[fnnzzz]

Summary

• agy's own --print-timeout watchdog exits 0 and writes ... with turn in progress; returning partial output to stderr when it fires mid-turn, regardless of whether anything was accomplished.
• Reproduced live against the real agy 1.2.0 binary. In production, agy-claude-opus-4-6-thinking and agy-gpt-oss-120b-medium each logged outcome:"success" after running ~486s (within seconds of the registry's --print-timeout 8m) with no file changed — the model had streamed some visible prose before the cutoff, and that non-blank stdout was enough for the existing empty-run guard to wave the run through.
• The failure_markers mechanism already exists for exactly this class (see kiro's "Monthly request limit reached"); the agy family's six registry entries never declared one. This adds failure_markers: ["with turn in progress; returning partial output"] to all twelve agy-* edit_exists/read_only transports.
• hasSubstantiveOutput still wins when real files changed before the cutoff, so a genuine partial edit is untouched — only an empty, cut-off turn is re-scored as a failure.
• No lib/dispatch.js changes — this is a registry-data-only fix using an existing, deliberate mechanism.

Test plan

• Added unit tests on declaredMarkerExitCode (marker + no files → failure; marker + real files → untouched)
• Added end-to-end runDispatch tests via a fake CLI reproducing agy's exact stderr line + partial stdout prose (with marker declared → failure; without → documents the pre-fix false success)
• node --test lib/*.test.js: 506 pass, 1 pre-existing skip, 0 fail
• agents.yaml re-parses cleanly (35 entries)
• Consensus gate: 3/4 reviewers responded (groq-gpt-oss-120b, qwencloud-qwen-flash-character, qwencloud-qwen-plus-character), all APPROVE, no critical issues raised

🤖 Generated with Claude Code