Skip to content

docs(spec): composer inline completion and source-following menus; Tokenizer splits on caller-listed delimiters (opt-in) - #10133

Draft
cixzhang wants to merge 11 commits into
mainfrom
spec/combobox-seams-delta
Draft

cixzhang wants to merge 11 commits into
mainfrom
spec/combobox-seams-delta

Conversation

@cixzhang

@cixzhang cixzhang commented Nov 5, 2026 •

Copy link
Copy Markdown
Contributor

Summary

People can't accept a suggested continuation in the chat composer. They don't see new commands that arrive while a / menu is open. When they paste a list of email addresses into a free-text Tokenizer, it becomes one broken token. spec:AST-075 and the merged useCombobox already cover editor-driven listboxes, inline result lists, and textarea suggestions. This draft covers the three gaps that remain.

Spec explainer (human view): kept internally with the owner review.

Owner decisions (draft until decided; parked, not yet queued)

  1. Inline completion (component:ChatComposerInput FR19, DEC-3). Recommended A: a composer seam for caller-supplied ghost text; Tab or ArrowRight accepts it, and it is exposed as a description. Alternatives: B, a one-row combobox; C, no seam.
  2. An open trigger menu follows its source (FR20, DEC-4). Recommended A: re-run the current query when the source changes, and keep the highlight on its item. Alternative: B, callers dispatch synthetic input.
  3. Delimiters and paste (component:Tokenizer FR29, DEC-5). This is the same question as Q6 on Cindy's selectors page (feat(Tokenizer): split delimited text into separate tokens on type and paste (#4175) #4329). Recommended A: an opt-in delimiters list splits typed and pasted text, because default splitting would change released behavior (spec:AST-017 FR5). Alternatives: B, split by default; C, Enter only.

Current-state impact

  • component:ChatComposerInput: FR19, FR20 (amended), the Public concepts row, ORD3, OQ1 (touch accept), two verification rows, and DEC-3 and DEC-4 with their Reference and Decider lines.
  • component:Tokenizer: FR29 (amended), Public concepts, the Contract at a glance rows, the Compatibility bullet, a verification row, and DEC-5 with its Reference and Decider lines.
  • No code changes.

Tokenizer FR29 and DEC-5 (delimiter splitting) are proposed and await the owner's decision; the existing owner rows do not cover them.

Writing rubric results

Run over component:ChatComposerInput and component:Tokenizer (this change's lines):

  • R1 narration: none.
  • R2 research in the record: none.
  • R3 evidence naming a change: none; no evidence, basis or verification cell names a pull request, issue or commit.
  • R4 rejected alternatives: one Rejected: line per alternative, inside DEC-3, DEC-4 and DEC-5.
  • R5 pull request or issue references remaining: none (searched # + digits, /pull/, /issues/); DEC-5's Decider reads pending.
  • R6 ruled open questions still listing options: none; OQ1 (touch accept) is unanswered.
  • R7 project status: FR19's verification cell states "FR19 gap: no touch or pointer accept (OQ1)". Tokenizer's existing "not yet one action group (FR27)" is unchanged text from main.

@vercel

vercel Bot commented Nov 5, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
astryx Ready Ready Preview Nov 21, 2026 3:33pm UTC

Request Review

@meta-cla meta-cla Bot added the CLA Signed This label is managed by the Meta Open Source bot. label Nov 5, 2026
@astracat-bot

astracat-bot Bot commented Nov 5, 2026 •

Copy link
Copy Markdown
Contributor

AI review status for this pull request.

Review status Updated
❓ Needs maintainer input (for maintainers only) Nov 21, 2026, 3:50 PM UTC

@github-actions github-actions Bot added the needs:spec-owner-review Current knowledge records await owner approval label Nov 5, 2026

@astracat-bot astracat-bot Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This pull request needs a maintainer check before review continues.

[Automated review]

…elta

# Conflicts:
#	docs/specs/components/Chat/ChatComposerInput.spec.md

@astracat-bot astracat-bot Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This pull request needs a maintainer check before review continues.

[Automated review]

…elta

# Conflicts:
#	docs/specs/components/Tokenizer/Tokenizer.spec.md
@github-actions

github-actions Bot commented Nov 8, 2026 •

Copy link
Copy Markdown
Contributor

PR Analysis Report

Current run: The current change has no visual or preview-producing scope. Visual checks are not applicable; preview links are not shown.


Generated by PR Enrichment workflow | View current CI run

@astracat-bot astracat-bot Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approval waits on an engineering owner's decision on three new requirements, and the records need the fixes below either way.

Owner decision: component:ChatComposerInput FR18 with DEC-3 (inline completion), FR19 with DEC-4 (an open trigger menu follows its source, for every trigger) and component:Tokenizer FR28 with DEC-5 (opt-in delimiters) add requirements that no current record states. An engineering owner records the decision by commenting /approve-spec <full head SHA> on the head that carries these fixes; the author may do so as an engineering owner.

