chore(computer-use): remove confirmed dead code in the computer-use and voice surface

[iamwhatever]

Dead-code audit of the computer-use / voice / connectors group (≈36k lines: computer_use/, browser/, stt/, apple_speech/, connections/, providers/, mcp_computer.py, transcribe.py, voice_reply.py, imaging.py, qr.py). One commit, 47 deletions, no behaviour change, no refactor, no reformatting.

Candidate generation

Three independent scanners; only the intersection was carried forward.

Pass	Findings
vulture --min-confidence 60	149 — dominated by TypedDict/dataclass fields and Protocol/base-class stubs, excluded by design
AST (defined names minus Load/Attribute/keyword names)	622 raw — too weak alone, any in-file self.x marks a symbol alive
tokenize (real NAME tokens only, repo-wide occurrence count == 1)	40 — excludes docstring/comment substring collisions
Confirmed after the full protocol	11

Each survivor then went through: repo-wide search across *.py *.ts *.tsx *.js *.md *.json *.yml *.yaml *.toml *.sh *.ps1 including tests/, docs/, system-specs/, builtin_skills/, .github/, baselines and pyproject.toml entry points (never truncated through head); a bare-string / value-dispatch sweep; a reverse-direction reachability proof; a public-surface check (__all__, docs, SKILL.md); the 30-day new-code gate; and a cross-platform reachability proof for every Windows-only and macOS-only candidate.

Deleted (11 symbols)

computer_use/types.py — observation-ceiling and refusal residue (9)

Root fact: gate.permitted_observation_channels() returns frozenset(ALL_OBSERVATION_CHANNELS) unconditionally and gate.apply_observation_ceiling() is return dict(payload) (gate.py:164-187) — no branch on platform, edition, config, session or policy, and neither consults the real GovernanceCeiling model in platform/governance.py. The layer was already a pass-through at the commit that introduced the package (verified with git show d6d8a0da4:.../gate.py), and the removal is documented in docs/system-specs/modules/computer-use.md:606-607, :2433-2437 and governance.md:2504-2510. These constants are residue of a documented non-feature, not an unfinished control.

Symbol	Evidence
OBS_VALUES_SUPPRESSED_NOTE	Zero refs; literal "Element values suppressed" only at its own def. The element_values deny path (tools.py:1039) drops detail via GOVERNED_VALUE_PLACEHOLDER and tools.py:1029-1032 states explicitly that ceiling notes are deliberately not appended.
OBS_TITLES_SUPPRESSED_NOTE	Zero refs; OBS_WINDOW_TITLES is never membership-tested, only declared into the pass-through gate (tools.py:1260).
OBS_TREE_SUPPRESSED_NOTE	Zero refs; OBS_A11Y_TREE never membership-tested (tools.py:1258 only).
OBS_PATHS_SUPPRESSED_NOTE	Zero refs; OBS_FILE_PATHS has no consumer at all beyond its def and the channel tuple.
GOVERNED_PATH_PLACEHOLDER	Zero refs; literal "<redacted:path>" appears nowhere else. Live sibling GOVERNED_VALUE_PLACEHOLDER is consumed at tools.py:1071 — kept.
PATH_SCRUB_PATTERNS	Zero refs; sole intended consumer _scrub_paths does not exist in the tree. No inlined duplicate of either regex.
SCREENSHOT_MIME	Zero refs. No image bytes cross a wire from this path: tool results are text-only, only the file path is relayed; the live-view frame uses screencast.FRAME_FORMAT and the browser hardcodes the data-URL prefix (useComputerUseFrame.ts:65). The four spool siblings are all alive — kept.
REFUSAL_UNATTENDED	Zero refs. The unattended rule is removed, stated in five source comments plus governance.md:1230-1233 ("Neither exists."). Not in error-code-baseline.json, no i18n catalog entry, no prose-enumerating test.
REFUSAL_GOVERNANCE	Zero refs. Unreachable by construction: require_computer_use is return None on every path (gate.py:88), so its one caller's if denial: branch can never be entered — a property test_computer_use_gate.py:286 pins.

