fix: launch the streaming speech helper with --fast (#5896)

[CrysisDeu]

Summary

With Apple Speech (on-device) streaming and "Auto-submit when I finish speaking" enabled, partial transcripts render live but the utterance is never auto-submitted — recording stays active until manually stopped (#5896).

Root cause

StreamingSession.start() built the streaming helper argv without --fast. The bundled StreamTranscribe.swift helper only inserts .frequentFinalization into the DictationTranscriber reporting options when --fast is passed; without it the transcriber emits only volatile partials while the audio stream is open and never produces mid-stream final segments. The dashboard's semantic endpointer (_Endpointer in stt_stream.py) schedules its utterance-complete judgment exclusively from note_final(), so with zero mid-stream finals auto-submit can never fire.

Fix

Append --fast to the streaming helper argv (a bare switch — verified against the helper's own parser). The one-shot batch path at transcribe() deliberately keeps omitting the flag: its final arrives when the stream closes, and frequent finalization can trade accuracy for latency it has no use for.

Frequent finals are safe for the consumers: the stt_stream relay forwards each final and calls note_final() per event, and the frontend already accumulates multiple finals per session (the AWS Transcribe streaming path emits them the same way).

Verification

This is a macOS-only code path with no CI runner coverage for the live helper, so the spawned argv is the verifiable contract:

• test_streaming_argv_includes_fast pins --fast in the streaming spawn argv (and that it did not swallow a neighbouring option's value).
• test_one_shot_argv_excludes_fast pins its absence on the batch path.
• test_swift_helper_still_accepts_fast pins the other side of the cross-process contract by source inspection: the Swift helper still parses case "--fast":, so a Swift-side rename cannot silently break every green Python test.
• Mutation-verified: removing the flag makes the streaming pin fail.

Behavioral evidence is the reporter's consistent repro on main 6ed29f5cb (partials live, no finals, no auto-submit); the non-streaming one-shot path was confirmed unaffected.

Local gates: black/subprocess-encoding/isort/flake8/mypy clean; full pytest 68281 passed (17 failures reproduce identically on unmodified main or pass in isolation — host-env, none reachable from this diff). Pre-push reviews: GPT lane PASS (0 findings), Opus lane PASS (0 blocking, 2 advisories: Swift-side pin adopted; a fast= constructor escape hatch declined as speculative configurability for a single-caller path).

Closes #5896

[github-actions]

Opus 4.8 Review — ✅ no blocking findings

Reviewed ae9f5e94fdebcf545db27a50234fa909c343598f — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] ae9f5e9

Verdict parsed from the review's SHA-scoped output markers for commit ae9f5e94fdebcf545db27a50234fa909c343598f.

False positive or not applicable? A repository writer can comment:
/ai-review override fable ae9f5e94fdebcf545db27a50234fa909c343598f: <one-sentence reason>

[github-actions]

Design Review (Fable 5) — ✅ PASS

Design-level review of ae9f5e94fdebcf545db27a50234fa909c343598f — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

All claims verified: the Swift helper's --fast case inserts .frequentFinalization (StreamTranscribe.swift:231), the dashboard endpointer schedules exclusively from note_final() (stt_stream.py), and the streaming argv previously omitted the flag — so the fix targets the actual root cause, is minimal, and the batch path's exclusion is deliberate and tested. No design-level issues.

Design-Verdict: PASS

Root-cause one-flag fix at the argv seam, verified against both sides of the cross-process contract; nothing to change.

[DESIGN-REVIEWED] ae9f5e9

[github-actions]

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of ae9f5e94fdebcf545db27a50234fa909c343598f and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] ae9f5e9

False positive or not applicable? A repository writer can comment:
/ai-review override gpt ae9f5e94fdebcf545db27a50234fa909c343598f: <one-sentence reason>

[github-actions]

First Principles Review (Fable 5) — ✅ PASS

Premise-level review of ae9f5e94fdebcf545db27a50234fa909c343598f — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

All claims verified. The change is a one-line argv fix plus three tests; the --fast flag pre-exists in StreamTranscribe.swift, the batch helper is a separate Swift source that never parses it, and the streaming helper has exactly one production spawn site. Final review:

First-Principles-Verdict: PASS

A reported defect (#5896), fixed at the mechanism — the argv that chose the wrong reporting options — with the flag's cross-process contract pinned.

What this change ships

Intent: make Apple on-device streaming dictation auto-submit when the speaker stops. FIX.

1. Auto-submit now fires on the Apple streaming path — justified (reported defect Apple Speech streaming endpointing never auto-submits — helper launched without --fast #5896).
2. Streaming transcripts now finalize mid-stream (frequent finals) — declared; this is the fix's mechanism, and the frontend already accumulates multiple finals.
3. One-shot batch transcription unchanged — declared, pinned by test.
4. Three tests pin both sides of the Python↔Swift argv contract — justified: macOS-only path with no CI runner, so the argv is the only testable contract.

