Skip to content

fix(daemon): cross-client sync follow-up cleanup (epoch-reset resync, approval-mode serialization, catch-up indicator)#4510

Merged
chiga0 merged 35 commits into
daemon_mode_b_mainfrom
fix/daemon-sync-followups-cleanup
May 28, 2026
Merged

fix(daemon): cross-client sync follow-up cleanup (epoch-reset resync, approval-mode serialization, catch-up indicator)#4510
chiga0 merged 35 commits into
daemon_mode_b_mainfrom
fix/daemon-sync-followups-cleanup

Conversation

@chiga0

@chiga0 chiga0 commented May 25, 2026

Copy link
Copy Markdown
Collaborator

What this does

Independent fixes and cleanup over the merged cross-client real-time sync base. Each item is self-contained.

Correctness

  • Epoch-reset resync. When a subscriber resumes with a Last-Event-ID past the bus high-water — which only happens after a daemon restart rebuilt the EventBus — the bus now emits state_resync_required{reason:'epoch_reset'} before replay. Previously the empty post-restart ring made the existing ring-evicted check a no-op, so the consumer got a bare replay_complete{replayedCount:0} — a false "you're caught up" while its reducer still held stale data from the dead epoch. Epoch reset also replays the whole fresh ring, so the new low-id events aren't filtered out by the stale cursor.
  • Approval-mode serialization. setSessionApprovalMode now runs through a per-session FIFO (mirroring the existing modelChangeQueue). Two concurrent approval-mode changes can no longer interleave their ACP roundtrips and publish an approval_mode_changed that disagrees with the mode the child actually settled on.

Cross-client UX

  • Catch-up indicator. DaemonSessionProvider consumes the replay_complete sentinel to expose connection.catchingUp, giving clients a deterministic "replaying history → live" transition instead of a spinner-timeout guess. Armed only on resume (the daemon emits replay_complete only for Last-Event-ID subscriptions).
  • One cancel per turn. prompt_cancelled is now broadcast at most once per active prompt, so a client that POSTs /cancel and then drops its socket no longer produces two cancel frames for the same turn.
  • No silent forward failures. The user echo is published before the prompt is forwarded; if the forward itself rejects (transport died / ACP error) without a user cancel, the bus now emits a compensating prompt_cancelled{reason:'forward_failed'} so peers see the turn end instead of waiting forever on the echoed input.

Wire / hardening

  • replay_complete now carries lastReplayedEventId (canonical) alongside a deprecated lastEventId alias; the SDK prefers the new name. Removes the naming collision with the SSE protocol's Last-Event-ID in raw traces. Additive and backward-compatible.
  • Documented the anonymous-prompt dedup precondition, added an Array.isArray guard on the echo path, and documented the deliberate _meta placement.

Benefits

  • Fixes a real correctness bug: daemon restart + client reconnect no longer silently leaves a client rendering stale state.
  • Removes two multi-client race/duplication artifacts (approval-mode drift, double cancel) and one silent-stall gap (failed forward).
  • Gives clients a first-class catch-up signal and a clearer replay wire contract.

Tests

acp-bridge 302 pass (new epoch-reset / approval-mode-serialization / cancel-dedup / forward-failed cases), SDK daemon 259 pass, webui daemon 30 pass (new catch-up case). Typecheck and lint clean.

doudouOUC and others added 18 commits May 19, 2026 13:40
* fix(serve): address qwen-latest review on merged #4291 (7 threads)

Seven post-merge findings from the qwen-latest review on #4291,
all real. Most are tightening fixes for issues introduced by the
earlier rounds of #4291 — the same security / DRY / observability
classes the original review surfaced, applied to surfaces that
weren't covered initially.

#1 (deviceFlow.ts:1179) — late-poll observer closure retained the
entire entry by reference (deviceCode/pkceVerifier BrandedSecrets +
cancelController) for the lifetime of the daemon if `provider.poll()`
never settled. Memory leak + indefinite secret retention. Destructure
the four fields the closure actually needs (deviceFlowId, providerId,
initiatorClientId, audit sink) so the entry is GC-eligible the
moment runPollTick returns.

#2 (server.ts) — `callerIsInitiator` was duplicated verbatim across
three locations: GET handler, toDeviceFlowStartResponseBody,
toDeviceFlowStateBody. The exact bug class #4291 was fixing was
"POST and GET diverged on the same redaction policy" — duplicating
the gate recreated the preconditions for divergence. Extracted to
shared `callerIsDeviceFlowInitiator(view, callerClientId)` helper
with the consolidated threat-model JSDoc. All three sites now call
the helper.