Blocking findings:

  • FR18 makes ArrowRight accept the completion, but in right-to-left text ArrowRight at the end of the text moves the caret back, so the accept key should be the inline-end arrow resolved from the editor's direction (design:bidirectionality DR6, DR11).
  • FR18 gives no touch or pointer way to accept a completion, which a phone keyboard without Tab or arrow keys needs (architecture:interaction-modality INV4).
  • Miss: FR19 keeps the highlight during a source-change re-run while FR13 makes Enter send the draft during a pending search, so Enter on the highlighted row can send a half-typed /sum; say whether rows stay pickable and reconcile FR13.
  • FR19 does not say what counts as a source change, and SearchSource has no change signal, so a trigger list built during render would re-search on every parent render.
  • Miss: FR19 does not say where the highlight goes when its item leaves the rows.
  • FR28 leaves the delimiters type, duplicate matching, IME composition behavior and the onChange report for a paste that adds several tokens (one batched report) unstated.
  • Miss: FR28 does not say a multi-token paste is announced once, as spec:AST-086 FR6 requires for one user action.
  • Miss: FR28's "stopping at maxEntries" drops the remaining pasted values without saying whether they stay in the input or how the person learns of it.
  • Miss: delimiters without hasCreate silently does nothing, where spec:AST-056 FR5 asks for a development warning.
  • Both records keep text the new rows contradict: ChatComposerInput's non-goal on new visuals, "decide a new visual state", "Released default preserved: yes", ORD3 key order and Contract at a glance, and Tokenizer's Public concepts and presentation scope paragraph.
  • Miss: Tokenizer's Contract at a glance and Compatibility and migration also omit delimiters, FR28 and DEC-5.
  • FR18 defines a second ghost-text look beside the dictation interim text the composer already renders, so use one path and style or say why they differ.
  • FR18's verification row should prove the description at the accessibility tree, and FR18 should say that hearing it depends on the screen reader (spec:AST-009 FR3, FR8).
  • DEC-5's Decider line cites a pull request and a working page, and DEC-3 to DEC-5 narrate the change ("Today", "now"), against architecture:knowledge-contracts INV10.
  • The description has no ## Writing rubric results, which every record pull request reports before review.
  • The title says Tokenizer "splits on common delimiters", but FR28 splits only on caller-listed characters and is opt-in.

Not blocking: the merge from main adds braces to a loop in apps/sandbox/perf/scenes.perf.spec.ts while the description says "No code changes"; drop it here or mention it.

[Automated review]

…elta

# Conflicts:
#	docs/specs/components/Chat/ChatComposerInput.spec.md

@astracat-bot astracat-bot Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approval waits on an engineering owner's decision on three new requirements, and the records need the fixes below either way. None of the earlier findings is addressed yet; the merge from main only renumbered the composer rows to FR19 and FR20.

Owner decision: component:ChatComposerInput FR19 with DEC-3 (inline completion), FR20 with DEC-4 (an open trigger menu follows its source, for every trigger) and component:Tokenizer FR28 with DEC-5 (opt-in delimiters) add requirements that no current record states. An engineering owner records the decision by commenting /approve-spec <full head SHA> on the head that carries these fixes; the author may do so as an engineering owner.

Blocking findings:

  • FR19 makes ArrowRight accept the completion, but in right-to-left text ArrowRight at the end of the text moves the caret back, so the accept key should be the inline-end arrow resolved from the editor's direction (design:bidirectionality DR6, DR11).
  • FR19 gives no touch or pointer way to accept a completion, which a phone keyboard without Tab or arrow keys needs (architecture:interaction-modality INV4).
  • Miss: FR19 shows a completion at any caret position but accepts it only at the end of the text, so a completion shown mid-text is one no key can take; show it only at the end of the text or accept it at the caret.
  • Miss: FR20 keeps the highlight during a source-change re-run while FR13 makes Enter send the draft during a pending search, so Enter on the highlighted row can send a half-typed /sum; say whether rows stay pickable and reconcile FR13.
  • FR20 does not say what counts as a source change, and SearchSource has no change signal, so a trigger list built during render would re-search on every parent render.
  • Miss: FR20 does not say where the highlight goes when its item leaves the rows.
  • FR28 leaves the delimiters type, duplicate matching, IME composition behavior and the onChange report for a paste that adds several tokens (one batched report) unstated.
  • Miss: FR28 does not say whether a paste over selected text replaces it.
  • Miss: FR28 does not say a multi-token paste is announced once, as spec:AST-086 FR6 requires for one user action.
  • Miss: FR28's "stopping at maxEntries" drops the remaining pasted values without saying whether they stay in the input or how the person learns of it.
  • Miss: delimiters without hasCreate silently does nothing, where spec:AST-056 FR5 asks for a development warning.
  • Both records keep text the new rows contradict: ChatComposerInput's non-goal on new visuals, "decide a new visual state", "Released default preserved: yes", ORD3 key order and Contract at a glance, and Tokenizer's Public concepts and presentation scope paragraph.
  • Miss: Tokenizer's Contract at a glance and Compatibility and migration also omit delimiters, FR28 and DEC-5.
  • FR19 defines a second ghost-text look beside the dictation interim text the composer already renders, so use one path and style or say why they differ.
  • FR19's verification row should prove the description at the accessibility tree, and FR19 should say that hearing it depends on the screen reader (spec:AST-009 FR3, FR8).
  • DEC-5's Decider line cites a pull request and a working page, and DEC-3 to DEC-5 and FR28 narrate the change ("Today", "now", "as today"), against architecture:knowledge-contracts INV10.
  • The description has no ## Writing rubric results, which every record pull request reports before review.
  • The title says Tokenizer "splits on common delimiters", but FR28 splits only on caller-listed characters and is opt-in.

