🤖 refactor: auto-cleanup

[mux-bot]

Summary

Long-lived auto-cleanup PR that accumulates low-risk, behavior-preserving refactors picked from recent main commits.

Changes

Collapse ReviewPanel selection-validity branches

ReviewPanel.tsx's post-filter selection-validity effect (touched by the immersive sidebar fix in #3249) split into two structurally identical arms: an isImmersive branch that validated selectedHunkId against hunks (the unfiltered diff) and a non-immersive branch that validated against filteredHunks. Both arms then ran the same setSelectedHunkId(filteredHunks[0].id) reset on miss, with the immersive arm needing its own early return to avoid double-running the reset. The duplication invited drift between the two modes — e.g. a future hide-read-style filter that adjusts the immersive reset would have to keep both arms in lockstep by hand.

Pick the validity list up front (const validityList = isImmersive ? hunks : filteredHunks;) and run a single selectedHunkId && validityList.some(...) check, with a short comment capturing the rationale so a future reader doesn't reintroduce the split. The early-return on filteredHunks.length === 0, the dependency list ([filteredHunks, hunks, isImmersive, selectedHunkId, setSelectedHunkId]), and the immersive-aware "only reset when the hunk has truly disappeared from the diff" semantics added in #3249 are all preserved.

Pure refactor — emitted JS, the selection-reset trigger conditions, and the existing 5-test ImmersiveReviewView suite (including the two immersive regression tests added in #3249) are all unchanged.

Previous cleanups

Extract renameAliasField helper for bash tool preprocess

The bash tool's z.preprocess shim (in src/common/utils/tools/toolDefinitions.ts) normalizes quirky model emissions to canonical field names. After the DeepSeek v4 fix in #3247 added a second description → display_name rename block (mirroring the existing command → script rename), the two blocks were structurally identical: skip when canonical is already a string, drop the alias via destructure, re-spread with the canonical name. Each alias still required its own as Record<string, unknown> & { <alias>: string } cast plus the same three-line destructure/spread.

Extract a private renameAliasField(obj, alias, canonical) helper that captures the rename pattern in one place, with the rationale for why aliases exist (and why they stay undocumented) noted on the helper. The call site collapses to two named lines that read as the intent (rename command to script, rename description to display_name) instead of fifteen lines of mostly-identical destructure/spread. A future alias (e.g. another quirky-model field) becomes a one-line addition.

Pure refactor — emitted JS, the canonical-field-wins precedence, the "no-op when neither alias nor canonical is a string" branches, and the existing 36-test toolDefinitions suite (including the four command/description alias precedence cases added in #3247) are all unchanged.

Derive TokenRecord from BrowserBridgeTokenPayload

BrowserBridgeTokenManager (in src/node/services/browser/BrowserBridgeTokenManager.ts, touched by the new other-workspace session picker in #3243) declared two structurally identical interfaces side by side: a private TokenRecord (the stored token state) and an exported BrowserBridgeTokenPayload (the validate() return shape), with TokenRecord differing only by the extra expiresAtMs: number deadline. validate() then rebuilt the payload by listing each of the four shared fields a third time. The #3243 commit made this drift surface especially visible — adding allowOtherWorkspaceSession required parallel edits in TokenRecord, BrowserBridgeTokenPayload, the mint() insert, and the validate() rebuild.

Collapse TokenRecord to extends BrowserBridgeTokenPayload (with the rationale captured in a short comment so a future reader doesn't reintroduce the duplicated field list) and rewrite the validate() rebuild as a rest-spread destructure (const { expiresAtMs, ...payload } = record; return payload;). The eslint config already enables ignoreRestSiblings: true, so the unused expiresAtMs binding doesn't trigger no-unused-vars. A future payload field (e.g. another scoping flag) now lives in exactly one place — the exported BrowserBridgeTokenPayload interface — and automatically flows through both the stored record shape and the validate-time rebuild.

Pure refactor — emitted JS, the TTL semantics, the null-on-expired path, and the existing 5-test BrowserBridgeTokenManager suite (including the new "preserves explicit other-workspace session scope" case) are all unchanged.