#3 (deviceFlow.ts:1110) — timeout callback constructed two separate
`DeviceFlowPollTimeoutError` instances (one for `signal.reason`, one
for the wrapper rejection). Each capture its own V8 stack trace,
and `signal.reason.stack` would diverge from the caught rejection's
stack — confusing for operators inspecting both. Build the sentinel
ONCE per timer fire and pass the same instance to both sites.

#4 (qwenDeviceFlowProvider.ts:273) — `Error.name` is a freely
assignable string property; a hostile fetch wrapper could set
`e.name = 'X\n[serve] FAKE LINE\x1b[31m'` to inject log lines or
ANSI sequences via the same vector we already closed for `oauthError`.
The non-OAuth catch path interpolated `${err.name}` raw. Apply the
same `sanitizeForStderr()` helper.

#5 (deviceFlow.ts:1551) — on the timeout path, `rawProviderError`
is undefined (deliberately, to skip the misleading
`provider.poll() threw (raw): ...` audit template), but that left
the audit hint field omitted entirely. Operators reading the
durable audit trail saw `errorKind: 'upstream_error'` with no signal
whether it was a hung IdP or a generic provider failure. Use
`result.hint` (which already carries the timeout-specific
`provider.poll() timed out after Nms; check IdP connectivity` text
built in the catch) so the audit matches the SSE event.

#6 (server.ts) — the `QWEN_SERVE_DEBUG` env-var check was inlined
in the GET route handler, duplicating the `isServeDebugMode()`
helper from `./debugMode.js` that workspaceAgents and
workspaceMemory already use. The inline copy also had a dead `?? ''`
fallback (the value is guaranteed truthy at that point per the
preceding check). Use the canonical helper.

#7 (deviceFlow.ts:1217) — late-rejection observer interpolated the
raw `lateErr.message` into the audit hint (truncated to 256 bytes,
but RFC 8628 `device_code` values fit comfortably in 256 bytes).
The provider's catch already uses the `name + length` redaction
pattern to prevent WAF-echoed `device_code`/PKCE leaks; the
registry layer was undoing that hardening because the same failure
settled late. Apply the same `name + length` pattern at the late-
rejection site.

Tests:
- Existing late-rejection test reseeded with a `device-code-secret-*`
  substring inside the long detail; hard-negative-asserts the seeded
  secret is absent from the audit + asserts the new
  `Error (message N bytes; raw suppressed)` shape.
- Existing poll-timeout test now also asserts: hint IS defined on
  the audit (not omitted), hint contains `'timed out after'` /
  `'check IdP connectivity'`, and `signal.reason instanceof
  DeviceFlowPollTimeoutError` (proves the single sentinel is
  shared between abort and reject).
- New `sanitizes control characters in attacker-controlled
  err.name` test in qwenDeviceFlowProvider.test.ts pins the round-4
  #4 fix with a hostile `e.name` containing `\n` + `\x1b[31m...`.

cli serve 702/702 (was 686, +16 — additional tests imported via
the acp-bridge package lift on main); sdk 421/421; typecheck clean
across all 4 workspaces; eslint --max-warnings 0 clean on touched
files.

Refs: #4175, #4255, #4291

🤖 Generated with [Qwen Code](https://github.com/QwenLM/qwen-code)

* fix(serve): address deepseek-v4-pro review on #4305 (4 threads)

Round-5 fold-in. Four findings from the deepseek-v4-pro review on
PR #4305 — all real, three are sister fixes for the same security
classes that #4305 already closed at adjacent surfaces.

#1 (deviceFlow.ts) — `pollTimedOut` race correctness. The flag was
set unconditionally inside the timer callback. If the provider
settled the wrapper at 29.9s, `finally` would call
`clearScheduled(pollTimer)` — but if the timer callback was already
queued for execution before the clear landed (a real possibility
in Node's event-loop ordering, even if not always observed in
practice), this branch could still run and incorrectly mark
`pollTimedOut`. Move the flag assignment to the catch block where
the settled cause is unambiguous via `instanceof
DeviceFlowPollTimeoutError`. New test pins the negative: provider
beats the timeout → no spurious `lost_late_poll_after_timeout`
audit even after ticking 2× the ceiling.

#2 (deviceFlow.ts) — late-rejection observer interpolated raw
`lateErr.name` into the audit hint without sanitization. Same
attacker-controlled vector closed at the provider layer for
`err.name` in round-4. Route through `sanitizeForStderr`.

#3 (deviceFlow.ts) — late-success observer interpolated
`latePollResult.kind` directly into the audit template. While the
typed shape is `'pending' | 'slow_down' | 'success' | 'error'`, a
non-conforming provider could return an arbitrary string. Same
log-injection vector. Route through `sanitizeForStderr`.

