Repository navigation

Commit c33e1aa

Browse files

yiliang114

and

qwencoder

authored

fix(cli): keep one-shot system-reminder prefixes out of shell mode (#12605)

* fix(cli): keep one-shot system-reminder prefixes out of shell mode Fixes #11626. Two gaps let `<system-reminder>` text reach bash or misroute queued input around shell mode: - handleFinalSubmit prepended the recovered-agents and the worktree-restore one-shot notices even in shell mode, where the leading `<system-reminder>` is a bash syntax error — and consuming the one-shot latch there dropped the notice before any model turn could see it. Both notices are now gated on `!shellModeActive` and stay armed for the next model-bound prompt. (The workflow-keyword reminder was already shell-gated.) - The queue drain routed submissions on the live shell-mode flag at drain time, so a flip while an entry waited in the queue could send a shell command to the model or a model prompt to bash. Shell intent is now recorded at submit time (addMessage/restoreMessages), travels with the queued entry, and the drain/submitQuery route on `shellModeIntent ?? shellModeActive`; batches stay intent-homogeneous so one routing decision never misroutes another entry. Producers that record no intent keep routing on the live flag, as before. Tests: packages/cli src/ui/AppContainer.test.tsx, hooks/use-llm-stream.test.tsx, hooks/useMessageQueue.test.ts — the new cases are red on the unfixed source (prefix reached shell-mode submissions; recorded intent ignored at drain) and green with the fix (544 passed, 0 failed). Co-authored-by: Qwen-Coder <[email protected]> Patrol-Run: qwen-issue-patrol/jmuf0dfzl0a * fix(cli): batch queued submissions by the head's intent run popNextSubmission selected every queue entry matching the first plain entry's recorded shell intent, so a later same-intent entry overtook an earlier different-intent one. Queueing a model prompt, a shell command and a second model prompt while a turn was responding drained both prompts as one batch and left the command queued behind them: the model answered the second prompt against the pre-command tree, and the command then ran with no prompt consuming its output. Take the contiguous same-intent run from the head instead. That keeps per-entry routing and submission order together; the only cost is one extra model turn when a shell command sits between two prompts, which no longer merge. batch[0] is still plainMessages[0], so the batch turnKey and the key peekNextUserBatchKey reserves for the Goal runtime are unchanged. The case that pinned the old reorder now asserts order preservation, and a new case pins that adjacent same-intent entries still aggregate. Review finding: R1-5 on PR 12605. Co-authored-by: Qwen-Coder <[email protected]> Patrol-Run: qwen-pr-closeout/jmufqt39629 * test(cli): pin the drain's shell-intent wire, correct stale gate comments The submitQuery metadata spread in the queued-submission drain is the only production hop that carries a queue entry's recorded shell intent to the router, and nothing pinned it: deleting those three lines left all three suites green, because the use-llm-stream cases inject the intent into submitQuery by hand and the existing drain assertions use objectContaining on a different key, which cannot fail on a missing one. Extend the existing persistent-admission-failure drain case instead of adding a second harness: its submission now records shellMode, the new assertion pins the metadata key itself, and the restore leg is asserted on intent rather than arity (third argument stays true, as that leg defers). Also correct two comments in this file that this change falsified: they said the other notices handleFinalSubmit prepends do not check shell mode and that the gate is load-bearing only through the workflow reminder's call site. Every notice the handler prepends is now gated at its own call site, and the two #11626 cases below depend on the recovered-agents and worktree gates rather than on this one. Review findings: R1-2 and R1-1 on PR 12605. Co-authored-by: Qwen-Coder <[email protected]> Patrol-Run: qwen-pr-closeout/jmufqt39629 * test(cli): pin the deferred leg's shell-intent argument Ctrl+Q is not shell-gated (InputPrompt.tsx binds QUEUE_MESSAGE after the `if (!shellModeActive)` block closes), so a shell-mode submission can take the deferUntilIdle branch of handleFinalSubmit. That leg's addMessage call was only asserted with the 4th argument false, which is indistinguishable from the parameter's default: replacing shellModeActive with false at AppContainer.tsx:3253 kept the whole suite green. Add the shell-mode arm so both directions of the argument are pinned. Co-authored-by: Qwen-Coder <[email protected]> Patrol-Run: qwen-pr-closeout/jmufv3et42h --------- Co-authored-by: yiliang114 <[email protected]> Co-authored-by: Qwen-Coder <[email protected]>

1 parent 177c2f6 commit c33e1aaCopy full SHA for c33e1aa

6 files changed

packages/cli/src/ui

Comments

(0)

File tree

0 commit comments