When the Hermes Web UI probe fails, the toolbar entry now opens an info
confirm dialog offering to run `hermes dashboard` in the user's preferred
terminal. Accepting spawns it via a temp bash/batch script; `hermes
dashboard` itself opens the browser once ready, so we do not poll.
The Memory panel and Health banner keep the existing toast behavior.
Also corrects the stale `hermes web` hint in the offline toast (the real
command is `hermes dashboard`) and reorders Linux terminal detection to
try `which` before stat'ing /usr/bin, /bin, /usr/local/bin.
useAutoCompact cached normalWidthRef = el.scrollWidth on every
non-compact resize, but per DOM spec scrollWidth === clientWidth
when content fits. Maximizing the window (content no longer
overflows) therefore wrote the container width into
normalWidthRef, making it impossible to re-enter compact when
the window was restored to its original size.
Move the assignment inside the overflow branch so the cache is
only written at the actual compact threshold, where scrollWidth
reflects the real content width.
Normalize all icon-only ghost buttons in the header toolbar to
32x32 (w-8 px-2). Previously Hermes/OpenClaw used default sm
padding (~40px) while Claude's Skills/Sessions used w-8 px-2,
so switching apps caused a width jump and put useAutoCompact's
cached normalWidthRef out of sync across apps.
Wire hermes through SkillApps struct, DAO SQL, command parser, and
SKILLS_APP_IDS. Add a Skills entry to the Hermes toolbar. Simplify
skill_sync test fixtures to use SkillApps::default().
Hermes 0.10.0 tightened custom_providers validation (commit 2cdae233):
invalid base_urls are rejected, unknown fields produce warnings, and
new fields (rate_limit_delay, bedrock_converse, key_env) landed.
- Add bedrock_converse to the api_mode selector (and i18n labels)
- Expose rate_limit_delay in a provider-level advanced panel
- Validate base_url client-side (URL shape, template-token friendly)
- Drop per-model max_tokens — not in _VALID_CUSTOM_PROVIDER_FIELDS
- Round-trip test asserts set_provider preserves rate_limit_delay /
key_env / any unknown forward-compat field
The model-mapping quick-set button referenced an undefined `reasoningModel`
prop and wrote `ANTHROPIC_REASONING_MODEL`, which the backend explicitly
marks as deprecated legacy (see services/provider/mod.rs and
services/proxy.rs). Remove all three references so typecheck passes and
the button matches the provider model schema.
Hermes' built-in api_mode detection only matches a handful of official
endpoints (api.openai.com, api.anthropic.com, api.x.ai, AWS Bedrock);
third-party / proxy endpoints silently fall back to chat_completions,
which causes opaque 401/404s on Anthropic-protocol or Codex-Responses
providers. The "Auto" option was misleading for the common third-party
case.
- Drop the "Auto" option from the API Mode dropdown; remove the
HermesApiModeChoice sentinel type so writes always emit api_mode.
- Default new providers and legacy entries lacking api_mode to
chat_completions (only persisted on user save).
- Deeplink imports now write api_mode: chat_completions explicitly
instead of relying on URL heuristics; test renamed accordingly.
- Rename the "Codex Responses (Copilot / OpenCode)" label to
"OpenAI Responses" to match OpenAI's /v1/responses naming.
After /simplify review of the P1-3 second wave, two small cleanups:
- Lift the `_cc_source` / `providers_dict` magic strings out of
ProviderCard into a shared helper (`isHermesReadOnlyProvider`) and
named constants in hermesProviderPresets.ts. Front-end and back-end
now document the same marker contract in two mirrored places
instead of drifting strings.
- Replace the duplicate `is_dict_only_provider` + `format!` branches
at the top of `set_provider` / `remove_provider` with a single
`ensure_provider_writable(config, name, verb)` guard. Future error
copy tweaks only have to happen once.
No behaviour change; all 52 hermes_config tests stay green.
ProviderCard now detects Hermes provider entries sourced from the
v12+ `providers:` dict via the `_cc_source` marker that the backend
injects, and renders a "Hermes Managed" badge beside the title.
ProviderActions receives an `isReadOnly` prop that disables the Edit
and Delete buttons (with a tooltip pointing the user at Hermes Web
UI) while keeping Switch and Duplicate enabled — switching only
touches `model.*`, and duplicate lets users fork the overlay into
their own `custom_providers:` list.
Three-locale i18n keys `provider.managedByHermes` /
`provider.managedByHermesHint` added.
Hermes v12+ migrated some provider entries from the `custom_providers:`
list into a `providers:` dict (keyed by id). CC Switch previously
ignored that source entirely, leaving users blind to providers they had
configured via Hermes' own Web UI; the only feedback was a generic
migration warning in the health banner.
`get_providers()` now unions both sources, matching upstream
`get_compatible_custom_providers` dedup order (list wins on name
collision). Entries coming from the dict carry a `_cc_source =
"providers_dict"` marker plus the original `provider_key`, which the
UI layer will use to render them read-only. `set_provider` and
`remove_provider` now refuse to touch dict-only entries, steering the
user to Hermes Web UI. `sanitize_hermes_provider_keys` strips the UI
markers on write so they never reach YAML.
The `schema_migrated_v12` health warning copy reframes the situation:
entries are shown read-only in CC Switch rather than invisible.
DeepLink Hermes import was emitting camelCase (baseUrl / apiKey /
apiMode) that the Hermes runtime does not recognise, poisoning
`custom_providers:` entries on activation. The MCP sync path was
also stripping `auth: oauth` on round-trip, silently downgrading
OAuth-type servers to unauthenticated calls.
The Hermes deeplink branch now emits snake_case via a dedicated
builder; `sanitize_hermes_provider_keys` runs on both `set_provider`
and `get_providers` so legacy DB records heal on next access.
`HERMES_EXTRA_FIELDS` preserves `auth`. The `api_mode` dropdown gains
`codex_responses` (Copilot / OpenCode), and the schema-migrated
warning copy no longer hard-codes "v12" (upstream `_config_version`
is now 19).
Replaces the greyed-out "Memory is disabled" banner with a real Switch
at the top of each memory tab. Users can now toggle Hermes' memory/user
blobs without leaving CC Switch; the underlying write goes through the
merge-aware `set_memory_enabled`, so budgets and external-provider
settings survive toggle operations. The new `useToggleHermesMemoryEnabled`
mutation invalidates the limits query so the Switch state and the
amber disabled-hint update in lockstep.
Reworks the `schema_migrated_v12` health banner copy to match the
simplified "CC Switch only manages custom_providers" posture — it now
tells users to reconcile migrated dict entries via Hermes Web UI,
instead of the earlier (and now inaccurate) "CC Switch reads both".
Adds a dedicated Hermes row to the directory-override settings so users
can point CC Switch at alternate Hermes config locations (e.g. a second
profile directory for work/personal split). `get_config_dir` on the
Rust side already supports hermes; this just wires up the frontend row.
