🤖 refactor: auto-cleanup

[mux-bot]

Summary

This is the long-lived auto-cleanup PR. Each run, the auto-cleanup agent reviews new commits merged to main, rebases onto the latest main, and applies at most one extremely low-risk, behavior-preserving cleanup. The branch accumulates a small stack of independent cleanups until it is merged.

Cleanups in this branch

Cleanups 1–78 (all previous runs; #33 and #43 were later dropped as superseded by `main`)

1. Dedupe memory sweep recordUsage callbacks (MemoryConsolidationService). The consolidation sweep and the harvest sweep each inlined the same 15-line callback that routes billed usage to the headless-usage sidecar and emits analyticsIngest. Extracted into a private makeSweepUsageRecorder(...) helper.

2. Dedupe the "memory scope is full" cap check (MemoryService). The create and saveFile (new-file) paths each inlined a byte-identical block that called store.listFiles(), compared the count against MEMORY_MAX_FILES_PER_SCOPE, and threw a MemoryCommandError with the same message. Extracted into a private assertScopeHasRoom(store, scope) helper.

3. Dedupe blockquote line formatting in the bash monitor wake prompt (bashMonitorWakeStore.ts). buildBashMonitorWakePrompt rendered both the matched-output lines and the lost-monitor script with the identical .map((line) => \> ${line}`).join("\n")blockquote pattern in two places. Extracted into a module-levelblockquoteLines(lines)` helper.

4. Dedupe the tool_search removal in prepareToolSearch (toolCatalog.ts). Both fallback branches (PTC enabled, and empty deferred catalog) inlined the identical { [TOOL_SEARCH_TOOL_NAME]: _removed, ...rest } destructure to drop the built-in tool_search entry from the tool record. Extracted into a module-level withoutToolSearch(tools) helper.

5. Dedupe the Anthropic cache-create token extraction in accumulateProviderMetadata (usageHelpers.ts). The function inlined the same verbose (metadata.anthropic as { cacheCreationInputTokens?: number } | undefined)?.cacheCreationInputTokens ?? 0 cast twice (once for the accumulated metadata, once for the current step). Extracted into a module-private getAnthropicCacheCreateTokens(metadata) helper.

6. Dedupe capability-model thinking-policy resolution (thinking/policy.ts). After 🤖 feat: integrate GPT-5.6 Sol/Terra/Luna with native max effort and pro-mode toggle #3708 taught the thinking policy to resolve mappedToModel aliases, both getThinkingPolicyForModel and hasExplicitThinkingPolicy inlined the identical getExplicitThinkingPolicy(resolveModelForMetadata(modelString, providersConfig ?? null)) call. Extracted into a private getExplicitThinkingPolicyForModel(modelString, providersConfig) helper.

7. Dedupe queue entry clear-callback projection (messageQueue.ts). After 🤖 feat: queue messages behind special sends instead of erroring (FIFO message queue) #3696 rewrote MessageQueue into FIFO QueueEntry items, both getClearCallbacks and removeWorkspaceTurn inlined the identical spread that builds a QueueClearCallbacks object from an entry's optional onCanceled / onAcceptedPreStreamFailure fields. Extracted into a private entryClearCallbacks(entry) helper.

8. Dedupe the OpenAI-origin model check in openaiExplicitPromptCachingAvailable (cacheStrategy.ts). After 🤖 feat: GPT-5.6 explicit prompt cache breakpoints for direct OpenAI #3712 added the GPT-5.6 explicit-prompt-caching eligibility gate, the function inlined the identical split(":", 2) + origin !== "openai" || !modelName check twice — once for the request model and once for the resolved capability target — and the destructured origin/modelName locals were unused past their guard in both places. Extracted into a module-private isOpenAIOriginModel(canonical) helper.

9. Dedupe the tool-call-execution-start emit in StreamManager (streamManager.ts). 🤖 fix: start tool elapsed timers when execute() actually runs #3716 introduced the ToolCallExecutionStartEvent, emitted from two places: applyToolExecutionStart (part already stored) and the "tool-call" case that consumes a pendingExecutionStart recorded before the part landed. Both inlined the byte-identical this.emit("tool-call-execution-start", { type, workspaceId, messageId, toolCallId, timestamp } satisfies ToolCallExecutionStartEvent) block, differing only in the toolCallId/timestamp source. Extracted into a private emitToolCallExecutionStart(workspaceId, streamInfo, toolCallId, timestamp) helper.

10. Dedupe model-parameter extras merge (aiService.ts). After 🤖 feat: apply mid-turn thinking-level changes at the next model step #3718 added mid-turn thinking-level rebuilds, the initial-model path and the fallback-model path each inlined a byte-identical closure (mergeModelParameterExtras / mergeNextModelParameterExtras) that folds providers.jsonc providerExtras UNDER the Mux-built provider-options namespace (short-circuiting when there are no extras, deep-merging via mergeProviderExtrasUnderMux when the namespace is a plain object). The two differed only in the namespace key and the overrides source. Extracted into a module-level makeModelParameterExtrasMerger(namespaceKey, providerExtras) factory that returns the merger closure.

11. Unify the legacy tool_search part-rename helper (toolCatalog.ts). 🤖 fix: avoid OpenAI tool search name collision #3719 renamed the built-in tool-search tool to tool_catalog_search and added request-time rewriting of historical tool_search call/result parts. It introduced two byte-identical helpers — renameLegacyToolSearchCallPart(part: ToolCallPart) and renameLegacyToolSearchResultPart(part: ToolResultPart) — that differ only in the part type; the rename body is identical. Collapsed both into a single generic renameLegacyToolSearchPart<T extends { toolName: string }>(part: T) and dropped the now-unused ToolCallPart / ToolResultPart imports.

12. Trim duplicated context-cap rationale comment (codexOAuth.ts). 🤖 fix: cap GPT-5.6 context over Codex OAuth #3724 added the GPT-5.6 family to CODEX_OAUTH_CONTEXT_WINDOW_OVERRIDES and rewrote the map's inline comment with a sentence that restated the rationale already given in the map's doc comment directly above it. Dropped the duplicated rationale sentence, keeping only the tier-specific explanation. Comment-only; no behavior change. (Re-applied on top of [openai] 🤖 fix: use 372K GPT-5.6 OAuth context #3730, which later rewrote the same inline comment and re-introduced the duplicate.)

13. Dedupe flat-section pinned block resolution (pinnedReorder.ts). 🤖 feat: project-less scratch chats #3723 added a scratch branch to locatePinnedBlock that renders scratch chats as one flat "Chats" section, mirroring the existing multi-project branch. Both branches inlined the byte-identical collectFlatSectionRows(...).filter(isWorkspacePinned).map((row) => row.id) projection, the if (!pinnedIds.includes(meta.id)) return null guard, and the return { fullOrder: pinnedIds, blockIds: pinnedIds } shape — differing only in the includeRow predicate ((row) => row.kind === "scratch" vs isMultiProject). Extracted into a private locateFlatSectionPinnedBlock(meta, sortedWorkspacesByProject, includeRow) helper.

14. Dedupe JSON-wrapped tool-output unwrap (workflowRunMessages.ts). 🤖 fix: stop terminal workflow await loops #3725 added isTerminalWorkflowRunToolOutput, which re-inlined the byte-identical output.type === "json" && "value" in output container check already used by stripWorkflowRunRecordForModel to detect the { type: "json", value } SDK/UI wrapper before recursing on the inner value. Extracted the check into a module-private isJsonWrappedOutput(output) helper that both functions call, moving the shared rationale into the helper's doc comment. No control flow or return-shape change.

15. Hoist errorType local in finalizeWorkspaceTurnFromStreamError (taskService.ts). 🤖 fix: keep workspace-turn handles running through auto-retryable stream errors #3729 reworked workspace-turn stream-error settlement, and the reworked function read event.errorType three times and repeated the event.errorType != null guard once for the explicitRecovery computation and once in the recovery if. Hoisted a single const errorType = event.errorType and routed all uses through it, deduplicating the repeated member access and null guard. ErrorEvent is a Zod-inferred plain data type, so the property read has no side effects; pure behavior-preserving simplification with no control-flow change.

16. Extract buildSkillDescriptor helper for skill discovery (common/orpc/schemas/agentSkill.ts + agent_skill_list.ts + agentSkillsService.ts). 🤖 feat: skills refresh — invocation control, $ARGUMENTS, dynamic context, .claude compat #3728 (skills refresh) added user-invocable / argument-hint / when_to_use frontmatter and normalized them via resolveSkillAdvertise / resolveSkillUserInvocable / resolveSkillWhenToUse. Both descriptor-building sites — readSkillDescriptor (the agent_skill_list tool) and readSkillDescriptorFromDir (agentSkillsService discovery) — then inlined the byte-identical 7-field object literal mapping parsed.frontmatter + scope into an AgentSkillDescriptor before AgentSkillDescriptorSchema.safeParse. Extracted the mapping into a shared buildSkillDescriptor(frontmatter, scope) in agentSkill.ts (co-located with the resolveSkill* helpers it calls) and dropped the now-unused resolveSkill* imports at both call sites. Callers still run safeParse themselves since they handle validation failure differently. No behavior change.

