Merged into release/v3.8.45. Delta do par #6246 (Test-All read-only + bulk enable/disable), reconciliado com o núcleo #6296 já mergeado. CHANGELOG restaurado (ambos bullets do #6246 coexistem). Reds pré-existentes: dast-smoke (infra) + file-size drift.
Merged into release/v3.8.45. Reds pré-existentes classificados: dast-smoke (infra), check:file-size (drift de baseline de god-files congelados), e 3 testes de contagem agentSkills stale (43→44 já corrigidos no tip pelo #6186 — somem no squash sobre o tip). Núcleo do fix#6246 (proxy IP-leak + over-deactivation).
The 'streaming no recognized content' branch added by #6216 marked ANY
stream that ended without content deltas as invalid — sweeping in two
regression-guarded pass-through contracts: an empty stream terminated by an
explicit 'data: [DONE]' (#3399 context-cache protection) and an incomplete
Claude lifecycle (ping only, no message_start; #3685 — stream-readiness
timeout territory, not failover). Both unit guards were red on the branch
and green on main (86/86 vs 84/86).
The branch now fires only for a truly EMPTY body (zero bytes — the Gemini
HTTP-200-empty case that motivated #6216), tracked via sawAnyBytes. New
guard: '#5976 truly EMPTY streaming body (zero bytes) -> invalid for combo
failover'. 87/87 across both files.
Also in this pre-flight batch:
- agentSkillTools-mcp: api.have upper bound 22 -> 23 (the #6186 catalog
addition updated the totals but missed this bound)
- delete tests/unit/free-provider-rankings-configured-filter.test.ts:
#6251 (server-side configuredOnly/availableOnly) superseded the #6245
client-side toggle it pinned; replacement declared in the test-masking
allowlist (tests/unit/freeProviderRankings-filters.test.ts, 11/11)
PR #6186 added omni-github-skills to the agent-skills catalog (API 22 -> 23)
but did not run the generator, so skills/omni-github-skills/SKILL.md never
existed and 6 integration assertions split between the old (42/43) and new
counts. Generated via scripts/skills/generate-agent-skills.mjs --apply and
aligned agent-skills-discovery to the real totals (43 = 23 API + 20 CLI;
handlers return 44 with config-codex-cli). 30/30 discovery+content tests green.
- executor-kiro/save-call-log/call-logs-correlation tests: replace 15 'as any'
casts with typed shapes (net-new no-explicit-any errors from #6213/#6216);
prune the now-empty suppression entries so the frozen baseline stays exact
- github-skills + usage/call-logs routes: raw toLowerCase().includes() search
replaced by matchesSearch() (no-restricted-syntax — Turkish-safe search,
behavior covered by tests/unit/call-logs-correlation-substring.test.ts and
tests/unit/github-collector.test.ts)
- validate-release-green.mjs: run ESLint with --suppressions-location (match
the npm run lint contract — frozen debt is not a release red) and raise the
lint timeout 15->30min (a full pass takes ~14min alone; the 15min ceiling
expired under concurrent suite load and surfaced as 'could not parse eslint
json')
- test-masking: allowlist the 4 verified-legitimate assert reductions of the
cycle (#6248 MiMo V2 removal, #6170 Kiro catalog correction, #6154 Copilot
catalog refresh) and register the #6164 AutoRoutingBanner test deletion with
a real replacement guard (tests/unit/home-no-autorouting-banner.test.ts —
asserts the banner stays out of home/page.tsx and the component stays deleted)
- docs-sync: executors count 68 -> 73 in ARCHITECTURE.md + CODEBASE_DOCUMENTATION.md
- env-doc-sync: document OMNIROUTE_NO_SUDO (#6249/#6122) in .env.example +
docs/reference/ENVIRONMENT.md
Turbopack (stable in Next 16) becomes the code default in the three entry
points that previously required an explicit OMNIROUTE_USE_TURBOPACK=1:
- scripts/build/build-next-isolated.mjs (production build)
- scripts/dev/run-next.mjs (dev server)
- scripts/dev/run-next-playwright.mjs (playwright dev runner)
OMNIROUTE_USE_TURBOPACK=0 remains the webpack escape hatch (Windows /
native-binding / bundler-compat issues), and only the documented '0'
opts out — junk values keep the default.
