Repository navigation
fix(cli): fence in-flight retractions across session re-keys and switches - #6070
Merged
Merged
Conversation
Rewinding to a turn that carried quotes used to fail closed with rewind_unsupported_quotes, because the TUI could only refill the human-facing text and the replacement submit would silently drop the turn's structured context (apache#5109). The runtime-host driver now returns the rewound turn's QuoteRefs verbatim and forwards quotes given to submitMessage through turn.message.submit, whose admission already accepts them (only session-context attachments are Host-owned). Attachments and directory references still fail closed, since the TUI cannot re-attach files. The TUI stages the restored quotes keyed to the branched session: the status line carries a quotes:<n> segment while staging is live, bare /quotes lists the staged excerpts, /quotes clear discards them, and the first admitted submit consumes the staging while a refusal or failure restages it for the retry. Part of apache#5109 Generated-by: GLM-5.3-Flash (ZCode)
assert.ok the recorded submit before reading its content, per the noUnsafeOptionalChaining lint rule. Part of apache#5109 Generated-by: GLM-5.3-Flash (ZCode)
…nly rewinds Two review findings on apache#5265: A failed or blocked admission restaged the rewound quotes tagged with the Session read at callback time, so a Session switch while the admission was in flight attached the old quotes to the next message of the wrong conversation. The originating Session and staging generation are now captured at dispatch, and the restore happens only when neither moved. A quote-only rewind refills an empty prompt, and the empty-text guards in submitPrompt, steerRunningTurn and Alt+Enter rejected it before the quote forwarding path could run. They now treat staged quotes as meaningful content; a cleared plate stays truly empty and keeps refusing. Part of apache#5109 Generated-by: GLM-5.3-Flash (ZCode)
The restage guard captured stagedGeneration before the dispatch's own clearStagedQuotes(), which bumps the generation, so the guard compared against a pre-clear value and a blocked or failed admission never restored the staged quotes to the draft. Read the generation after the clear instead, and route every staged-quote write through one setter that bumps the generation, so a re-rewind landing while an admission is in flight is never overwritten by the older failure's restage (apache#5109 review). Generated-by: GLM-5.3-Flash (ZCode)
… transcript Third-round review items on the rewind quote staging: - A session change now clears the staged quotes outright instead of only hiding them while the user is elsewhere: keying alone let a silent resurrection re-arm the quotes on return, potentially many turns later. The rewind re-stages its own quotes after the switch settles. - /quotes clear distinguishes the nothing-staged case (including quotes that already left on an in-flight submit) instead of always claiming a discard. - A quote-only submit stored no text, so the replacement message left no trace in the transcript — an answer to an invisible prompt. The durable user entry now carries the restored-quote count and renders a trace line for it. - Coverage: the blocked-disposition restage, two-quote ordering across the status line, /quotes listing, and the submit, and /quotes routing mid-turn. Generated-by: GLM-5.3-Flash (ZCode)
# Conflicts: # packages/cli/src/session-driver.ts
The runtime-host PTY close-wait timeout fired on a merge head whose runtime-host tree is identical to green upstream; the PR's delta is confined to packages/cli. Local pi-tui + transcript suites pass on the merge head.