Wiring it through `useDirectorySettings` revealed a scaling problem:
every supported app required five parallel ternary chains across
`computeDefaultConfigDir`, `updateDirectory`, `browseDirectory`,
`resetDirectory`, and `updateDirectoryState`. Replaces those with two
lookup tables (`APP_DIRECTORY_META`, `DIRECTORY_KEY_TO_SETTINGS_FIELD`)
so adding the next app is two entries, not fifteen edit sites.
Drive-by cleanup from the same touch:
* `resetAllDirectories` takes a `ResolvedAppDirectoryOverrides` object
instead of five positional optional strings.
* `setResolvedDirs` returns the same reference when the sanitized
value is unchanged, so no-op edits don't cascade renders.
Also lands all i18n updates for this series (`hermesConfigDir` and
placeholder, Memory section's enable/disable/toggleFailed copy, and
the reworded `schemaMigratedV12` warning) in zh/en/ja together.
Hermes has no slash-prompt concept (templates live as Skills), so the
Prompts tab for the Hermes app was always empty. Swap the toolbar Book
button for a Brain button that opens a new Memory panel editing
~/.hermes/memories/{MEMORY,USER}.md — Hermes' first-class memory store
which its Web UI exposes only as on/off toggles, never as an editor.
The panel shows each file in its own tab with a character-budget bar
read from config.yaml's nested memory.* section (memory_char_limit /
user_char_limit, default 2200 / 1375). Edits are written atomically;
Hermes picks them up on the next session start per MemoryStore.
Also extract useDarkMode to src/hooks/useDarkMode.ts — the codebase
already repeats the same MutationObserver pattern in 12+ places; this
PR introduces the shared hook and uses it once, leaving the migration
of the other copies to a follow-up.
Slim the Hermes surface in CC Switch to match its core positioning —
cross-client provider switching and shared MCP/prompts/skills — and
delegate deep configuration (model, agent, env, skills, cron, logs)
to the Hermes Web UI at http://127.0.0.1:9119.
- Drop AgentPanel/EnvPanel/ModelPanel and their mutation commands,
hooks, types, and i18n keys across zh/en/ja.
- Add open_hermes_web_ui Tauri command that probes /api/status and
launches the URL in the system browser. Hermes injects its own
session token into the returned HTML, so CC Switch doesn't need
to touch auth.
- Surface the launcher from the Hermes toolbar and the health banner
via a shared useOpenHermesWebUI() hook; the offline error code is
defined once per side and referenced across the contract.
- Keep read-only access to model.provider so ProviderList can still
highlight the active supplier; apply_switch_defaults continues to
write the top-level model section when switching providers.
Net diff: +152 / -1253.
Migrate 18 Hermes provider presets from anthropic_messages to
chat_completions to sidestep known upstream Hermes bugs (model-name
dot-mangling in normalize_model_name, api_mode drop after v11->v12
migration, and auxiliary_client OpenAI hardcode).
Native providers now target each vendor's official OpenAI-compatible
endpoint with correct model IDs: Kimi (kimi-k2.5-preview on /v1),
Bailian (compatible-mode/v1 with Qwen3 defaults), Xiaomi MiMo, Longcat
(/openai/v1), Zhipu GLM (/api/paas/v4), ModelScope, MiniMax, SiliconFlow,
and Novita (/v3/openai).
Aggregators (Shengsuanyun, AiHubMix, DMXAPI, Compshare, TheRouter)
default to GPT-5.4 on chat_completions, mirroring the Codex preset
lineup. TheRouter omits gpt-5.4-pro since that variant is Responses-only
and Hermes implements only chat_completions. OpenRouter's existing
openai/gpt-5 entry is bumped to openai/gpt-5.4.
Claude-only proxies are left on anthropic_messages; their Codex
counterparts use wire_api=responses, so there is no evidence their
chat_completions endpoints serve OpenAI models.
Remove the Anthropic, OpenAI, and Google AI presets from the Hermes
preset list. They were placeholder samples introduced when the Hermes
module first landed and do not match the actual user paths in
CC Switch (Claude / Codex go through OAuth, Gemini Native is its own
adapter), and the upstream endpoints are not reachable for most of
the target users anyway.
Fix the Nous Research preset: its base_url was a fabricated domain
(inference.nous.hermes.dev) that has never resolved. Point it at the
real Nous Portal endpoint (inference-api.nousresearch.com/v1) and
add apiKeyUrl so users can jump straight to portal.nousresearch.com
to provision a key.
Drop the now-orphan providerForm.presets.{anthropic,openai,googleai}
i18n keys from zh / en / ja since no preset references them anymore.
Import 38 Claude presets into Hermes by mapping env-style
ANTHROPIC_BASE_URL/AUTH_TOKEN to flat base_url/api_key, deriving
api_mode from apiFormat (anthropic_messages or chat_completions),
deduping ANTHROPIC_*_MODEL into models[], and pointing
suggestedDefaults at ANTHROPIC_MODEL. Skip OAuth-only presets
(Codex, Copilot), Bedrock SigV4, Gemini Native, and the three
already shipped on the Hermes side (OpenRouter, Anthropic,
DeepSeek). Place Shengsuanyun at the head of the Hermes array so
the partner shows first in the preset panel.
In the Claude preset list, restore Shengsuanyun back ahead of
Gemini Native. The Gemini Native preset (#1918) was inserted
between Claude Official and Shengsuanyun, which made the
third_party category register first in the reduce-based grouping
and pushed the aggregator block (and Shengsuanyun) behind it.
Backfill the missing providerForm.presets translations across zh,
en, and ja (openrouter, anthropic, openai, googleai, deepseek,
together; plus shengsuanyun for en and ja) so existing Hermes
preset names no longer render literal i18n keys.
Switching a Hermes provider previously only fired a toast because the frontend treated Hermes as non-additive (unlike backend AppType::is_additive_mode, which lists OpenCode | OpenClaw | Hermes) and relied on the unused is_current DB flag for highlighting. Align the UI model with the backend:
- Include Hermes in ProviderActions' isAdditiveMode so the main button switches between "Add" and "Remove".
- Drive the "current" highlight from model.provider (via useHermesModelConfig) instead of the DB is_current field; model.provider is Hermes's real SSOT for the active provider.
- Reuse OpenClaw's set-as-default button slot to expose an "Enable" action on Hermes that calls switchProvider, so providers already in config can be activated without re-adding. switch_normal + apply_switch_defaults already atomically update custom_providers and model.provider, so no backend change is needed.
- Invalidate liveProviderIds + modelConfig + health in parallel after add/update/delete/switch via a new invalidateHermesProviderCaches helper, replacing four copies of three sequential awaits.