computer_use/overlay.py (1)

points_payload — zero refs. Its docstring claimed "used by tests and diagnostics"; that is false, there were zero test refs and zero diagnostic callers at the parent commit. Superseded by the inline comprehension in _move_command (overlay.py:423), which is the actual wire path. Not in overlay.py.__all__. Removing it left Sequence unused, so that import went too (no other use remains in the file).

computer_use/tools.py (1)

_KEYBOARD_TOOLS — zero refs, a dead duplicate. Superseded by _ELEMENT_REQUIRED_TOOLS (tools.py:244-251), a strict superset whose own comment says so ("keyboard tools included"). The live "keyboard tool refused without an index, so a keystroke cannot land on a password box" control is carried by _ELEMENT_REQUIRED_TOOLS plus _SECURE_TARGET_TOOLS (tools.py:268, used at :715) — both contain TOOL_TYPE_TEXT and TOOL_PRESS_KEY, both active. No guard is weakened. (One blank line was restored under the removed block to satisfy E302; that is the single + line in the diff.)

Post-deletion: zero residual references for all 11 across every scanned file type.

Deferred (13)

30-day new-code gate (10). All from f4e31653b (#4530, 10d) and 14775e490 (#3260, 13d):

• _SMOOTHING_NONE (capture_windows.py:98) — dead on evidence including on Windows: its consumer GdipSetSmoothingMode is not bound in the FFI table (capture_windows.py:264-265), and the dynamic getattr(lib, symbol) loop iterates that same table.
• ACTION_COLLAPSE (windows_driver.py:243) — local alias only; WINDOWS_ACTION_COLLAPSE is alive and the value "collapse" still reaches the driver by string via SUPPORTED_ACTIONS.
• registry_slugs (connections/tool_aliases.py:183) — dead on evidence (13d); every real consumer calls get_all_registry_providers() directly.
• 7 × Win32 FFI enum mirrors (windows_ffi.py): TreeScope_Element, TreeScope_Descendants, ExpandCollapseState_PartiallyExpanded, ScrollAmount_SmallDecrement, DPI_AWARENESS_CONTEXT_{UNAWARE,SYSTEM_AWARE,PER_MONITOR_AWARE}. Held for a second reason too: each is one member of a faithful mirror of a documented UIA/Win32 enum whose other members are load-bearing, and deleting a middle ordinal makes the block read as a mistranscription — the exact failure the UIA_IsPasswordPropertyId comment (windows_ffi.py:151-157) documents at length.

Test-referenced with no production caller — refactor, not deletion (3): exposed_server_keys (tool_aliases.py:270, superseded by a direct _parse_tool_refs call at :321), unsupported_snapshot (backend.py:525, its stated consumer never materialized), char_keystroke (keymap.py:391, its per-character path was deliberately rejected for Unicode key events at macos_driver.py:495-499).

Rejected as alive — worth recording

• DriftVerdict.prior_state_loaded / _recorded_at / _discarded (connections/l0_drift.py:58-60) — DriftVerdict is a TypedDict, so the field names are the runtime dict keys, written literally at l0_drift.py:158-160 and asserted in test_connections_l0_probe.py. The scanners saw one identifier occurrence because the values flow through StateLoad while the keys are string literals. This is the expected false-positive class for serialized shapes.
• _packaged_ffmpeg_version_probe (transcribe.py:586) — imported at build time by packaging/build-desktop.sh:184 and pinned by website/electron/test/packaging.test.js:342-343.
• 26 further vulture findings across alias_record.py, mint.py, ownership.py, status.py, command_bus.py, enable_state.py, stt/engine.py, voice_reply.py, transcribe.py, qr.py, imaging.py, mcp_computer.py, backend.py — all have production callers, mostly in dashboard/handlers/connections.py, dashboard/handlers/messaging.py, agent.py, or a CLI verb (run_mcp_server is the kirocrew mcp-computer verb at cli.py:2675).