Not blocking: apps/sandbox/perf/scenes.perf.spec.ts gains loop braces while the description says "No code changes"; drop it here or mention it.

[Automated review]

…elta

# Conflicts:
#	docs/specs/components/Chat/ChatComposerInput.spec.md

@astracat-bot astracat-bot Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approval waits on an engineering owner's decision on three new requirements, and the records need the fixes below either way. None of the earlier findings is addressed yet; this head only merges main.

Owner decision: component:ChatComposerInput FR19 with DEC-3 (inline completion), FR20 with DEC-4 (an open trigger menu follows its source, for every trigger) and component:Tokenizer FR28 with DEC-5 (opt-in delimiters) add requirements that no current record states. An engineering owner records the decision by commenting /approve-spec <full head SHA> on the head that carries these fixes; the author may do so as an engineering owner.

Blocking findings:

  • FR19 makes ArrowRight accept the completion, but in right-to-left text ArrowRight at the end of the text moves the caret back, so the accept key should be the inline-end arrow resolved from the editor's direction (design:bidirectionality DR6, DR11).
  • FR19 gives no touch or pointer way to accept a completion, which a phone keyboard without Tab or arrow keys needs (architecture:interaction-modality INV4).
  • Miss: FR19 shows a completion at any caret position but accepts it only at the end of the text, so a completion shown mid-text is one no key can take; show it only at the end of the text or accept it at the caret.
  • Miss: FR20's re-run is a pending search, during which FR13 makes Enter send the draft and the menu shows its loading row, so FR20 must say whether rows stay shown and pickable during the re-run and reconcile FR13.
  • FR20 does not say what counts as a source change, and SearchSource has no change signal, so a trigger list built during render would re-search on every parent render.
  • Miss: FR20 does not say where the highlight goes when its item leaves the rows.
  • FR28 leaves the delimiters type, duplicate matching, IME composition behavior and the onChange report for a paste that adds several tokens (one batched report) unstated.
  • Miss: FR28 does not say whether a paste over selected text replaces it.
  • Miss: FR28 does not say a multi-token paste is announced once, as spec:AST-086 FR6 requires for one user action.
  • Miss: FR28's "stopping at maxEntries" drops the remaining pasted values without saying whether they stay in the input or how the person learns of it.
  • Miss: delimiters without hasCreate silently does nothing, where spec:AST-056 FR5 asks for a development warning.
  • Both records keep text the new rows contradict: ChatComposerInput's non-goal on new visuals, "decide a new visual state", "Released default preserved: yes", ORD3 key order and Contract at a glance, and Tokenizer's Public concepts and presentation scope paragraph.
  • Miss: Tokenizer's Contract at a glance and Compatibility and migration also omit delimiters, FR28 and DEC-5.
  • FR19 defines a second ghost-text look beside the dictation interim text the composer already renders, so use one path and style or say why they differ.
  • FR19's verification row should prove the description at the accessibility tree, and FR19 should say that hearing it depends on the screen reader (spec:AST-009 FR3, FR8).
  • DEC-5's Decider line cites a pull request and a working page, and DEC-3 to DEC-5 and FR28 narrate the change ("Today", "now", "as today"), against architecture:knowledge-contracts INV10.
  • The description has no ## Writing rubric results, which every record pull request reports before review.
  • The title says Tokenizer "splits on common delimiters", but FR28 splits only on caller-listed characters and is opt-in.

Not blocking: the new rows' basis cells say "Proposed, pending" where DEC-3, DEC-4 and DEC-5 would name the deciding fact; apps/sandbox/perf/scenes.perf.spec.ts gains loop braces while the description says "No code changes".

[Automated review]

@astracat-bot astracat-bot Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approval waits on an engineering owner's decision on three new requirements, and the records need the fixes below either way (the Tokenizer row is now FR29 after the merge from main).

Owner decision: component:ChatComposerInput FR19 with DEC-3 (inline completion), FR20 with DEC-4 (an open trigger menu follows its source, for every trigger) and component:Tokenizer FR29 with DEC-5 (opt-in delimiters) add requirements that no current record states. An engineering owner records the decision by commenting /approve-spec <full head SHA> on the head that carries these fixes; the author may do so as an engineering owner.