The Hermes provider form previously only exposed Base URL and API Key,
forcing users to drop into the Model panel to hand-edit model IDs after
adding a provider. Following OpenClaw's shape, the form now carries:
- An API Mode selector (auto-detect / chat_completions / anthropic_messages).
"auto" is a UI-only sentinel — selecting it deletes api_mode from the
config so the YAML doesn't leak a redundant field.
- A model list editor where the first row is badged as the default and
each row has a collapsible Advanced panel for context_length and
max_tokens. Adding/removing rows uses a UUID-keyed ref so typing in
one input doesn't drop focus when another row is added.
- A Fetch Models button that pulls /v1/models from the configured
endpoint and exposes the catalog in a per-row dropdown, identical to
OpenClaw's flow. The vendor grouping is memoized so keystrokes don't
trigger a reduce+sort per model row — Radix DropdownMenuContent does
not lazy-mount, so the inner JSX evaluates on every render regardless
of whether the menu is open.
Three-locale i18n keys are added together (zh/en/ja).
Hermes custom_providers entries now carry an ordered models array
(id / context_length / max_tokens) plus suggestedDefaults. The backend
serializes the array to the YAML dict shape Hermes expects on write and
inverts it on read, preserving insertion order via the preserve_order
feature on serde_json.
When a user switches providers, switch_normal calls apply_switch_defaults
so the top-level model.default / model.provider follow the selected
provider's first model. Previously switching a Hermes provider only
shuffled custom_providers[] and left Hermes pointing at whatever
model.provider was set before.
Seven existing Hermes presets now ship with a curated models list so
switching lands on a working default without a detour through the
Model panel.
The previous Hermes icon was an inline purple "H" SVG unrelated to the
real Hermes Agent brand. Replace it with the official Nous Research
avatar sourced from hermes-agent/landingpage/icon-512.png, routed
through iconUrls as a Vite URL import. The PNG is post-processed to
strip the 4px black rectangle border (cropped inward by 6px and pasted
back onto a 512x512 transparent canvas). Also switch defaultColor to
black to match the monochrome artwork when the asset falls back to an
initial glyph.
- OpenClaw: replace opus-4-6 with opus-4-7 across 17 aggregator presets
(id, name, primary, modelCatalog); AWS Bedrock entry rewritten to new
SKU anthropic.claude-opus-4-7 (drops -v1 and dated suffix per official
4.7 model card) and pricing corrected to $5/$25/$0.50/$6.25 during the
SKU swap, aligning with schema.rs source of truth
- OpenCode: same replacement for 13 aggregators plus
OPENCODE_PRESET_MODEL_VARIANTS entries for @ai-sdk/amazon-bedrock and
@ai-sdk/anthropic, plus AWS Bedrock provider models map
- OpenRouter / TheRouter / GitHub Copilot in claudeProviderPresets use
dot-style id; update to anthropic/claude-opus-4.7 (missed by 509d2250)
- omo: switch agent/category recommended to opus-4-7; replace key in
OMO_BACKGROUND_TASK_PLACEHOLDER priority map
- hermes_config.rs: update doc comments and test fixtures to opus-4-7;
Hermes ModelPanel placeholder and i18n defaultHint examples follow
- i18n unspecifiedHigh category description bumped to 'Claude Opus 4.7
max variant' to match omo recommended
- Test fixtures updated: therouter preset assertion and opencode Bedrock
variant lookup now check for opus-4-7
- Sonnet 4.6 / Haiku 4.5 untouched - no official 4.7 release for them
- Seed claude-opus-4-7 pricing (same tier as 4.6: $5 / $25 / $0.50 /
$6.25 per million tokens). Relies on incremental INSERT OR IGNORE
seeding; no SCHEMA_VERSION bump needed.
- Whitelist opus-4-7 in thinking optimizer so it uses adaptive
thinking + max effort + 1M context beta, matching 4.6 behavior.
- Bump default OPUS model in PIPELLM and AWS Bedrock (AKSK / API Key)
presets to 4.7. Bedrock SKU drops the -v1 suffix per the official
4.7 model card (anthropic.claude-opus-4-7 and
global.anthropic.claude-opus-4-7).
Align the New API usage query template with the GENERAL template by
including "User-Agent: cc-switch/1.0" in its request headers, so
cc-switch requests are identifiable in provider server logs and less
likely to be blocked by UA-based rate limiting on some New API
deployments.
Per Anthropic's effort parameter docs, "high" is the API default and
setting effortLevel="high" is equivalent to omitting the field entirely.
The toggle previously produced no effect.
Claude Opus 4.6, Sonnet 4.6, and Opus 4.7 now support a "max" level
that enables unconstrained reasoning. Rename the checkbox (effortHigh
-> effortMax) and write effortLevel="max" when toggled on. Existing
"high" values in user configs are left untouched.
Updates zh/en/ja locales and user-manual entries accordingly.
- Add "hermes" to AppId union type and all exhaustive Record<AppId>
- Add HermesModelConfig, HermesAgentConfig, HermesEnvConfig types
- Add hermes field to VisibleApps, McpApps, ProxyTakeoverStatus
- Create src/lib/api/hermes.ts with Tauri invoke wrappers
- Create src/hooks/useHermes.ts with 5 query + 3 mutation hooks
- Register hermes in APP_IDS, APP_ICON_MAP (violet color scheme)
- Split MCP_SKILLS_APP_IDS into MCP_APP_IDS (includes hermes) and
SKILLS_APP_IDS (excludes hermes, since Hermes has no Skills support)
- Wire hermes additive-mode into App.tsx (remove/duplicate handlers),
ProviderList.tsx (live provider ID query + In Config badge),
mutations.ts (cache invalidation on switch/add/delete)
- Add Hermes checkbox to McpFormModal
- Add basic hermes i18n keys (en/zh/ja)
The takeover.failed i18n template lacked the {{detail}} placeholder
and three useProxyStatus onError callbacks omitted the detail variable,
so proxy start/stop/takeover failures all displayed a generic message
regardless of the underlying cause.
ANTHROPIC_REASONING_MODEL was a non-official env var that forced all
requests with thinking params to use a single "reasoning model",
overriding the user's /model selection. Since new Claude Code versions
send adaptive thinking by default, this caused /model to silently fail.
- Remove reasoning_model field and has_thinking_enabled() from model_mapper
- Simplify map_model() to pure type-based matching (haiku/sonnet/opus)
- Remove reasoning model UI field from provider form
- Retain ANTHROPIC_REASONING_MODEL in ENV_EXCLUDES and override-key
cleanup lists so legacy configs don't leak into common config
* fix(claude-plugin): sync current provider config to settings.json on toggle enable
- Extract syncClaudePluginIfChanged to share logic between autoSaveSettings and saveSettings
- Fix P1: enableClaudePluginIntegration toggle in General tab now actually syncs ~/.claude/settings.json
- Fix P2: check syncCurrentProvidersLiveSafe() return value and show toast on failure
- Fix P3: sync providers on both enable and disable, not just enable
- Fix P4: avoid double syncCurrentProvidersLiveSafe when plugin toggle + dir change happen together
- Remove duplicate comment
- Add missing providersApi.getCurrent/getAll mocks in tests
* style: reformat after rebase onto main
Prettier flagged a line-break introduced by the openclaw directory
change (from main) after rebase.