17. Hoist duplicated Date.parse(record.createdAt) in the bash monitor delivery gate (workspaceService.ts). 🤖 fix: defer bash monitor wakes during task_await #3732 (defer bash monitor wakes during task_await) reworked the delivery gate in drainBashMonitorWakes so a match is re-checked against the shown frontier while pinned to its originating process instance via Date.parse(record.createdAt). The new non-blocking getMonitorWakeDeliveryState branch and the fallback getSettledShownThroughOffset branch each inlined the identical Date.parse(record.createdAt) call as the originNotAfterMs argument. Hoisted a single const originNotAfterMs = Date.parse(record.createdAt) before the branches (with a clarifying comment on why the origin timestamp pins the check) and routed both calls through it. Date.parse is pure, so the hoist is behavior-preserving.

18. Dedupe the "wait for any in-flight load" block in DevToolsService (devToolsService.ts). 🤖 fix: clean up devtools.jsonl on archive/remove and reap orphaned session dirs #3733 added removeWorkspaceData (archive/remove DevTools cleanup) directly beside the existing clear; both inlined the byte-identical const pendingLoad = this.loadingPromises.get(workspaceId); if (pendingLoad) { await pendingLoad; } guard that drains any in-flight loadFromDisk before mutating in-memory state so a late load cannot repopulate stale data after the mutation. Extracted into a private awaitPendingLoad(workspaceId) helper with the shared rationale in its doc comment; both call sites keep their situational one-line comment. No control-flow change.

19. Dedupe MCP OAuth redirect URI resolution (router.ts). Both the global (mcpOauth.startServerFlow) and per-project (projects.mcpOauth.startServerFlow) handlers inlined the byte-identical block that derives the OAuth callback redirectUri from request headers — preferring the Origin header (used verbatim when it parses as a URL), then falling back to x-forwarded-host/host with the forwarded proto (defaulting to http), and returning Err("Missing Host header") when no usable Host header exists. Extracted into a module-level resolveMcpOauthRedirectUri(headers) helper that returns the resolved URI or undefined; each handler now maps undefined to the same Err. startServerFlow is async (its returned promise is passed through unawaited), so moving the call out of the origin-branch try cannot change behavior — the try only ever guarded the synchronous new URL(...) construction. No header semantics or return-shape change.

20. Extract getTotalTokens helper for total-token sums (usageAggregator.ts + 6 call sites). 🤖 feat: show per-model cost breakdown in workspace Costs tab #3739 (per-model cost breakdown in the Costs tab) added a fifth+ copy of the "sum every usage component" expression — input + cached + cacheCreate + output + reasoning .tokens — already inlined byte-for-byte in CostsTab (session model rows), WorkspaceStore (session totalTokens), tokenMeterUtils (calculateTokenMeterData total), sessionUsageService (per-model totalTokens accumulation), and twice in cli/run.ts (budget hasTokens gates). Added a getTotalTokens(usage) helper in usageAggregator.ts, co-located with and mirroring the existing getTotalCost (same five-component iteration; undefined → 0), and routed all six sites through it. The four-component sum in cli/debug/costs.ts (which omits cacheCreate) was intentionally left untouched to preserve its existing behavior.

21. Hoist the duplicated dedupeKeys snapshot in removeByDedupeKeyPrefix (messageQueue.ts). 🤖 refactor: support incremental subagent reports #3714 (incremental sub-agent reports) added MessageQueue.removeByDedupeKeyPrefix, which spread the entry's dedupeKeys Set into an array once for the matchingKeys prefix filter and then re-spread the same Set inside the entry.messages.filter(...) callback — once per message iteration — to map each message index back to its dedupe key. The Set is not mutated until after keptMessages is computed, so both reads observe the same ordered snapshot. Hoisted a single const dedupeKeyList = [...entry.dedupeKeys] before the filter and routed both reads through it, eliminating the per-message re-spread. Pure behavior-preserving simplification with no control-flow change.

22. Dedupe the settled workspace-turn reconciliation guard (taskService.ts). 🤖 fix: report live workspace-turn state from task_await instead of stale settlements #3738 (report live workspace-turn state from task_await) added two read-time reconciliation helpers — persistRepairedSettledWorkspaceTurn and reviveRetryingWorkspaceTurn — that each open their settlement-lock body with the byte-identical guard: reload the handle via getWorkspaceTurn and return current unless it is still the exact record we reconciled against (current != null && current.status === record.status && current.updatedAt === record.updatedAt). Extracted the condition into a module-level isReconciledWorkspaceTurnUnchanged(current, record) type guard, co-located with 🤖 fix: report live workspace-turn state from task_await instead of stale settlements #3738's own isSelfHealEligibleSettledWorkspaceTurn, so the generic "compare updatedAt too, not just status" rationale lives in one doc comment while each call site keeps its situational note. The current is WorkspaceTurnTaskHandleRecord return type preserves the non-null narrowing that reviveRetryingWorkspaceTurn relies on after the guard. No control-flow or return-shape change.

23. Dedupe the fire-and-forget archive-all catch (TaskGroupListItem.tsx). 🤖 fix: archive all sidebar variants #3741 (archive all sidebar variants) added an onArchiveAll prop invoked from two places — the archive keyboard-shortcut branch in onKeyDown and the Archive all variants context-menu item's onClick. Both inlined the byte-identical fire-and-forget props.onArchiveAll(...).catch(() => { /* the sidebar owner surfaces archive failures through its shared error UI */ }) block, differing only in optional-call syntax (inert because the prop is defined in both branches). Extracted a local archiveAll(buttonElement) helper so the swallow-and-surface rationale lives in one place. No control flow, arguments, or error-handling change.

24. Extract someDescendantAgentTaskWorkspace helper for sticky-descendant queries (taskService.ts). [tasks] 🤖 feat: support sticky subagents #3744 (sticky subagents) added two adjacent query methods — hasStickyDescendants and hasUnarchivedStickyDescendants — that each rebuilt the agent-task index the same way (loadConfigOrDefault() → buildAgentTaskIndex(cfg) → listDescendantAgentTaskIdsFromIndex(index, workspaceId).some(...)) and differed only in the .some() predicate. Extracted a private someDescendantAgentTaskWorkspace(workspaceId, predicate) helper that resolves each descendant entry and threads it through the predicate (keeping .some() short-circuiting); the two public methods now just supply their predicate and keep their own assert. The helper's descendant != null && predicate(descendant) guard is equivalent to the prior index.byId.get(descendantId)?.taskSticky === true form, so no behavior changes.

25. Drop the duplicated Kimi K3 max-effort rationale (providerOptions.ts). 🤖 feat: add native Kimi K3 support via a new Moonshot AI provider #3737 (native Kimi K3 via a new Moonshot AI provider) added the isKimiK3Model predicate, whose docstring is the authoritative statement that K3 always reasons and supports only the max reasoning effort and that the provider-options branches key off it. Both the Moonshot and OpenRouter branches of buildProviderOptions then restated that same lead sentence verbatim, so the duplicated sentence was trimmed from each while keeping only the branch-specific "send it explicitly" rationale (Moonshot: don't rely on the API default; OpenRouter: enabled: true alone falls back to the unsupported default medium effort). Comment-only; behavior-preserving.

26. Drop the redundant structuredOutput guard at subagent report call sites (taskService.ts). 🤖 feat: present subagent reports in chat #3742 (present subagent reports in chat) extracted formatSubagentReportUserMessage, which already omits structuredOutput from the report envelope when it is undefined (its internal !== undefined conditional spread). Both call sites — the incremental in_progress progress report in the agent_report path and the terminal completed report in deliverReportToParentUnlocked — nonetheless re-implemented that exact ...(report.structuredOutput !== undefined ? { structuredOutput: report.structuredOutput } : {}) guard before handing the value to the helper. Each now forwards report.structuredOutput directly, and the helper documents that it owns the omission. Behavior-preserving: the helper's internal guard yields byte-identical envelope output whether the key is absent or passed explicitly as undefined.

27. Extract isZipMediaType helper for staged attachment media-type checks (supportedAttachmentMediaTypes.ts). 🤖 feat: stage arbitrary pasted/dropped files into the workspace from chat #3746 (stage arbitrary pasted/dropped files) generalized the ZIP-only staged-attachment pipeline to arbitrary files, and in doing so the cast-laden ZIP membership check ZIP_MEDIA_TYPES.includes(normalized as (typeof ZIP_MEDIA_TYPES)[number]) was inlined byte-identically in both isSupportedStagedAttachmentMediaType and getSupportedStagedAttachmentMediaType. Extracted a module-private isZipMediaType(normalized) helper so the as const tuple cast lives in one place; both call sites now read isZipMediaType(normalized). Behavior-preserving.