Extract detachLanguageModelCleanup helper in languageModelCleanup

moveLanguageModelCleanup and runLanguageModelCleanup (in the new src/node/services/languageModelCleanup.ts from the OpenAI WebSocket transport opt-in PR #3241) both implemented the same "look up the attached cleanup, delete the symbol slot, return it" sequence inline. The duplication is structurally identical — same LanguageModelWithCleanup cast, same symbol read, same delete — so both call sites were one stray edit away from drifting on whether the slot gets cleared before or after the cleanup runs (which matters because attachLanguageModelCleanup asserts the slot is empty).

Extract a private detachLanguageModelCleanup helper that does the single-shot pop in one place and returns the cleanup (or undefined). Both surviving public functions reduce to their intent: move re-attaches the popped cleanup to the target, and run invokes it inside the existing try/catch that swallows + logs failures. A short comment captures the rationale on the helper itself so the next caller (e.g. a future "cancel without invoking" path) can't accidentally leak a slot.

Pure refactor — emitted JS, the symbol's single-shot semantics, and the existing 6-test languageModelCleanup suite are all unchanged.

Extract isAgentTaskActiveStatus predicate in task_await

Three call sites in src/node/services/tools/task_await.ts gated on the active agent-task subset ("queued" | "running" | "awaiting_report") by inlining the three-arm equality check. The new task_await elapsed-timing commit (#3234) extended both the timeoutMs === 0 branch and the timed out catch-branch symmetrically with ...getAgentTaskElapsedField(taskId), making the two structurally identical { status, taskId, ...elapsed } returns sit a handful of lines from a third copy in the ForegroundWaitBackgroundedError branch that picks between the same three values to coalesce against a "running" fallback.

Add a local AgentTaskActiveStatus subset alias plus an isAgentTaskActiveStatus type predicate near the existing coerceTimeoutMs / parseTimestampMs helpers, and collapse all three inline checks to a single call. The predicate narrows the nullable AgentTaskStatus | null return of getAgentTaskStatus to the active subset, so the surviving { status, taskId, ... } returns keep their narrowed status field with no as const gymnastics. The new comment captures the rationale so a future field added to one of the active-status branches won't silently drift away from the others.

Pure refactor — emitted return shapes (including elapsed_ms), narrowed status literals, and the existing 17-test task_await suite are all unchanged.

Drop dead length guard in parseBedrockModelName secondPart

The early dotParts.length < 2 return at the top of parseBedrockModelName (in src/common/utils/ai/modelDisplay.ts) already guarantees dotParts.length >= 2 by the time secondPart is computed, which makes the dotParts.length > 1 ? dotParts[1].toLowerCase() : "" ternary's empty-string branch unreachable. The DeepSeek V4 commit (#3237) extended the surrounding formatter without touching this site, but the new DeepSeek branch made the asymmetry more obvious — the line directly above (firstPart) already accesses dotParts[0].toLowerCase() without a guard, so the ternary on secondPart was the odd one out.

Inline to a direct dotParts[1].toLowerCase() access (matching firstPart's shape) and capture the rationale in a one-line comment so a future reader doesn't reintroduce the guard. Pure dead-code cleanup — emitted JS, runtime behavior, and the existing 18-test modelDisplay suite (including the new DeepSeek + existing Bedrock cases) are all unchanged.

Drop unused workspaceName param from parseRuntimeString

The second argument to parseRuntimeString (in src/browser/utils/chatCommands.ts) was named _workspaceName (underscore-prefixed = unused) when it was introduced in the original SSH runtime PR, and has never been referenced by the function body — error messages are all generic and don't include any workspace-specific context.

The /new-mirrors-/fork simplification (#3230) made the noise especially visible: the new createNewWorkspace call site had to pass options.workspaceName ?? "(auto-generated)" purely to satisfy the signature, and added a comment claiming parseRuntimeString only uses the name for error reporting context — but the function never reads it. Mobile already had the cleaner shape (parseRuntimeStringForMobile(runtime?: string)).