#4 (qwenDeviceFlowProvider.ts → deviceFlow.ts) —
`sanitizeForStderr` only stripped ASCII C0/C1 + DEL; bypass via
Unicode lookalikes:
  - U+2028/U+2029: LINE/PARAGRAPH SEPARATOR (newline-equivalent in
    most Unicode-aware terminals — most direct log-forging vector)
  - U+200B–U+200F: zero-width chars + LRM/RLM
  - U+202A–U+202E: bidirectional override controls
  - U+FEFF: BOM / ZWNBSP

A malicious IdP returning `slow_down
[serve] FAKE` in
`oauthError` would otherwise still forge log lines.

Architectural change: `sanitizeForStderr` was previously private to
`qwenDeviceFlowProvider.ts`. To address #2/#3, the registry layer
needs to call it too. Lifted into `deviceFlow.ts` (the foundation
module) and re-imported from the provider. Single source of truth;
the regex is now a module-level constant compiled once with explicit
`\uXXXX` escapes (via `String.raw` so the source is greppable, not
literal-Unicode-laden).

Tests:
- `does NOT attach late-poll observer when the provider beats the
  timeout` — N1 race regression
- `sanitizes hostile latePollResult.kind in late-observer audit` — N3
- `sanitizes hostile lateErr.name in late-rejection observer audit` — N2
- `sanitizes Unicode lookalike controls (U+2028 LINE SEPARATOR,
  bidi, ZWNBSP) in oauthError` — N4

cli serve 706/706 (was 702, +4 — all new round-5 tests); sdk
421/421; typecheck clean; eslint --max-warnings 0 clean on touched
files.

Refs: #4175, #4255, #4291, #4305

🤖 Generated with [Qwen Code](https://github.com/QwenLM/qwen-code)

* fix(serve): address gpt-5.5 + qwen-latest review on #4305 round-5 (5 threads)

Round-6 fold-in. Five findings split between maintainability,
security hardening, and a real defensive bug.

#1 (qwenDeviceFlowProvider.test.ts) — gpt-5.5: round-5 #4 test
embedded U+2028 / U+200E / U+FEFF as literal characters in source.
Invisible in GitHub diffs / most editors; the negative
`not.toContain('')` looked like an empty-string check. Rewrote
the payload + assertions to use named `\uXXXX`-bound constants.
Also added a companion test exercising U+2066–U+2069 (round-6 #5
below).

#2 (deviceFlow.ts) — qwen-latest: the late-poll observer's
`void tracked.then(...)` was missing a terminal `.catch(() => {})`.
A synchronous throw inside either handler (e.g., a misbehaving
`audit.record`: backpressure, malformed payload, sink out-of-disk)
would reject the derived promise unhandled. On Node 22's default
`--unhandled-rejections=throw`, that crashes the daemon. Added the
terminal `.catch(() => {})` matching the persist-tracker pattern.
New test injects a poison audit sink that throws specifically on
the `lost_late_poll_after_timeout` call; asserts `flushAsync()`
resolves cleanly.

#3 (deviceFlow.ts) — qwen-latest: the `case 'error'` audit-record
hint interpolated `rawProviderError` (raw `err.message`) without
`sanitizeForStderr`. Per ES2019+ `JSON.stringify` no longer escapes
U+2028/U+2029 — those would still forge log lines downstream
through file/stdout audit sinks. Apply the same sanitizer used on
every other provider-controlled audit path. New test pins a hostile
provider message containing U+2028 + ANSI escape and asserts
neither survives.

#4 (deviceFlow.ts) — qwen-latest: the round-5 #1 comment claimed
"`DeviceFlowPollTimeoutError` isn't exported as a public DeviceFlow
contract", but it IS `export class` (the test file constructs it
directly for fixtures). With `pollTimedOut = true` keyed solely on
`instanceof`, a future provider that imports + throws the class
would spoof the registry's "I caused the timeout" signal —
attaching a phantom late-poll observer.

Fix: introduce a runtime brand `_isRegistryTimeout: boolean` on the
class (default `false`) plus an internal-only
`makeRegistryPollTimeoutError(ms)` helper that sets the brand to
`true`. The brand is set ONLY at the registry's race-timer
construction site. Both gates updated:
  - `if (err instanceof X && err._isRegistryTimeout === true)` in
    the catch (for `pollTimedOut`)
  - `if (lateErr instanceof X && lateErr._isRegistryTimeout === true)`
    in the late-rejection self-filter

A provider-thrown brand-false instance now flows through the
generic provider-throw audit path — correctly auditing the misuse
rather than silently swallowing it. Repurposed the original "no
double-audit when registry's own DeviceFlowPollTimeoutError is
late-rejected" test (which was actually exercising the brand-false
path) into the inverted assertion: brand-false provider throw IS
audited as a real failure. Removed the orphaned old assertion; the
brand-true happy path is implicitly covered by the hanging-provider
test (which exercises the registry-built timeout end-to-end).

