🤖 refactor: auto-cleanup (coder#2942)

## Summary

Periodic auto-cleanup: removes the dead `setPRStatusStoreInstance`
export from `PRStatusStore.ts`.

## Background

The function was exported but never imported or called anywhere in the
codebase. The getter (`getPRStatusStoreInstance`) creates the singleton
on demand; no test or production code ever needed to inject a custom
instance via the setter. It is the only `set*StoreInstance` pattern
across all stores, so there is no convention to maintain.

## Validation

- `make typecheck` — passes
- `make lint` — passes
- `make fmt-check` — passes
- `bun test src/browser/stores/PRStatusStore` — 12/12 pass
- Grep confirms zero references outside the definition site

Auto-cleanup checkpoint: ff743d1

---

_Generated with `mux` • Model: `anthropic:claude-opus-4-6` • Thinking:
`xhigh` • Cost: `$0.00`_

<!-- mux-attribution: model=anthropic:claude-opus-4-6 thinking=xhigh
costs=0.00 -->

Co-authored-by: mux-bot[bot] <264182336+mux-bot[bot]@users.noreply.github.com>

release: v0.20.0

🤖 feat: add per-model provider modelParameters overrides (coder#2743)

## Summary
Add config-driven per-model parameter overrides from
`~/.mux/providers.jsonc` and thread them through the chat streaming path
so standard call settings (for example `max_output_tokens`,
`temperature`, and `top_p`) are applied at request time, while
provider-specific extra keys are safely merged into provider options.

Closes coder#2574

## Background
Users need a way to tune model request parameters without code changes
or UI additions. This adds a providers-config surface that supports
exact model entries plus wildcard fallbacks and keeps runtime behavior
defensive for hand-edited config.

## Implementation
- Added shared schemas and types for `modelParameters` in provider
config, including a single snake_case→AI SDK call-settings mapping
source of truth.
- Added `resolveModelParameterOverrides(...)` with precedence:
  1. effective model id
  2. canonical model id
  3. `"*"` wildcard
- Split resolved values into:
  - `standard` call settings (mapped to SDK keys)
  - `providerExtras` (unknown keys)
- Integrated override resolution in `AIService.streamMessage()` and:
  - deep-merged `providerExtras` under the provider namespace
  - preserved Mux-built provider options as conflict winners
- passed resolved standard settings through `startStream(...,
callSettingsOverrides)`
- Threaded call settings through StreamManager request construction and
into `streamText(...)`, with explicit precedence for `maxOutputTokens`:
  1. explicit per-request value
  2. config override
  3. model stats fallback
- Added debug logs for resolved overrides under
`MUX_DEBUG_LLM_REQUEST=1`.

## Validation
- `make static-check`
- `bun test src/common/config/schemas/providersConfig.test.ts
src/common/utils/ai/modelParameterOverrides.test.ts
src/node/services/streamManager.test.ts
src/node/services/aiService.test.ts`

## Risks
- Medium/contained: the change touches the main chat stream assembly
path and provider option composition.
- Mitigations:
- provider extras are namespace-scoped and merged under existing Mux
options
- standard keys are type-constrained and defensively validated at
resolution time
  - precedence behavior is covered by focused unit tests

---

_Generated with `mux` • Model: `openai:gpt-5.3-codex` • Thinking:
`xhigh` • Cost: `$13.96`_

<!-- mux-attribution: model=openai:gpt-5.3-codex thinking=xhigh
costs=13.96 -->

🤖 feat: merge Costs + Stats into unified Stats tab with sub-tabs (cod…

…er#2729)

## Summary

Merge the standalone "Costs" and "Stats" top-level tabs into a single
unified "Stats" tab with sub-tabs (Cost, Timing, Models).

## Background

Previously the right sidebar had separate "Costs" and "Stats" tabs. The
Stats tab was feature-flagged and duplicated some navigation patterns.
This consolidates them into a single top-level "Stats" tab with a
ToggleGroup for switching between sub-views.

## Implementation

- **Internal key preserved**: The stored tab key remains `"costs"` to
avoid persisted layout migrations. Only the display name changes to
"Stats".
- **Sub-tab container** (`StatsContainer.tsx`): Uses `ToggleGroup` with
"Cost", "Timing", and "Models" options. Timing/Models only appear when
the `statsTab` feature flag is enabled.
- **Extracted panels**: `TimingPanel` and `ModelBreakdownPanel`
extracted from `StatsTab.tsx` via shared `useStatsData` hook.
- **Unified tab label**: Merged `CostsTabLabel` + `StatsTabLabel` into
one component showing both cost and duration badges.
- **Layout migration**: `stripLegacyStatsTab` runs as a pre-parse step
in `parseRightSidebarLayoutState`, stripping `"stats"` from persisted
layouts before `isTabType` validation. This preserves user layout
configuration during upgrade.