Drop the parameter from the signature, drop the placeholder + misleading comment at the only non-test caller (createNewWorkspace), and drop the workspaceName = "test-workspace" constant + 23 second-arguments in chatCommands.test.ts. Pure dead-parameter cleanup — emitted JS, error messages, and runtime behavior are all identical, and the desktop signature now matches mobile's.

Drop redundant isPlanHandoffAgent boolean in streamContextBuilder

The isPlanHandoffAgent boolean in buildPlanInstructions was extracted when the gate was effectiveAgentId === "exec" || effectiveAgentId === "orchestrator". After #3224 ripped out the Orchestrator agent, the boolean collapsed to a single equality check (effectiveAgentId === "exec"), and the trailing else if (isPlanHandoffAgent && chatHasStartHerePlanSummary) redundantly re-evaluated the same gate just to log a debug line.

Replace the flat if/else if with a nested if (effectiveAgentId === "exec") { … } that tests the Start Here summary inside, removing the duplicate gate re-check and the now-meaningless boolean. A short comment captures the rationale so a future reader doesn't reintroduce the alias.

Pure refactor — emitted control flow, the debug log, and planContentForTransition assignments are identical.

Extract seedScrollDirectionBaseline helper in useAutoScroll

useAutoScroll (touched by #3226) seeds lastScrollTopRef from two call sites — jumpToBottom and disableAutoScroll — to keep handleScroll's released-branch direction check (currentScrollTop > previousScrollTop) honest after a programmatic ownership transfer that may not produce a follow-up scroll event. Both sites repeated the same write (lastScrollTopRef.current = contentRef.current?.scrollTop ?? 0) and a multi-line block comment re-explaining the same shared rationale.

Extract a seedScrollDirectionBaseline useCallback with the shared rationale captured once on the helper itself. Each call site reduces to a single call plus a one-line comment naming the site-specific reason it doesn't get a free scroll-event refresh (stickToBottom skips the write at max; disableAutoScroll never fires a scroll event itself). The dependency arrays for jumpToBottom and disableAutoScroll add the new helper, which has empty deps (useCallback(..., [])) so its identity is stable across renders — no extra re-creations of the surrounding callbacks.

This shrinks the surface area for future drift: a third path that needs to flip autoScrollRef/programmaticDisableRef without a guaranteed scroll-event tail (e.g. the deferred workspace-switch hydration race called out in #3226's "Pains") can call the helper instead of duplicating the rationale a third time. Pure refactor — emitted writes, write order, and ref values are identical, and the existing 25-test useAutoScroll suite (including the 6 new regression tests added in #3226) passes unchanged.

Extract pushStreamErrorRow helper in StreamingMessageAggregator

buildDisplayedMessagesForMessage now has two branches that synthesize a stream-error DisplayedMessage: the existing message.metadata?.error path and the finishReason === "length" path added in #3223. Both pushes were structurally identical, differing only in id suffix, error string, and errorType — the parent-message-derived fields (historyId, historySequence, model, routedThroughGateway, timestamp) were duplicated across both call sites.

Extract a local pushStreamErrorRow closure that captures the shared fields once. Each branch reduces to a single call passing the three differing values. The model access switches from message.metadata.model (which relied on the outer if (message.metadata?.error) narrowing) to message.metadata?.model, which is functionally identical when metadata is defined and falls through to undefined otherwise — same emitted value either way.

This shrinks the surface area for future drift: the next branch added (e.g. a different finish-reason synthesis) can't accidentally pass a stale model or forget routedThroughGateway. Pure refactor — emitted DisplayedMessage objects are identical, and the existing 77-test SMA suite (including the 6 new max-tokens tests) passes unchanged.