* fix(claude-plugin): read prev enabled state from live cache to avoid stale closure
syncClaudePluginIfChanged compared enabled against data?.enableClaudePluginIntegration
captured in a useCallback closure. After invalidateQueries + refetch, the React
Query cache is up to date, but the consuming hook's closure does not see the new
value until React re-renders. Quick on->off toggles could therefore skip
applyClaudePluginConfig, leaving ~/.claude/config.json in the previously enabled
state even though settings.json was persisted as disabled.
Read the previous value synchronously from queryClient.getQueryData(["settings"])
before saveMutation.mutateAsync(), then pass it to the helper as prevEnabled.
getQueryData bypasses the closure and reflects the live cache at call time.
Test covers the race: closure data stays at false while the cache reports true;
the helper must still call applyClaudePluginConfig({ official: true }).
---------
Co-authored-by: Jason <farion1231@gmail.com>
* feat(ui): add quick-set button for model mapping fields
Add a "一键设置" (Quick Set) button in the model mapping section to
simplify provider configuration. When users enter a model name in any
of the five model fields, they can now click this button to populate
all fields with that same value.
This addresses the UX friction of manually filling all five model
mapping fields (主模型, 推理模型, Haiku, Sonnet, Opus) when the
provider uses the same model name across all request types.
Implementation:
- Add Wand2 icon import from lucide-react
- Insert quick-set button alongside existing fetch-models button
- Logic picks first non-empty model value and applies to all fields
- Show success toast after applying
- Disabled state when all model fields are empty
- Add i18n strings for zh, en, ja locales
Relates to user feedback about tedious model configuration workflow.
* style(ui): format ClaudeFormFields component code
Apply consistent code formatting to ClaudeFormFields.tsx following
project linting rules. Includes multi-line import statements and
improved readability for conditional expressions.
* feat(copilot): add GitHub Enterprise Server support
* fix(copilot): address GHES PR review findings (P1 + 2×P2)
- P1: Use composite account ID (domain:user_id) for GHES to prevent
cross-instance ID collisions; github.com keeps plain numeric ID for
backward compatibilit
- P2-a: Use get_api_endpoint() for model list URL with automatic
fallback to static URL when dynamic endpoint resolution fails
- P2-b: Add normalize_github_domain() as backend SSOT for domain
normalization (lowercase, strip protocol/path/query, reject userinfo)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
* refactor(proxy): extract take_sse_block helper with CRLF delimiter support
Replace inline `buffer.find("\n\n")` SSE splitting logic across streaming,
streaming_responses, response_handler, and response_processor with a shared
`take_sse_block` function that handles both `\n\n` and `\r\n\r\n` delimiters.
* feat(proxy): add Gemini Native URL builder and full-URL resolver
Introduce gemini_url module that normalizes legacy Gemini/OpenAI-compatible
base URLs into canonical models/*:generateContent endpoints. Supports both
structured Gemini URLs (auto-normalized) and opaque relay URLs (pass-through
with query params only).
* feat(proxy): add Gemini Native schema, shadow store, transform, and streaming
- gemini_schema: Gemini generateContent request/response type definitions
- gemini_shadow: session-scoped shadow store for thinking signature and
tool-call state replay across streaming chunks
- transform_gemini: bidirectional Anthropic Messages ↔ Gemini Native
request/response conversion with thinking block and tool-use support
- streaming_gemini: Gemini SSE → Anthropic SSE streaming adapter with
incremental thinking/text/tool_use delta emission
* feat(proxy): wire Gemini Native format into proxy core and Claude adapter
Integrate gemini_native api_format throughout the proxy pipeline:
- ClaudeAdapter: detect Gemini provider type, Google/GoogleOAuth auth
strategies, and suppress Anthropic-specific headers for Gemini targets
- Forwarder: Gemini URL resolution, shadow store threading, endpoint
rewriting to models/*:generateContent with stream/non-stream variants
- Handlers: route Gemini streaming through streaming_gemini adapter and
non-streaming through transform_gemini converter
- Server/State: add GeminiShadowStore to shared ProxyState
- StreamCheck: support gemini_native health check with proper auth headers
* feat(ui): add Gemini Native provider preset and api format option
- Add gemini_native to ClaudeApiFormat type and ProviderMeta.apiFormat
- Add "Gemini Native" provider preset with default Google AI endpoints
- Show Gemini-specific endpoint hints and full-URL mode guidance
- Add gemini_native option to API format selector in ClaudeFormFields
- Add i18n strings for zh/en/ja
* feat(proxy): add Gemini Native tool argument rectification
* feat(proxy): update Gemini streaming and transformation logic
* fix(proxy): align shadow turns to tail on client history truncation
* fix: revert unrelated cache_key change in claude proxy transform
Restore .unwrap_or(&provider.id) fallback for cache_key to match main
branch behavior. Only gemini_native related changes should be in this branch.
* Prevent Gemini review regressions in streaming and tool rectification
PR #1918 review feedback exposed two correctness issues in the Gemini Native adapter path. Gemini SSE buffering was still using lossy UTF-8 decoding, which could corrupt split multibyte payloads and drop streamed output. Tool arg rectification also removed top-level parameters eagerly, which broke tools that legitimately define a parameters field.
This change moves Gemini SSE buffering onto the existing append_utf8_safe path and makes parameters flattening conditional on the schema actually expecting nested extraction. The old Skill rectification path stays intact, and new regression tests cover both the preserved parameters case and UTF-8-split JSON payloads.
Constraint: Existing PR #1918 review feedback must be fixed without staging unrelated local docs and artifact files
Rejected: Keep String::from_utf8_lossy in Gemini SSE buffering | corrupts split multibyte payloads and can drop JSON chunks
Rejected: Always preserve the parameters wrapper | regresses the existing nested-parameters rectification path for Skill-style tools
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep Gemini SSE buffering on the UTF-8-safe accumulator path and only unwrap parameters when the target schema does not declare it as a legitimate field
Tested: cargo fmt --manifest-path src-tauri/Cargo.toml --all; cargo test --manifest-path src-tauri/Cargo.toml preserves_utf8_boundaries_when_json_payload_spans_chunks; cargo test --manifest-path src-tauri/Cargo.toml gemini_to_anthropic_rectifies_tool_args_from_schema_hints; cargo test --manifest-path src-tauri/Cargo.toml rectifies_streamed_skill_args_from_nested_parameters; cargo test --manifest-path src-tauri/Cargo.toml gemini_to_anthropic_preserves_legitimate_parameters_arg
Not-tested: Full src-tauri test suite; live end-to-end Gemini relay traffic against upstream services
* Keep Gemini tool replay stable across Claude request boundaries
Claude Code follow-up requests were still falling back to locally reconstructed functionCall parts, which dropped Gemini thought signatures and triggered INVALID_ARGUMENT errors from the official Gemini API. The replay path needed to survive real Claude request boundaries, not just idealized in-process test flows.