Blocking findings:

  • FR19 makes ArrowRight accept the completion, but in right-to-left text ArrowRight at the end of the text moves the caret back, so the accept key should be the inline-end arrow resolved from the editor's direction (design:bidirectionality DR6, DR11).
  • FR19 gives no touch or pointer way to accept a completion, which a phone keyboard without Tab or arrow keys needs (architecture:interaction-modality INV4).
  • Miss: FR19 shows a completion at any caret position but accepts it only at the end of the text, so a completion shown mid-text is one no key can take; show it only at the end of the text or accept it at the caret.
  • Miss: FR20's re-run is a pending search, during which FR13 makes Enter send the draft and the menu shows its loading row, so FR20 must say whether rows stay shown and pickable during the re-run and reconcile FR13.
  • FR20 does not say what counts as a source change, and SearchSource has no change signal, so a trigger list built during render would re-search on every parent render.
  • Miss: FR20 does not say where the highlight goes when its item leaves the rows.
  • Tokenizer FR29 leaves the delimiters type, duplicate matching, IME composition behavior and the onChange report for a paste that adds several tokens (one batched report) unstated.
  • Miss: FR29 does not say whether a paste over selected text replaces it.
  • Miss: FR29 does not say a multi-token paste is announced once, as spec:AST-086 FR6 requires for one user action.
  • Miss: FR29's "stopping at maxEntries" drops the remaining pasted values without saying whether they stay in the input or how the person learns of it.
  • Miss: delimiters without hasCreate silently does nothing, where spec:AST-056 FR5 allows one switch and asks for a development warning.
  • Both records keep text the new rows contradict: ChatComposerInput's non-goal on new visuals, "decide a new visual state", "Released default preserved: yes", ORD3 key order and Contract at a glance, and Tokenizer's Public concepts and the presentation scope text that never places FR29.
  • Miss: Tokenizer's Contract at a glance and Compatibility and migration also omit delimiters, FR29 and DEC-5.
  • FR19 defines a second ghost-text look beside the dictation interim text the composer already renders, so use one path and style or say why they differ.
  • FR19's verification row should prove the description at the accessibility tree, and FR19 should say that hearing it depends on the screen reader (spec:AST-009 FR3, FR8).
  • DEC-5's Decider line cites a pull request and a working page, and DEC-3 to DEC-5 and FR29 narrate the change ("Today", "now", "as today"), against architecture:knowledge-contracts INV10.
  • The description has no ## Writing rubric results, which every record pull request reports before review.
  • The title says Tokenizer "splits on common delimiters", but FR29 splits only on caller-listed characters and is opt-in.

Not blocking: the new rows' basis cells say "Proposed, pending" where DEC-3, DEC-4 and DEC-5 would name the deciding fact, and Tokenizer's FR29 verification row sits above FR1–FR3 instead of in ID order.

[Automated review]

…elta

# Conflicts:
#	docs/specs/components/Tokenizer/Tokenizer.spec.md

@astracat-bot astracat-bot Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approval waits on an engineering owner's decision on three new requirements, and the records need the fixes below either way.

Owner decision: component:ChatComposerInput FR19 with DEC-3 (inline completion), FR20 with DEC-4 (an open trigger menu follows its source, for every trigger) and component:Tokenizer FR29 with DEC-5 (opt-in delimiters) add requirements that no current record states. An engineering owner records the decision by commenting /approve-spec <full head SHA> on the head that carries these fixes; the author may do so as an engineering owner.

Blocking findings:

  • FR19 makes ArrowRight accept the completion, but in right-to-left text ArrowRight at the end of the text moves the caret back, so the accept key should be the inline-end arrow resolved from the editor's direction (design:bidirectionality DR6, DR11).
  • FR19 gives no touch or pointer way to accept a completion, which a phone keyboard without Tab or arrow keys needs (architecture:interaction-modality INV4).
  • Miss: FR19 shows a completion at any caret position but accepts it only at the end of the text, so a completion shown mid-text is one no key can take; show it only at the end of the text or accept it at the caret.
  • Miss: FR20's re-run is a pending search, during which FR13 makes Enter send the draft and the menu shows its loading row, so FR20 must say whether rows stay shown and pickable during the re-run and reconcile FR13.
  • FR20 does not say what counts as a source change, and SearchSource has no change signal, so a trigger list built during render would re-search on every parent render.
  • Miss: FR20 does not say where the highlight goes when its item leaves the rows.
  • Tokenizer FR29 leaves the delimiters type, duplicate matching, IME composition behavior and the onChange report for a paste that adds several tokens (one batched report) unstated.
  • Miss: FR29 does not say whether a paste over selected text replaces it.
  • Miss: FR29 does not say a multi-token paste is announced once, as spec:AST-086 FR6 requires for one user action.
  • Miss: FR29's "stopping at maxEntries" drops the remaining pasted values without saying whether they stay in the input or how the person learns of it.
  • Miss: delimiters without hasCreate silently does nothing, where spec:AST-056 FR5 allows one switch and asks for a development warning.
  • Both records keep text the new rows contradict: ChatComposerInput's non-goal on new visuals, "decide a new visual state", "Released default preserved: yes", ORD3 key order and Contract at a glance, and Tokenizer's Public concepts and the presentation scope text that never places FR29.
  • Miss: Tokenizer's Contract at a glance and Compatibility and migration also omit delimiters, FR29 and DEC-5.
  • FR19 defines a second ghost-text look beside the dictation interim text the composer already renders, so use one path and style or say why they differ.
  • FR19's verification row should prove the description at the accessibility tree, and FR19 should say that hearing it depends on the screen reader (spec:AST-009 FR3, FR8).
  • DEC-5's Decider line cites a pull request and a working page, and DEC-3 to DEC-5 and FR29 narrate the change ("Today", "now", "as today"), against architecture:knowledge-contracts INV10.
  • The description has no ## Writing rubric results, which every record pull request reports before review.
  • The title says Tokenizer "splits on common delimiters", but FR29 splits only on caller-listed characters and is opt-in.

