Join Codex and Claude Code logs into Agents run conversations

[VibeSWE]

Stacked on #33. Review that first; this PR's diff is against its branch.

Summary

#33 builds Codex and Claude Code runs from their traces. The prompts, responses and tool results are exported separately, as log records, so the Conversation tab stayed empty. Both harnesses stamp each log record with the trace and span it happened in, so the join can be automatic.

What changes

• Logs dataset pairing (lib/agentHarnessLogs.ts): when the selected traces dataset holds Codex or Claude Code traces, Agents scans the workspace once for logs datasets of either harness (Rescan datasets repeats the scan; Refresh doesn't). Schemas that return 403/404 are skipped; other failures are shown with Retry instead of reporting that no logs dataset exists. It recognizes Codex logs by conversation.id + originator, and Claude Code logs by prompt.id + event.sequence, each with trace_id and event.name. It picks the one named like the traces dataset (codex-traces → codex-logs). A Conversation logs selector shows the choice and can turn the join off (logs=none in the URL). The choice is cleared when switching traces datasets, and a logs= value that isn't a candidate for the current dataset is ignored. A traces dataset with both harnesses offers logs of either and joins the selected one with its own harness.

• Query: one extra query fetches only conversation events (event.name IN …) for the loaded runs' trace IDs, within the same time bounds and record limit, newest first, so the most recent runs keep their events. A note says when the limit was reached.

• Attachment, in the order things happened:

  	Codex	Claude Code
  User prompt	turn span: the turn whose recorded parent is the codex.turn_input span the prompt is logged on, else the turn containing it in time; ambiguous prompts stay unattached	interaction span
  Assistant response	not exported by Codex	its model call, by request_id
  Tool input	arguments on the tool span, by call_id within the trace	tool_input on the tool span, by tool_use_id within the trace
  Tool output	output on the tool span	the tool span's tool.output trace event (shown even without a logs dataset)
  Approval decision	tool span, by call_id	tool span, by tool_use_id

• Redacted content keeps its recorded value and gets a "(redacted)" label, so it's clear the exporter withheld it rather than the dashboard.

• Unloaded runs: records for runs outside the current view are ignored, never guessed. Codex's code-mode child tool calls run in a separate trace with no turn, so they're skipped; the outer exec result already includes their output.

• Conversation tab and search: the tab lists spans in time order, and run search now matches conversation text.

• Setup guide: says to export logs to their own dataset as well.

Evidence

Captured with content flags on for a probe run: codex exec -c otel.log_user_prompt=true, and Claude Code with OTEL_LOG_USER_PROMPTS, OTEL_LOG_ASSISTANT_RESPONSES, OTEL_LOG_TOOL_DETAILS and OTEL_LOG_TOOL_CONTENT. Running this branch's discovery, queries and attachment against the local codex-traces/codex-logs and claude-code-traces/claude-code-logs produced:

Codex
  [session_task.turn] User prompt: "Run: cat sample.txt  then reply with just the file contents."
  [exec] Tool arguments · exec: "const r = await tools.exec_command({cmd:\"cat sample.txt\", …
  [exec] Tool output · exec: "Script completed\nWall time 0.2 seconds\nOutput:\n\nhello\n"
Codex (earlier turn, prompt logging off)
  [session_task.turn] User prompt (redacted): "[REDACTED]"
Claude Code
  [claude_code.interaction] User prompt: "Run cat sample.txt with the Bash tool, then reply …"
  [claude_code.tool] Tool decision · Bash: "accept (config)"
  [claude_code.tool] Tool input · Bash: "{\"command\":\"cat sample.txt\", …}"
  [claude_code.tool] Tool output · Bash: "hello"
  [claude_code.llm_request] Assistant response: "hello"

Testing

• npm run format:check, tsc --noEmit, vite build
• npm test: 2,389 passed, 14 skipped. New agentHarnessLogs.test.ts, with fixtures trimmed from the real logs, covers detection, discovery (skip, unchecked and rethrow paths), pairing, quoted newest-first SQL, Codex/Claude attachment and ordering, per-turn prompt attribution, trace-scoped tool calls, repeated events, redaction labels, trace-only Claude tool output, and ignoring unloaded runs.
• Playwright on a Blacksmith testbox, agents.spec.ts + agent-setup.spec.ts: 32 passed. New specs: the paired logs dataset (selector default, Conversation tab, log query, search, turning logs off), switching traces datasets with logs set, Refresh without rediscovery and Rescan picking up a new logs dataset, the logs record-limit note, a failed logs scan with Retry, and a mixed traces dataset.
  • In local runs of all five Agents spec files with one worker, a few different tests fail each run, from "Target crashed" or a LogsPage dynamic import failing while Vite re-optimizes dependencies. origin/main fails the same way on this machine (2 failed / 42 passed, including a "Target crashed"), and the failing tests pass on rerun. I've treated this as local dev-server instability, not a regression; CI should confirm.

Limitations

• Codex doesn't export assistant reply text, so Codex conversations show the prompt and tool activity only.
• What appears depends on each CLI's content settings. With defaults, prompts and responses are redacted. Codex exports tool arguments and output even with log_user_prompt = false.
• Logs exported after the page loaded appear on refresh.

🤖 Generated with Claude Code

---

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

[comment-clanker]

Reviewed 64925c231ffdc0df6334cf653b57a960a11cb027 (the diff against #33's 0febb045) 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: generic GenAI datasets and demo mode send no extra requests, the SQL quoting is sound, stale responses can't overwrite newer ones, and a failed logs query leaves runs rendering. agentHarnessLogs.test.ts passes (8/8). The edge cases were reproduced with scratch scripts against the PR's functions; Playwright and live ingestion were not rerun.