feat(run_code): pass verifier records to the script as input["verifiers"]

[tejas-latupalli]

What and why

A run_code script receives {"args", "results"}, where results is context.metadata["script_results"]. Verifier steps record their outcome under context.metadata["verifications"], which a script can't see. A pipeline that wants to act on a verifier's findings, such as repairing a task after a Lighthouse audit, could only do it if each verifier also copied its record into script_results.

RunCodeTaskStep._build_input now also passes "verifiers": context.metadata.get("verifications", {}), and the module docstring documents the new key. Only the script input key is named verifiers; the context metadata key stays verifications. This change is additive: existing scripts that read only args and results behave the same. The new key holds an empty dict when no verifier has run.

The PR also fixes a race in tst/unit/store/mongo_document_store_test.py::test_a_unique_index_is_awaited_however_long_it_builds, which failed CI on this PR (assert 0.2989 >= 0.3). The test started the 0.3s release timer before reading time.monotonic(), so on a slow runner the measured wait could come in just under 0.3s. The test now reads the clock first.

Related: FD-3919. This replaces scaleapi/scaleapi#167303, which mirrored the Lighthouse audit record into script_results. Once this releases, the SDK only needs its agentenv-framework floor raised.

No size cap is added for the staged input.json. The verifier records already live in the persisted context, so they are JSON-serializable and bounded by what the context already carries.

How it was tested

• pytest tst/unit/task_step/test_run_code.py tst/unit/store/mongo_document_store_test.py: 87 passed. Two new tests use the real runner path (_real_run):
  • test_script_reads_earlier_verifier_records: the script reads inp["verifiers"]["audit"]["state"], and inp["results"] still comes through.
  • test_script_gets_empty_verifiers_when_no_verifier_ran: inp["verifiers"] == {}.
• I ran the fixed index-wait test 15 times in a row, and it passed every time.
• make unit-test on the first commit: 7657 passed, 7 skipped, 1 failed. The failure is tst/unit/store/local_grant_server_test.py::test_the_platform_default_is_served_on_loopback_too_where_a_vm_on_this_machine_reaches_it, with httpx.ConnectError: Connection reset by peer. It fails the same way on main without this change; my sandbox's network proxy blocks that loopback request.

🤖 Generated with Claude Code

https://claude.ai/code/session_01D6YRwGQbj6RUzhErtTpXVV

---

Generated by Claude Code

Confidence Score: 5/5

The PR appears safe to merge; no actionable issues remain.

Summary

run_code scripts now receive verifier records through input["verifiers"].

• Run-code scripts can read verifier records alongside their other inputs.
• The Mongo index test starts its release timer after its stopwatch.

_{Reviews (2) · Last reviewed commit: "fix(run_code): name the script input key..." · Reviewed by Greptile}

[agutmanstein-scale]

Breaks run_code design

[tejas-latupalli]

Closing in favor of the SDK-side fix (scaleapi/scaleapi#167303, reopened). The Lighthouse step will mirror its record into script_results, so run_code stays unchanged here.

---

Generated by Claude Code