#5 (deviceFlow.ts) — qwen-latest: `sanitizeForStderr` regex covered
U+202A–U+202E (bidi embedding/override) but missed U+2066–U+2069
(LRI/RLI/FSI/PDI). These are the primary CVE-2021-42574
("Trojan Source") attack vectors — a hostile IdP swapping U+2066
for U+202D achieves the same visual reordering and would have
bypassed the round-5 filter entirely. Extended the regex range and
JSDoc; new test exercises U+2066/U+2068/U+2069 in `oauthError` and
asserts none survive while substantive ASCII parts remain.

cli serve 713/713 (was 710, +3 round-6 tests + the round-5 #4
rewrite + the round-6 #5 companion); typecheck clean across all 4
workspaces; eslint --max-warnings 0 clean on touched files.

Refs: #4175, #4255, #4291, #4305

🤖 Generated with [Qwen Code](https://github.com/QwenLM/qwen-code)

* fix(serve): replace literal U+2028 with explicit 
 escape in round-6 #3 test

PR #4312 review (Copilot): the round-6 #3 test (sanitizes
rawProviderError) regressed back to embedding a literal U+2028
character in source via `const U_2028 = ' '`. That's the same
maintainability anti-pattern round-6 #1 was fixing in the sister
test. Internal-consistency fix: switch to the explicit `
`
escape so the constant is greppable and reviewable in GitHub diffs.

Refs: #4291, #4305, #4312

🤖 Generated with [Qwen Code](https://github.com/QwenLM/qwen-code)
* fix(serve): post-merge P2 corrections from Codex review on #4282

Follow-up to PR #4282 (Wave 4 PR 17) addressing four P2 issues
flagged by Codex's `/review` after the squash-merge to main:

P2-1 — Read the workspace context filename for init
  `qwen serve` parent never goes through `loadCliConfig`, so the
  process-global `getCurrentGeminiMdFilename()` stays on the default
  `QWEN.md` even when the workspace configures
  `context.fileName: 'AGENTS.md'`. `runQwenServe` now snapshots the
  workspace's merged setting at boot and forwards via
  `BridgeOptions.contextFilename`, so init writes the same file the
  ACP child reads.

P2-2 — Restart MCP servers with a fresh disabledTools snapshot
  `Config.disabledTools` was frozen at construction time;
  `setWorkspaceToolEnabled` only updated settings.json. The
  documented "toggle + restart" workflow re-registered just-disabled
  tools because rediscovery still saw the bootstrap snapshot. Added
  `Config.setDisabledTools()` plus a re-read at the ACP restart
  handler so `discoverMcpToolsForServer` honors the latest set.

P2-3 — Match the SDK timeout to the daemon's restart budget
  Bridge waits up to 300s for stdio MCP discovery; SDK helper used
  the client-wide 30s default and aborted valid slow restarts.
  Added a per-call `timeoutMs` plumbed through `fetchWithTimeout`,
  defaulting `restartMcpServer` to 5 minutes.

P2-4 — Reject symlinked parent directories before init writes
  `lstat(target)` only checked the final component; a symlinked
  parent (e.g. `docs -> /tmp` with `context.fileName:
  'docs/QWEN.md'`) would let `writeFile` follow the link and create
  / truncate outside `boundWorkspace`. Added
  `canonicalizeExistingAncestor` (walks up through ENOENT to the
  deepest extant ancestor, then `realpath`s) and verifies the
  canonical parent stays within the canonical workspace.

5 new tests (4 bridge / 2 SDK):
- contextFilename snapshot honored
- parent-symlink escape rejected
- nested real subdir accepted
- restartMcpServer survives 1.2s response with 1s default timeout
- restartMcpServer honors a 50ms caller override

Typecheck clean across cli / sdk-typescript / core.
1604/1604 unit tests pass.

* fix(serve): fold-in 1 — address 16:32:44-round review on #4282

Follow-up addressing the 8 unresolved review threads opened on PR
shipping in this same #4297; addresses correctness gaps + missing
test coverage that would otherwise let regressions ride into main.