Drop redundant GuardAnchors type alias in file_edit_insert

In src/node/services/tools/file_edit_insert.ts, GuardAnchors was defined as Pick<InsertContentOptions, "insert_before" | "insert_after">, but InsertContentOptions itself is already Pick<FileEditInsertToolArgs, "insert_before" | "insert_after"> after the .nullish() strict-mode refactor in #2250 stripped the InsertContentOptions interface down to those same two fields. The two aliases became structurally identical, so GuardAnchors is dead.

Drop the alias and use InsertContentOptions for the two callers (insertWithGuards, resolveGuardAnchor). Both names were file-local; no exports change. The function names already convey "guard" context, so the parameter type doesn't need to repeat it. This noise was especially visible to the new guardless-empty-file path added in #3220 since it sits next to insertContent which uses InsertContentOptions.

Pure type-level cleanup — emitted JS, runtime behavior, and the public tool surface are all unchanged.

Extract ResolvedWorkspaceAiSettings type alias in taskService

The shape { model: string; thinkingLevel?: ThinkingLevel } (used internally to read per-agent AI settings off partial workspace metadata) was duplicated 7 times across the parameter and return types of resolveWorkspaceAISettings, resolveTaskAISettings, and resolveParentAutoResumeOptions in src/node/services/taskService.ts. The new sub-agent defaults split (#3215) made this duplication especially visible because it added a third method copying the same inline shape.

A private ResolvedWorkspaceAiSettings interface at module scope replaces all 7 inline occurrences. Pure type-level cleanup — emitted JS, runtime behavior, and the schema-derived WorkspaceAISettings type (where thinkingLevel is required) are all unchanged. The new interface is intentionally local to this file (not exported) since it captures the looser internal-reader shape, distinct from the canonical schema-derived type.

Extract shared ServiceTier type from ServiceTierSchema

The OpenAI service-tier literal union ("auto" | "default" | "flex" | "priority") was duplicated in three places: the CLI options interface (src/cli/run.ts), the providerModelFactory cast at the providers.jsonc boundary (src/node/services/providerModelFactory.ts), and a local OpenAIServiceTier alias in ProvidersSection.tsx. ServiceTierSchema (src/common/config/schemas/providersConfig.ts) already defines the same enum as the runtime source of truth, so this change derives a TypeScript ServiceTier alias via z.infer once and imports it at each site.

The Settings UI keeps its OpenAIServiceTier local alias (used by OpenAIServiceTierSelectValue and the isOpenAIServiceTier type guard) but it is now type OpenAIServiceTier = ServiceTier, so the disambiguating name and call sites stay intact.

Drop unused appendSpace literals on skill/model alias suggestions

In src/browser/utils/slashCommands/suggestions.ts, the SuggestionDefinition literals built for agent skills and model alias one-shots set appendSpace: true, but the build callbacks for those two paths hardcode the trailing space and never read definition.appendSpace. Only the top-level command and subcommand build callbacks consult the field. Remove the no-op appendSpace: true from the skill and model alias mappings and leave a short comment near each so a future reader doesn't assume the field is consulted on these paths.

Behavior is preserved — the field is unread on these paths, and the existing suggestions.test.ts / inlineSkillSuggestions.test.ts / suggestionMatching.test.ts suites all still pass.

Route ThemeContext color-scheme through isLightThemeMode

Replace the inline theme === "light" || theme === "flexoki-light" check in getColorScheme (in src/browser/contexts/ThemeContext.tsx) with the shared isLightThemeMode helper from src/browser/utils/highlighting/shiki-shared.ts.

The helper was just introduced as the single source of truth for the -light suffix → light-theme mapping, and the highlighting call sites (MarkdownComponents, HighlightedCode, highlightDiffChunk) already use it. This change extends that convention to ThemeContext so a future palette like solarized-light would automatically map to colorScheme: "light" without revisiting this site.

Behavior is preserved: for the four valid ThemeMode values ("light", "dark", "flexoki-light", "flexoki-dark"), isLightThemeMode returns the same result as the previous explicit comparison.