This change makes Claude requests reuse X-Claude-Code-Session-Id as the shadow session key, records streamed Gemini tool turns before tool_use events are fully drained, and matches assistant tool_use turns to shadow state by tool_use id and normalized tool name before positional fallback. Together these fixes keep thoughtSignature-bearing Gemini tool calls available for the next request in the loop.
Constraint: Claude Code sends a stable X-Claude-Code-Session-Id header while metadata.session_id may be absent on follow-up requests
Rejected: Rely on metadata-only Claude session extraction | generated fresh session ids and broke cross-request shadow replay
Rejected: Record Gemini shadow only after streaming completes | loses the race when the client sends the next request immediately after tool_use
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Preserve Gemini shadow continuity across requests by keying Claude sessions from the header first and persisting tool-call shadow before yielding tool_use events downstream
Tested: cargo fmt --manifest-path src-tauri/Cargo.toml --all; cargo test --manifest-path src-tauri/Cargo.toml test_extract_session_from_claude_header; cargo test --manifest-path src-tauri/Cargo.toml test_extract_session_from_claude_header_precedes_metadata; cargo test --manifest-path src-tauri/Cargo.toml stores_tool_shadow_before_tool_use_events_are_fully_drained; cargo test --manifest-path src-tauri/Cargo.toml shadow_replay_matches_tool_use_turn_by_id_when_position_drifts; cargo test --manifest-path src-tauri/Cargo.toml shadow_replay_aligns_to_latest_turns_after_client_truncation
Not-tested: Full src-tauri test suite without test filters; live end-to-end Gemini relay after this exact commit hash
* style: apply cargo fmt to pass Backend Checks CI
Wrap prompt_cache_key chained call across lines per rustfmt default
formatting. Pure formatting change, no behavior difference.
* fix(proxy/gemini): synthesize unique ids for no-id tool calls + enforce object params schema
P1 — Parallel tool calls without Gemini-assigned ids no longer collapse.
Gemini 2.x native parallel `functionCall` entries may omit the `id` field.
The previous `merge_tool_call_snapshots` fell back to matching by `name`,
which silently merged two parallel calls to the same function into one
entry — dropping the first call's args. The non-streaming path and shadow
store further bottlenecked on empty-string ids: multiple `tool_use` blocks
shared the same id, and `tool_name_by_id.get("")` could only return one
mapping, causing later `tool_result` round-trips to fail with
`Unable to resolve Gemini functionResponse.name` or bind to the wrong tool.
Fix: introduce `synthesize_tool_call_id()` producing `gemini_synth_<uuid>`.
Both streaming and non-streaming response paths now guarantee every
Anthropic-visible tool_use carries a unique id. `merge_tool_call_snapshots`
matches by id first, falling back to the `parts` array position (for the
cumulative-streaming case) while preserving the synthesized id across
chunks. `convert_message_content_to_parts` detects the synthetic prefix
and strips the id from outbound `functionCall`/`functionResponse` so the
internal identifier never leaks upstream. `shadow_parts` performs the
same strip when replaying a recorded assistant turn.
P2 — Vertex AI rejects empty `parameters` schemas. When an Anthropic tool
arrives with missing or empty `input_schema`, the proxy used to emit
`"parameters": {}` (no `type`), which fails Vertex AI validation with
`functionDeclaration parameters schema should be of type OBJECT`.
Contrary to the automated-review suggestion, the fix is not to omit
`parameters` (that too is rejected) but to normalize to the canonical
empty-object form `{type: "object", properties: {}}`.
Refs: google-gemini/generative-ai-python#423, BerriAI/litellm#5055.
Fix: new `ensure_object_schema` helper in `gemini_schema` promotes
missing `type` to `"object"` and adds empty `properties` when absent,
while leaving atomic (non-object) schemas untouched.
Tests: seven new regressions covering parallel no-id calls, cumulative
chunk id reuse, synthetic-id round-trip both directions, shadow replay
id stripping, and the three Vertex-AI schema shapes.
The two existing wrapper functions (`gemini_to_anthropic` and
`gemini_to_anthropic_with_shadow`) gain `#[allow(dead_code)]` to clear
a pre-existing clippy -D warnings failure — they are part of the public
transform API surface and intentionally kept for future callers.
Addresses Codex review P1/P2 on #1918.
* fix(proxy/gemini): narrow URL normalization + guard empty OAuth access_token
P2a — Preserve opaque relay URLs that contain `/v1/models/` prefixes.
`should_normalize_gemini_full_url` previously flagged any full URL whose
path merely contained `/v1beta/models/` or `/v1/models/` as a structured
Gemini endpoint, forcing rewrite to `.../v1beta/models/{model}:method`.
This silently dropped legitimate relay route segments (e.g.
`https://relay.example/v1/models/invoke` → `.../v1beta/models/...:generateContent`,
losing `/invoke`) and sent traffic to the wrong upstream path.
Replace the bare `contains(...)` checks with
`matches_structured_gemini_models_path`, which requires the
`/models/` segment to be followed by a canonical Gemini method call
(`*:generateContent` or `*:streamGenerateContent`). The
`matches_bare_gemini_models_path` helper is generalized (and renamed) to
handle both `/v1beta/models/` and `/v1/models/` alongside the original
bare `/models/` shape.
P2b — Reject empty Gemini OAuth access_tokens before they reach the
bearer header.
`GeminiAdapter::parse_oauth_credentials` accepts refresh-token-only JSON
(and surfaces `{"access_token": "", ...}` for expired credentials) with
`access_token` defaulting to `""`. The Claude adapter's GeminiCli branch
then called `AuthInfo::with_access_token(key, creds.access_token)`
unconditionally, so the bearer-header builder at
`AuthStrategy::GoogleOAuth` resolved to `Authorization: Bearer ` — a
deterministic 401 from upstream.
CC Switch does not currently exchange the refresh_token for a fresh
access_token (`OAuthCredentials::needs_refresh` / `can_refresh` are
annotated `#[allow(dead_code)]`). Until that exists, only attach
`access_token` when it is non-empty; fall back to plain GoogleOAuth
strategy with the raw key and log a warn pointing users at
`~/.gemini/oauth_creds.json` so the failure mode is observable.
Tests:
- gemini_url.rs: three new regressions — opaque `/v1/models/invoke`,
opaque `/v1beta/models/route`, and the positive counter-case where a
structured `/v1/models/...:generateContent` path still normalizes.