Behavior fix:
- broadcastWorkspaceEvent gains a `skipSessionId` parameter; when
  `setSessionApprovalMode` runs with `persist:true`, the broadcast
  skips the requesting session so it doesn't receive the same
  `approval_mode_changed` event twice (once via session-scoped
  publish + once via broadcast). The SDK reducer's
  `approvalModeChangedCount` now increments by 1, not 2, on the
  requesting client (peers still see 1 via the broadcast).
  Addresses #3260501134.

Observability + posture:
- broadcastWorkspaceEvent now mirrors PR 16's publishWorkspaceEvent
  member: per-entry success/failure accounting + an "ALL buses
  dropped" stderr elevation. The previous local helper silently
  swallowed every publish failure. Addresses #3260501126.
- WorkspaceInitPathEscapeError + WorkspaceInitSymlinkError typed
  classes for the two boundary guards in initWorkspace, mapped to
  HTTP 400 by sendBridgeError. Previous generic `Error` fell
  through to the 500 handler, telling operators "daemon broken"
  when the actual fix was workspace-config correction. Addresses
  #3260501161.

Public surface symmetry:
- Re-export McpServerNotFoundError, McpServerRestartFailedError,
  WorkspaceInitPathEscapeError, WorkspaceInitSymlinkError from the
  serve barrel. External embeds matching these via `instanceof`
  no longer need deep imports. Addresses #3260501163.

Test coverage:
- restartMcpServer bridge tests (5): success + event broadcast,
  soft-skip + refused event, McpServerNotFoundError translation,
  McpServerRestartFailedError translation, originator clientId
  stamping. Addresses #3260501141.
- sendBridgeError mapping tests (4): McpServerNotFoundError → 404,
  McpServerRestartFailedError → 502, WorkspaceInitPathEscapeError
  → 400, WorkspaceInitSymlinkError → 400. Addresses #3260501148.
- initWorkspace boundary guard tests (2 added): symlink-at-target
  rejected, contextFilename '../outside.md' rejected. Addresses
  #3260501157.
- TrustGateError tests assert the typed class via `.toThrow(TrustGateError)`,
  not just message text. Addresses #3260501165.

Also updates the existing fold-in 4 S2 broadcast test to reflect
the new no-duplicate semantics on the requesting session.

Typecheck clean across cli / sdk-typescript / core.
1615/1615 unit tests pass.

* fix(serve): fold-in 2 — copilot + wenshao review on #4297

Round-2 reviewer adoption on the same PR:

Critical fixes:
- `restartMcpServer` JSDoc documents `timeoutMs: 0` as "disable the
  timeout entirely", but the `> 0` guard in `fetchWithTimeout`
  rejected `0` and silently fell back to the 30s client default.
  Loosened the guard to `>= 0` so `0` flows through to the
  no-timeout branch via the existing truthiness check; NaN /
  negative inputs still coerce to the client default. Addresses
  duplicate reports from copilot (#3260577538) and wenshao
  (#3260661833).
- TS2322 in the slow-fetch test stub: `resolveResponse` was typed
  against `import('undici-types').Response` but assigned a
  `(v: Response) => void`. Re-typed against the global `Response`
  throughout. Caught only by tsc runs that include the test
  files. Addresses #3260663072.

Test fidelity:
- Slow-fetch stub now observes `init.signal` and rejects on abort,
  so a regression that drops the per-call `timeoutMs` override
  will reliably fail the test instead of resolving after the
  timer fired (false-negative coverage). Addresses #3260577600.
- New test pinning the `timeoutMs: 0` semantics: 1ms client
  default + a stub that resolves after 50ms. Without the `>= 0`
  fix, the call would abort at 1ms; with it, the explicit
  `0` disables the timer and the call completes.

Bug fixes:
- `runQwenServe.contextFilenameForInit` previously called
  `String(arr[0])` on the array branch, producing a literal
  `"[object Object]"` filename for hand-edited bad data. Now
  validates each element with `typeof === 'string'` and falls
  back to `undefined` (so the bridge uses its
  `getCurrentGeminiMdFilename()` default) when no string is
  found. Addresses #3260577641.

Documentation drift:
- `Config.getDisabledTools()` JSDoc rewritten to describe the
  mutable-via-`setDisabledTools()` semantics introduced by P2-2,
  and the "registration-time only / no retroactive unregister"
  contract that pairs with it. Old comment claimed the set was
  frozen at construction. Addresses #3260577677.

Observability:
- `acpAgent` MCP-restart `loadSettings` failure now surfaces a
  stderr line naming the server + the underlying error, instead
  of silently swallowing it. The documented "toggle + restart"
  workflow used to break with zero diagnostic when settings.json
  was corrupted or unreadable. Addresses #3260663303.