Not blocking: the new rows' basis cells say "Proposed, pending" where DEC-3, DEC-4 and DEC-5 would name the deciding fact.

Automated review (maintainer link)

@cixzhang

Copy link
Copy Markdown
Contributor Author

/approve-spec 34cae66

@github-actions github-actions Bot removed the needs:spec-owner-review Current knowledge records await owner approval label Nov 21, 2026
@github-actions github-actions Bot added the needs:spec-owner-review Current knowledge records await owner approval label Nov 21, 2026
@cixzhang

Copy link
Copy Markdown
Contributor Author

/approve-spec ef4562f

@github-actions github-actions Bot removed the needs:spec-owner-review Current knowledge records await owner approval label Nov 21, 2026

@astracat-bot astracat-bot Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Changes requested: an engineering owner approved the three new requirements at this head, so the direction is settled, but the records need the fixes below.

Blocking findings:

  • FR19 makes ArrowRight accept the completion, but in right-to-left text ArrowRight at the end of the text moves the caret back, so the accept key should be the inline-end arrow resolved from the editor's direction (design:bidirectionality DR6, DR11).
  • FR19 gives no touch or pointer way to accept a completion, which a phone keyboard without Tab or arrow keys needs (architecture:interaction-modality INV4).
  • Miss: FR19 shows a completion at any caret position but accepts it only at the end of the text, so a completion shown mid-text is one no key can take; show it only at the end of the text or accept it at the caret.
  • Miss: FR20's re-run is a pending search, during which FR13 makes Enter send the draft and the menu shows its loading row, so FR20 must say whether rows stay shown and pickable during the re-run and reconcile FR13.
  • FR20 does not say what counts as a source change, and SearchSource has no change signal, so a trigger list built during render would re-search on every parent render.
  • Miss: FR20 does not say where the highlight goes when its item leaves the rows.
  • Tokenizer FR29 leaves the delimiters type, duplicate matching, IME composition behavior and the onChange report for a paste that adds several tokens (one batched report) unstated.
  • Miss: FR29 does not say whether a paste over selected text replaces it.
  • Miss: FR29 does not say a multi-token paste is announced once, as spec:AST-086 FR6 requires for one user action.
  • Miss: FR29's "stopping at maxEntries" drops the remaining pasted values without saying whether they stay in the input or how the person learns of it.
  • Miss: delimiters without hasCreate silently does nothing, a configuration spec:AST-056 FR5 forbids, so state a development warning or let delimiters alone enable creation.
  • Both records keep text the new rows contradict: ChatComposerInput's non-goal on new visuals, "decide a new visual state", "Released default preserved: yes", ORD3 key order and Contract at a glance, and Tokenizer's Public concepts and the presentation scope text that never places FR29.
  • Miss: Tokenizer's Contract at a glance and Compatibility and migration also omit delimiters, FR29 and DEC-5.
  • FR19 defines a second ghost-text look beside the dictation interim text the composer already renders, so use one path and style or say why they differ.
  • FR19's verification row should prove the description at the accessibility tree, and FR19 should say that hearing it depends on the screen reader (spec:AST-009 FR3, FR8).
  • DEC-5's Decider line cites a pull request and a working page, and DEC-3 to DEC-5 and FR29 narrate the change ("Today", "now", "as today"), against architecture:knowledge-contracts INV10.
  • The description has no ## Writing rubric results, which every record pull request reports before review.
  • The title says Tokenizer "splits on common delimiters", but FR29 splits only on caller-listed characters and is opt-in.

Not blocking: the new rows' basis and status cells say "Proposed, pending" where DEC-3, DEC-4 and DEC-5 would name the deciding fact. A head with the fixes needs a fresh /approve-spec <full head SHA> from an engineering owner, which the author may give.