Benchmarked on this codebase (same tree, Next 16.2.9): webpack 1035s vs
Turbopack 539s on a 32-core box; ~20min vs 6min59s on ubuntu-latest.
Artifact validated end-to-end (standalone smoke + e2e/package-artifact/
electron-package-smoke CI jobs, Docker amd64+arm64 builds clean with the
v3.8.27 ImportTracer panic gone on 16.2.9).
TDD: tests/unit/build-bundler-default-turbopack.test.ts (new) +
run-next-playwright.test.ts extended with the unset-env default case;
both red before the flip, green after. ENVIRONMENT.md updated.
feat(providers): add Yuanbao (web) cookie-session provider (#6196). TDD-covered; base-reds only (dast-smoke #6228, docs version-drift, executor-kiro anys — #6145 guard fixed on tip via #6270). Integrated into release/v3.8.45.
Realign the #6145 onboarding-href guard to the #6166 helper refactor (buildProviderDetailsHref). Test-only; unblocks the fast-path unit job across the open PR queue. Base-reds only (dast-smoke #6228, docs version-drift, executor-kiro anys). Integrated into release/v3.8.45.
refactor(dashboard): extract tested buildProviderDetailsHref helper for onboarding wizard (#6166). Behavioral #6144 fix already on tip via #6145; this lands the tested-helper hardening. Thanks @KooshaPari. Integrated into release/v3.8.45.
fix(antigravity): strip trailing assistant prefill for Vertex Claude models (#6114). TDD-covered (6/6), merged on TDD strength per owner. Reds are pre-existing base-red drift on release/v3.8.45. Thanks @anki1kr. Integrated into release/v3.8.45.
fix(api): count tool_use/tool_result/thinking blocks in count_tokens estimate (#6221, port from 9router#2337). TDD-covered (6/6); typed the test casts to clear no-new-eslint. Reds are pre-existing base-red drift on release/v3.8.45 (dast-smoke #6228; executor-kiro.test.ts eslint anys; changelog/package.json version drift) — none introduced by this PR. Thanks @luweiCN. Integrated into release/v3.8.45.
#6078 registered zed in the OAuth PROVIDERS registry but did not add it to the
constants PROVIDERS id map nor the oauth-providers-config enumeration test, leaving
that test red on the release tip (getProvider enumeration vs EXPECTED mismatch).
Adds ZED to the id constants + zed to EXPECTED_PROVIDER_KEYS/EXPECTED_CONFIG_BY_PROVIDER.
Registers a minimal import_token entry for the existing Zed IDE keychain-import
provider so getProvider("zed") no longer throws "Unknown provider: zed" when the
UI probes the OAuth capability endpoint; generateAuthData returns { supported: false }.
Test runner fix: the regression test imported from "vitest" but lives in tests/unit/
(node:test territory, outside the vitest include globs) — it ran in no runner. Converted
to node:test + node:assert so it actually executes (8/8 green).
Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com>
* fix(doubao-web): switch provider to Dola global
* fix(doubao-web): use .dola.com cookie domain for s_v_web_id + rebaseline test
The Dola switch left s_v_web_id with a host-only "www.dola.com" domain, which fails
the token-source contract (domain must start with "." or "http") — the sibling
sessionid/ttwid cookies and the canonical cookieDomain already use ".dola.com", which
also matches www.dola.com. Also rebaselines web-cookie-providers-new.test.ts (850->890)
for the provider-switch regression cases.
Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com>
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
* fix(doctor): resolve two false-positive WARNs (#6162)
The `omniroute doctor` command reported two warnings on healthy installs
even though the underlying checks actually passed. Both came from the
doctor probing state that already worked; they looked like bugs but users
couldn't tell without manual digging.
Issue 1 — Server liveness HTTP 401
/api/health and /api/health/degradation both require the management
token. Doctor called them without auth → 401 → WARN, even when the
Next.js server was clearly alive and listening.
Fix: probe the configured health endpoint first; on 401/403, fall
back to a publicly served static asset (/favicon.ico) to confirm the
server is alive. WARN now only fires when both probes fail.
