commit 92bb7d00361aafbc40523b8a74a22f41d828b937 from: Isaac Meerleo date: Fri Jul 17 20:57:47 2026 UTC Complete Queue and Steer contract ticket commit - 0716373ef1fefafca2d72d1ba8a4be8189504a0b commit + 92bb7d00361aafbc40523b8a74a22f41d828b937 blob - 82c1a8f57f5c15803d56f7ff8330a75203da3161 blob + 475f401bf966efa87f71fac05c0bac94cf09808f --- .scratch/prompt-admission-semantics/issues/02-harden-queue-and-steer-contract.md +++ .scratch/prompt-admission-semantics/issues/02-harden-queue-and-steer-contract.md @@ -1,6 +1,7 @@ # Harden the Queue and Steer contract and regressions Status: ready-for-agent +Resolution: completed Type: task ## What to build @@ -61,24 +62,24 @@ post-stream boundary deterministically. ## Acceptance criteria -- [ ] `handoff/behavior.md` defines admission, promotion, dispatch, provisional +- [x] `handoff/behavior.md` defines admission, promotion, dispatch, provisional Journal append, and Turn commit as separate boundaries. -- [ ] The behavior contract says one ordinary Queue entry is removed per +- [x] The behavior contract says one ordinary Queue entry is removed per `UI_END` and completes its whole Turn before the next removal. -- [ ] The behavior contract defines ordered Steer batching, final-response +- [x] The behavior contract defines ordered Steer batching, final-response Steer priority, abandonment, and non-durability of undispatched Queue state. -- [ ] The 50-Tool-batch rule explicitly says Steer delivery resets the +- [x] The 50-Tool-batch rule explicitly says Steer delivery resets the canonical Turn counter zero times. -- [ ] `handoff/verification.md` names the corresponding fixtures and exact +- [x] `handoff/verification.md` names the corresponding fixtures and exact assertions. -- [ ] PTY coverage proves multiple-Steer ordering and committed replay. -- [ ] PTY coverage proves FIFO one-at-a-time ordinary Queue dispatch through a +- [x] PTY coverage proves multiple-Steer ordering and committed replay. +- [x] PTY coverage proves FIFO one-at-a-time ordinary Queue dispatch through a multi-Generation Turn and interleaved Steer priority. -- [ ] PTY coverage proves the final-response, failure/cancellation, +- [x] PTY coverage proves the final-response, failure/cancellation, session-switch, and resume/loss cases above. -- [ ] A pure Turn regression proves one or many Steers do not reset the Tool +- [x] A pure Turn regression proves one or many Steers do not reset the Tool allowance. -- [ ] Existing Queue, Steer, Turn, Journal, and full clean-room regressions +- [x] Existing Queue, Steer, Turn, Journal, and full clean-room regressions remain green. ## Non-goals @@ -107,3 +108,10 @@ None. The source comparison and policy decision are co 2026-07-17 research follow-up: OpenCode V2 was compared at pinned commit `0df2f6245a9cd966c0912e12db2c9d809e0c589f`. The investigation found contract and regression gaps but no confirmed Fugu runtime defect. + +2026-07-17 implementation: commits `bfd5adf` through `0716373` document the +five admission boundaries and add exact pure-Turn and staged PTY coverage for +ordered Steers, FIFO dispatch, final-response priority, abandonment, and +process-local loss. No production runtime change was needed. Focused suites +passed with 150 Turn-mechanics and 292 PTY checks; normal and `FUGU_DEBUG=1` +clean-room gates each passed all 37 suites and 12,501 checks.