[abrichr]

Stacked on #1299. Merge that first.

What this adds

openadapt_flow.integrations.harness2.traces writes the filesystem that Meta-Harness gives its proposer: every candidate's harness, plus each rollout's full execution trace and its reward. Meta-Harness reports that full traces beat summaries, 56.7 to 38.7.

• Full trace: every step is kept in full: the screen, the action, the agent's stated reason, and any error.
• Identity at each write: was the patient on screen the requested one (verified / mismatch / unreadable)?
• Read-back: the records added, whose they are, the outcome, the reward, and the receipt id.
• Two sources:
  • from_rollout reads an agent rollout from feat(flow): Harness² benchmark adapter scored by read-back #1299.
  • from_flow_run reads a compiled program's report.json and keeps the runtime's own identity-gate verdicts and effect checks. The repair search in the next PR reads halted runs this way.
• Layout: INDEX.md, index.jsonl (best reward first), candidates/<label>/harness/, and traces/<id>/trace.{md,json}.
• Privacy: screen text goes through privacy.scrub_text, the same as REPORT.md. Identity verdicts are computed before scrubbing, so they survive it.
• Wiring: run_task writes the export for every task. In readback mode, the identity verdict also appears in the trajectory that Harness² proposers read.

Tests

tests/test_harness2_traces.py covers:

• a wrong-patient write produces a mismatch, with a read-back showing one stray record;
• a right-patient write is verified;
• the export layout and its ordering;
• scrubbed text keeps the verdict;
• report.json keeps the runtime identity.

The adapter test now also checks the export and the identity line in the proposer's context. Both files pass locally (27 tests).

🤖 Generated with Claude Code