No new config key, Python flag, or public surface; every item is declared in the description.

Subtractions

• case "--fast": in StreamTranscribe.swift is now a knob every caller turns on: the helper has exactly 1 production spawn site (grep stream_helper_path → apple_speech/__init__.py:813), which always passes it, and the batch path uses a different source (AppleTranscribe.swift, 0 --fast hits). Insert .frequentFinalization unconditionally in the streaming helper and delete the flag, the new argv element, and test_swift_helper_still_accepts_fast — the cross-process contract that test exists to protect disappears with the flag.

[FIRST-PRINCIPLES-REVIEWED] ae9f5e9

[CrysisDeu]

🤖 Kiro Crew Auto-Pipeline [operator: CrysisDeu#0c98c3a2]

Status: functionally review-ready. Everything this PR owns is green; the remaining reds are inherited from main and cannot be fixed from this branch.

• All 52 owned checks green; all 5 AI review lanes PASS (GPT 5.6, Opus 4.8, UX, Design, First Principles); MERGEABLE.
• The 6 red backend shards are exactly the two known main-side breakages, with zero overlap with this diff (which touches only apple_speech/__init__.py + test/test_apple_speech.py):
  • shards 2 (3.10/3.12/Windows): test/test_kiro_usage_api.py::TestWindowsCliStore::test_local_appdata_store_token_is_trusted_without_arn — main red: TestWindowsCliStore.test_local_appdata_store_token_is_trusted_without_arn fails on all backend shards #5879, fixed by open PR fix(test): accept endpoint kwarg in local-appdata trust test stub (#5879) #5887
  • shards 4 (3.10/3.12/Windows): test/test_trust_reads.py::TestIsReadOnlyBash::test_a_word_bash_deletes_cannot_forge_a_read_mode — main red: #5521 raw-scan refuses all input redirects, contradicting #5541 pinned redirect-read allowance #5846, fixed by open PR fix(trust): allow input redirects into read verbs at raw scan (#5846) #5880
• Both failures reproduce on unmodified main (local probe at base 6ed29f5cb; main's own ci.yml runs at 6ed29f5cb and b65222c54 are red on the same classes). Coverage Gate and PR Readiness reds are the cascade from those shards.
• fix(test): accept endpoint kwarg in local-appdata trust test stub (#5879) #5887 and fix(trust): allow input redirects into read verbs at raw scan (#5846) #5880 are cross-deadlocked (each inherits the red the other fixes); merging either greens both, and a rebase of this PR after they land clears all inherited reds here.

No action needed on this branch until the unblockers merge; a rebase + re-run then makes the board fully green.

[NicholasRBowers]

🤖 Kiro Crew [operator: NicholasRBowers#a942f9ca]: This PR has been inactive for 7+ days with failing CI. I've assessed the blockers and they appear resolvable — I'll push fixes directly to this branch as a co-author.

Assessment: All owned checks and all 5 review lanes are green; the only reds (Backend Tests shards 2 & 4 + Coverage Gate cascade) are inherited main breakages (#5879, #5846) that are now fixed on main — a rebase onto current green main + re-push clears them. No code changes expected.

If you'd prefer I don't touch this PR, add the pr-no-autofix label.

[NicholasRBowers]

🤖 Kiro Crew [operator: NicholasRBowers#a942f9ca]

Rebase-only push: d8abad090 → ae9f5e94f

The only red checks on this PR (Backend Tests shards 2 & 4 on 3.10/3.12/Windows, plus the downstream Coverage Gate / PR Readiness cascade) were inherited from two main-side breakages — #5879 (test_kiro_usage_api.py::TestWindowsCliStore) and #5846 (test_trust_reads.py::TestIsReadOnlyBash) — both verified to reproduce on unmodified main and both now fixed upstream. Main's last four completed ci.yml runs are green.

What this push changes:

• Rebased the branch onto current green origin/main (was 679 commits behind). Zero conflicts; the diff is byte-identical in scope — only src/kiro_crew/apple_speech/__init__.py and test/test_apple_speech.py, exactly as before.
• No source changes. A fresh push rebuilds refs/pull/5907/merge and re-dispatches CI cleanly (a bare gh run rerun --failed would reuse the stale pinned merge commit and re-fail).
• Commit identity hygiene: the author field carried a non-public internal email; amended to the author's GitHub noreply identity (Zezhen Xu <32421101+CrysisDeu@users.noreply.github.com>) with name credit preserved in the commit body, plus a Co-authored-by: Kiro Crew trailer for the rebase mechanics.

Local gates before pushing (all green): isort, flake8, mypy (1177 files, no issues), black on the two changed files, and pytest test/test_apple_speech.py (68 passed, 3 skipped).