28. Hoist the duplicated /goal-bypass-for-attachments check in the ChatInput send handler (ChatInput/index.tsx). 🤖 feat: stage arbitrary files from the creation composer #3748 (stage arbitrary files from the creation composer) computed goalCommandBypassedForAttachments (parsed?.type === "goal-set" && attachments.length > 0) verbatim in two mutually-exclusive branches of the send handler: the creation-variant route and the workspace send path (the latter carrying a "mirror the creation-composer bypass" comment). Both branches resolve parsed and attachments identically, so the boolean is now computed once above the routing and both inline copies dropped. Pure, behavior-preserving.

29. Dedupe the anchored Anthropic model-id regex construction (ai/models.ts). The ANTHROPIC_NATIVE_1M_PATTERNS / ANTHROPIC_BETA_1M_PATTERNS lists that back getAnthropic1MContextMode each spelled out new RegExp(`^<id>${OPTIONAL_VERSION_SUFFIX}$`, "i") per entry — ten near-identical copies restating the anchoring, the optional dated-snapshot suffix interpolation, and the case-insensitive flag, so every new model (Opus 5 in 🤖 feat: add support for Claude Opus 5 #3750 being the latest) had to repeat the whole construction. Extracted a module-private anthropicModelIdPattern(baseModelId) helper and reduced both lists to base model-id strings mapped through it. Generated regex sources and flags are byte-identical to before, and base ids are literal model names with no regex metacharacters, so matching is unchanged.

30. Dedupe MCP header telemetry flag derivation (orpc/router.ts). Every mcp_server_config_changed capture site recomputed the has_headers / uses_secret_headers payload flags inline — eight copies across the global and per-project MCP routers (add, remove, setEnabled, setToolAllowlist). Two variants existed: an input-based one (Boolean(input.headers && Object.keys(input.headers).length > 0) plus the "secret" in v scan) and a server-based one that prefixed both with a server.transport !== "stdio" guard, needed because headers only exists on the HTTP-ish arm of the MCPServerInfo union. Extracted describeMcpHeaderTelemetry(headers) and a thin describeMcpServerHeaderTelemetry(server) wrapper that returns { hasHeaders: false, usesSecretHeaders: false } for stdio — exactly what transport !== "stdio" && … already evaluated to — so the "what counts as a secret header" rationale lives in one doc comment. Behavior-preserving; ~80 lines removed.

31. Extract _child_dirs helper for job folder discovery (benchmarks/terminal_bench/prepare_leaderboard_submission.py). find_job_folders walked directory trees with three copies of the same "iterate a directory, keep only the subdirectories" pattern: two nested for item in <dir>.iterdir(): if item.is_dir(): job_folders.append(item) loops (the direct jobs/ branch and the per-artifact branch) plus a for artifact_dir in artifacts_dir.iterdir(): if not artifact_dir.is_dir(): continue skip-guard at the top of the per-artifact scan. Extracted a module-level _child_dirs(path) helper returning [child for child in path.iterdir() if child.is_dir()] and rewrote all three sites in terms of it. iterdir() ordering and the resulting job_folders ordering are unchanged; behavior-preserving, 9 lines removed.

32. Drop the redundant "exec" fallback duplication for normalizeAgentId (workspaceModeAi.ts, WorkspaceModeAISync.tsx, ChatInput/index.tsx). normalizeAgentId(value, fallback) in common/utils/agentIds.ts already declares fallback: string = WORKSPACE_DEFAULTS.agentId, and WORKSPACE_DEFAULTS.agentId is "exec". Four call sites in the per-agent workspace AI settings paths nonetheless passed the bare literal "exec", re-hardcoding the centralized default that the signature supplies — exactly the kind of duplicated constant that goes stale if the default agent ever changes (every other "workspace default agent" call site either omits the argument or passes WORKSPACE_DEFAULTS.agentId). All four now omit the argument. With the literal gone, workspaceModeAi.ts's module-private normalizeAgentId(agentId) wrapper existed only to supply that fallback, so it and its aliased normalizeAgentId as normalizeWorkspaceAgentId import were removed in favor of importing normalizeAgentId directly. Behavior-preserving: the omitted argument resolves to the identical string.

33. Name the digest truncation bounds in the timeline mapper (timelineMapper.ts). Dropped — superseded upstream. This cleanup introduced module-private DIGEST_MAX_LENGTH / DIGEST_ELLIPSIS constants so that truncateDigest's 117 no longer silently encoded 120 - "...".length. main then solved the same problem better in 🤖 fix: dedupe timeline preview text and sub-agent rows, tint agent event categories #3861, which lifted the bound into an exported TIMELINE_ROW_DIGEST_MAX_LENGTH = 120 in common/orpc/schemas/timeline.ts and shares it between the mapper and TimelinePanel (which compares text.length === TIMELINE_ROW_DIGEST_MAX_LENGTH to detect truncated rows). Keeping this commit would have re-introduced a module-private duplicate of a constant that is now shared, so the commit was dropped during the rebase rather than force-resolved; the file keeps main's version verbatim. Numbering is left intact so earlier entries still line up.

34. Dedupe the defensive unknown-field reads in the timeline mapper (timelineMapper.ts). 🤖 feat: classify machine-authored turns on the workspace timeline #3756 (machine-authored turn classification) added readMonitorWakeProcesses, which pulls records off muxMetadata and then a displayName off each record. Because muxMetadata crosses the oRPC boundary as any, both reads spelled out the same defensive guard inline — typeof x === "object" && x !== null ? (x as Record<string, unknown>)[field] : undefined — and the pre-existing readMuxMetadataField in the same file carried a third copy of it as an early return, so one file held three hand-rolled versions of "index a field off a value that might not be an object". Extracted a module-private readObjectField(value, field): unknown and rewrote all three sites in terms of it, following the precedent already set by getWorkflowResultField in common/utils/workflowRunMessages.ts and readPreviewText in timelineService.ts. Behavior-preserving: all three guards admitted exactly the same shapes (non-null objects, arrays included, functions excluded by typeof), and each caller still applies its own narrowing afterwards (typeof value === "string" for the metadata fields, Array.isArray for records), so every input maps to the same result as before. 14 lines removed, 11 added; no new exports.

35. Share the mobile-touch media query constant (constants/layout.ts, App.tsx, WorkspaceMenuBar.tsx, WorkspaceShell.tsx, ChatInput/index.tsx, UserMessage.tsx). 🤖 feat: redesign workspace chrome (footer info bar, title header, creation hero, composer) #3753 (workspace chrome redesign) leaned further on Mux's mobile-affordance gate, and the string that defines it — (max-width: 768px) and (pointer: coarse) — was copied verbatim into seven window.matchMedia(...) call sites across five renderer files (the sidebar width override, both handleOpenTerminal popout branches, the menu bar's isTouchMobileScreen, UserMessage's isMobileTouch, and the composer's useState initializer plus its change-listener effect). Each copy was independently responsible for staying in sync with the matching @media block in globals.css, which is the actual source of truth for the styles these branches mirror. Hoisted to an exported MOBILE_TOUCH_MEDIA_QUERY in src/constants/layout.ts — directly above MOBILE_TOUCH_TARGET_PX, which already documents this same coarse-pointer environment — and rewrote all seven call sites to use it. Behavior-preserving: every call site passed a byte-identical literal (verified by an exact-match grep over src/), so each matchMedia call receives precisely the value it did before; the only other diff is Prettier rejoining four now-shorter expressions onto single lines. Four of the five consumers already imported from @/constants/layout, so this adds just one new import statement.

36. Share the docked toast overlay placement class (constants/layout.ts, ConnectionStatusToast.tsx, ChatInputToast.tsx, ChatInput/index.tsx). Three toast hosts each hard-coded the same absolute overlay box — pointer-events-none absolute right-[15px] bottom-full left-[15px] z-[1000] mb-2 [&>*]:pointer-events-auto — as two identically-named local wrapperClassName constants plus one inline className on the composer's toast stack. ConnectionStatusToast's own doc comment asserts that it "uses the same overlay placement as ChatInputToast", so the invariant was real but enforced only by copy-paste, and the composer renders both components with wrap={false} under a third copy of the box — so a drifting inset would misalign a toast depending on which host happened to wrap it. Hoisted to CHAT_DOCK_TOAST_OVERLAY_CLASS in src/constants/layout.ts, directly below CHAT_DOCK_GUTTER_CLASS whose 15px inset it mirrors. Behavior-preserving: the two wrapperClassName sites now reference the identical string they previously defined locally, and the composer's stack composes it as cn(CHAT_DOCK_TOAST_OVERLAY_CLASS, "flex flex-col gap-2") — the same utility set with no conflicting utilities, so tailwind-merge yields the same computed styles (only the class attribute's token order shifts). The value stays a literal string in layout.ts because Tailwind scans source text, the same constraint already documented on CHAT_DOCK_GUTTER_CLASS.