Validation

• bun test src/browser/features/RightSidebar/CodeReview/ImmersiveReviewView.test.tsx — 5/5 pass (includes both immersive regression tests added in 🤖 fix: keep immersive sidebar review click on its hunk #3249, which together exercise the immersive validity-list branch and the non-immersive validity-list branch end-to-end through the parent harness).
• make static-check — eslint, tsgo (×2), prettier all pass; only shfmt fails because the binary is unavailable in this environment, and no shell files are touched.

Risks

Minimal — purely a local effect-body simplification in ReviewPanel.tsx. The isImmersive ? hunks : filteredHunks selection produces the identical validity list each branch was already using, the surviving selectedHunkId && validityList.some(...) check is a literal hoist of the previous duplicated expression, and setSelectedHunkId(filteredHunks[0].id) runs in exactly the same conditions (selection missing or stale against the chosen list, after the existing filteredHunks.length === 0 early-return). The dependency list is byte-identical to the post-#3249 version, so React's effect-rerun cadence is unchanged. There's no IPC, persisted-state, or rendered-output coupling.

---

Auto-cleanup checkpoint: 555621d

---

Generated with mux • Model: anthropic:claude-opus-4-7 • Thinking: xhigh

[mux-bot]

⚠️ Auto-fixup could not resolve CI failure: flaky tests, not caused by the cleanup commit.

Failed job: Test / Integration (run 25244368462)

Failures (both in tests/ui/compaction/compaction.test.ts):

1. force compaction triggers during streaming — timed out at 120000 ms waiting for "Compaction boundary" transcript text.
2. /compact with Ctrl+Enter (turn-end) does NOT auto-background foreground bash — waitFor in tests/ui/helpers.ts:144 could not locate the workspace in the sidebar.

What I checked:

• The PR diff is type-only refactors (extract ResolvedWorkspaceAiSettings / ServiceTier aliases, drop unused GuardAnchors / appendSpace literals, route ThemeContext color-scheme through isLightThemeMode). None of it touches compaction, streaming, the workspace consumer manager, or the test harness.
• Ran both failing tests locally on 1312f5a68:
  • force compaction triggers during streaming → PASS in 6.7s.
  • /compact with Ctrl+Enter ... → PASS in 4.7s.
• The streaming-compaction test has a 120s budget and clearly hit a slow/contended runner; the workspace-not-found timeout looks like the usual sidebar race in the harness.

No fix pushed. A re-run of Test / Integration should clear it; if it recurs on this PR or others, the test harness itself likely needs a stability pass (separate from this cleanup PR).

[mux-bot]

⚠️ Auto-fixup: CI failure looks like a flaky integration test, not caused by the cleanup commits.

Failed job: Test / Integration (run 25400653486)

Failing test: tests/ui/compaction/compaction.test.ts › force compaction triggers during streaming — timed out at 120s waiting for "Compaction boundary" to appear in the transcript. The next test in the same file cascaded into "Project not found in sidebar" (broken state from the prior timeout).

Why this looks flaky, not regression-caused:

• None of the files touched by this PR are in the compaction code path. The diff is pure type/dead-code cleanup in ThemeContext, useAutoScroll, StreamingMessageAggregator (stream-error row helper extract — no behavioral change), streamContextBuilder (drop unused isPlanHandoffAgent bool), taskService, file_edit_insert, provider config, slash command suggestions, chatCommands, and cli/run.ts.
• This branch's prior PR-workflow runs (May 1, May 2, May 3) all passed with the same nine refactor commits already on the branch. The only commit added since the last green run is 5f8c84bab refactor: drop dead length guard in parseBedrockModelName secondPart, which is unrelated to compaction.
• tests/ui/compaction/compaction.test.ts has a documented history of stabilization patches (88c99f08c, 5dc32b039, 355e0007c).
• The failure mode is a 120s timeout, typical of CI runner contention rather than a logic regression.

No fix pushed — recommend re-running the failed job (gh run rerun 25400653486 --failed). If it reproduces deterministically, it's worth investigating whether one of today's main-branch merges (DeepSeek V4 support, the perf revert, or the Harbor CI change) destabilized this scenario, since the auto-cleanup branch picked those up via rebase.