…the quote trace Two findings from the adversarial re-review at a127b13: - An outcome_unknown submit (or an interruption after the dispatch went out) resolves without a receipt, and the dispatch had already consumed the quote staging, so the quotes vanished silently with no restage and no test coverage. Restage them when the receipt is absent and surface a notice naming the uncertainty: admission cannot be proven either way, and losing the user's explicit context to an unproven outcome is worse than a visible duplicate ride (status line shows the restore; /quotes clear discards it). - The durable user-entry trace said "restored quote(s)" for every message carrying quotes, but StoredMessage.quotes also carries plain desktop quotes (including edit-restaged ones), so any quoted message resurfaced in the TUI mislabeled itself as rewind-restored. Word the trace neutrally ("· N quote(s)"). New UnknownOutcomeSubmitDriver pins the resolve-undefined path the previous tests only exercised through rejection. Generated-by: GLM-5.3-Flash (ZCode)
…talog The notice introduced for the unknown submit outcome was a visible literal in pi-tui-runner, which check:tui-copy correctly rejects — TUI copy must go through the localized catalog. Add quotesRestoredUnknown to the rewind catalog (en / zh-CN / zh-TW) and reference it. Generated-by: GLM-5.3-Flash (ZCode)
Generated-by: GLM-5.3-Flash (ZCode)
An ordinary submit that carries no quotes and an explicit /quotes clear with nothing staged leave the staging untouched, so neither advanced the generation — an in-flight submit's failure callback could then re-arm its QuoteRefs onto a conversation the user had already moved past. Advance the generation on both, making the guard airtight by construction rather than by the reader's memory of which writes bump it. Carries the apache#5109 review finding on the runner. Generated-by: GLM-5.3-Flash (ZCode)
Third-round review (apache#5109): a queued or steered submit carries the staged quotes, but taking the message back (Alt+↑ or Esc) lost them. The Host's `queue.retract` returns the full `MessageContent` — quotes included — while `retractQueued()` kept only the text and message ids, and `acceptRetraction` refilled only the text. A resubmit then went out without the quotes, and a quote-only message left an empty editor. `MakaRetractedMessages` now carries the retracted entries' quotes and `acceptRetraction` restages them on the current session through the generation-guarded setter, so the restoration obeys the same veto rules as every other staging write. Two runner tests cover one and several retracted messages. Also drop the unused `ShellRunSnapshotResult` import. Generated-by: GLM-5.3-Flash (ZCode)
The previous commit added the quotes field to MakaRetractedMessages but missed this deepStrictEqual expectation. Generated-by: GLM-5.3-Flash (ZCode)
A mid-turn `/session` can land while Alt+Up's `queue.retract` is still in flight (switchAwayMidTurn does not wait for it). The retraction response then restored its text and restaged its quotes under the session the driver had moved to, where the next submit could carry the abandoned context into the wrong conversation (apache#5109 review). Capture the owning session before the first await and discard the response when the driver has re-keyed; the switched-to session keeps its own empty staging. Covered by a delayed-retract/session-switch regression. Generated-by: GLM-5.3-Flash (ZCode)
…sion Two gaps in the retraction fence (apache#5109 review): - The session check ran after `retractQueued()`. A parked enqueue inside `settlePendingEnqueues()` let a mid-turn `/session` complete during the wait, so the Host call read the driver's *new* session and retracted its queued messages — which the mismatch fence then discarded silently. The fence now runs before the call: a moved-past retraction never touches the new session's queue. - The interrupt path still called `driver.stop()` unconditionally, so a Ctrl+C aimed at the abandoned session aborted the switched-to session's running turn. The stop now sits inside the fence; the switch flow owns stopping the session it left. Both regressions drive the retraction through a held gate while the session switch completes: no `retractQueued` call, and no stop on the landed-on session. Generated-by: GLM-5.3-Flash (ZCode)
Fifth-round review (apache#5109 review): the retraction fence discarded the retracted payload whenever the driver re-keyed mid-flight, and every switch entry point had its own gap — the idle `/session` path, the activity-lease wait inside `runControl`, the detach toggle, and Alt+Up's own `settlePendingEnqueues` wait all allowed a switch to land between the Host retraction and its response. Serialize instead: both retraction paths register on a retraction task set, and `switchSession`/`switchAwayMidTurn` await `settleRetractions()` at their entry. The retracted text and quotes therefore always land in the session they were asked for (via the existing acceptRetraction restore) before any switch re-keys the driver; the parked-enqueue regression now fires Alt+Up while the enqueue is held, so it exercises the real race, and a new regression pins the switch-behind-retraction ordering. Generated-by: GLM-5.3-Flash (ZCode)