Issue 2 — CLI Tools '@/shared' import
tool-detector.ts (and 3 other cli-helper files) import @/shared/...
aliases that resolve via tsconfig.json paths. The CLI ships raw TS
source (no compile step) and runs through tsx, but tsx does not honor
tsconfig paths at runtime, and tsconfig-paths only hooks CJS
Module._resolveFilename while doctor uses ESM `import()`.
Fix: replace @/shared/... with relative imports in the 4 cli-helper
files. This is the same pattern these files already use for ./config-
generator/* imports. No new dependency, no architectural change, and
the fix doesn't regress Next.js itself which keeps using @/shared.
Verified on v3.8.43 (Node v24.17, Windows 11):
Before: 7 ok, 2 warning(s), 0 failure(s)
After: 8 ok, N warning(s), 0 failure(s)
where N accurately reflects which CLI tools are installed and
configured for OmniRoute (e.g. Hermes Agent installed but not
pointed at 20128 → 2 real warnings, not 1 false-positive).
Refs #6162
* fix(doctor): derive fallback URL from primary URL via new URL()
Per Gemini code-assist review feedback: the previous fallback constructed
the /favicon.ico URL from defaults (127.0.0.1:PORT) which ignored custom
host/port/protocol configurations supplied via:
- OMNIROUTE_DOCTOR_LIVENESS_URL
- OMNIROUTE_DOCTOR_HOST
- --liveness-url / --host CLI flags
Parse the primary URL with new URL() to preserve protocol, host, port, and
subpaths. The previous default-based fallback remains as a catch-all for
invalid primary URLs.
* test(doctor): add regression tests for #6162 fixes
Two new test files lock the fix and satisfy the PR Test Policy gate
("production code change without tests"):
- tests/unit/cli-helper-tool-detector-paths-6162.test.ts
Locks the @/shared → relative imports fix across all 4 cli-helper
files. Asserts (a) no @/shared alias remains in the cli-helper
sources, and (b) each file is importable at runtime via tsx/ESM,
which would have thrown "Cannot find package '@/shared'" before
the fix.
- tests/unit/cli-doctor-liveness-fallback-6162.test.ts
Locks the /favicon.ico fallback in doctor.mjs. Asserts the
fallback probe exists, derives its URL from the primary URL via
new URL() (per Gemini review feedback), and that the buggy
'Server responded with HTTP 401' WARN path is gone.
Both tests use only node:test + node:assert/strict so they slot into
the existing 'test' and 'test:unit' scripts with no extra config.
* test(doctor): fix primary.ok regex in fallback test
The earlier regex /primary\.ok\s*\?/ required a '?' immediately after,
but the actual doctor.mjs code uses a multi-line if-block:
if (primary.ok) {
return ok(...);
}
Use /\bprimary\.ok\b/ instead so the assertion matches the existing
branching.
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
* fix(providers): refresh github copilot catalog
Limit GitHub Copilot discovery to the curated supported model set and keep the provider cooldown panel client-safe by moving countdown formatting out of localDb.
* chore(quality): rebaseline providerPageHelpers.ts file-size (+13, #6154 copilot catalog)
The GitHub Copilot catalog refresh grows the provider-page model-section helper
(1021->1034). Fast-path PR->release skips check:file-size, so the bump lands with
the PR. Justification recorded in file-size-baseline.json.
Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com>
---------
Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com>
* fix(chatcore): exempt opencode client from the default 128-tool truncation
The default MAX_TOOLS_LIMIT (128) cap made truncateToolList blind-slice
tools.slice(0, 128), dropping opencode's built-in task tool and part of
its MCP tools when the inbound list exceeded 128 — so models routed
through OmniRoute could not launch subagents or reach all their tools.
Detect the opencode client (any x-opencode-* header, or 'opencode' in
the user-agent) and bypass ONLY the speculative 128 default. A known
provider ceiling (proactive PROVIDER_TOOL_LIMITS or a detected limit)
always wins and still truncates, even for opencode, so upstreams with
real hard limits (e.g. grok-cli 200) keep their 400-avoidance guard.
Non-opencode clients are unchanged.
- requestFormat.ts: add isOpencodeClient(headers, userAgent) + expose it
on resolveChatCoreRequestFormat.
