Recognize native Codex and Claude Code telemetry in Agents

[VibeSWE]

Summary

Codex and Claude Code are the coding agents people most often point at Parseable, but both export OTLP traces with their own attribute names:

• Codex: the dataset never appeared in Agents. Its spans have no gen_ai.agent.*, gen_ai.operation.name, gen_ai.request.model or gen_ai.tool.name, so discovery skipped it.
• Claude Code: runs appeared because model-call spans carry gen_ai.request.model. Tool calls had no name, token usage showed as not recorded, and the run was named after the raw span (claude_code.interaction).

This PR teaches the Agents view to recognize both harnesses and read their native fields.

What changes

• Detection (lib/agentHarnesses.ts): a dataset is recognized from schema fields specific to each harness (codex.turn.token_usage.input_tokens / codex.turn.phase, interaction.sequence / llm_request.context), together with span_name. One dataset can hold both harnesses; each row is mapped by its own harness (Codex rows are those whose service.name starts with codex). Logs datasets are still not listed.
• Mapping: derives the GenAI fields the view already reads.
  • Codex: session_task.turn becomes the agent span, try_run_sampling_request the model calls, and spans with tool_name the tool calls. Per-request usage is lifted from handle_responses to its model call.
  • Claude Code: claude_code.interaction becomes the agent span, claude_code.llm_request the model calls (input usage = input_tokens + cache_read_tokens + cache_creation_tokens, left unknown when input_tokens is missing or malformed), and claude_code.tool the tool calls (tool_name).
  • Recorded values are never replaced, and absent usage stays unknown.
• Query: Codex traces a turn with about 475 internal spans (file reads, rollout writes…), which would use up the 1,000-record limit after about two runs. For Codex services the run query keeps only turn, model-call, stream, usage and tool spans, plus any span that already carries GenAI agent fields; rows from other services are untouched. The record-limit banner reflects the queried row count, not the rows left after mapping. Runs are seeded from Codex turns.
• Tree: Codex's intermediate spans aren't loaded, so model and tool calls are attached to their nearest loaded ancestor, or to the turn whose time range contains their start. A call that no single loaded turn contains (for example one from a turn that is still running) is left alone and keeps the missing-parent warning. Whichever parent field a row uses (span_parent_span_id or parent_span_id) is rewritten.
• Details panel: shows a "Mapped from native … telemetry" note on spans that had fields derived or their parent rewritten. Recorded fields and Parent span ID still show exactly what the harness stored.
• Setup guide and agent brief: no longer tell users to map these two harnesses by hand.

Evidence

The fixtures are trimmed from real exports into a local Parseable: Codex CLI 0.162.1 (app server and codex exec) and Claude Code 2.1.296 with CLAUDE_CODE_ENHANCED_TELEMETRY_BETA=1. Running this branch's query and mapping against that data produced:

Dataset	Before	After
codex-traces	not discovered	4 runs named Codex: turn → model calls + exec tool; tokens match Codex's own turn totals (e.g. 32,558 in / 38 out); 28 rows loaded instead of thousands
claude-code-traces	runs without tools or usage	Claude Code run: 2 model calls + Bash tool; 42,554 in / 154 out including cache tokens
codex-logs	not discovered	not discovered (logs aren't runs)

Testing

• npm run format:check, tsc --noEmit, vite build
• npm test: 2,350 passed, 14 skipped. agentHarnesses.test.ts covers detection (including mixed datasets), SQL (including that absent columns are never referenced and GenAI-field spans are kept), Codex usage lifting, time-range turn attachment across several turns, cross-service parents, both parent fields, Claude cache-inclusive and unknown usage, the mapping note, and recorded fields staying unchanged.
• Playwright agents.spec.ts + agent-setup.spec.ts on a Blacksmith testbox: 26 passed. New specs: native Codex traces → discovered dataset, run row, tree, mapping note and recorded parent ID; record-limit banner when Codex mapping drops queried rows. The other Agents specs (agents-navigation, agents-review-*) also passed; one test failed once on cold-start Vite re-optimization and passed on rerun.

Limitations

• Only traces are mapped. Codex/Claude Code logs (prompts, tool results) are not joined into the run conversation yet.
• Claude Code's standalone session-title request has no interaction span, so it shows as its own small run named claude_code.llm_request.
• Codex's nested code-mode tool executions are exported in a separate trace with no turn span. They don't appear as runs; the exec call inside the turn represents them.
• Field names are tied to the versions above; a harness that renames its spans would fall back to the current generic behaviour.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[comment-clanker]

Reviewed 0febb045707ee9be014ee8ab71cdf2f0850d1b51 with two GPT-6.1-Sol and two Claude Opus 5.5 reviewers, each covering a different area. Actionable findings are in the threads below.

What holds up: datasets with only generic OTel GenAI telemetry get byte-identical SQL and untouched rows. Codex usage lifting from handle_responses matches the turn totals, recorded values are never overwritten, "Recorded fields" and "Parent span ID" read the original rows, and the turn total isn't double-counted. agentHarnesses.test.ts passes. The edge cases were reproduced with scratch scripts against the PR's functions; Playwright and live ingestion were not rerun.