Sixth-round review (apache#5265 review) flagged the idle `/session` path: the retraction guard ran before `runControl`, but `runControl` parks on `activities.acquire` before reaching the switch, and Alt+Up stays live during that wait - a retraction started there could land after the released lease let the switch re-key, and the switched-session fence would discard the text and quotes the Host had already taken back. The serialization at the switch boundary (`switchSession` awaits `settleRetractions()` before re-keying) already closes that race on this head and supersedes the reviewed branch's flag-based guard. Pin the reversed order with a regression: hold the session's activity lease, start the idle `/session` so `runControl` parks on the lease, fire Alt+Up while it waits, then release - the retracted text and quotes must reach the editor before the switch re-keys. With the boundary await removed the test fails (the switch re-keys before the retraction lands, `switch:...` then `retract-done:...`); with it, the four focused retraction tests pass individually. Generated-by: GLM-5.3-Flash (ZCode)
…window Seventh-round review (apache#5265 review) flagged that the idle switch's `settleRetractions()` drain only covers retractions asked for before the switch starts. `driver.switchSession()` is asynchronous - it stops user commands and opens the target Session channel before it re-keys - and Alt+Up stays live through that whole window. A retraction pressed there captures the old session id, and when the driver re-keys before the retract response arrives, the switched-session fence discards the returned text and quotes after the Host already removed the queued entries: a real loss. Count every runner-level switch (idle `/session`, mid-turn detach, and rewind's branch-and-switch) in a switch window and make `retractQueuedMessages` drop keypresses while one is open. Swallowing the keypress loses nothing: the entries stay queued on the Host and Alt+Up works once the switch lands, retracting from the session that is then current. The regression test parks the fake driver inside `switchSession`, before its re-key, and fails on the previous head with a `queue.retract` crossing the window (`retractCalls` 1 !== 0); it passes with the gate and fails again with the gate removed (ablation). The four previous focused retraction tests, the rewind in-flight tests, and the full runner suite (241 tests, only the two known pre-existing SIGTERM failures) pass individually. Generated-by: GLM-5.3-Flash (ZCode)
The side-conversation open and close re-key the driver through the same internal switchSession path as `/session` (Runtime Host openSideConversation switches onto the forked side Session, closeSideConversation switches back onto the parent), but neither entered the sessionSwitchesInFlight window, so Alt+Up stayed live across the re-key. A retraction asked inside that window removes the parent's queued entries via queue.retract while the driver is re-keying, and the response-time session fence then discards the returned text and quotes — the same silent input loss already fixed for `/session`, mid-turn detach, and rewind at the current head. Hold the same window around both switch-owning operations: the open wraps its adopt (both the idle runControl path and the mid-turn detach path) and the close wraps its whole body behind the existing closeSideConversation entry points. The window swallows the keypress without loss: the entries stay on the Host queue and Alt+Up works once the re-key lands. Regression tests park a gated fake driver inside the open's and close's re-key, fire Alt+Up there, and assert no queue.retract crosses (retractCalls stays 0) while the queued entry survives untouched. Both fail on the unfixed head (retractCalls 1 !== 0) and pass with the guard; removing only the two new holdSwitchWindow wrappers brings both back to red. Generated-by: GLM-5.3-Flash (ZCode)
holdSwitchWindow only blocks Alt+Up requests made inside a switch window; a retraction already in flight when /side, a side close, or a rewind started could still resolve after the driver re-keyed. The Host removes the queued entries as the retraction resolves, and the response-time session fence then discards the returned text and quotes - reachable by queuing a follow-up, pressing Alt+Up (retraction pending), and running the command before it settles; the reviewer reproduced the loss through real TUI key/command paths (apache#5265 review, P2). Copy the `/session` pattern (drain pendingRetractionTasks inside the held window, before the driver re-key starts) into the three remaining re-keying paths: the side open's adopt (idle runControl path and mid-turn detach path alike), the side close's runCloseSideConversation, and the rewind's branch-and-switch. Rewind had the same disease: every rewind re-keys onto a fresh branch Session, so a pending retraction crosses a session change and the branch fence discards its response exactly like the side fences. Tests park a gated retraction on the fake driver's Host call, start the open / close / rewind behind it, and assert the re-key waits: retract-done lands before the re-key and the retracted text reaches the editor. The rewind driver mints a fresh branch id per rewind so the second rewind crosses a session change the way the real driver does. All four tests fail on the unfixed runner (re-key observed while the retraction is still pending), pass with the fix, and fail again with only the three settleRetractions calls removed. Generated-by: GLM-5.3-Flash (ZCode)