Needs a decision (not in this diff)

1. PAYLOAD_SCREENSHOT_META (gate.py:54, in __all__) — a real correctness finding rather than dead code. gate.py:50-53 says these payload keys are named constants "because tools and render both build and destructure these dicts, and a typo would silently drop a field rather than fail" — yet tools.py:1123-1125 builds all three keys with bare string literals. Worse, the three keys are write-only: _render_snapshot's rebuild (tools.py:1132-1145) reads back only window-title/elements/screenshot and takes width and height from snap, never from shaped. A typo is silently harmless today and silently lossy the day the ceiling stops being a pass-through. Fix is either to wire tools.py:1123-1125 through the constant, or to delete the constant and the three write-only keys together.
2. is_mutating_action (gate.py:155, in __all__, named in governance.md:2441) — retained-by-design accessor. The spec is accurate and does not claim a consumer: "it currently has no caller in the package: it is retained as the accessor an edition would use." The "one live consumer" is hooks, which reads the SSOT table directly (hooks.py:1409). Not an inlined duplicate — the predicates genuinely differ, since end_turn is CU_CLASS_CONTROL, making is_mutating_action("end_turn") False while hooks' predicate is True. Deleting it would require changing a spec paragraph in the same commit.
3. Stale comments naming a function that does not exist. tools.py:1014-1037 describes an active file-path scrub via _scrub_paths (absent from the tree) and claims permitted_observation_channels "returns the empty set on an error" — it cannot, there is no try/except and the return is unconditional. Pre-existing drift on main; left untouched because this PR makes no prose or behaviour changes.
4. expire_dead_mints (connections/warm.py:106) — half-wired, documented deferral. The withdrawal half is wired (dashboard/handlers/connections.py:411) but the loop skips every row without entry.get("shared"), and nothing in the repo ever writes shared: True — the only occurrence is the TypedDict field declaration (mint.py:113). warm.py:19-21 and docs/architecture/design-notes/connections-warm-table.md both say this is deferred to slice N2b. Not deletable: it answers correctly the moment N2b starts filling the table.
5. BrowserCommandBus.is_registered (browser/command_bus.py:189) — half-wired read accessor. The write half _register_locked is called from drain (:350) and complete (:403); the public query half has zero production readers. The endpoint that logically wants it re-derives liveness inline (submit → _panel_alive_locked, :232), and dashboard/handlers/messaging.py:2963-2965's docstring claims that check is the registration when it is actually the private predicate. No unregister counterpart exists at all — deregistration is TTL-only (:81).

Also noted, no action here: find_brew (transcribe.py:758) has zero production callers while the one site that resolves brew uses bare shutil.which("brew") (cloud/ssm.py:231) with no fixed-prefix fallback — exactly the GUI-launched-gateway failure _BREW_CANDIDATE_PATHS exists to prevent. A hardened resolver that was never adopted; keep it.

Verification

• flake8, isort --check, black --check clean on all three touched files.
• Affected-scope pytest: 681 passed (test_computer_use_gate.py, test_computer_use_overlay.py, test_mcp_computer.py, test_computer_use_snapshot.py, test_computer_use_backend.py, test_computer_use_api.py, test_computer_use_capture.py, test_computer_use_unsupported.py, test_computer_use_registration.py). No test was deleted or adjusted — nothing referenced any deleted symbol.
• Two local adversarial review lanes on the diff before push, both no findings: one on collateral damage (residual refs, value-dispatch reachability, guard weakening, import/whitespace, doc drift), one on the security angle (whether the observation gate is genuinely unconditional, whether any edition/CPP/policy seam reaches the deleted constants, whether any baseline or i18n catalog enumerates the two refusal strings, and whether deleting the strings erases the only evidence a control is missing — position: no, the authoritative record lives in the retained spec and the retained reintroduction seam).

[github-actions]

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of 712b93b3180db3ea1338df5d968601e0dd9fe9ce and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 712b93b

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 712b93b3180db3ea1338df5d968601e0dd9fe9ce: <one-sentence reason>

[github-actions]

Design Review (Fable 5) — ✅ PASS

Design-level review of 712b93b3180db3ea1338df5d968601e0dd9fe9ce — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

Confirmed-dead residue of a documented non-feature, removed with zero behavior change; every deletion claim verified against the tree, guards demonstrably carried by live supersets.

Grep confirms zero residual references to all 11 deleted symbols; _ELEMENT_REQUIRED_TOOLS (containing TOOL_TYPE_TEXT, TOOL_PRESS_KEY) is live at the dispatch check, so the secure-field keystroke guard is intact; the unattended rule's removal is independently documented in retained source comments and spec, so the deleted refusal strings erase no evidence. Reversible via git, spec unchanged because documented behavior is unchanged, and the description's deferred/needs-decision ledger (notably the write-only PAYLOAD_SCREENSHOT_META keys and the stale _scrub_paths comment at tools.py:1016) is the right disposal path for what this PR correctly left alone.