---

_Generated with `mux` • Model: `anthropic:claude-opus-4-6` • Thinking:
`xhigh` • Cost: `$6.61`_

<!-- mux-attribution: model=anthropic:claude-opus-4-6 thinking=xhigh
costs=6.61 -->

release: v0.19.0

🤖 feat: add Debugging docs for Output tab (coder#2645)

## Summary

Add documentation for Mux's Output-tab debugging workflow.

- Added a new Reference page: `docs/reference/debugging.mdx`
- Documented command-palette command mode (`>`) before running **Show
Output**/**Hide Output**
- Documented log-level filtering, clearing logs, and opening the raw log
file
- Added navigation entry in `docs/docs.json` under **Reference**
- Added redirect: `/debugging` → `/reference/debugging`
- Regenerated `src/node/builtinSkills/mux-docs.md` via `make fmt`

## Validation

- `make fmt`
- `make static-check`

## Risks

Low risk. Changes are documentation-only plus the expected generated
mux-docs index sync.

---

_Generated with `mux` • Model: `openai:gpt-5.3-codex` • Thinking:
`xhigh` • Cost: `$1.41`_

<!-- mux-attribution: model=openai:gpt-5.3-codex thinking=xhigh
costs=1.41 -->

🤖 fix: eliminate dev-server startup crash from tsgo/tsc-alias race (c…

…oder#2612)

## Summary

Fix the `make dev-server` startup crash (`Cannot find module
'@/node/config'`) that occurs on every cold start on non-Windows
platforms.

## Background

The non-Windows `dev-server` target launches `tsgo -w` and `tsc-alias
-w` as **independent parallel watchers**. On startup, tsgo immediately
rebuilds `dist/` with unresolved `@/` path aliases, and nodemon restarts
the server after its 500ms delay — before tsc-alias finishes rewriting
all ~415 files. This produces a transient `MODULE_NOT_FOUND` crash every
time.

The Windows `dev-server` target already avoids this by using
`scripts/build-main-watch.js`, which runs tsgo → tsc-alias
**sequentially**.

## Implementation

Replace the parallel watcher pair with a source-watching nodemon that
executes the existing sequential `build-main-watch.js` script on each
change — the same approach Windows already uses. This is a one-line
Makefile change (plus a comment).

## Validation

- `make build-main` passes
- `make dev-server` starts without the `Cannot find module
'@/node/config'` crash
- Source file changes trigger sequential rebuild → clean server restart

---

_Generated with `mux` • Model: `anthropic:claude-opus-4-6` • Thinking:
`xhigh` • Cost: `$2.23`_

<!-- mux-attribution: model=anthropic:claude-opus-4-6 thinking=xhigh
costs=2.23 -->

🤖 refactor: apply react-effects guidance to runtime settings (coder#2570

)

## Summary

Apply `react-effects` skill guidance to eliminate 4 anti-pattern
useEffects and tighten 1 more in the runtime settings UI code. Net
removal of ~90 lines of effect/ref machinery, replaced by render-time
derivation and event-handler colocation.

## Background

An audit of the runtime settings code (`useDraftWorkspaceSettings.ts`,
`RuntimesSection.tsx`) flagged several effects that violate the
project's `react-effects` skill guidelines:

1. **Mode-switch restoration** lived in a `useEffect` watching
`selectedRuntime.mode` via a prev-ref, but mode switches are always
user-initiated events
2. **Project override sync** stored 3 derived values in state and synced
them via a large effect, when they're deterministic functions of
existing data
3. **Scope validity guard** used an effect to reset scope when a project
disappears, when the valid scope can be derived during render
4. **Project/default reset** used two prev-refs to detect project or
default-mode changes and imperatively reset state, solvable with React's
`key` pattern
5. **Default seeding** re-ran on every dep change instead of once per
project mount

## Implementation

### `RuntimesSection.tsx`

- **Scope validity:** Removed the effect. Added `effectiveScope`
computed during render that falls back to `ALL_SCOPE_VALUE` when the
selected project isn't in the projects map. UI binds to
`effectiveScope`.
- **Project override sync:** Removed 3 `useState` + the sync effect.
Added `deriveProjectOverrideState()` pure function called during render.
Handlers (`handleOverrideToggle`, `handleRuntimeToggle`) now rely solely
on the existing cache + version-bump mechanism for optimistic updates.

### `useDraftWorkspaceSettings.ts`

- **Mode-switch restore:** Extracted `mergeRememberedRuntimeConfig()`
pure helper. `setSelectedRuntime` merges remembered fields synchronously
before committing state. Deleted the 50-line effect +
`prevSelectedRuntimeModeRef`.
- **Project reset:** Updated `ChatInput` key in `ProjectPage.tsx` to
include `projectPath`, so project switches trigger natural React
remount. Removed `prevProjectPathRef`. Simplified default-mode tracking
to `appliedDefaultRuntimeChoiceRef`.
- **Default seeding:** Added `seededProjectPathRef` so seeding runs once
per project mount.