Claims and files checked

Review mode: full; reason: recipe-changed (many rules)

  • Claim: the composer cannot accept a suggested continuation, an open command menu misses commands that arrive while it is open, and a pasted list becomes one token: verified; the composer has no completion seam, the menu searches only on input, and the token field creates a token only on Enter.
  • Claim: the existing owned-field combobox record and the shipped combobox hook already cover editor-driven lists and textarea suggestions: verified; that record is current and covers textarea and contenteditable fields.
  • Claim: each recommended option matches the requirement rows added: verified for the completion, source-follow and opt-in delimiter rows.
  • Claim: splitting by default would change released behavior, so delimiters are opt-in: verified against the compatibility record, which makes only additive options nonbreaking.
  • Claim: the base branch added its own rows first, and the impact list names the rows, verification rows and decisions added: verified.
  • Claim: no code changes: verified; only two Markdown records change.
  • Claim: no open pull request claims these lines: verified for the record lines; the related open delimiter pull request changes code and edits no record.
  • Title claim that Tokenizer splits on common delimiters: false (blocking finding above).
  • File docs/specs/components/Chat/ChatComposerInput.spec.md: the FR19 and FR20 findings, the unamended-text, ghost-text, description-evidence and narration findings, and the status note.
  • File docs/specs/components/Tokenizer/Tokenizer.spec.md: the FR29 findings, the unamended-text and narration findings, and the status note.

Automated review (maintainer link)

@cixzhang

Copy link
Copy Markdown
Contributor Author

/revoke-spec ef4562f

@github-actions github-actions Bot added the needs:spec-owner-review Current knowledge records await owner approval label Nov 21, 2026
@cixzhang cixzhang changed the title docs(spec): composer inline completion and live trigger menus; Tokenizer splits on common delimiters docs(spec): composer inline completion and source-following menus; Tokenizer splits on caller-listed delimiters (opt-in) Nov 21, 2026

@astracat-bot astracat-bot Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approval waits on an engineering owner's decision on three new requirements, and the records need the fixes below either way, though most earlier findings are fixed at this head.

Owner decision: component:ChatComposerInput FR19 with DEC-3 (inline completion), FR20 with DEC-4 (an open trigger menu follows its source, for every trigger) and component:Tokenizer FR29 with DEC-5 (opt-in delimiters) add requirements the current records do not state, and this head carries no approval since the earlier one was revoked. An engineering owner records the decision by commenting /approve-spec <full head SHA> on the head that carries the fixes; the author may do so if they are an engineering owner.

Blocking findings:

  • FR19 still shows the completion on a touch keyboard, which has neither Tab nor an arrow key, so phone users see a suggestion they cannot take (architecture:interaction-modality INV4); name a tap or pointer accept, or show the completion only where a key can take it, and close OQ1.
  • FR19's ghost text takes dictation's interim style, the disabled text colour at 70% opacity, which is about 1.7:1 on a white surface, below WCAG 1.4.3's 4.5:1, so name a style both ghost texts share that meets it.
  • FR20 is classed patch-compatible with "Migration decision: none" while it requires callers to keep each trigger's source stable, which a caller that builds an async source during render must change to keep Enter picking rows (spec:AST-017 FR1, architecture:public-component-api INV9), so classify it as breaking or make it opt-in, and name the requirement under Builder impact.
  • FR20's highlight fallback cites "the menu's rule for a new result list (DEC-4)", but DEC-4 and the rest of the record state no such rule, so state the fallback in FR20.
  • FR29 keeps values past maxEntries in the input "so the person sees what was not added", but Tokenizer FR1 lets the field hide that input at the limit and the one announcement names only what was added, so say how the person learns what was left out, and join leftovers with a character a single-line input keeps.
  • The description's owner-decision list still says Tab or ArrowRight accepts the completion, and its R7 rubric line places the FR19 gap in a verification cell and omits FR29's "Proposed, pending" cells, so correct both lines.

Not blocking: FR29's basis and status cells still read "Proposed, pending", and the composer's impact list in the description omits four changed sections.

Claims and files checked

Review mode: full; reason: spec-record

  • Claim: the composer cannot accept a suggested continuation, an open command menu misses commands that arrive while it is open, and a pasted list becomes one token: verified; the composer has no completion seam, the menu searches only on input, and the token field creates a token only on Enter.
  • Claim: the existing owned-field combobox record and the shipped combobox hook already cover editor-driven lists, inline results and textarea suggestions: verified; the record is current and the hook implements it.
  • Claim: a human-view explainer is kept with the owner review: not checkable from the repository.
  • Claim: the recommended inline-completion option accepts with Tab or ArrowRight: false; the record now accepts with the inline-end arrow (blocking finding above).
  • Claim: the recommended source-following option re-runs the query and keeps the highlight on its item: verified.
  • Claim: delimiter splitting is the same question an open contribution raises, and splitting by default would change released behavior: verified; the open contribution splits by default, and the compatibility record makes only additive options nonbreaking.
  • Claim: the current-state impact lists name what each record changes: verified for Tokenizer; the composer list is true but omits four changed sections (not blocking).
  • Claim: no code changes: verified; only two Markdown records change.
  • Claim: Tokenizer FR29 and DEC-5 await the owner's decision: verified; its decider is pending and no approval covers this head.
  • Claim: writing rubric results R1 to R6: verified; R7: false in its location and incomplete (blocking finding above).
  • Title claim that Tokenizer splits on caller-listed delimiters, opt-in: verified.
  • File docs/specs/components/Chat/ChatComposerInput.spec.md: the FR19 touch and contrast findings, the FR20 compatibility and highlight findings.
  • File docs/specs/components/Tokenizer/Tokenizer.spec.md: the FR29 left-over values finding and the status note.

