature
    3bca2e1
  2. Re-tag missed user-simulator chats via model-mismatch detection; filter then catches them automatically

    Verified signature
    67f59b4
  3. Schema: v1-style native struct attributes + status struct; OpenAI envelope unwrapped; data_page_size=4MB

    Verified signature
    fc9e394
  4. Initial: full corpus (typed list<struct> spans, optimized for load_dataset)

    Verified signature
    6a19522