- claude.rs: three new `test_extract_auth_gemini_cli_*` tests covering
refresh-only JSON, empty-string access_token JSON, and the valid-JSON
pass-through.
All 839 lib tests pass; cargo fmt + clippy -D warnings clean.
Addresses Codex review P2 findings on #1918.
* fix(proxy/gemini): treat empty-string functionCall id as missing in streaming path
Follow-up to the earlier P1 fix: some Gemini relays serialize an absent
functionCall id as `"id": ""` instead of omitting the field. The
non-streaming `extract_tool_call_meta` already filters these via
`.filter(|s| !s.is_empty())`, but the streaming counterpart
`extract_tool_calls` passed the empty string straight through
`function_call.get("id").and_then(|v| v.as_str())` into
`GeminiToolCallMeta::new`, producing a `Some("")` id.
Downstream, `merge_tool_call_snapshots` would then match two parallel
no-id calls against each other on their shared empty-string id,
collapsing them into a single snapshot (silent data loss for the first
call) and emitting an Anthropic `tool_use.id: ""` that breaks tool_result
correlation on the Claude Code client.
Fix:
- `extract_tool_calls`: apply the same `filter(|s| !s.is_empty())` guard
used in the non-streaming path so empty strings become `None` before
reaching the shadow meta.
- `merge_tool_call_snapshots`: defensively collapse any incoming
`Some("")` to `None` up front — keeps the "missing vs present" invariant
local to the merge step for future callers that might build
`GeminiToolCallMeta` by hand.
Tests (2 new, both in streaming_gemini):
- `parallel_empty_string_id_calls_are_treated_as_missing_and_preserved`
covers two parallel calls with explicit `"id": ""` — asserts both
surface, no empty tool_use id leaks, and each gets a unique
`gemini_synth_` id.
- `single_empty_string_id_tool_call_gets_synthesized_id` covers the
non-parallel degraded-relay case.
All 841 lib tests pass; cargo fmt + clippy -D warnings clean.
Addresses Codex follow-up P1 on #1918.
* fix(proxy/gemini): gate generic REST path suffixes behind Google host whitelist
`should_normalize_gemini_full_url` previously treated any full URL whose
path ends with `/v1`, `/v1/models`, `/models`, `/v1/openai`, or `/openai`
as a structured Gemini endpoint and rewrote it to
`/v1beta/models/{model}:generateContent`. These are ubiquitous REST
conventions — opaque relays such as `https://relay.example/custom/v1`
legitimately use them for fixed endpoints — so the rewrite silently
routed traffic to the wrong upstream path.
Split the predicate into two layers:
- **Unconditional**: `matches_structured_gemini_models_path` (i.e. a
`/models/...:generateContent` method call anywhere in the path), the
Google-specific `/v1beta*` family, and the deep OpenAI-compat paths
(`/v1beta/openai/chat/completions`, `/openai/chat/completions`, and
their `responses` siblings). These remain host-agnostic because the
path grammar itself is Gemini-specific.
- **Google-host gated**: `/v1`, `/v1/models`, `/models`, `/v1/openai`,
`/openai`. Only normalized when the host is one of
`generativelanguage.googleapis.com`, `aiplatform.googleapis.com`, or a
real `*-aiplatform.googleapis.com` Vertex regional endpoint. The match
is exact/suffix (not `contains`), so lookalike hosts like
`aiplatform.example.com` are correctly treated as opaque relays.
Tests (8 new in `gemini_url::tests`):
- Four opaque-relay cases: `/custom/v1`, `/custom/models`,
`/custom/v1/models`, `/custom/openai` — all preserved as-is.
- Three Google-host counter-cases: `/v1`, `/models`, and
`us-central1-aiplatform.googleapis.com/v1` still normalize.
- One lookalike safety case: `aiplatform.example.com/v1` is NOT
treated as Google.
All 849 lib tests pass; cargo fmt + clippy -D warnings clean.
Addresses Codex review P2 on #1918.
* fix(proxy/gemini): align shadow id with client-visible id in non-streaming path
When Gemini returns a `functionCall` without an id (common in 2.x
parallel calls), `gemini_to_anthropic_with_shadow_and_hints` previously
generated TWO independent synthesized UUIDs:
1. Line 186-197 — synthesized id `A` used for the Anthropic-visible
`content[tool_use].id` returned to the client.
2. Line 850-881 — `extract_tool_call_meta` independently synthesized
id `B ≠ A`, which populated `shadow_turn.tool_calls[i].id`.
`shadow_content` (line 225-228, cloned from `rectified_parts`) retained
the original missing/empty id. Result: the client sees id `A`, the
shadow store holds id `B`.
On the next turn, `convert_messages_to_contents` builds
`tool_name_by_id` from `build_tool_name_map_from_shadow_turns`, which
uses `tool_calls[i].id` — so the map contains `B → name` but not
`A → name`. When the client sends back `tool_result(tool_use_id=A)`,
resolution fails with:
Unable to resolve Gemini functionResponse.name for tool_use_id `A`
This affects both truncated histories (client sends only the
tool_result) and full histories (shadow-replay branch at line 342-354
skips `convert_message_content_to_parts`, so the assistant tool_use
block never registers id `A` itself).
Fix: make `rectified_parts` the single source of truth. After
`rectify_tool_call_parts`, run a pre-pass that writes
`synthesize_tool_call_id()` back into any `functionCall` that lacks a
non-empty id. All three readers — the content builder (186-197), the
shadow_content clone (225-228), and `extract_tool_call_meta` — then
observe the same id. `shadow_parts()` already strips synthesized ids on
replay (line 616-628), so the internal identifier never leaks to
Gemini upstream.
This mirrors the streaming path, which already has single-source-of-
truth semantics via `tool_call_snapshots` in `streaming_gemini.rs` —
no change needed there.
Tests (5 new in `transform_gemini::tests`):
- `non_stream_shadow_id_matches_client_visible_id`: asserts
`response.content[0].id == shadow.tool_calls[0].id ==
shadow.assistant_content.parts[0].functionCall.id`.
- `non_stream_missing_id_scenario_a_truncated_history_resolves`: turn 2
sends only `[tool_result(id=A)]`; resolution must succeed.
- `non_stream_missing_id_scenario_b_full_history_replay_resolves`: turn 2
sends `[assistant(tool_use=A), tool_result(A)]`; shadow-replay branch
strips the synth id from outgoing `functionCall` while still
resolving the subsequent `tool_result`.
- `non_stream_preserves_original_gemini_id_when_present`: regression —
genuine Gemini ids flow through unchanged.
- `non_stream_synthesized_id_not_leaked_to_gemini_via_shadow_replay`:
defensive — shadow-replay path must strip synth ids from both
`functionCall.id` and `functionResponse.id`.
All 854 lib tests pass; cargo fmt + clippy -D warnings clean.
Addresses Codex follow-up P1 on #1918.