Code organization:
- Moved `canonicalizeExistingAncestor` after `describeStatKind` so
  the latter's JSDoc is no longer orphaned (TypeScript only
  associates the last `/** ... */` block before a declaration).
  Addresses #3260668618.

Typecheck clean across cli / sdk-typescript / core.
1616/1616 unit tests pass.

* fix(serve): fold-in 3 — read merged scope on MCP restart refresh

Critical bug from wenshao review (#3260725526) on PR #4297:
the P2-2 acpAgent re-read narrowed `Config.disabledTools` to
`SettingScope.Workspace` alone, dropping User / System scope
entries. The bootstrap Config received `merged.tools?.disabled`
(union of all scopes), so user-level / system-level disables
worked at boot — but the first `mcp restart` would replace the
in-memory set with the workspace scope alone, silently re-enabling
any tool that was disabled at a higher scope but absent from the
workspace file.

The asymmetry vs. the persist-write path is deliberate and
documented:
- Reads (here): merged — match the bootstrap Config snapshot,
  preserve user/system policy.
- Writes (`runQwenServe.persistDisabledTools`): workspace scope —
  don't bake higher-scope entries into the workspace file
  (per-#4282 fold-in 1 H2 fix).

Two paths look alike but answer different questions.

Typecheck clean across cli / sdk-typescript / core.
1616/1616 unit tests pass.

* fix(test): fold-in 4 — wire timeoutMs:0 stub to init.signal

Critical follow-up from wenshao (#3260810242) on PR #4297:
the new `timeoutMs: 0` regression test (added in fold-in 2)
inherited the same flaw it was meant to prevent — the slow-fetch
stub didn't observe `init.signal`, so a regression that ignored
the `0` override would fire the AbortController at the 1ms client
default but the stub would keep the promise pending. The 50ms
`resolveResponse` would win, the test would still pass, and the
documented "0 disables timeout" contract would be unprotected.

Mirrored the listener pattern already used by the two sibling
tests in fold-in 2 — `init.signal.addEventListener('abort', () =>
reject(...))`. Now a regression that re-rejects `0` triggers the
abort, the stub rejects, the test fails.

8/8 restartMcpServer SDK tests pass; SDK typecheck clean.

* fix(serve): fold-in 5 — TOCTOU + setDisabledTools coverage

Two new critical reviews from wenshao on PR #4297:

C1 — TOCTOU between lstat and writeFile (#3260836305):
The `lstat(target)` symlink check and the subsequent `writeFile`
were two separate syscalls, leaving a race window where a local
attacker with workspace write access could substitute a symlink
between them. With `force: true`, `writeFile` would follow the
link and truncate an external target.

The `action === 'created'` path now uses `fs.open(target, 'wx')`
(O_WRONLY|O_CREAT|O_EXCL), which atomically refuses any
pre-existing inode (regular file, dir, OR symlink) at the target
path. EEXIST after the absence check most plausibly means a
race-created symlink, so we throw `WorkspaceInitSymlinkError(kind:
'target')` — same typed class the route maps to 400.

The `force: true` overwrite path retains the existing TOCTOU as a
documented limitation; closing it requires `O_NOFOLLOW`-aware open
which the post-PR18 `WorkspaceFileSystem` migration will provide.

C2 — P2-2 zero test coverage (#3260836302):
The `setDisabledTools` runtime sync was the only Wave-4 P2 fix
without a dedicated test. Added 5 Config-level tests:
- Initializes from `disabledTools` ConfigParameters
- Defaults to empty set when omitted
- `setDisabledTools` replaces the live snapshot
- Defensive copy: caller-set mutations don't leak into the live snapshot
- Accepts an empty set (clears live snapshot)

Plus a TOCTOU regression test in httpAcpBridge.test.ts that
spies fs.lstat / fs.readFile to simulate the race window:
pre-creates a symlink, makes lstat lie about it, asserts the
'wx' open catches the racing inode and throws the typed
`WorkspaceInitSymlinkError(kind: 'target')`.

1622/1622 unit tests pass; typecheck clean across cli /
sdk-typescript / core.

* fix(serve): fold-in 6 — count actual skips in broadcast alarm

DeepSeek review on #4297 (#3261079572):
`broadcastWorkspaceEvent` unconditionally subtracted 1 from the
`eligible` recipient count whenever `skipSessionId` was set, even
when the id matched zero live sessions (caller mistake, stale id,
or the matching session was just torn down between resolution and
broadcast). In a single-session workspace that's the difference
between `eligible = 0` (alarm suppressed) and `eligible = 1`
(alarm fires when the publish failed) — silently losing the
all-dropped breadcrumb the telemetry was meant to surface.