Eighth-round review (apache#5265 review) confirmed the drains order the Host retraction before each re-key but flagged two inline P2s in what happens to the restored payload afterwards: - Side open: after the drain, acceptRetraction() has already put the parent's recovered text into the shared editor; the open then recorded it as parentDraft but never swapped the editor to the side view's own (empty) draft, so the parent's message sat in the side editor where Enter would resubmit it inside the side conversation. The open now switches the editor to the side draft after capturing parentDraft, per the established per-view draft contract. - Side close: the Ctrl+C close guard admits the close against an empty editor, and the drain then restores the side's queued text into it — which the following editor.setText(parentDraft) overwrote, losing the recovered message. The close now aborts with a notice when the editor is non-empty after the drain: closing is only admitted from an empty draft, so anything there came from the drain (or typing under it), and a later Ctrl+C clears it like any draft. Both behaviors were red under the reviewer's own reproductions (parent text visible in the new side editor; side text absent after close) and green with the guards; the three updated regressions also failed with only the guards removed. Full runner suite 245/247 - the two failures are the known pre-existing SIGTERM tests. check:tui-copy (new notice literal registered) and check:locale-hygiene pass. Generated-by: GLM-5.3-Flash (ZCode)
…m dying in the switch paths Ninth-round review (apache#5265 review) flagged three P2s and three P3s in what happens to a drained retraction's payload after the drain: - Quotes lost across sessions (P2): every switch path cleared the staging a drained retraction had just staged, and the Host already removed the queue entries the quotes came from - the TUI copy was the only one, so side open, /session, a quote-only side close, and a rewind all lost them. Retraction-restored quotes now ride the recovered draft text, wherever it goes: applySwitchResult still clears turn-staged quotes outright (the apache#5109 rule stands) but spares the draft-riding lane; the side conversation's draft slots carry quotes alongside their text (parked on toggle/open, staged back for the owning view on return); a rewind that branches without its own quotes keeps them; a rewind that carries its own quotes still replaces them (apache#5109 semantics, now visible and pinned by a test with distinct quotes per side). - Ctrl+/ lost the recovered text (P2): the toggle captured the draft before waiting a pending retraction out, then wrote the stale draft back over what the drain had restored - side editor, parent editor, and Host queue all empty with no notice. The capture moved after the switch's drain, like the side open. - The close window overwrote live typing (P2): a close admitted against an empty editor left input unlocked while its driver call was in flight, then setText(parentDraft) discarded whatever was typed during the wait. The live editor content now outranks the parked parent draft (kept, parent draft appended below when both exist). - Quote-only close (P3): the abort guard only looked at editor text, so a retraction returning quotes with no text still closed and dropped the staging; staged quotes now count as a keep-open reason. - Notice wording (P3): the kept-open notice no longer claims a retracted message for what may be the user's own typing - two neutral literals replace the old one (check:tui-copy registry updated). - Rewind test (P3): the drain-before-rewind test minted the same quote on both sides, so the designed replacement could not be distinguished from survival; each rewind now carries a distinct quote and the resubmit asserts the rewound turn's quote won. Every finding verified red under the reviewer's reproduction on the pre-fix runner and green after; new coverage asserts what the TUI actually submitted (driver.submittedQuotes) across the side round trip, the /session switch, and the quote-less rewind. Full pi-tui-runner suite: 252 pass / 2 known pre-existing SIGTERM failures. check:tui-copy, check:locale-hygiene (vs upstream/main), check:asf-headers, and biome on the touched files pass; merge-tree against upstream/main is clean. Generated-by: GLM-5.3-Flash (ZCode)
The last open P2 from the ninth-round review (apache#5265 review) is an ordering bug in the submit path's staging bookkeeping: - A submit captured the staged quotes' draft-following provenance AFTER its own clear: clearStagedQuotes() resets the flag, so for every quote-carrying submit the capture always read false. When the admission then failed (or resolved without a receipt), restageForRetry() rebuilt the staging as ordinary turn-staged quotes, and the next /session or side-view switch cleared them in applySwitchResult() - stranding the recovered text without its context, the exact loss the draft-following lane exists to prevent. The read now happens before the dispatch's clear, so a failed restage re-arms the quotes in the lane they were staged in and a later switch carries them with the draft. The generation guard keeps its post-clear read: the apache#5109 restage-fence semantics are untouched. New coverage pins the combination path no prior test reached: a retraction-restored quote submit whose admission fails, then a /session switch - the resubmit after the switch must still carry the quotes. Red under the reviewer's ordering on the pre-fix runner, green with the moved read, and red again with only the read moved back (ablation). Focused quote/retraction/switch/side pattern 47/47; full pi-tui-runner suite 253/255 with the two failures being the known pre-existing SIGTERM tests; biome on the touched files, check:tui-copy, and check:asf-headers pass. Generated-by: GLM-5.3-Flash (ZCode)