- toolLimitDetector.ts: add getKnownToolLimit(); getEffectiveToolLimit
becomes getKnownToolLimit(provider) ?? DEFAULT_LIMIT (byte-identical
for existing callers).
- upstreamBody.ts: truncateToolList takes bypassDefaultToolLimit and
encodes the precedence; fix cosmetic debug-log count.
- chatCore.ts: thread the flag into prepareUpstreamBody.
- tests: extend tool-limit-detector unit tests.
* refactor(tools): accept nullable provider in tool-limit resolvers
Address PR review: widen getKnownToolLimit / getEffectiveToolLimit to
(provider: string | null | undefined) to match the call sites in
truncateToolList, and add unit assertions covering null/undefined
providers (getKnownToolLimit -> null, getEffectiveToolLimit -> 128).
---------
Co-authored-by: DKotsyuba <16292493+DKotsyuba@users.noreply.github.com>
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
Kiro/CodeWhisperer streams Claude's reasoning as native `reasoningContentEvent`
frames when adaptive thinking is enabled, but the Kiro executor had no handler
for them, so `reasoning_effort` requests returned no reasoning. Wire it end to
end:
- translator (openai-to-kiro): enable Kiro thinking when the request carries
`reasoning_effort`, Anthropic `output_config.effort`, or a `thinking` block
(`{type:"enabled",budget_tokens}` mapped to a level; `{type:"adaptive"}`
defaults to `high`, matching Anthropic's documented default). Prepends the
Kiro `<thinking_mode>`/`<max_thinking_length>` prompt directive and sets
top-level `additionalModelRequestFields` ({output_config.effort,
thinking:{type:"adaptive"}, max_tokens}). Gated on `supportsReasoning`; drops
non-default temperature/top_p (rejected by adaptive-only Claude models).
- executor transformRequest: forward `additionalModelRequestFields` to AWS
(previously dropped by the strict top-level allowlist).
- executor stream loop: parse `reasoningContentEvent` (and reasoningText
variants) into the OpenAI reasoning_content channel.
Verified against the live CodeWhisperer stream: reasoningContentEvent frames are
returned, and larger effort/budget measurably deepens reasoning up to the model
cap. Unit tests cover the effort sources, forwarding, temp/top_p stripping, and
native reasoning-frame parsing.
POSIX shells (bash/zsh) always set HOSTNAME to the machine name. The
.env loader uses first-wins semantics, so HOSTNAME=0.0.0.0 in .env is
silently ignored. This causes the server to bind to the LAN hostname
instead of 0.0.0.0, breaking localhost access and all internal
self-requests (ModelSync, HealthCheck, cloud sync).
The fix compares process.env.HOSTNAME against os.hostname(): when they
match, it's the POSIX auto-set signature and HOSTNAME is ignored.
OMNIROUTE_SERVER_HOST takes precedence as the dedicated escape hatch.
Backward compatibility is preserved: users who set HOSTNAME to a value
that doesn't match the machine name (e.g. Windows CMD/PowerShell users
with HOSTNAME in .env) will still have their value honoured.
Closes#6194
* feat(provider): add Claude 5 Sonnet to Claude Web provider (#6200)
* test(providers): guard claude-web claude-sonnet-5 registry entry (#6209)
Adds the missing regression test the PR-test-policy gate requires: asserts the
claude-web registry exposes claude-sonnet-5 (Claude 5 Sonnet web) alongside the
existing 4.6 Sonnet / 4.5 Haiku entries. Fails on the release base (no entry).
Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com>
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
NVIDIA NIM API (nvidia/* models) silently truncates the tool list to 128
(the default MAX_TOOLS_LIMIT) because nvidia is not in PROVIDER_TOOL_LIMITS.
Tools beyond index 127 are dropped, causing agents to lose access to
critical tools like task, read, or high-index MCP tools.
Verified that NVIDIA NIM API supports up to 1536 tools by direct testing.
End-to-end confirmed: 198 tools sent, model successfully called tools at
indices 193, 195, and 197 (previously dropped by truncation to 128).
Follows the same pattern as #5563 (grok-cli: 200), integrated in v3.8.43.