[DESIGN-REVIEWED] 712b93b

[github-actions]

Opus 4.8 Review — ✅ no blocking findings

Reviewed 712b93b3180db3ea1338df5d968601e0dd9fe9ce — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] 712b93b

Verdict parsed from the review's SHA-scoped output markers for commit 712b93b3180db3ea1338df5d968601e0dd9fe9ce.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 712b93b3180db3ea1338df5d968601e0dd9fe9ce: <one-sentence reason>

[github-actions]

First Principles Review (Fable 5) — ✅ PASS

Premise-level review of 712b93b3180db3ea1338df5d968601e0dd9fe9ce — 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.

Verification complete. All 11 deleted symbols have zero remaining references anywhere in the tree (one grep across all deleted names returned a single hit — a docstring mention of _scrub_paths at tools.py:1016). The live secure-keystroke guard is intact: _ELEMENT_REQUIRED_TOOLS (tools.py:244) and _SECURE_TARGET_TOOLS (tools.py:268) both still contain TOOL_TYPE_TEXT and TOOL_PRESS_KEY, so the deleted _KEYBOARD_TOOLS duplicate carried no enforcement. Sequence is genuinely unused in overlay.py after points_payload went.

First-Principles-Verdict: PASS

Pure subtraction: eleven zero-reference symbols deleted, every claim in the description independently re-verified against the tree, no guard weakened, nothing rides along.

What this change ships

Intent: shrink the computer-use surface by deleting confirmed-dead symbols — a FIX (removing residue), shipping only deletions.

1. Four never-displayed "policy suppressed X" notes deleted — justified, zero refs (grepped, 0 hits)
2. Unused path-redaction placeholder and scrub-pattern table deleted — justified; sole intended consumer _scrub_paths never existed
3. Unused screenshot MIME constant deleted — justified, zero refs
4. Two unreachable refusal messages (unattended, governance) deleted — justified; gate.py returns None unconditionally, test_computer_use_gate.py pins it
5. Unused points_payload helper plus its now-orphaned Sequence import deleted — justified, zero refs; live wire path is the inline comprehension in _move_command
6. Dead _KEYBOARD_TOOLS set deleted — justified duplicate; _ELEMENT_REQUIRED_TOOLS and _SECURE_TARGET_TOOLS both still carry TOOL_TYPE_TEXT/TOOL_PRESS_KEY, verified at tools.py:244-276

Watch

One residue of the same root cause (the never-built path-scrub channel) survives: the docstring at src/kiro_crew/computer_use/tools.py:1016 still says "the channel _scrub_paths already understands" — a function that does not exist and whose pattern table this PR just deleted. Grepped _scrub_paths: 1 hit, that comment.

Subtractions

• Delete the parenthetical "(the channel _scrub_paths already understands)" from the tools.py:1016 docstring — it names a nonexistent function whose supporting constant this PR removes.

[FIRST-PRINCIPLES-REVIEWED] 712b93b

[bolichen97]

Tier 1 auto-approve: chore (3 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: removal of confirmed dead code in computer_use (overlay/tools/types) — no runtime behaviour change.

[bolichen97]

Tier 1 auto-approve: chore (3 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: pure dead-code deletion in the computer-use surface (points_payload, _KEYBOARD_TOOLS, unused refusal/suppression constants and scrub patterns); zero references repo-wide verified by grep against main tip d7b7d65, no behaviour change. Auto-merge arming is unavailable to this pipeline agent (command denied by its hardened allowlist), so merging is deferred to a human maintainer with branch protection as the gate.

[NicholasRBowers]

Tier 1 auto-approve: chore (3 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: dead-code removal in computer_use/{overlay,tools,types}.py -- removed symbols verified unreferenced outside their defining files, no runtime impact.

[bolichen97]

Independently verified every removed symbol (grep across full repo + PR head, string/dynamic/registry lookup checks) has zero remaining references anywhere -- no dangling handler/registry entry (the #6876 defect class). CI all green. Approving.