Today's call sites pass real session ids so the bug doesn't
manifest in practice, but the defensive shape is small: track
`skippedCount` inside the loop and subtract that, so the alarm
condition is self-consistent regardless of how the caller mis-uses
the param.

162/162 bridge tests pass; CLI typecheck clean.

* fix(serve): fold-in 7 — close overwrite TOCTOU, harden boot + diagnostics

Round-7 review on PR #4297. Three critical fixes + one suggestion
test, plus a regression test for the overwrite TOCTOU close.

C1 — force:true overwrite TOCTOU (#3262615446):
The fold-in 5 fix only closed the `'created'` action via 'wx';
the `'overwrote'` branch still used plain `fs.writeFile`, so a
local writer could swap the verified regular file to a symlink
between the lstat/readFile checks and the write and have the
forced overwrite truncate an external target. Switched to
`fs.open(target, O_WRONLY | O_TRUNC | O_NOFOLLOW)` — `O_NOFOLLOW`
makes open() fail with ELOOP on a symlink at the final component
even under race. ELOOP / ENOENT (race-deleted) translate to
`WorkspaceInitSymlinkError(kind: 'target')` so the route still
maps to a structured 400 instead of a generic 500.

C2 — settings.json corrupt blocks daemon boot (#3262625091):
`loadSettings(boundWorkspace)` at boot had no try/catch — a
corrupted, malformed, or temporarily unreadable settings file
threw synchronously and prevented daemon startup. Pre-PR this
never happened because settings were read lazily inside request
handlers. Wrapped in try/catch with stderr fallback so the daemon
keeps booting (with the bridge's default context filename) when
the file is broken.

C3 — malformed `tools.disabled` clears policy silently (#3262625101):
When `merged.tools?.disabled` is present but not an array
(boolean / string / object from a hand-edited settings.json), the
ternary `Array.isArray(...) ? ... : []` substituted an empty list
without firing the surrounding catch block. After an MCP restart
every disabled tool would silently re-register. Added an explicit
`!Array.isArray && !== undefined` check that stderr-logs the
malformed type before clearing — operators see the
misconfiguration instead of a stealth re-enable.

S1 — contextFilename extraction tested (#3262690842):
Lifted the inline `firstStringInArray` + branching into an
exported `extractContextFilename(value: unknown)` helper and
added `runQwenServe.test.ts` with 5 tests covering the four
branches the suggestion called out: non-empty string, array with
strings, array with no strings, non-string non-array.

Plus a TOCTOU regression test for the overwrite path that
verifies `O_NOFOLLOW` returns `WorkspaceInitSymlinkError(kind:
'target')` when the file is race-substituted with a symlink
behind the lstat/readFile mocks.

S2 (acpAgent restart-handler integration test #3262690845) is
deferred — Config-level coverage of `setDisabledTools` already
locks the load-bearing surface (5 tests in fold-in 5), and
adding a full acpAgent integration test requires heavy ext-method
plumbing. The new C3 stderr diagnostic plus existing tests give
us the regression signal we need without that scaffolding.

1627/1627 unit tests pass; typecheck clean across cli /
sdk-typescript / core / acp-bridge.

* fix(serve): fold-in 8 — split ELOOP / ENOENT diagnostic in overwrite path

qwen-latest review on PR #4297 (#3262861754):
The fold-in 7 ELOOP/ENOENT branch shared one error message that
said "swapped to a symlink." That's accurate for ELOOP (genuine
O_NOFOLLOW rejection — likely an attack race) but misleading for
ENOENT in the overwrite path: there `readFile` just succeeded
proving the file existed, so ENOENT means the file was DELETED
between the content check and the open — a benign race with a
concurrent writer (git checkout, editor save, lockfile rename),
NOT a symlink swap. An operator seeing the symlink language for
a benign delete would `ls -la`, see no symlink, and waste time
hunting an attack that didn't happen.

Split into two messages:
- ELOOP: "swapped to a symlink between the content check and the
  overwrite — refusing to follow it"
- ENOENT: "deleted between the content check and the overwrite
  (likely a concurrent writer) — refusing to recreate blindly"

Both still surface as `WorkspaceInitSymlinkError(kind: 'target')`
so the route maps to a structured 400; the class doubles as the
workspace-init race-condition bucket with kind='target' meaning
"target inode misbehaved at write time" generally.

Updated the existing fold-in 7 TOCTOU test to assert the ELOOP
message specifically, and added a new ENOENT race-delete test
that mocks lstat/readFile to land on the overwrote action against
a non-existent path — verifies the message says "deleted" and
NOT "swapped to a symlink."