* fix(translator): strip reasoning param for nvidia z-ai/glm-5.2
NVIDIA NIM OpenAI-compatible wrapper rejects the reasoning body field
and returns HTTP 400 "Unsupported parameter(s): `reasoning`".
Add a StripRule scoped to provider=nvidia + model /z-ai\/glm-5\.2/i.
Mirrors PR #6102 drop pattern (minimax-m2.7 thinking).
* docs(translator): tighten nvidia glm-5.2 strip-rule comment
* fix(translator): anchor glm-5.2 strip rule with word boundary
* fix(dashboard): providers page data-timeout guard + live-ws standalone wiring
Captura de trabalho em progresso: timeout de dados na página de providers,
ajuste em ProviderLimits e instrumentation-node, com testes novos
(providers-page-data-timeout, live-ws-standalone-wiring).
* chore(quality): rebaseline ProviderLimits/index.tsx file-size (+6, #6211 data-timeout guard)
Cohesive fix growth from PR #6211's data-timeout guard on the quota page's two
first-paint fetches (1121->1127). The fast-path PR->release skips check:file-size,
so the bump lands with the PR. Justification recorded in file-size-baseline.json.
When ChatGPT Web generates an image as an image_asset_pointer but the pointer
fails to resolve to a downloadable URL (unknown asset scheme, download 403/
expired, oversize), resolveImagePointers returned [] — indistinguishable from
'no image produced' — so the image-generation handler reported the misleading
502 'completed without returning image markdown'. The image genuinely existed
upstream; OmniRoute dropped it silently.
Fix: the executor flags x_image_resolution_failed when a pointer existed but
none resolved (and logs the unresolved asset scheme for follow-up), and the
handler surfaces a truthful 'generated but not retrievable' 502 instead of
'no image markdown'. Adds executorFactory DI for unit testing.
TDD: tests/unit/chatgpt-web-image-silentdrop.test.ts (red -> green), plus the
existing chatgpt-web / image-generation-handler suites stay green.
Reported via community triage (mesh escalated backlog).
* fix(providers): correct Kiro model catalog to real upstream ids
Kiro's API (generateAssistantResponse) returns 400 "Invalid model. Please
select a different model" for any id it does not recognize. The registry
exposed fabricated ids (copied from OmniRoute's own Anthropic catalog) that
Kiro never serves, so every call to them 400'd. Live-verified on the VPS:
Removed (400 Invalid model):
- auto-kiro (no "auto" model id — was sent verbatim upstream)
- claude-fable-5 (Kiro offers no Fable)
- claude-opus-4.8/4.7/4.6 (Kiro offers no Opus)
Corrected:
- claude-sonnet-4.6 -> claude-sonnet-4.5 (Kiro's Sonnet is 4.5; 4.5 -> 200)
Kept:
- claude-sonnet-5 (real Kiro model, plan-gated per account)
- claude-haiku-4.5, deepseek-3.2, glm-5, minimax-m2.5/m2.1,
qwen3-coder-next (all proven 200 on the VPS)
Aligns the free-model catalog and drops the orphaned auto-kiro price key.
Regression guard: tests/unit/kiro-catalog-real-models.test.ts (3/3).
Kiro cluster #6112/#6113/#6099.
* test(providers): align stale Kiro-catalog tests to the corrected upstream ids
The fabricated Kiro ids removed in the parent commit (claude-fable-5,
claude-opus-4.8/4.7/4.6, claude-sonnet-4.6) were still asserted as present by
three pre-existing tests, which encoded the bug:
- catalog-updates-v3x: now asserts Kiro does NOT expose Fable 5 / Opus (kept the
legit cc exposure) and guards the real claude-sonnet-4.5 pricing.
- model-family-fallback-notation: the dot-notation example moves from kiro/ to
anthropic/ (which genuinely serves Opus/Fable in dot notation) — coverage kept.
- provider-models-route: the Kiro local-catalog assertion now expects the real
Sonnet 5 / Sonnet 4.5 set and negatively guards the fabricated ids.
Co-authored-by: diegosouzapw <diegosouza.pw@gmail.com>
* fix(cline): force upstream streaming for Cline/ClinePass (streaming-only API)
Cline's API (api.cline.bot) only implements streaming (streamText). A
non-streaming request returns HTTP 500 "generateText is not implemented" (Claude
models) or HTTP 502 "empty response" (others). Live-verified on the VPS:
stream:true → works (STREAM_OK), stream:false → fails. This is why testing a Cline
model in the dashboard (the test button sends stream:false) failed.
