…lict

The message.includes('409') fallback made ALL 409 errors appear
transient, preventing deterministic conflict classification. Only actual
transient indicators (lock/contention/conflict) should trigger it.

- geminiChat.ts: delegate shouldRetryOnError to classifyError for 408/409/network errors
- retry.ts: deduplicate defaultShouldRetry → classifyError (single source of truth)
- retry.ts: update isTransientCapacityError for persistent mode (408 + network errors)
- retry.test.ts: add 15 integration tests for retryWithBackoff (408/409/network paths)
- retry.test.ts: add 6 tests for updated isTransientCapacityError
- geminiChat.test.ts: add 4 tests (408 retry, 409 transient retry, 409 deterministic no-retry, network retry)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Remove `await` from `expect(promise).rejects.toThrow()` calls that are
followed by `vi.runAllTimersAsync()`. In vitest 3.2.4, awaiting the
assertion promise before advancing fake timers creates a deadlock: the
inner promise can't reject until timers advance, but timers can't advance
while the microtask queue is blocked by the await.

Also add explicit `initialDelayMs: 10` to tests using default options
(7 attempts × 1500ms default = 76s of fake time, triggering the 5s
test timeout), and reduce abort-signal test delay from 10s to 100ms.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Remove await from expect(promise).rejects.toThrow() calls that are
followed by vi.runAllTimersAsync(). In vitest 3.2.4, awaiting the
assertion promise before advancing fake timers creates a deadlock:
the inner promise can't reject until timers advance, but timers
can't advance while the microtask queue is blocked.

Also add explicit initialDelayMs: 10 to tests using default options
(7 attempts x 1500ms default = 76s of fake time, triggering the 5s
test timeout), and reduce abort-signal test delay from 10s to 100ms.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Remove 'conflict' from isTransientConflict keywords (too broad: matches
  standard HTTP 409 reason phrase "Conflict"). Keep only 'lock' and 'contention'.
- Fix 409 test to use proper Error instance instead of plain object
  (getErrorMessage returns "[object Object]" for plain objects, masking the bug).
- Scope isTransientCapacityError to 429/529 only. 408 and network errors are
  no longer persistent-retryable since they can indicate permanent config issues
  (wrong endpoint, firewall) that would hang unattended jobs for hours.
- Update isTransientCapacityError tests to match new scope.
- Document accepted retryable categories in geminiChat.ts shouldRetryOnError.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Adds an in-flight discovery guard to McpClientManager.discoverMcpToolsForServer()
that prevents concurrent re-discovery of the same server. Without this guard,
overlapping calls (e.g. from health-check reconnect + incremental discovery)
could spawn duplicate MCP child processes because the disconnect→connect
sequence is not atomic.

Changes:
- Add inFlightDiscoveries Set to track per-server discovery in progress
- Early-return if discovery is already in flight for the same server
- Clean up failed clients in catch block so subsequent retries can proceed
- Stop health check timer during re-discovery disconnect path
- Clear inFlightDiscoveries in stop() for completeness
- Add 2 tests: concurrent discovery deduplication + in-flight cleanup

Closes QwenLM#3817

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

retry.test.ts:
- Remove erroneous await before expect().rejects.toThrow() in ~15 fake-timer
  test locations that caused deadlocks (await blocked before runAllTimersAsync)
- Update 408/409/network-error tests to use explicit classifyError-based
  shouldRetryOnError instead of relying on defaultShouldRetry (see below)

retry.ts:
- Narrow defaultShouldRetry to only retry 429/5xx, matching original behavior.
  Previously it delegated to classifyError which also retried 408, transient
  409, and network errors — silently expanding retry behavior for all callers
  using the default (baseLlmClient.generateJson, client.generateContent).
  Callers that want broader retry (geminiChat) now use classifyError directly
  via a custom shouldRetryOnError.

mcp-client-manager.ts:
- Fix TOCTOU race: move inFlightDiscoveries.add() to immediately after the
  has() guard and before await disconnect(), preventing concurrent calls from
  slipping through during the async gap.
- Wrap new McpClient() + connect/discover in try/finally so discoveryDone()
  runs even if the constructor throws, preventing permanent entry leaks.
- Clean inFlightDiscoveries in disconnectServer() to prevent stale entries
  from blocking future discoveries.
- Guard reconnectServer() against running when a discovery is already in
  flight, preventing false "Successfully reconnected" logs.

geminiChat.ts:
- Update comment to clarify isSchemaDepthError/isInvalidArgumentError are
  independent safety nets not covered by classifyError.
- Remove redundant if (status === 400) return false (classifyError handles it).
- Remove unused getErrorStatus import.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

… tests

Remove await from const assertionPromise = expect(promise).rejects.toThrow()
in ~15 locations where the assertion is followed by vi.runAllTimersAsync().
The await blocks the microtask queue before timers can advance, causing
deadlocks in vitest 3.2.4.

Add eslint-disable-next-line vitest/valid-expect to prevent the pre-commit
hook from re-adding the await (the rule doesn't account for this fake-timer
pattern where await must come after runAllTimersAsync).

All 105 retry tests pass.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- removeServer() now cleans inFlightDiscoveries and isReconnecting
- reconnectServer() no longer has racy in-flight check; verifies
  server is actually connected before reporting success
- Add test for cleanup state after server removal

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

- retry.ts: remove overly-broad ENOTFOUND/EHOSTUNREACH/"network error"
  substring from isRetryableNetworkError; fix isTransientConflict to use
  word-boundary regex preventing false positives on "blocked"/"clock"/"flock"
- retry.test.ts: update tests to reflect removed retryable codes
- geminiChat.test.ts: add test confirming invalid argument errors are not retried
- mcp-client-manager.ts: add inFlightDiscoveries guard in reconnectServer()
  after reconnect delay to prevent unnecessary reconnects

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Implement Option A from PR QwenLM#3799 suggestion to support Anthropic
model listing via the official /v1/models endpoint.

Changes:
- Add fetchAnthropicModels() in commands/auth/anthropic.ts
  - Uses x-api-key + anthropic-version headers per API docs
  - Returns same { data: [...] } shape as OpenAI-compatible
- Update manageModels.ts to support 'anthropic' source
- Add Anthropic tab to ManageModelsDialog.tsx
- Fix type-safety by removing dangerous casts

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>