Automated review (maintainer link)

| FR16 | Escape on a visible menu dismisses it, and the menu stays hidden while the same query keeps growing from the same trigger. While open, the menu stays open for a press on the editor or in the menu and closes for a press elsewhere (`architecture:layer-runtime`). | DEC-1. |
| FR17 | The editor carries `enterkeyhint`: the `enterKeyHint` value when one is set, otherwise `send`, because Enter sends (FR2). `enterKeyHint` takes the closed union `EnterKeyHint` (`component:TextInput` FR16), so TypeScript rejects any other value. The hint only labels the key: FR2 and FR3 decide what Enter and Shift+Enter do whatever it says, so a composer whose `onKeyDown` makes Enter insert a line break passes `enter`. ChatComposer's default input carries the same default. | DEC-2; `component:TextInput` DEC-1 |
| FR18 | Backspace with the caret in or right after a token's trailing space, when no text follows that space, removes the token and that space in one step, and leaves the text as it was before the token was inserted, with no extra line break; an emptied last line keeps the browser's placeholder. | Current source and tests |
| FR19 | A caller may supply inline completion text. It shows only while the caret is collapsed at the end of the text and no trigger menu is open, as ghost text after the caret in the style the editor already uses for dictation's interim text; it is not part of the value. It is exposed to assistive technology as a description of the editor (`aria-describedby`), not as typed text; whether a person hears that description depends on the screen reader (`spec:AST-009` FR3, FR8). Tab, or the inline-end arrow key resolved from the editor's direction (ArrowRight in left-to-right text, ArrowLeft in right-to-left text; `design:bidirectionality` DR6, DR11), accepts it into the value; any other key, a caret move, or new text dismisses it. IME composition never shows or accepts it. FR19 gap: no touch or pointer accept (OQ1). | DEC-3 |

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A phone keyboard has neither Tab nor an arrow key, so this completion shows where nothing can take it, the case the FR19 verification row lists as a failure (architecture:interaction-modality INV4). Also, dictation's interim style is the disabled text colour at 70% opacity, about 1.7:1 on a white surface, below WCAG 1.4.3's 4.5:1. Suggest a tap or pointer accept (or no completion where no key can take it) and a ghost-text style both paths share that meets 4.5:1.

- Released default preserved: yes
- Compatibility class: patch-compatible; one additive optional trigger field, one additive optional prop, `enterKeyHint`, whose `send` default is a touch-keyboard hint that leaves what Enter does unchanged (patch `[feat]`, `spec:AST-017` FR5, FR7; `enterKeyHint="enter"` keeps the plain Return key), and trigger-key corrections for every trigger
- Released default preserved: yes, except FR20: an open trigger menu re-runs its query when its trigger's source changes, where it kept its rows before
- Compatibility class: patch-compatible; one additive optional trigger field, one additive optional prop, `enterKeyHint`, whose `send` default is a touch-keyboard hint that leaves what Enter does unchanged (patch `[feat]`, `spec:AST-017` FR5, FR7; `enterKeyHint="enter"` keeps the plain Return key), and trigger-key corrections for every trigger; FR19's inline completion is an additive optional input, and FR20 is a released behavior change for triggers whose source changes while their menu is open (`architecture:public-component-api` INV9), with no migration needed beyond keeping a source stable across renders

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

"No migration needed beyond keeping a source stable across renders" is a migration: a caller that builds an async source during render must change it to keep Enter picking rows, which spec:AST-017 FR1 counts as breaking (architecture:public-component-api INV9). Suggest classifying FR20 as breaking for a planned minor, or making it opt-in beside the old behavior, and naming the requirement under Builder impact and Migration decision.

| FR17 | The editor carries `enterkeyhint`: the `enterKeyHint` value when one is set, otherwise `send`, because Enter sends (FR2). `enterKeyHint` takes the closed union `EnterKeyHint` (`component:TextInput` FR16), so TypeScript rejects any other value. The hint only labels the key: FR2 and FR3 decide what Enter and Shift+Enter do whatever it says, so a composer whose `onKeyDown` makes Enter insert a line break passes `enter`. ChatComposer's default input carries the same default. | DEC-2; `component:TextInput` DEC-1 |
| FR18 | Backspace with the caret in or right after a token's trailing space, when no text follows that space, removes the token and that space in one step, and leaves the text as it was before the token was inserted, with no extra line break; an emptied last line keeps the browser's placeholder. | Current source and tests |
| FR19 | A caller may supply inline completion text. It shows only while the caret is collapsed at the end of the text and no trigger menu is open, as ghost text after the caret in the style the editor already uses for dictation's interim text; it is not part of the value. It is exposed to assistive technology as a description of the editor (`aria-describedby`), not as typed text; whether a person hears that description depends on the screen reader (`spec:AST-009` FR3, FR8). Tab, or the inline-end arrow key resolved from the editor's direction (ArrowRight in left-to-right text, ArrowLeft in right-to-left text; `design:bidirectionality` DR6, DR11), accepts it into the value; any other key, a caret move, or new text dismisses it. IME composition never shows or accepts it. FR19 gap: no touch or pointer accept (OQ1). | DEC-3 |
| FR20 | While a trigger menu is open, a change to that trigger's search source re-runs the current query and updates the open menu's rows without the person typing. A source change is a change in the identity of the source the trigger definition passes, so a caller keeps that source stable across renders; a source rebuilt on every render re-runs the query on every render. The re-run is a pending search for the active query, so FR13 governs it until its results arrive. The highlighted row stays on the same item when it is still present; otherwise the highlight follows the menu's rule for a new result list (DEC-4). | DEC-4 |