## Validation

- `make static-check` passes
- 16/16 focused tests pass (`useDraftWorkspaceSettings` +
`useCreationWorkspace`)

## Risks

- **Key-based remount** (`ProjectPage.tsx`): Adding `projectPath` to the
`ChatInput` key means project switches fully remount the input. This is
the intended behavior (reset draft state), but worth verifying that no
expensive initialization (branch fetching, etc.) causes a visible delay
during project navigation.
- **Optimistic override UI**: Handlers now rely entirely on cache writes
+ version bumps for re-derivation. Verified that both
`handleOverrideToggle` and `handleRuntimeToggle` already write to
`overrideCacheRef` and bump the version counter.

---

_Generated with `mux` • Model: `anthropic:claude-opus-4-6` • Thinking:
`xhigh` • Cost: `$3.22`_

<!-- mux-attribution: model=anthropic:claude-opus-4-6 thinking=xhigh
costs=3.22 -->

🤖 fix: unify SSH clone prompt mediation for clone flows (coder#2544)

## Summary

Adds a unified SSH prompt mediation pipeline for the git clone flow so
that OpenSSH prompts (host-key verification, credential/passphrase) are
surfaced in-app instead of causing cryptic clone failures.

Most of the diff is tests.

## Background

When cloning via SSH (especially with GitHub shorthand like
`owner/repo`), OpenSSH prompts were not visible in the Electron app.
This caused silent hangs and opaque error messages. The fix introduces a
mediated askpass architecture that intercepts SSH prompts, routes them
through `SshPromptService` to the frontend via oRPC subscriptions, and
classifies errors deterministically from structured prompt outcomes.

## Implementation

**Prompt mediation pipeline:**
- `SshPromptService` — manages pending prompt lifecycle, deduplication
by SSH endpoint, timeout, and responder gating
- `openSshPromptMediation.ts` — `createMediatedAskpassSession()` bridges
askpass ↔ `SshPromptService`, classifies prompts (host-key vs
credential), records `MediatedPromptOutcome` for downstream error
classification
- `sshAskpass.ts` — spawns a local HTTP server as `SSH_ASKPASS` target
so OpenSSH can route prompts back to the app
- `SshPromptDialog.tsx` — generic in-app dialog for both host-key and
credential SSH prompts

**Error classification (outcome-first architecture):**
- `sshCloneFailure.ts` — dedicated module with
`classifySshCloneFailure()` that derives error codes from structured
prompt outcomes (timeout, rejection, cancellation) rather than fragile
stderr regex matching; stderr heuristics remain as fallback only
- `summarizeCloneStderr()` — returns last 3 meaningful stderr lines
instead of just the final line, preventing truncated diagnostic messages
like "and the repository exists."

**Canonical SSH transport detection:**
- Introduced `parseCloneTransport()` — a single parser/classifier for
clone URL transport semantics that all SSH-related decisions consume
(askpass enablement + prompt deduplication)
- Recognizes SCP-style (`[user@]host:path`), `ssh://`, `git+ssh://`, and
`ssh+git://` as SSH transports
- `isSshCloneUrl()` and `deriveSshClonePromptDedupeKey()` both delegate
to `parseCloneTransport()` — no duplicated protocol logic
- Previously, `git+ssh://` and `ssh+git://` URLs bypassed askpass setup,
causing opaque clone failures

**Schema/contract tightening:**
- Clone error `code` is now required (not optional) in the oRPC schema,
matching the backend behavior that always emits a code
- Frontend `formatCloneError` accepts required `code`, providing
deterministic error-to-message mapping

**Clone URL handling:**
- `parseScpStyleSshUrl()` handles `[user@]host:path` SCP-style URLs
without requiring a username, with a protocol-URL exclusion guard to
prevent matching `https://...`

## Validation

- 36 projectService tests pass including: host-key accept/reject,
endpoint dedupe for concurrent clones, timeout classification (50ms),
ambiguous transport → `clone_failed`, credential success, credential
cancel → `ssh_credential_cancelled`, multi-line stderr summary,
HTTPS/git:// no-askpass, ssh://+git+ssh://+ssh+git:// askpass injection
- 21 sshPromptService tests pass
- 8 SshPromptDialog tests pass
- `make static-check` passes (typecheck, lint, format, docs link check)

## Risks

- **Low:** SSH askpass HTTP server binds to localhost on a random port
per clone; if the clone process outlives cleanup, the port is leaked
until process exit (mitigated by `using` disposable pattern)
- **Low:** Stderr regex fallbacks could mis-classify novel SSH error
formats; mitigated by preferring structured prompt outcomes when
available

---

_Generated with `mux` • Model: `anthropic:claude-opus-4-6` • Thinking:
`xhigh`_

<!-- mux-attribution: model=anthropic:claude-opus-4-6 thinking=xhigh -->