37. Share the primary mouse button guard (browser/utils/events.ts, ChatPane.tsx, DiffRenderer.tsx). 🤖 fix: keep iPad composer clicks from selecting the whole transcript #3759's new composer-dock mousedown handler gated on the bare magic number event.button !== 0 — the same primary-button guard the diff review drag-select handler already spelled out inline. Named the check once as isPrimaryMouseButton(event) in browser/utils/events.ts, next to the existing isEventFromDialogPortal / stopKeyboardPropagation event helpers, and pointed both call sites at it so each reads as intent instead of a DOM constant. The helper accepts both React synthetic and native mouse events.

38. Name the ModelSelector row's selection/highlight state (ModelSelector.tsx). 🤖 fix: align composer pickers and size local workers by memory #3760 reworked the model dropdown option row to share composerPickerOptionClass with AgentModePicker and to accent the selected row. In the process the row grew to recompute value === model four separate times — for the option class's isSelected, for aria-selected, for the ProviderIcon's text-accent/text-muted ternary, and for the model-name span's accent — plus index === highlightedIndex twice (data-highlighted and the option class's isHighlighted). Hoisted both into isSelected / isHighlighted locals at the top of the map callback, matching the naming the sibling AgentModePicker already uses for the same two states, so the row's state is named once and the four accent/ARIA consumers cannot drift apart.

39. Share the workspace footer pill class (WorkspaceFooterBar.tsx). 🤖 feat: link the footer GitHub slug to the repository #3762 turned the footer's GitHub owner/repo slug into a link and — per its own PR description — styled it "to match the sibling Last prompt pill", which in practice meant copying that pill's 13-utility Tailwind string verbatim into the new <a>. The two strings then differed only by the three <button> resets (cursor-pointer, border-0, bg-transparent), so any future restyle of one pill would silently drift from the other. Extracted the shared styling into a module-level FOOTER_PILL_CLASS: the anchor consumes it directly, and the button composes it as cn(FOOTER_PILL_CLASS, "cursor-pointer border-0 bg-transparent").

40. Share the safe inactive-animation pause install (browser/utils/inactiveAnimations.ts, main.tsx, terminal-window.tsx). 🤖 perf: reduce idle dev CPU usage #3768 (reduce idle dev CPU usage) added installInactiveAnimationPause and wired it into both renderer entrypoints. Because the pause is a pure optimization that must not be able to take startup down with it — AGENTS.md's "startup-time initialization must never crash the app" rule — each entrypoint wrapped the call in its own five-line try/catch, and the two blocks were byte-identical down to the comment (// Animation throttling is an optimization and must never block renderer startup.). Both also discard the disposer the installer returns, so the duplication was pure ceremony repeated per entrypoint rather than anything either window customized. Extracted installInactiveAnimationPauseSafely() into the installer's own module, directly beneath installInactiveAnimationPause, so the swallow policy is stated once where the risk lives and any future entrypoint (or a third window) inherits it by calling one function. The doc comment records both the "never fail startup" rationale and the deliberate disposer drop, which was previously implicit in the call sites.

41. Share the bash monitor wake message predicate (utils/messages/messageUtils.ts, ChatPane.tsx, MessageRenderer.tsx). 🤖 fix: quiet monitor wake events in chat #3779 gave background monitor wakes their own quiet transcript presentation, which split one concept — "this persisted user turn is a machine-authored monitor event, not a human prompt" — across two files that each re-derived it inline. MessageRenderer routes on message.bashMonitorWake != null to pick BashMonitorWakeMessage over UserMessage; ChatPane's userMessageNavigationByHistoryId memo independently filters on message.bashMonitorWake == null so the prev/next prompt arrows skip wakes. The two tests are the same classification written twice with opposite polarity, and they have to stay in agreement: if only one is updated when the wake representation changes, the transcript renders a quiet event that the navigation arrows still count as a prompt (or vice versa), which is a silent UX bug rather than a type error. Extracted isBashMonitorWakeMessage(message: DisplayedUserMessage) into messageUtils, alongside the existing DisplayedMessage predicates (shouldShowInterruptedBarrier, computeBashOutputGroupInfos), and pointed both call sites at it. Both files already imported from messageUtils, so this adds no new module edge.

42. Dedupe the monitor disposition branch in terminate() (backgroundProcessManager.ts). 🤖 fix: cancel stale background monitor wakes #3776 (cancel stale background monitor wakes) gave BackgroundProcessManager.terminate() a new options.monitorDisposition parameter and, to honour it, inlined the same five-line if (shouldFlushMonitor) { this.stopMonitor(proc, true); } else { this.cancelMonitor(proc); } block at both of the method's two monitor-retirement points: the idempotent already-terminated shortcut and the live-kill path inside the try. The two copies are byte-identical and must stay that way — they answer the same question ("does this caller still want a wake?") for the same process. Replaced both with a resolveMonitorForTermination(proc, shouldFlush) private helper placed next to stopMonitor / cancelMonitor, so the disposition rule lives in one place and the two exit paths cannot drift apart.

43. Share the queued-message action button classes (QueuedMessage.tsx). Dropped during an earlier run's rebase — superseded upstream. 🤖 feat: refine transient transcript interactions #3790 rewrote this action row wholesale: Edit shrank to an h-6/px-1.5 button and Send now moved into a new queue-status dropdown, so both copies of the deduped class string are gone from main and the shared constant had no second caller left. The file now matches main byte-for-byte. Original rationale: 🤖 fix: restore queued message text hierarchy #3781 restored the queued draft's text hierarchy by dropping the Edit and Send now labels from text-xs to text-[11px] — and had to make that one-token edit twice, because both buttons inlined a byte-identical flex h-7 items-center gap-1.5 rounded-md px-2.5 text-[11px] font-medium transition-colors run of geometry/typography utilities and differed only in their colour treatment (text-muted + hover for Edit; bg-pending/10 + disabled states for Send now). Hoisted the shared half into a QUEUED_ACTION_BUTTON_CLASSNAME constant and composed each button's colours on top with cn(...), so the next typography tweak lands in one place instead of drifting between the two.

44. Extract parseSubagentReportFromMessage helper for report-envelope history scans (subagentReportEnvelope.ts + 4 call sites). Subagent report envelopes reach history as synthetic user messages, so every scanner that wants the parsed envelope must first reconstruct the message text. Four sites inlined the byte-identical projection — message.parts.filter((part): part is Extract<typeof part, { type: "text" }> => part.type === "text").map((part) => part.text).join("\n") followed by parseSubagentReportEnvelope(text). 🤖 fix: avoid duplicate subagent completion responses #3783 (avoid duplicate subagent completion responses) added the third and fourth copies in TaskService.findProgressRespondedTaskIds and TaskService.hasAcceptedSubagentProgressReport, joining the pre-existing copies in TaskService's sibling synthetic-report discovery and AgentSession.isVisibleCompletedSubagentReportMessage. Extracted the projection into parseSubagentReportFromMessage(message) in subagentReportEnvelope.ts, co-located with the string parser it wraps, and routed all four sites through it. The MuxMessage import is import type, so the module stays runtime-dependency-free; parseSubagentReportEnvelope remains exported for the callers that already hold a text string (timelineMapper.ts, tests). (Since reduced to two live call sites: 🤖 fix: resume parents directly from sub-agent reports #3816 rewrote TaskService.findProgressRespondedTaskIds and deleted hasAcceptedSubagentProgressReport, so this run's rebase dropped those two hunks and kept main's replacement code verbatim.)

45. Drop stale userOverridable references from the experiments UI (ExperimentsSection.tsx, useExperiments.ts). Two comments still described a userOverridable experiment flag that no longer exists anywhere in src/: the Settings list actually filters on showInSettings !== false, and isExperimentEnabled returns undefined whenever no explicit localStorage override exists, not only for "user-overridable" experiments. Comment-only change.

46. Reuse the isTaskAwaitMessage predicate in the transcript projection (transcriptRenderProjection.ts). 🤖 fix: refine task wait transcript presentation #3788 added task_await poll grouping along with a small family of helpers, three of which re-inlined the same message.type === "tool" && message.toolName === "task_await" shape test that isTaskAwaitMessage already performs — getTaskAwaitResultEntries, hasTaskAwaitCallFailure, and hasTaskAwaitCallInterruption. Widening isTaskAwaitMessage into a type predicate (message is Extract<DisplayedMessage, { type: "tool" }>) lets those three delegate to it and still read the tool-only status / result fields, so the shape test now lives in exactly one place.

47. Dedupe the non-blank string coercion in taskReportLinking (taskReportLinking.ts). Every field this module pulls out of persisted tool args/results is typed unknown, so each read hand-rolled the same three-part check: typeof x === "string", x.trim().length > 0, then use x.trim(). 🤖 feat: expose sub-agent model and thinking level in task schedule and report #3789 and 🤖 feat: show task kind and spawn intent in single-task task_await summary #3793 grew that from four copies to seven (getTaskIdsFromToolResult ×3, getTitleFromTaskToolArgs, the new getAgentTypeFromTaskToolArgs, getBashSpawnTaskId, and getBashSpawnInfoFromArgs). All seven now call one file-local coerceNonBlankString helper.

48. Drop the redundant agentType re-check in task_await awaited rows (TaskToolCall.tsx). While task_await is still in flight, TaskAwaitToolCall builds one awaitedRows entry per pending task and resolves each row's agent type with resolvePersistedAgentId(metadata, "") — an intentionally empty fallback. The very next line already collapses that empty string to undefined (resolvedAgentType.length > 0 ? resolvedAgentType : undefined), but the awaitedRows.push({ ... }) literal then re-tested the same value with agentType && agentType.length > 0 ? agentType : undefined. That second guard cannot change anything: undefined short-circuits to the undefined arm, and any string that survived the first line is non-empty by construction. Replaced with the agentType, shorthand, and left a comment on the normalizing line recording why the empty fallback exists so the intent survives the removed guard.

49. Hoist the duplicated isGrok45Model check in xAI model creation (providerModelFactory.ts). 🤖 fix: honor mapped Grok 4.5 aliases #3804 (mapped Grok 4.5 aliases) introduced a capabilityModel local from resolveModelForMetadata and called isGrok45Model(capabilityModel) to choose between provider.responses(modelId) and provider.chat(modelId). 🤖 feat: default Grok Responses to store=false for ZDR parity #3807 (the store=false ZDR default) then added a second, byte-identical isGrok45Model(capabilityModel) call three lines below, to gate injectGrok45StoreDefault. Both read the same const, and isGrok45Model is a pure regex test over a prefix-stripped string, so the two evaluations are guaranteed to agree. Folding them into a single const isGrok45 drops the redundant call and lets the ternary collapse onto one line. It also restores consistency with buildProviderOptions, which already keeps exactly such an isGrok45 local for the same predicate. Net +2 −3.

50. Share the bash-monitor-wake metadata type guard (messageQueue.ts, agentSession.ts). 🤖 fix: keep workspace turns alive across bash-monitor-wake queue cuts #3797 needed to answer "is this unknown muxMetadata a bash-monitor wake?" in two places — MessageQueue.isNextEntryBashMonitorWake (queue head) and AgentSession.hasPendingBashMonitorWakeContinuation (mid-dispatch, dequeued but not yet streaming) — and each site hand-rolled the check with a different unsound cast: (muxMetadata as Record<string, unknown>).type === "bash-monitor-wake" in one, this.dispatchingQueuedEntryMuxMetadata as MuxMessageMetadata | undefined in the other. The second is the worse of the two: it asserts a unknown field into the full MuxMessageMetadata union purely to reach ?.type. messageQueue.ts already keeps a family of narrow, module-local metadata guards (isCompactionMetadata, isAgentSkillMetadata, isWorkspaceTurnMetadata, hasReviews) built on exactly this shape, so the new code was the odd one out in its own file. Extracted isBashMonitorWakeMetadata alongside those siblings, exported it, and reused it at both call sites. The declared BashMonitorWakeMetadata interface carries only type, matching what the guard actually validates — neither caller reads records, so widening the claim would be dishonest. Net effect: two type assertions removed, one definition of "what a wake looks like" instead of two that can drift, and the file's established guard pattern restored.

51. Share the persisted tool-error message extraction (Shared/toolUtils.tsx, TimelineEventToolCall.tsx, AgentSkillListToolCall.tsx). feat: add a timeline_event transcript card #3814's new timeline card added a private extractErrorMessage, which resolves the two error shapes a transcript card can be handed: the standard { success: false, error } that top-level tools persist, and the bare { error } with no success flag that displayedMessageBuilder reconstructs for a failure inside a nested code_execution/PTC call. AgentSkillListToolCall.toSkillListView already inlined the same resolution — same isToolErrorResult first branch, same !("success" in x) + typeof error === "string" fallback — and both files carried the same rationale comment. Extracted into extractToolErrorMessage(result) in Shared/toolUtils.tsx, co-located with the isToolErrorResult / isFailedToolOutput family it belongs to, and routed both cards through it, dropping their now-unused isToolErrorResult imports.

52. Share the terminal-attention message text join (taskService.ts). 🤖 fix: resume parents directly from sub-agent reports #3816 rewrote sub-agent terminal handoff so the parent resumes directly from an injected report, and in doing so gave TaskService two history scans that both classify rows by parsing a sub-agent report/failure envelope: ensureAgentTerminalMessages (the repair pass that appends a missing report before the wake-up) and consumeRespondedAgentTerminalAttention (the responded-scan that decides whether a pending wake is already answered). Each inlined the identical message.parts flatten — same Extract<…, { type: "text" }> predicate, same .map((part) => part.text), same "\n" separator — before handing the string to parseTerminalSubagentTaskId. Extracted joinMessageText(message) next to that parser so both scans share one definition. These two genuinely must agree: if one drifted on the separator or started admitting a non-text part, the repair pass could append a report the responded-scan can't match, and the parent would be woken to integrate a report already sitting in its history.

53. Share the agent-plugin error diagnostic push (agentPlugins/discovery.ts). 🤖 feat: experimental Agent Plugins 1.0.0 support (skills + MCP) #3815 landed Agent Plugins 1.0.0 discovery, whose §11.3 failure-isolation contract means practically every failure path ends in the same two-step move: log.warn with the owning plugin directory as prefix, then push a severity: "error" diagnostic describing the same problem. That pair was hand-written three times across two functions. Extracted pushErrorDiagnostic so the log prefix, the severity, and the diagnostic shape live in one place.

54. Share the creation project-picker option list (ChatInput/CreationProjectSelect.tsx, CreationControls.tsx, ChatInput/index.tsx). 🤖 feat: add project switcher to the scratch creation page #3817 extracted CreationProjectSelect so the scratch page could reuse the project picker, but it left the options construction inlined at both call sites: each independently did Array.from(userProjects.keys()).map((path) => ({ value: path, label: formatProjectHierarchyLabel(path, userProjects) })). Extracted projectSelectOptions(userProjects) next to the component that consumes it, and gave the option shape a name (CreationProjectOption) so the prop type and the builder agree. These two copies have to stay in lockstep: the trigger renders selectedLabel as an explicit child rather than letting Radix mirror the matched <SelectItem/>, so if one call site's label formatting drifted from the other's, the picker would show a label that never matches any item in its own dropdown.

55. Share the MCP raw-tool wrapper across both startup paths (mcpServerManager.ts). 🤖 feat: migrate MCP client to official TS SDK v2 with 2026-07-28 spec support #3822's SDK v2 migration gave 2026-07-28 connections an on-demand refreshTools, and because refreshing re-wraps a freshly listed tool set, each startup path needed its wrapping options in a reusable place. Both paths solved it the same way and landed on byte-identical local lambdas — startStdioInstance at the old L2191 and startRemoteInstance at the old L2436, each wrapMCPTools(raw, { onActivity, onClosed: () => { if (instanceRef.current) instanceRef.current.isClosed = true; } }), closing over identically-typed onActivity: () => void and instanceRef: { current: MCPServerInstance | null } locals. Hoisted to a module-level createRawToolWrapper(onActivity, instanceRef) returning that same closure, so both call sites collapse to const wrapRawTools = createRawToolWrapper(onActivity, instanceRef).

56. Share the augmented exec-failure error construction (disposableExec.ts). 🤖 feat: back up Mux settings to a git repository #3767 needed execFileAsync to grow abort signals, timeouts, output caps, and process-tree kills, and that growth left the module building its rejection value in three separate places: the new pre-execution abort guard, plus the close handlers of execAsync and execFileAsync. Each one hand-wrote the same inline as Error & { code, signal, stdout, stderr } cast and then assigned the same four properties. Extracted createExecError(message, outcome) behind a named ExecError interface so the rejection contract is declared once instead of three times.

57. Dedupe the failed project-result construction in applyProjectPatch (tools/task_apply_git_patch.ts). The multi-project patch-apply path grew one guard at a time — path-component validation, artifact resolution, expected-HEAD checks, dry-run worktrees, dirty-overlap rejection, conflict-recovery notes — and every new guard hand-wrote another { success: false, projectResult: { projectPath, projectName, status: "failed", ... } } literal, reaching ten copies of the same six-line preamble inside a single function. Extracted a local failed(details) helper whose parameter is Omit<TaskApplyGitPatchProjectResult, "projectPath" | "projectName" | "status">, so the project identity and the "failed" status cannot be overridden from a call site and each branch now spells out only the error / conflictPaths / failedPatchSubject / note that actually vary. Net −59 lines; the two success returns (status: "applied") are deliberately left inline, since they share only two of the fields.

58. Dedupe the popover-error anchor math (hooks/usePopoverError.ts + ProjectSidebar.tsx, WorkspaceMenuBar.tsx, ArchivedWorkspaces.tsx). Ten call sites each re-derived the same anchor before handing it to usePopoverError's showError: convert the trigger's getBoundingClientRect() into document space via rect.top + window.scrollY, then offset a bare 10 past rect.right. The clones had drifted into four different spellings of one expression — a let anchor + if (buttonElement) block (×3 in ProjectSidebar), an IIFE inside a ternary, a rect ? {…} : undefined inline ternary (×2 in WorkspaceMenuBar), and an if (anchorEl) showError(…, {…}) else showError(…) if/else in ArchivedWorkspaces — which is exactly the shape that silently disagrees once someone tunes the gutter. Extract resolvePopoverErrorAnchor into usePopoverError, which already owns the anchor type and the showError signature that consumes it, name the gutter as POPOVER_ERROR_ANCHOR_GAP_PX, and give the repeated inline { top: number; left: number } shape a PopoverErrorAnchor alias.

59. Dedupe the notification settings panel (WorkspaceMenuBar.tsx). The notify-on-response control set — the two Checkbox labels plus the "Agents can also notify on specific events" docs paragraph — was inlined twice verbatim: once inside the hover TooltipContent and once inside the click PopoverContent of the same Popover. The two copies were 34 lines each and byte-identical once indentation is stripped. Extracted to a single notificationSettingsContent element rendered in both slots. Net −28 lines.

60. Dedupe the persisted-choice normalization (Analytics/AnalyticsDashboard.tsx). The dashboard header persists three independent selections — time range, timing metric, and (as of 🤖 feat: add analytics timezone mode #3839) timezone mode — and each had its own normalize* function with an identical body: a typeof value === "string" guard, a Set membership test, a cast to the union, and a literal fallback. A single generic normalizePersistedChoice<T extends string>(value, validValues, fallback) now covers all three, so the next persisted dropdown adds a call instead of a fourth copy of the same three lines.

61. Extract mergeFetchHeaders for provider fetch wrappers (services/providerModelFactory.ts). Three of the file's fetch wrappers — the DevTools-header stripper in getProviderFetch, the Copilot wrapper, and the Coder wrapper added by 🤖 feat: add Coder provider with "Login with Coder" OAuth #3835 — each opened with the same six-line block to compute the request's effective headers: seed a Headers from input.headers when input instanceof Request, then overlay init.headers so per-call values win. One local helper now serves all three.

62. Dedupe unknown-to-record narrowing in PRStatusStore (browser/stores/PRStatusStore.ts). 🤖 feat: stack-aware PR indicator with gh-stack dropdown #3840 (stack-aware PR indicator) added parseStackViewOutput and mergeStackPullRequestMetadata, which walk untrusted gh stack view / gh api graphql output field by field. Between them they introduced seven new copies of the same two-step narrowing — an if (typeof x !== "object" || x === null) return <bail>; guard immediately followed by an x as Record<string, unknown> cast — and mergeStackPullRequestMetadata re-inlined the three-level data → repository GraphQL envelope descent that fetchMergeQueueEntry already had. Extracted a module-private asRecord(value) returning Record<string, unknown> | null (matching the existing per-module isRecord convention in stableStringify.ts, toolOutputUiOnly.ts, transcriptShare.ts, and workflowReportPayload.ts rather than adding a cross-layer import), plus a graphqlRepositoryFields(raw) that composes it for the data.repository descent. Routed all eleven sites in the file through them, including the four that predate 🤖 feat: stack-aware PR indicator with gh-stack dropdown #3840 (summarizeStatusCheckRollup, parseMergeQueueEntry, and the two levels inside fetchMergeQueueEntry). asRecord reproduces the original predicate exactly — notably it still lets arrays through, since typeof [] === "object" — so every call site keeps rejecting arrays via the field checks that follow it rather than at the guard. Net −17 lines.

63. Dedupe the native-xhigh label branches in getThinkingDisplayLabel (common/types/thinking.ts). 🤖 feat: make Grok 4.6 the default Grok model #3842 (Grok 4.6 as the default Grok) taught the label helper that Grok 4.6 exposes xhigh as a real reasoning effort, and did so by appending a second if (level === "xhigh" && <predicate>(modelString)) return "XHIGH"; line directly beneath the existing Anthropic Opus 4.7+ one. Both guards repeat the same level === "xhigh" test and both branches return the same literal, so they collapse into one condition with an || chain over the two model predicates, and the two one-line situational comments merge into a single explanation of what the branch is for. anthropicSupportsNativeXhigh and isGrok46Model are pure regex tests over the prefix-stripped model string, and || short-circuits in the original Anthropic-first order, so the fall-through to THINKING_DISPLAY_LABELS is unchanged.

64. Extract rebuild_native_module in the postinstall script (scripts/postinstall.sh). The script's last two sections rebuilt node-pty and DuckDB for Electron's ABI with structurally identical blocks: the same stamp-file short-circuit and "already rebuilt … – skipping" log, the same $REBUILD_CMD @electron/rebuild -f -m <path> invocation, the same four-line non-fatal failure handler (warn, point at make rebuild-native, exit 0), the same touch of the stamp, and the same "rebuilt successfully (cached at …)" log. They differed only in three values — display label, module path, and stamp file — so they collapse into one rebuild_native_module label module_path stamp_file helper called twice, leaving each call site's else branch ("package missing – skipping") in place. The failure path deliberately keeps exit 0 inside the function rather than return-ing: a rebuild failure must abort the whole script with success (a broken rebuild degrades desktop terminal/DB features but must not fail bun install), and that intent is now stated in a comment above the helper.

65. Dedupe the Gemini Flash thinking-level tier prefixes (common/utils/thinking/policy.ts). isGeminiFlashThinkingLevelModelName answers "does this bare model ID use Google's thinkingLevel config instead of Gemini 2.x thinkingBudget?", and it had grown one hand-written normalized.startsWith(tier) && !normalized.startsWith(tier + "-lite") clause per Flash release — 3.5, then 3.6, and now 3.7 in 🤖 feat: make Gemini 3.7 Flash the default Gemini Flash model #3845. The three dotted tiers now live in a GEMINI_FLASH_THINKING_LEVEL_TIERS list consumed by a single matchesFlashTierExcludingLite matcher, so the next Flash repoint is a one-line list addition rather than another copy of the prefix/-lite pair. The dashless gemini-3-flash tier is deliberately not folded into the list: it uses a stricter exact-or-dash boundary check (=== "gemini-3-flash" || startsWith("gemini-3-flash-")), and generalizing that to the dotted tiers — or relaxing it to a plain prefix match — would be a logic change rather than a dedupe, so it stays spelled out with a comment saying why.

66. Dedupe the task-group running-work predicate (browser/components/ProjectSidebar/TaskGroupListItem.tsx). TaskGroupListItem spelled out props.isRunActive === true || props.runningCount > 0 in two places: once inside getAggregateVisualState to decide whether the aggregate status dot renders "active", and once as the hasRunningWork local that drives the row's data-running attribute, the green (text-content-success) kind glyph, and the brighter text-content-primary title. Extracted a module-level groupHasRunningWork(props) predicate and routed both sites through it, so the status dot and the icon/title running treatment cannot drift apart if the definition of "working" changes again.

67. Dedupe the plural-suffix logic in the sidebar activity labels (browser/components/AgentListItem/AgentListItem.tsx). 🤖 feat: add setting to hide sub-agents in the left sidebar #3847 (hide sub-agents in the sidebar) added formatHiddenSubAgentsPresentation, which brought two fresh copies of the count === 1 ? "" : "s" ternary into a file that already inlined it in formatSubAgentCount — three hand-written copies of the same regular-noun pluralization rule, each spelled against a differently-named count expression (count, agentCount, summary.subAgentCount), which is what kept them from reading as duplication. Extracted a module-level pluralSuffix(count) helper and routed all three sites through it.

68. Extract splitCoderGatewayModelId for the Coder gateway canonical-identity resolvers (common/constants/coderOAuth.ts). 🤖 feat: generalize Coder AI Gateway integration to arbitrary provider instances #3844 generalized the Coder AI Gateway to arbitrary provider instances, and its two canonical-identity resolvers — resolveCoderWireCanonicalModel (wire origin) and resolveCoderMetadataCanonicalModel (pricing/capability identity) — each opened with a byte-identical 13-line prelude: scan for the <instance>/<model> separator, reject IDs with an empty instance or model half, resolve the instance name through resolveCoderGatewayProvider over both parsed provider lists, and bail when no provider matches. Extracted into a module-private splitCoderGatewayModelId(gatewayModelId, metadata) returning { provider, modelId } or null, and routed both resolvers through it.

69. Dedupe the timeline rule-vs-event row dispatch (browser/features/RightSidebar/Timeline/TimelinePanel.tsx). 🤖 feat: collapse scheduler machinery noise in the timeline panel #3858 taught expanded machinery groups to render the rule rows they absorb, which added a second copy of a branch the flat day list already had: test isRuleKind(getTimelineEventKind(event)), then render either TimelineRuleRow or TimelineEventRow with a byte-identical key / event / selected / onSelect prop set. The two components have had identical prop signatures all along, so the branch was pure dispatch. Extracted a TimelineRow component that performs it once and routed both call sites through it.