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

DEC-4 states no rule for a new result list, and no other clause mentions the highlight. Suggest stating the fallback here, for example "otherwise the first row", which is what the menu does today.

| FR26 | One open sheet edits one draft. The draft starts as `value` when the sheet opens and takes every row toggle, create, and Clear all (FR19, FR22, FR24); `value` and `onChange` are untouched while the sheet is open. Save closes the sheet; with an unchanged draft it reports nothing, and with a changed one it reports the whole draft once through `onChange(items, change)`, where `change` names the last edit that still holds: the last toggle, create, or clear when its item is in the draft and was not selected at open (`add`, or `create` for an item created in this sheet) or was selected at open and is not in the draft (`remove`); otherwise, when the last edit was a removal and some item selected at open is gone, a `remove` naming one such item; otherwise an `add` (`create` when the last edit created that item) naming an item in the draft that was not selected at open; otherwise a `remove` naming an item gone since open. A field that is disabled when Save fires reports nothing. A toggle is announced when it happens (AR6), not again on Save. | DEC-4 | Verified current behavior. |
| FR27 | Swipe, scrim, Escape, and platform close request dismissal of the sheet close it directly while the draft equals the selection at open. While the draft differs, they open a discard confirmation inside the sheet — "Discard changes?", "Your changes to {label} haven't been saved." — with Discard changes as the destructive action and Keep editing as the safe one, which is focused first; Discard changes closes the sheet without reporting the draft, and Keep editing closes the confirmation and moves focus to the search field so editing resumes where the sheet's own reading order starts (`spec:AST-085/FR4`). While the discard confirmation is open, the sheet waits at its peek (`component:BottomSheet` FR21–FR24): after Keep editing it returns to its detent, and after Discard changes it leaves from the peek. The sheet's confirmations (this one and Clear all's, FR24) are one modal dialog shape: a title, a description, and two full-width large actions stacked in reading order — the destructive action, then the safe action, which takes focus first; the two actions are one action group (`family:action-groups`) while keeping their separated, full-width, stacked look; scrim and Escape take the safe action. A close the field forces (becoming disabled, FR17) asks nothing and reports nothing: the draft is dropped. | DEC-4; `component:BottomSheet` DEC-5 | Verified, with two implementation gaps: the sheet springs back under the confirmation rather than waiting at the peek; its two actions are not one action group. |
| FR28 | In the inline presentation, `menu`, `onOpenChange` and the handle's `open()` and `close()` reach BaseTypeahead, so the menu's listbox and dialog roles, its open handle and its focus return follow `component:BaseTypeahead` FR13–FR16. A token press that calls `open()` with a dialog-role menu opens the dialog, and Escape returns focus to that token: focus a press puts on a token's own control, and the menu's focus return, stay there instead of moving to the input. The sheet presentation keeps its own sheet: it ignores `open()` and `close()`, and `menu` or `onOpenChange` draws its development warning. | `component:BaseTypeahead` FR13–FR16 | Verified current behavior |
| FR29 | With `hasCreate` and `delimiters` (a list of single characters), typing a listed delimiter after text, outside IME composition, creates a token from the text before it; a delimiter typed during composition stays text until composition ends. Pasting text that contains listed delimiters replaces any selected text in the input and creates one token per non-empty trimmed value, in order, skipping a value whose label matches a selected token or an earlier value in the same paste under FR22's case-insensitive matching. Values past `maxEntries` are not created; they stay in the input as text, joined by the first listed delimiter, so the person sees what was not added. One paste reports once through `onChange` with every new token, its `change` naming the last token added, and is announced once, as one message that names how many tokens were added (`spec:AST-086` FR6). In the sheet presentation the same values join the draft through FR22's create path. `delimiters` without `hasCreate` creates nothing and draws one development warning (`spec:AST-056` FR5). Without `delimiters`, only Enter creates a token (DEC-5). | Proposed, pending (DEC-5) | Proposed, pending |

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

At maxEntries FR1 lets the field make its input zero-width and transparent, and it does, so the left-over values are not seen, and the one announcement names only what was added. A line-break delimiter listed first also runs the leftovers together in a single-line input. Suggest saying how the person learns what was left out and joining with a character the input keeps.

This branch was successfully deployed

1 active deployment
Preview — a4299dad Deployed Nov 21, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Meta Open Source bot. needs:spec-owner-review Current knowledge records await owner approval

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant