fix(agentic): verdict extractor records NONE when the token is a markdown heading

[rickylabs]

Summary

The OpenHands verdict extractor records verdict: NONE when the agent's summary emits its verdict token as a markdown heading (## OPENHANDS_VERDICT: PASS) rather than a bare line. The evaluation is complete and unambiguous; only the machine-readable marker is wrong.

Reproduction

PLAN-EVAL rerun on PR #1558 (issue #1459), run 31593309658, model openrouter/minimax/minimax-m3.

The posted summary contains, as its final line:

## OPENHANDS_VERDICT: PASS

and its own body states Verdict: **PASS**. The comment's machine marker nevertheless reads:

<!-- openhands-run: {"run_id":31593309658,...,"verdict":"NONE","verdict_source":"none"} -->

and the rendered header prints OPENHANDS_VERDICT: NONE.

Why it matters

verdict is what automation and a supervisor key off. A PASS recorded as NONE:

• makes a completed evaluation look like it produced nothing, which is indistinguishable from a failed or hung run;
• pressures the supervisor into re-running an evaluator purely to obtain a parseable token, which is duplicate spend the phase-eval policy explicitly forbids;
• would, in an automated gate, block a PR whose evaluation actually passed.

The same class as #1561: the tooling's failure mode is to report nothing useful rather than to report the problem.

Acceptance criteria

• A verdict token is extracted whether it appears as a bare line or with leading markdown heading markers / emphasis
• When extraction genuinely finds no token, the summary says so explicitly and distinguishes "agent emitted no verdict" from "verdict could not be parsed"
• A test covers the heading-prefixed form, since that is how at least one approved evaluator model formats it

Provenance

0.0.6 runtime lane reopen, 2026-08-12. PR #1558 run 31593309658. The supervisor accepted the body verdict rather than re-running, and recorded the discrepancy instead.

[github-actions]

Acceptance evidence mirrored from #1644.

• "A verdict token is extracted whether it appears as a bare line or with leading markdown heading markers / emphasis" — Agentic/workflow wrapped-token fixtures; S2 Tier-A PASS comment fix(tooling): make evidence and verdict failures actionable #1644 (comment)
• "When extraction genuinely finds no token, the summary says so explicitly and distinguishes "agent emitted no verdict" from "verdict could not be parsed"" — parsed/absent/unparseable provenance fixtures in agentic-lib_test.ts and phase-eval-workflow_test.ts
• "A test covers the heading-prefixed form, since that is how at least one approved evaluator model formats it" — Heading-prefixed workflow fixture included in the 81/81 focused S2 receipt

Provenance: updatedAt=2026-08-14T23:18:14Z; bodySha256=14bef11490b61371b50df1a98c8db8c5562244a94c5562bee4aada2a793f1d91