70. Extract buildCommandReplacement for slash suggestion trailing spaces (browser/utils/slashCommands/suggestions.ts). Two of the three suggestion builders in this file inlined the same two-step expression to decide whether a completion inserts a trailing space: const appendSpace = definition.appendSpace ?? true; followed by `${base}${appendSpace ? " " : ""}`. buildTopLevelSuggestions used it for top-level commands, and buildSubcommandSuggestions used it for subcommands — the only difference being the replacement base (/${key} vs. the joined /${prefixTokens} ${key} path). Extracted a module-level buildCommandReplacement(base, definition) that owns the appendSpace ?? true default, and moved the rationale for that default (the caret should land where arguments go; appendSpace: false opts out for commands that are complete on their own) into its doc comment. The top-level callback also had `/${definition.key}` written twice — once for display, once as the replacement base — so it now hoists a single display local and passes it as the base, which is the same string it always built.

71. Dedupe the pending patch-artifact seed in GitPatchArtifactService.generate() (node/services/gitPatchArtifactService.ts). Four writers inside generate() each inlined a byte-identical 13-line expression to resolve "the artifact already on disk, or a freshly built pending one": the three workspace-shape guards (workspacePath missing, runtimeConfig missing, workspaceName missing) and the per-project ensureProjectArtifact updater. The three guards additionally repeated the whole surrounding updateArtifact((existing) => failPendingProjectArtifacts({ artifact: …, error: …, updatedAtMs: nowMs })) call, differing only in the error string. Extracted a seedPendingArtifact(existing) closure (defined immediately after const ws = entry.workspace;, so it captures the narrowed entry/ws the seed needs) plus a failGeneration(error) wrapper, collapsing each guard from 20 lines to one. The !entry guard and the outer catch intentionally keep their inline projectArtifacts: [] seeds — neither has a workspace entry in scope to enumerate project repos from, and the outer catch also stamps updatedAtMs: Date.now() rather than nowMs.

72. Dedupe the blob content-address computation in BlobStore (node/utils/journal/blobStore.ts). 🤖 feat: add shared agent-foundation layer (event spine, journal kit, sandbox host, capability grants) #3865 (shared agent-foundation layer) added the journal kit's content-addressed BlobStore, whose put() and get() each inlined the same byte-identical address derivation (sha256 hex digest of a Buffer, prefixed with sha256:) and differed only in what they did with the result: put() names new content before writing it, while get() re-derives the address of the bytes just read and compares it against the requested ref to detect corruption. Extracted a blobRefFor(bytes) helper so the write path and the verification path cannot drift — if the two ever disagreed, every blob would fail its own integrity check on read.

73. Dedupe the sub-agent status presentations into one outcome-keyed table (browser/components/SubAgentTasksDecoration/SubAgentTasksDecoration.tsx). 🤖 feat: port anti-slop lint rules and enable switch exhaustiveness checking #3868 rewrote getSubAgentStatusPresentation so the optional taskExecutionStatus is narrowed before its switch, leaving two sibling switches consulted in sequence. Four presentations — Queued, Running, Completed, Interrupted — were spelled out twice, once per switch, each repeating its label, its lucide-react icon, and its icon class. Both switches satisfy exhaustiveness independently, so the duplication was invisible to the compiler. Extracted a module-level SUB_AGENT_STATUS_PRESENTATIONS table keyed by outcome rather than status (the two status unions are not parallel: taskExecutionStatus: "starting" | "running" both render Running, whereas taskStatus: "starting" renders a distinct Starting), plus a named SubAgentStatusPresentation interface reused as the function's return type.

74. Dedupe the replay/lookup/watermark prelude in SessionTimingService (node/services/sessionTimingService.ts). Five of the eight stream-lifecycle handlers — handleStreamDelta, handleReasoningDelta, handleToolCallStart, handleToolCallDelta, handleToolCallEnd — each opened with the same four lines encoding three invariants: replayed events must never re-time a stream, an event for a workspace with no active stream has nothing to update, and anything surviving both still counts as stream activity. Extracted a private touchActiveStream(data) returning the resolved ActiveStreamState or null, typed Pick<StreamDeltaEvent, "replay" | "workspaceId" | "timestamp"> so all five event shapes match structurally. Named "touch" because the lookup deliberately advances the watermark. handleStreamStart (creates state rather than resolving it) and handleStreamAbort/handleStreamEnd (never had the guard or the bump) were deliberately left alone.

75. Dedupe the analytics date-filter epoch round-trip (browser/hooks/useAnalytics.ts). All nine analytics data-fetching hooks re-derived the same Date -> epoch-ms -> Date round-trip, 18 call sites in total, which exists so each hook's useEffect dependency array holds primitives rather than a Date reference that is fresh on every render. Extracted module-level toEpochMs/fromEpochMs helpers as literal extractions of the duplicated expressions, and recorded that previously-unwritten rationale in a comment.

76. Dedupe the budget-exceeded reporting in the run CLI (cli/run.ts). xum run --budget enforced its limit from three independent points in the chat listener (stream-end, session-usage-delta, and the pre-send precheck), and each one inlined the same three statements: set the budgetExceeded flag, emit the machine-readable budget-exceeded JSON line, and write the yellow human-readable line — including three copies of the format string that defines the user-visible message. Extracted one reportBudgetExceeded(cost, budgetLimit) closure beside the budgetExceeded declaration it writes to; each site becomes a single call. The stream-end site's deliberate omission of session.interruptStream(...) stays at the call site, and the similar-looking budget-error preludes were left alone because their messages genuinely differ.

77. Dedupe the oversized pass-through gate in transformMCPResult (node/services/mcpResultTransform.ts). 🤖 fix: cap MCP tool result text output before it enters history #3891 capped MCP tool result text at 64KB and inlined the same measure-compare-log prelude in both non-standard pass-through branches (toolResult passthrough and content-less result objects); the copies differed only in the log.warn message and the notice wrapper. Extracted oversizedPassthroughBytes(result, logMessage): number | null beside jsonByteLength, returning the measured size so each caller can still build its own notice.

78. Replace agentSession.ts's three open-coded errno checks with the centralized isErrnoWithCode helper (node/services/agentSession.ts). 🤖 fix: re-enable retry when provider config changes after credential errors #3907 added a third copy of the same typeof error === "object" && error !== null && "code" in error narrow-and-compare, joining pre-existing copies in loadAutoRetryEnabledPreference() and persistAutoRetryState(). node/utils/fs.ts already exports isErrnoWithCode for exactly this pattern and it is used at ~20 node-layer call sites, so agentSession.ts was converted rather than growing a fourth local variant.

This run

Cleanup #79 — dedupe the canonical/legacy plan-path resolution in the runtime helpers (src/node/utils/runtime/helpers.ts).

readPlanFile() and hasNonEmptyPlanFile() each opened with the same three lines:

const xumHome = runtime.getXumHome();
const planPath = getPlanFilePath(workspaceName, projectName, xumHome);
const legacyPath = getLegacyPlanFilePath(workspaceId, xumHome);

#3905 (make .xum canonical for project metadata) edited both copies in the same commit — renaming the muxHome local to xumHome in each — which is exactly the "same thing had to be edited twice" signal this agent looks for.

The duplication also encodes a real invariant. hasNonEmptyPlanFile() is the cheap existence probe that decides whether plan-mode UI advertises a plan; readPlanFile() is the reader that serves the content and transparently migrates the legacy by-workspace-id file onto the canonical per-project path. If the two ever disagreed about where a plan lives, the probe would advertise a plan the reader could not find, or vice versa.

Extracted a module-private getPlanFilePaths(runtime, workspaceName, projectName, workspaceId) that returns { planPath, legacyPath }, with a short comment recording why both readers must resolve through one place. Both call sites now destructure from it. Net: 6 lines removed, 12 added (the helper plus its doc comment); the two function bodies each lose their three-line prelude.

Equivalence. The helper makes the same three calls, in the same order, with the same arguments, and is invoked at the same point in each function. In hasNonEmptyPlanFile() it still runs after the defensive empty-identifier guard, so runtime.getXumHome() is still not called for invalid inputs. Neither function referenced the xumHome local anywhere else, and no exported signature, log message, or returned value changes.

movePlanFile() in the same file was deliberately left alone: it resolves two canonical paths (old and new workspace name) and no legacy path, so it does not fit this shape and folding it in would mean a helper with an unused return field.