Fix (reuses the existing isClaudeCodeCompatible mechanism, no new handler):
- Flag `cline` and `clinepass` registry entries with `forceStream: true`.
- In chatCore, OR `providerRequiresStreaming` into `upstreamStream` (line 1591)
so the upstream request always streams for these providers, while the client's
original `stream` intent still drives the response format. The existing
non-streaming branch (parseNonStreamingResponseBody) already accumulates the
upstream SSE and converts it back to JSON for stream:false clients — the same
path Claude-Code-compatible providers already use.
Tests (Rule #18): tests/unit/cline-force-stream.test.ts pins the registry flags +
resolveStreamFlag forcing behavior. Live VPS before/after recorded on the PR.
* fix(sse): cline forceStream must stream upstream only, keep client JSON
The #2081 wiring fed providerRequiresStreaming into resolveStreamFlag,
forcing the client-facing stream flag to true for forceStream providers.
That skips the if(!stream) branch that drains a forced upstream SSE and
converts it back to JSON, so a stream:false caller (model-test button,
plain JSON API) got STREAM_EARLY_EOF instead of a JSON body.
Keep providerRequiresStreaming only on upstreamStream (force upstream to
stream); leave the client-facing stream as the client sent it, so
readNonStreamingResponseBody accumulates the SSE into JSON. The promised
handleForcedSSEToJson (#2081 comment) was never implemented — this uses
the existing non-streaming SSE-buffering path (same as isClaudeCodeCompatible).
Live-verified on VPS: cline stream:true worked, stream:false failed.
Clicking 'test' on a provider model (e.g. a ClinePass flash model) could freeze
the entire dashboard. Root cause: POST /api/models/test returned an OBJECT in
`error` on the Zod-validation and invalid-JSON paths (`validation.error.format()`
/ a details object). The client does `notify.error(data.error)`, and
NotificationToast renders the message directly as a React child — an object throws
React #31 ('Objects are not valid as a React child'), crashing the tree = frozen
page instead of a toast.
Fixed in three layers (defense in depth):
1. Server (root cause): /api/models/test now returns a STRING `error` on every
path — flattens Zod issues to text, returns 'Invalid JSON body' for bad JSON.
2. Client: onTestModel funnels the response through extractApiErrorMessage() so any
object-shaped error is coerced to a string before notify.error.
3. Toast: NotificationToast coerces title/message via toToastText() — a resilient
catch-all so no future caller can freeze the page with a non-string.
Tests (Rule #18, both node:test / blocking suite):
- tests/unit/models-test-error-shape.test.ts — asserts STRING error on Zod-fail,
missing-field, and invalid-JSON (fails on the pre-fix route: 3/3 red -> green).
- tests/unit/notification-toast-coercion.test.ts — toToastText coercion matrix.
`src/app/api/oauth/[provider]/[action]/route.ts` grew to 959 lines, past its
frozen cap of 924 (`check:file-size` → Fast Quality Gates red on release/v3.8.44).
The growth came from #6054 (graceful 400 for keychain-import-only providers / zed):
a doc block, two Sets (KEYCHAIN_IMPORT_ONLY_PROVIDERS, OAUTH_FLOW_ACTIONS) and a
keychainImportOnlyResponse() helper, plus two duplicated guard blocks in GET/POST.
That is a cohesive, self-contained leaf, so extract it to a new
`keychainImportOnly.ts` exposing `keychainImportOnlyGuard(provider, action)`
(returns the 400 NextResponse or null). The two route callsites collapse to a
2-line guard each. route.ts: 959 -> 918 (< 924, freeze restored). No behavior
change.
Tests (Rule #8/#18):
- Existing tests/unit/oauth-keychain-import-only-6041.test.ts (route-level GET/POST
zed 400) still pass unchanged — behavior preserved.
- New tests/unit/oauth-keychain-import-only-guard.test.ts pins the extracted guard
in isolation (zed+flow -> 400, normal provider -> null, zed+non-flow -> null).