All five AI review lanes passed on the identical diff at d8abad090; they will re-run on this head. Monitoring CI to green.

[NicholasRBowers]

🤖 Kiro Crew [operator: NicholasRBowers#a942f9ca]

Flake triage — re-ran Backend Tests (3.12, 1), no code change

The round's single root failure was test/test_cli_logging.py::TestDrainBeforeHardExit::test_no_listener_is_a_silent_no_op (asserts a QueueListener is None; 1 failed / 18,918 passed on the shard). Coverage Gate and PR Readiness reds are downstream cascade of that one shard.

Evidence it is cross-test global-state pollution, not this PR's diff:

• The file is untouched by this PR (diff is apple_speech/__init__.py + test/test_apple_speech.py only).
• The full test_cli_logging.py file passes locally on this exact head (33 passed).
• The identical shard passed on 3.10 and on Windows in the same round.
• The test is green on main's last four completed CI runs.

Dispatched gh run rerun --failed on run 33330084606 (same merge ref — freshly built from current main ~35 min ago, so no stale-base concern). Monitoring continues.

[NicholasRBowers]

🤖 Kiro Crew [operator: NicholasRBowers#a942f9ca]

Review-ready ✅ — head ae9f5e94f, 68/68 checks green

Drive summary:

• Rebased the branch onto current green main (was 679 commits behind); zero conflicts, no source changes — the diff remains exactly src/kiro_crew/apple_speech/__init__.py + test/test_apple_speech.py as the author wrote it.
• Commit hygiene: author identity moved to GitHub noreply form (internal email removed), name credit in the body, Co-authored-by: Kiro Crew trailer.
• One CI flake weathered: test_cli_logging.py::TestDrainBeforeHardExit::test_no_listener_is_a_silent_no_op failed once on Backend Tests (3.12, 1) — cross-test QueueListener pollution in a file this PR doesn't touch (passed locally, on 3.10, on Windows, and on main). Cleared on a single job re-run with no code change.
• All five AI review lanes PASS on this head (GPT 5.6, Opus 4.8, Design, UX, First Principles), PR Readiness passed, MERGEABLE, 0 unresolved threads. Closes Apple Speech streaming endpointing never auto-submits — helper launched without --fast #5896 on merge.

Auto-merge is deliberately NOT armed — merge decision is left to maintainers. Thanks to Zezhen Xu (@CrysisDeu) for the original fix.

[NicholasRBowers]

Tier 1 auto-approve: fix (2 files). Criteria: no conflict, no requested changes, security path denylist clean, design-doc gate clean, SAST annotations clean, security checklist all-NO, AI reviewers green. Category: fix with clear root cause -- streaming helper argv gains --fast so mid-stream FINAL segments reach the endpointer's note_final(), enabling auto-submit (#5896); argv contract pinned by new tests on both streaming and batch paths.

[bolichen97]

Readiness passed on the current head, all checks green, mergeable against main, and no outstanding change requests from any reviewer. Approving as part of a maintainer sweep of fix-type PRs.

[bolichen97]

Tier 1 auto-approve: fix (2 files). Criteria: no conflict, no requested changes, security path denylist clean, design-doc gate clean, SAST annotations clean, security checklist all-NO, AI reviewers green. Category: fix streaming speech auto-submit by passing --fast to the streaming helper argv (batch path deliberately excluded), pinned by argv + Swift-source tests.

[bolichen97]

Tier 1 auto-approve: fix (2 files). Criteria: no conflict, no requested changes, security path denylist clean, design-doc gate clean, SAST annotations clean, security checklist all-NO, AI reviewers green. Category: fix with a clear root cause — the default DictationTranscriber emits only volatile partials so the endpointer's note_final()-driven auto-submit never fires; adds the static --fast switch to the already-sandboxed streaming-helper argv to enable .frequentFinalization (#5896).

[bolichen97]

Tier 1 auto-approve: fix (2 files). Criteria: no conflict, no requested changes, security path denylist clean, design-doc gate clean, SAST annotations clean, security checklist all-NO, AI reviewers green. Category: clear root cause for #5896 -- the streaming helper was spawned without --fast, so the transcriber never inserted .frequentFinalization and emitted only volatile partials until the stream closed; the dashboard endpointer schedules its utterance-complete judgment solely from note_final(), so auto-submit could never fire. The production change is one bare switch added to the existing sandboxed argv; the sandbox wrapper itself is untouched and the one-shot batch path deliberately keeps the flag off.

[bolichen97]

Tier 1 auto-approve: fix (2 files). Criteria: no conflict, no requested changes, security path denylist clean, design-doc gate clean, SAST annotations clean, security checklist all-NO, AI reviewers green. Category: clear root cause (#5896) -- the streaming speech helper omitted --fast so no mid-stream finals were emitted and auto-submit could never fire; flag added with argv-pin tests on both paths.