The mid-turn `/session` detach set `detaching` only after awaiting the retraction drain, while every other navigation guard (`goToSession`, the side toggle, a mid-turn side open) reads `detaching` to refuse concurrent switches. With a retraction held in flight, two `/session` commands (or a command and a toggle) could both pass their guard while waiting on the same drain and then call `driver.switchSession` concurrently: the Runtime Host driver does not serialize those calls, so the later navigation can be lost to the channel-generation fence, and concurrent side toggles can apply the draft handoff twice. Reserve the detach before the drain so the existing guard holds through adoption, and hoist the `goToSession` guard above the idle branch: the running Turn can end while the drain is pending, which makes the idle path reachable mid-drain. Both new regressions drive two navigations across one held retraction and assert a single `driver.switchSession`; both fail on the previous head with two concurrent re-keys. Generated-by: GLM-5.3-Flash (ZCode)
…rain window for /side and /new Main apache#5910 removed the Alt+Up retract chord and rebound retraction to Shift+Left, so on the merge ref a `\x1b[1;3A` send no longer retracts: this PR's retraction tests sat waiting for a refill that never came and the CI `test` job timed out. Point all 22 staged sends at `\x1b[1;2D` and rename the Alt+Up comments and test titles to Shift+Left, matching the binding apache#5910 left behind. Merging main also completes the drain-window picture the previous commit shored up for `/session`: with the Turn ended while a mid-turn detach is still parked inside its retraction drain, an idle `/side` — and `/new`, which bypasses the busy gate by design (apache#3210 review) — could pass unguarded and re-key the driver concurrently with the in-flight switch's own re-key. Return early on `detaching` at the top of `openSideConversation` and `newSession`. Both new regressions drive the second navigation across one held retraction and assert it never reaches the driver; both fail on the previous head with the concurrent re-key. Generated-by: GLM-5.3-Flash (ZCode)
apache#5265 mixed two themes: staging rewind/retraction quotes into the replacement submit, and retraction/switch fencing for pre-existing plain-text races. Keep only the fencing half here — settleRetractions, holdSwitchWindow, the detaching guards, the side-close abort, and the retraction-fenced interrupt — so each half can be reviewed and reverted on its own. The quote half (QuoteRef plumbing through the driver, protocol, and transcript, quote-only rewind, the /quotes command, and every test asserting on quotes) stays on apache#5265. The fencing tests move over rewritten against plain text only: their setup no longer stages quotes through /rewind, and the four "drains a retraction before X" and three "blocks Shift+Left while X" cases are each consolidated into one table-driven test.
/new is the only re-keying path outside the switch window: a Shift+Left retraction still in flight across /new was discarded after the Host had already removed the entries — the fresh-session fence drops any response landing past startNewSession's re-key (apache#5265 review). Wrap /new in the same holdSwitchWindow + settleRetractions pair /session uses: the pending retraction lands, its text reaching the editor, before the re-key, and retractions asked inside the window are blocked until it settles.
4 of 6 tasks
Astro-Han
reviewed
Oct 11, 2026
Astro-Han
left a comment
Astro-Han
left a comment
Contributor
There was a problem hiding this comment.
Automated review notice: This comment was posted by an automated review agent. It is not an independent human review and does not replace one.
I found no P0–P2 issue at 586759fc707879be75f70f893680f811f47b2e42.
The retraction drain runs before session adoption, draft capture, side-session closure, rewind, and /new. The switch window prevents a new Shift+Left retraction while a session is being re-keyed, and the origin-session checks prevent an interrupt from stopping a different session. Side-session closure preserves restored text and text typed while the drain is pending.
I ran four complete CLI test files: 463 passed, with three Windows-only skips and no failures. CLI compilation/type checking, formatting, and the copy checker passed. Removing only the new /new drain from the generated runner makes the pending-retraction test fail; restoring it passes. An additional held-/new probe confirms Shift+Left cannot start a retraction inside that switch window. These checks use the real runner with controlled terminal/driver fixtures, not a live provider or interactive desktop session. The broader root build has the inherited Desktop fixture type error described on #5265; I do not claim a clean full-repository build.
Two points remain:
- P3 — contribution disclosure: Please add the required explicit generative-tool usage declaration to the PR description. The history includes GLM-generated commits, but the current description does not state whether tools were used and which ones. Keep the applicable commit trailers as well.
