Second harness defect from the Vinci Worker soak gate (distinct from #5's no-progress loop guard). Observed live: task msg_9b20e235, 2026-08-27, vinci 0.0.50, openrouter/deepseek-v4-flash-0731, unattended vinci -p under the worker daemon.
What happened: the agent wrote the requested script, staged it, and then had git commit refused SIX times with: Vinci reserved the remaining actions for verification or the final answer. (One earlier, separate refusal: Vinci blocked a shell-based file write; use the structured write or edit tool instead. — that one is correct behavior.) The agent gave an honest final report; the run ended exit 0, work staged but uncommitted, evidence gate classified UNVERIFIED.
Defect: the action-budget reserve treats git commit as an ordinary expendable action. In worker mode the commit IS the deliverable — reserving the tail of the budget for 'verification or the final answer' makes task completion structurally impossible once the reserve is reached, and no user exists to grant more actions.
Suggested direction: in non-interactive mode, either exempt commit-shaped actions from the reserve, count the reserve against the envelope's own limits instead, or surface the reserve state as a BLOCKED outcome so the daemon classifies it honestly (same classification ask as #5).
Evidence: session JSONL sessions/msg_9b20e235/ on the worker box (6× refusal tool-results); soak ledger row 11. Both DeepSeek-lane failures today (07 = #5's latch, 11 = this reserve) are harness stops, not model failures — the same tasks completed fine on the GLM lane only because those runs used fewer actions.
🤖 Generated with Claude Code
Second harness defect from the Vinci Worker soak gate (distinct from #5's no-progress loop guard). Observed live: task msg_9b20e235, 2026-08-27, vinci 0.0.50, openrouter/deepseek-v4-flash-0731, unattended
vinci -punder the worker daemon.What happened: the agent wrote the requested script, staged it, and then had
git commitrefused SIX times with:Vinci reserved the remaining actions for verification or the final answer.(One earlier, separate refusal:Vinci blocked a shell-based file write; use the structured write or edit tool instead.— that one is correct behavior.) The agent gave an honest final report; the run ended exit 0, work staged but uncommitted, evidence gate classified UNVERIFIED.Defect: the action-budget reserve treats
git commitas an ordinary expendable action. In worker mode the commit IS the deliverable — reserving the tail of the budget for 'verification or the final answer' makes task completion structurally impossible once the reserve is reached, and no user exists to grant more actions.Suggested direction: in non-interactive mode, either exempt commit-shaped actions from the reserve, count the reserve against the envelope's own limits instead, or surface the reserve state as a BLOCKED outcome so the daemon classifies it honestly (same classification ask as #5).
Evidence: session JSONL sessions/msg_9b20e235/ on the worker box (6× refusal tool-results); soak ledger row 11. Both DeepSeek-lane failures today (07 = #5's latch, 11 = this reserve) are harness stops, not model failures — the same tasks completed fine on the GLM lane only because those runs used fewer actions.
🤖 Generated with Claude Code