* refactor(proxy/gemini): share build_anthropic_usage between stream and non-stream paths
`streaming_gemini::anthropic_usage_from_gemini` and
`transform_gemini::build_anthropic_usage` were byte-for-byte identical
(32 lines each) — both converting Gemini `usageMetadata` into the
Anthropic `usage` shape including `cache_read_input_tokens` mapping.
Promote the non-streaming version to `pub(crate)` and reuse it from the
streaming SSE converter. Removes ~30 lines of duplication and guarantees
the two paths cannot drift apart.
No behavioral change; all 854 lib tests pass; cargo fmt + clippy -D
warnings clean.
* fix(proxy/gemini): gate /v1beta behind Google host + normalize models/ model id prefix
Two related P2 corrections to the Gemini Native URL surface, both
folding into the existing Google-host-whitelist architecture.
## P2a — `/v1beta` suffix should not unconditionally trigger rewrite
`should_normalize_gemini_full_url` placed `/v1beta` and `/v1beta/models`
in the unconditional layer on the reasoning that `/v1beta` is
Google-specific. In practice an opaque relay fronting a non-Gemini
service at `https://relay.example/custom/v1beta` would still be
silently rewritten to `/v1beta/models/{model}:generateContent`,
breaking the deployment.
Move `/v1beta`, `/v1beta/models`, and `/v1beta/openai` into the
Google-host gated layer alongside `/v1`, `/models`, and friends. The
unconditional layer now only accepts paths whose grammar is
intrinsically Gemini — `/models/...:generateContent` method calls and
the deep OpenAI-compat endpoints like `/openai/chat/completions` and
`/openai/responses`. Pasted AI-Studio URLs such as
`https://generativelanguage.googleapis.com/v1beta` still normalize
because the host matches the whitelist.
## P2b — `model: "models/gemini-2.5-pro"` produced doubled path prefix
Gemini SDKs (and the official `list_models` response) commonly surface
model ids in resource-name form `models/gemini-2.5-pro`. Raw
interpolation into `format!("/v1beta/models/{model}:...")` produced
`/v1beta/models/models/gemini-2.5-pro:streamGenerateContent` which
upstream rejects — yielding false-negative health checks for otherwise
valid provider configs.
Introduce `normalize_gemini_model_id(&str) -> &str` in `gemini_url`
as the single source of truth: strips an optional leading `/` then an
optional `models/` prefix, leaving bare ids untouched. Apply in the
three call sites that build a Gemini method URL:
- `services/stream_check.rs::resolve_claude_stream_url` (unified path)
- `services/stream_check.rs::check_gemini_stream` (Gemini-only path)
- `proxy/forwarder.rs::rewrite_claude_transform_endpoint` (production)
Tests (9 new):
- `gemini_url`: 3 regressions for opaque vs Google-host `/v1beta*`
handling + 5 unit tests pinning `normalize_gemini_model_id` behavior
(strip prefix, leave bare id, preserve nested slashes past the one
stripped prefix, tolerate leading slash, pass through empty input).
- `stream_check`: one end-to-end regression confirming
`models/gemini-2.5-pro` collapses to the expected single-prefix URL.
- `forwarder`: one end-to-end regression on the production rewrite
path.
All 864 lib tests pass; cargo fmt + clippy -D warnings clean.
Addresses Codex P2 feedback on #1918.
* fix(proxy/gemini): trim API key before provider-type detection and OAuth parsing
Leading whitespace on a copied oauth_creds.json (e.g. trailing newline
when the user copies the file content as-is) would slip past the
`starts_with("ya29.") || starts_with('{')` prefix check in
`ClaudeAdapter::provider_type`, causing the provider to be misclassified
as raw-API-key Gemini and fall back to `x-goog-api-key` with the raw
JSON as the key — which upstream rejects with 401.
The frontend's `handleApiKeyChange` already trims on keystrokes but
deep-link imports, the JSON editor, and live-config backfill all bypass
that path. Trim at every backend extraction point so the coverage is
uniform:
- `ClaudeAdapter::extract_key` (5 env / fallback branches) gets
`.map(str::trim)` before `.filter(|s| !s.is_empty())` so that
whitespace-only values are also treated as missing.
- `GeminiAdapter::extract_key_raw` gets the same chain (including
the `.filter` it was missing before).
- `GeminiAdapter::parse_oauth_credentials` gets a defensive
`let key = key.trim();` at the entry as a belt-and-suspenders guard.
Adds two regression tests covering JSON and bare `ya29.` keys with
leading newline/space.
* fix(proxy/gemini): gate generic REST suffix stripping behind Google host in non-full-URL mode
`build_gemini_native_url` unconditionally stripped `/v1`, `/v1beta`,
`/models`, and `/openai` suffixes from the base path regardless of
host. This worked for Google's own endpoints but silently rewrote
third-party relay URLs like `https://relay.example/custom/v1` to
`.../custom/v1beta/models/...`, breaking any relay that mounts its
Gemini-compatible namespace under a versioned prefix.
The result was also asymmetric with the previously-fixed full-URL
branch: toggling the "full URL" switch changed the outbound URL for
the same base_url, which is exactly the kind of invisible behavior
that makes debugging proxy deployments painful.
Align `normalize_gemini_base_path` with
`should_normalize_gemini_full_url`'s layered model:
- Unconditional: `/models/...:method` structured paths and deep
OpenAI-compat endpoints (`/openai/chat/completions`,
`/openai/responses` and their versioned variants) — these are
unambiguous Gemini-specific grammar on any host.
- Google-host gated: generic `/v1`, `/v1beta`, `/models`, `/openai`
suffixes only get stripped on `generativelanguage.googleapis.com`,
`aiplatform.googleapis.com`, or `*-aiplatform.googleapis.com`.
Other hosts preserve the prefix verbatim so relays keep their
intended routing.
Adds seven regression tests for the non-full-URL flow: opaque relay
preservation (v1 / v1beta / models / openai suffix variants), Google
host normalization (counter-case), and boundary cases (structured
method path and deep OpenAI-compat endpoint stripped regardless of
host).
Test count: 864 -> 873.
* Revert "fix(proxy/gemini): gate generic REST suffix stripping behind Google host in non-full-URL mode"
This reverts commit d19ff09cb7.
* test(proxy/gemini): pin non-full-URL versioned relay base stripping
Adds two regression tests that lock in the intentional asymmetry
between full-URL and non-full-URL modes:
- Full-URL mode: opaque base path (e.g. `https://relay.example/custom/v1beta`)
is preserved verbatim. Already covered by
`preserves_opaque_full_url_with_bare_v1beta_suffix`.
- Non-full-URL mode: base path MUST strip `/v1`, `/v1beta`, etc. so the
standard `/v1beta/models/{model}:method` endpoint can be appended
without producing a doubled `/v1beta/v1beta/models/...` path.