Considered and rejected

• The two further copies in workspaceService.ts (~L3745 getPlanPathsForWorkspace-style filtering and ~L9660 plan cleanup) — also renamed muxHome -> xumHome by 🤖 refactor: make .xum canonical for project metadata #3905, but the first site additionally needs the home value itself for its xumHome.startsWith("~") tilde-expansion branch, so a shared helper would have to return xumHome as a third field that the other three call sites ignore. Deduping across two modules with a widened return shape is more than one extremely-low-risk step; deferred rather than forced.
• listProjectMetadataRelativePaths / getCanonicalProjectMetadataRelativePath (common/compat/legacyMux.ts, 🤖 refactor: make .xum canonical for project metadata #3905) — the two share a one-line relativePath.length > 0 ? ... : "" suffix computation. Extracting a helper for a single ternary would add indirection without removing meaningful duplication.
• 🤖 fix: preserve auto-compaction follow-ups #3912 (preserve auto-compaction follow-ups) — the fix threads an existing follow-up through compactionHandler.ts/agentSession.ts and lands with 88 lines of new tests. The changed branches are single-use; there is no repeated block to extract.
• 🤖 refactor: make .xum canonical for project metadata #3905's renamed modules (xum_agents_*, xum_config_*, xumignore.ts) — the legacy/canonical fan-out is already funneled through the shared PROJECT_METADATA_DIR_NAMES / listProjectMetadataRelativePaths contracts in common/compat/legacyMux.ts. Nothing left worth a follow-up extraction.
• 🤖 ci: remove unsigned Windows signing fallback #3916 (remove unsigned Windows signing fallback) and 🤖 ci: skip existing xvfb installation #3913 (skip existing xvfb installation) — CI/workflow-only changes, out of scope per the auto-cleanup constraints.

Validation

make static-check passes end to end: ESLint, both tsgo projects, Prettier, shell/Python formatting, shellcheck, hadolint, the generated-file and docs-sync checks, and the code-to-docs link check.

bun test src/node/utils/runtime/helpers.test.ts passes (8 tests, 28 assertions). That suite covers both converted functions' path resolution directly, including the cases that pin the invariant this cleanup protects: readPlanFile falling back to the local, SSH, and Docker runtime-home legacy paths "not a hardcoded ~/.xum root", and the resolved-before-quoting migration command.

No test was added. The change moves three lines into a private helper with no behavioral delta; a test asserting that helpers.ts calls getPlanFilePaths would be tautological.

Rebase note

This run rebased the 74-commit stack onto cabfd0a0 (current main), picking up #3912, #3913, #3905, and #3916. The rebase was clean — no conflicts, no cleanup commits dropped.

Risks

Very low for this run's change: one file, one module-private helper, no control-flow change, no exported API change, and no user-visible string touched. The extracted lines are three pure path-construction calls with no I/O.

The affected product area is plan-file resolution (plan-mode content reads, the legacy-to-canonical plan migration, and the has-a-plan probe behind plan-mode UI). A regression would require the helper to resolve a different path than the inlined code, which is ruled out by the call-for-call equivalence above and pinned by the runtime-home fallback tests.

The broader risk profile of the stack is unchanged: 79 behavior-preserving refactors, intended to be reviewed and squash-merged as one. The branch sits directly on current main.

Auto-cleanup checkpoint: cabfd0a

---

Generated with xum • Model: anthropic:claude-opus-5 • Thinking: xhigh

[mux-bot]

@codex review

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.

[mux-bot]

⚠️ Auto-fixup: CI failure appears to be infrastructure/flaky, not caused by this cleanup commit — no code pushed.

Root cause: The Static Checks job failed with the runner annotation "The self-hosted runner lost communication with the server" after running the full 10m, and its log blob was never uploaded (BlobNotFound). Test / Integration was cancelled and Required failure are both downstream of that (Required only aggregates results).

Verification: Ran make static-check locally on this PR's HEAD (a2454d31b). ESLint, both TypeScript configs (tsgo --noEmit ×2), and Prettier all pass. The only local miss is fmt-shell-check because shfmt isn't installed in the fixup sandbox — and this PR changes no shell files, so it's irrelevant.

Recommendation: Re-run the failed CI jobs. No code change needed.

[mux-bot]

@codex review

Latest push rebases onto origin/main (through #3698) and adds one low-risk, behavior-preserving cleanup: extract the duplicated per-scope "memory scope is full" cap check in MemoryService.create / saveFile into a private assertScopeHasRoom(store, scope) helper. No control flow, thresholds, error types, or events change. make static-check (ESLint + both tsconfigs + Prettier) and bun test memoryService.test.ts (75/0) pass locally.

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.

[mux-bot]

@codex review

New in this run: cleanup #3 — deduped the identical blockquote line-prefixing (> per line, joined by newlines) in buildBashMonitorWakePrompt into a module-level blockquoteLines(lines) helper. Behavior-preserving; validated by the existing buildBashMonitorWakePrompt output-format tests (21/21 pass).

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.

[mux-bot]

@codex review

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.

[mux-bot]

@codex review

Added auto-cleanup #5: getCodexOauthContextLimit now reuses the already-resolved, non-null compatibilityModelId with the isCodexOauthAllowedModelId / isCodexOauthRequiredModelId variants instead of re-invoking isCodexOauthAllowedModel / isCodexOauthRequiredModel (which re-derive the same id and re-scan providersConfig). Behavior-preserving; targeted tests + make static-check (minus sandbox-only shfmt) pass.

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.

[mux-bot]

@codex review

Added cleanup #6: extracted a module-private getAnthropicCacheCreateTokens(metadata) helper in usageHelpers.ts to dedupe two byte-identical Anthropic cache-create token reads inside accumulateProviderMetadata. Behavior-preserving; rebased onto latest main (checkpoint ec47caf).

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.

[mux-bot]

@codex review

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.

[mux-bot]

@codex review

Added cleanup #7: deduped the byte-identical QueueClearCallbacks projection in MessageQueue.getClearCallbacks and removeWorkspaceTurn into a private entryClearCallbacks(entry) helper (behavior-preserving). Rebased onto latest main and advanced the checkpoint to 956ac533e.

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.

[mux-bot]

@codex review

This run adds one behavior-preserving cleanup (#8 in the branch list): dedupes the two identical openai-origin model checks in openaiExplicitPromptCachingAvailable (cacheStrategy.ts, introduced by #3712) into a module-private isOpenAIOriginModel(canonical) helper. The destructured origin/modelName locals were unused past their guard. Rebased onto origin/main (48722b9); make static-check passes (except fmt-shell-check, which needs shfmt unavailable in this env) and the cacheStrategy/providerOptions test suites pass (173).

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.

[mux-bot]

@codex review

This run rebased the stack onto current main (cabfd0a0, picking up #3912, #3913, #3905, #3916 — clean, no conflicts) and added cleanup #79: readPlanFile() and hasNonEmptyPlanFile() in src/node/utils/runtime/helpers.ts shared an identical three-line canonical+legacy plan-path prelude (both edited by #3905 in the same commit), now extracted into a module-private getPlanFilePaths(). Behavior-preserving: same calls, same order, same arguments, same position relative to hasNonEmptyPlanFile()'s defensive guard.

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.

[mux-bot]

Auto-cleanup run: rebased onto cabfd0a0 + cleanup #79 pushed — CI green, Codex gate still unavailable

Commits considered. Checkpoint was 174e8554; origin/main had four new commits: #3912 (preserve auto-compaction follow-ups), #3913 (CI: skip existing xvfb installation), #3905 (make .xum canonical for project metadata), #3916 (CI: remove unsigned Windows signing fallback). Checkpoint advanced to cabfd0a0.

Rebase. The 74-commit stack replayed onto cabfd0a0 cleanly — no conflicts, no cleanup commits dropped.

Cleanup #79 — dedupe the canonical/legacy plan-path resolution (src/node/utils/runtime/helpers.ts, 1 file, +12/−6).

readPlanFile() and hasNonEmptyPlanFile() each opened with the identical three-line prelude resolving the runtime home plus the canonical and legacy plan paths. #3905 had to edit both copies in the same commit (renaming muxHome → xumHome in each). The duplication also encodes a real invariant: hasNonEmptyPlanFile() is the existence probe behind plan-mode UI and readPlanFile() is the reader that migrates the legacy file onto the canonical path, so the two must agree on where a plan lives. Extracted a module-private getPlanFilePaths() returning { planPath, legacyPath }.

Behavior-preserving: same three calls, same order, same arguments, invoked at the same point in each function — in hasNonEmptyPlanFile() still after the defensive empty-identifier guard, so runtime.getXumHome() is still skipped for invalid inputs. movePlanFile() was left alone (two canonical paths, no legacy path). The two further copies in workspaceService.ts were deferred: one of them also needs the home value itself for a tilde-expansion branch, so sharing would mean widening the return shape across two modules — more than one extremely-low-risk step.

Validation. make static-check green end to end locally; bun test src/node/utils/runtime/helpers.test.ts passes (8 tests, 28 assertions), covering the local/SSH/Docker runtime-home legacy fallbacks for both converted paths. All non-visual-review CI checks pass on this push (Static Checks, Test / Unit, Test / Integration, Test / Storybook, Test / E2E (linux, macos), Test / Windows, all Build jobs, Smoke / Server, Required).

Codex gate. Still blocked outside this agent's control: @codex review returns "To use Codex here, create a Codex account and connect to github." No Codex review threads are outstanding (0 unresolved). The PR therefore cannot reach the Codex-approval condition until that connector is configured.

---

Generated with xum • Model: anthropic:claude-opus-5 • Thinking: xhigh