- #5265 and this branch cannot be merged together unchanged: their exact heads conflict in
pi-tui-runner.ts, its tests, andcheck-tui-copy.mjs. After either lands, rebase the other and retain both quote and plain-text protections, then rerun the combined side-close, draft, retraction, and session-switch tests. I have not tested a conflict-resolved combination.
The merge with current main is clean and passes the epoch-217 guard. The exact-head CI run 38148780022 is successful.
Member
|
Reviewed exact head One thing before I approve: please add the AI-use declaration to the PR description, as CONTRIBUTING requires (the history includes GLM-generated commits), and keep the commit trailers. Merge order: I'd land this first, then rebase #5265 on it, keeping both its quote-only side-close guard and this branch's plain-text protections. |
Contributor
Author
|
@liugddx Added the AI-use declaration to the PR description (GLM-5.3-Flash (ZCode), under human direction and review), plus the template checklist. The |
liugddx
approved these changes
Oct 12, 2026
liugddx
left a comment
liugddx
left a comment
Member
There was a problem hiding this comment.
Approving exact head 586759fc7. Everything I checked holds:
- Scope: the diff is fencing-only. The
quotehits are re-indentation of existingrewind_unsupported_quoteshandling. /new: it drains in-flight retractions (runNewSessioncallssettleRetractions()beforestartNewSession) and holds the switch window like/session.- Disclosure: the AI-use declaration is now in the description, with the
Generated-by: GLM-5.3-Flash (ZCode)trailers kept. Please make sure the squash commit carries that trailer. - Verification: Astro-Han found no P0–P2 at this head, and his mutation (removing the
/newdrain fails the pending-retraction test) shows the new test guards it. CI is green, and the branch merges cleanly with currentmain(6b9457f5b).
After this lands, #5265 rebases on it. Your note there matters: because each branch is "merge base minus the other half", a plain rebase silently drops the QuoteRef pipeline files without conflict markers. When I review the rebased #5265, I'll check those five files are present and that the driver's rewind_unsupported_quotes path yields to the quote-side restage.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Split from #5265 per @liugddx's scope review: the retraction/switch fencing for pre-existing plain-text races lives on its own branch, so each half can be reviewed and reverted independently.
What's here
settleRetractions/trackRetractionand the retraction fences keyed on the owning sessionholdSwitchWindowacross session re-keys:/session, side open/close,/rewind, and now/newdetachingpre-guards ongoToSession/openSideConversation/newSessionrunSessionSwitch/runMidTurnSwitch/rewind/openSide/closeSide/newparks behind an in-flight retraction and holds the switch window the way/sessiondoes — the P3 from the fix(cli): stage a rewound turn's quotes into the replacement submit #5265 review of4d7764420(586759fc7)The quote-staging half stays in #5265; this branch removes it, so the net diff here is fencing-only (runner + its tests + one
check-tui-copy.mjswhitelist line). Roughly 19 plain-text race tests, none involving quotes. Per the review suggestion, the four "drains a retraction before X" tests and the three "blocks Shift+Left while X" tests are consolidated into two table-driven tests (assertion semantics unchanged).Relationship to #5265
check-tui-copy.mjswhitelist: each PR adds its own line.Verification
4d7764420the suite was 278 with 2 Windows-only SIGTERM fails noted by the reviewer; not reproduced locally — the 17 quote tests moved to fix(cli): stage a rewound turn's quotes into the replacement submit #5265, 9 tests consolidated into 2 table-driven ones, +1 new/newtest)biome checkon the changed files clean;check:asf-headerspass;check-tui-copyokdrains a retraction started before /new re-keysruns red before586759fc7(/newexecuted immediately: 1 !== 0) and green afterTwo test assertions were re-graded to plain-text equivalents when the quote half left:
discards a retraction after a mid-turn switchnow asserts the editor does not contain the restored text (instead of a status-line quote count), andserializes a mid-turn switchassertsretractedTexts.at(-1)instead of a quote payload. Happy to tighten either back up.cc @liugddx for the scope re-check.
AI use
Select exactly one:
Tool(s) and scope: GLM-5.3-Flash (ZCode) implemented the retraction/switch fencing changes and their tests under human direction and review. The affected commits carry
Generated-by: GLM-5.3-Flash (ZCode)trailers, and the final squash commit should keep that trailer.Checklist
Does this PR entail a change in behavior?
/session,/new, side open/close,/rewind), so their text is no longer lost or applied to the wrong session.