170/170 bridge tests pass; CLI typecheck clean.

* fix(serve): fold-in 9 — route MCP restart through registry cleanup wrapper

gpt-5.5 critical review on PR #4297 (#3263088414):

The fold-in 5 P2-2 fix refreshed `Config.disabledTools` from merged
settings, but then called `manager.discoverMcpToolsForServer()`
directly — bypassing the `ToolRegistry.discoverToolsForServer`
wrapper that PURGES the server's existing `DiscoveredMCPTool`
entries (and `revealedDeferred` markers) plus its prompts before
rediscovery. Without the cleanup, `registerTool` only consulted
the refreshed `disabledTools` set for NEWLY-discovered tools —
entries already in the registry from the prior MCP boot kept
serving requests. Net effect: toggle-disable-then-restart
silently left the disabled tool live, breaking the documented
"toggle + restart" workflow that P2-2 was meant to fix.

Routed through `toolRegistry.discoverToolsForServer(serverName)`
which:
1. Removes existing `DiscoveredMCPTool` entries for this server
2. Drops their `revealedDeferred` reveal state
3. Removes the server's prompts via `removePromptsByServer`
4. THEN delegates to `manager.discoverMcpToolsForServer` for the
   actual reconnect + rediscover

The pre-discovery budget / in-flight checks still go through the
`manager` reference (which is the same object the registry
wrapper would forward to) — so soft-skip semantics for
`budget_would_exceed`, `in_flight`, `disabled` are preserved.

CLI typecheck clean; 403/403 server + bridge tests pass.

* fix(serve): fold-in 10 — qwen-latest 05:45-round review on #4297

5 review threads from qwen-latest's late round on PR #4297 (now closed
in favor of #4313 against `daemon_mode_b_main`). 1 critical + 4
suggestions, all adopted.

C1 — extractContextFilename / getCurrentGeminiMdFilename divergence
(#3263954685): with `context.fileName: ['  ', 'AGENTS.md']`, the
daemon parent's `extractContextFilename` (which skips empty entries)
wrote `AGENTS.md`, but the ACP child's `getCurrentGeminiMdFilename`
(which returned `arr[0]` unconditionally) read `''`. The init'd file
was orphaned. Aligned `getCurrentGeminiMdFilename` to skip empty
entries with the same semantics, falling back to
`DEFAULT_CONTEXT_FILENAME` when all entries are empty.

S2 — WorkspaceInitSymlinkError reused for non-symlink races
(#3263954690): the EEXIST race-create and ENOENT race-delete cases
were surfacing as `code: 'workspace_init_symlink'`, misleading
operators into hunting symlink attacks for benign concurrent-
modification windows. Split into a sibling `WorkspaceInitRaceError`
class (`kind: 'eexist' | 'enoent'`, HTTP code
`workspace_init_race`). The genuine symlink class stays for ELOOP,
lstat-detected target symlinks, and parent-realpath escapes.

S3 — fsConstants.O_NOFOLLOW defensive `?? 0` (#3263954697): matches
the existing codebase convention in
`core/src/utils/{sessionStorageUtils,gitDiff}.ts` and
`cli/src/ui/utils/customBanner.ts`. Functionally a no-op (JS
bitwise coerces undefined to 0) but consistent.

S5 — Parent-directory TOCTOU still open (#3263954707): O_NOFOLLOW
only protects the final path component; a local writer could swap
a real parent dir for a symlink between
`canonicalizeExistingAncestor` and `fs.open`. Added
`verifyParentWithinWorkspace` post-open helper that re-realpaths
`path.dirname(target)` and refuses with
`WorkspaceInitSymlinkError(kind: 'parent')` if the parent moved.
On the create path (where we just opened with `'wx'`), the failure
also unlinks the file we just made best-effort. Residual race
window narrowed from "between pre-check and open" to "between
post-open realpath and writeFile" — sub-millisecond, documented as
accepted Stage-1 trust posture.

S4 — broadcastWorkspaceEvent vs publishWorkspaceEvent stale comment
(#3263954688): the "now removed" comment was inaccurate (5 call
sites still use the closure). Replaced with an accurate
description of why both coexist (factory closure can't `this`-call
proxy member; closure also takes `skipSessionId` for persisted
approval-mode mirror) and a TODO marker for future helper extraction.

Two existing tests updated to assert the new `WorkspaceInitRaceError`
class for EEXIST / ENOENT scenarios (the symlink-class assertions
are preserved for ELOOP / lstat / parent cases).

1759/1759 unit tests pass; typecheck clean across all 4 packages.