The non-full-URL contract is "base URL + cc-switch appends the
canonical Gemini endpoint". A user who needs a relay's custom
namespace (e.g. `/v1/models/...`) must use full-URL mode and paste
the complete method path. This commit adds regression coverage so a
future attempt to mirror full-URL's host-whitelist gating into
`normalize_gemini_base_path` will fail the test suite immediately.
* chore(lint): address clippy 1.95 findings in existing modules
CI upgraded to Rust 1.95 and flagged ten pre-existing warnings that
older toolchains did not enforce. None relate to the Gemini proxy
integration PR itself but they block CI on the feature branch, so
clean them up here as a separate commit for easy review:
collapsible_match:
- proxy/providers/gemini_schema.rs: `"items" if value.is_object()`
match guard instead of nested if.
- proxy/providers/transform_responses.rs: fold
`map_responses_stop_reason`'s `"completed"` / `"incomplete"` arms
into match guards, relying on the existing `_ => "end_turn"` fall-
through for non-matching guard conditions (semantics preserved).
- services/session_usage_codex.rs: fold
`"session_meta" if state.session_id.is_none()` guard, relying on
the existing `_ => {}` fall-through.
unnecessary_sort_by:
- services/provider/endpoints.rs: `sort_by_key(|ep| Reverse(ep.added_at))`.
- services/skill.rs (backup list): same Reverse idiom on `created_at`.
- services/skill.rs (skill listings x2): `sort_by_key(|s| s.name.to_lowercase())`.
useless_conversion:
- services/skill.rs: drop the explicit `.into_iter()` on `zip`'s argument.
while_let_loop:
- services/webdav_auto_sync.rs: `while let Some(wait_for) = ...`
instead of `loop { let Some(...) = ... else { break }; ... }`.
All changes are mechanical and preserve behavior. `cargo test --lib`
remains green (868 passed).
* fix(proxy/gemini): reconcile synthesized tool-call ids with later real ids + preserve thoughtSignature
Three related findings on `streaming_gemini.rs` for Gemini's cumulative
`streamGenerateContent` stream, all centered on `merge_tool_call_snapshots`:
1. (P1) Match upgraded tool-call IDs by position.
When Gemini delivers a `functionCall` without an id on chunk 1
(cc-switch synthesizes `gemini_synth_*`) and then upgrades it to a
real id on chunk 2, the `Some(incoming_id)` branch only matched by
id and missed the existing synthesized snapshot. A second entry
would be pushed, yielding duplicate `tool_use` content blocks at
stream end — one with the synthesized id, one with the real id —
which could trigger duplicate tool execution and break tool_result
correlation. Add a positional fallback: when no id match exists but
the same-position slot holds a synthesized id, merge into it.
`or(preserved_id)` already lets the real id win the merge.
2. (P2) Preserve prior thoughtSignature when merging snapshots.
`tool_call_snapshots[index] = tool_call` overwrote the slot
entirely, dropping any `thoughtSignature` captured on an earlier
chunk if the current cumulative snapshot omitted it. Since
`build_shadow_assistant_parts` writes `thoughtSignature` into the
shadow turn from `tool_call.thought_signature`, a dropped signature
would cause later replay requests to Gemini to be rejected with
invalid-signature errors. Preserve the existing signature when the
incoming chunk does not carry one.
3. (P2) Document the part-order streaming trade-off.
All `tool_use` content blocks are emitted after the final text
`content_block_stop`, so interleaved [text, functionCall, text,
functionCall] parts arrive at the Anthropic client as [text(concat),
tool_use, tool_use] — different from the non-streaming transformer,
which preserves part order. This is intentional given the cumulative
snapshot model and the consumers we target (claude-code-like clients
don't depend on strict interleaving for tool execution correctness).
Add a block comment at the flush site describing the trade-off and
what a strict-order fix would entail, so this isn't rediscovered as
a bug later.
Regression tests:
- upgraded_real_id_merges_into_existing_synthesized_snapshot
- thought_signature_preserved_when_later_chunk_omits_it
Test count: 868 -> 870. clippy 1.95 clean. fmt clean.
* fix(proxy/gemini): prefer exact tool-call id over normalized-name fallback
The shadow-turn matcher used a three-branch `||` chain (id / full name /
normalized name). When two tools share a suffix (e.g. `server_a:search`
and `server_b:search`), the normalized-name clause could short-circuit
on an earlier turn whose id is actually wrong for the incoming tool_use,
mis-routing replay state (functionCall id / thoughtSignature) for later
tool_result resolution.
Split matching into two layers: when the incoming message carries any
tool_use ids, run id-based lookup first and return on the earliest hit.
Only fall back to full-name / normalized-name matching when the incoming
ids are absent or none of them resolve.
Add two regressions:
- shadow_replay_prefers_exact_id_match_over_normalized_name_collision
Two shadow turns with colliding normalized names and two assistant
messages whose ids cross the positional order; asserts each message
replays the id-correct shadow turn (including thoughtSignature).
- shadow_replay_falls_back_to_name_when_ids_absent
Shadow turn with no id and incoming tool_use with an empty id;
asserts the name fallback still populates the replayed part.
---------
Co-authored-by: Jason <farion1231@gmail.com>
* feat(usage): enhance usage stats backend and query hooks
* feat(usage): redesign calendar date range picker with auto-switch and simplified layout
* refactor(usage): streamline dashboard layout and stats components
* refactor(usage): compact request log table with merged cache/multiplier columns and centered layout
* feat(i18n): add cache short labels and usage stats translations for zh/en/ja
* Align usage dashboard stats with range boundaries
The usage dashboard mixed second-precision detail rows with day-level rollups, which caused custom half-day ranges to overcount historical rollup data and left the request log paginator on stale pages after top-level filter changes.
This change limits rollups to fully covered local days, aligns multi-day trend buckets with natural local days, and resets request log pagination when the dashboard range or app filter changes.
Constraint: usage_daily_rollups stores only daily aggregates after pruning old detail rows
Rejected: Include partial boundary rollups proportionally | historical intra-day detail is unavailable after pruning
Rejected: Force RequestLogTable remount on range change | would discard local draft filters unnecessarily
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep summary, trends, provider stats, and model stats on the same rollup-boundary rules
Tested: cargo test --manifest-path src-tauri/Cargo.toml usage_stats
Tested: pnpm exec vitest run tests/components/RequestLogTable.test.tsx
Tested: pnpm typecheck
Not-tested: Manual UI validation in the Tauri app
* Preserve full-day usage filters at minute precision
The latest review surfaced two interaction bugs in the usage dashboard: rollup-backed stats undercounted end days selected via the minute-precision picker, and immediate select changes accidentally applied unsubmitted text drafts from the request log filters.
This change treats 23:59 as a fully selected local end day for rollup inclusion and narrows select-side state syncing so app/status updates do not commit provider/model drafts.
Constraint: The custom range picker emits minute-precision timestamps, while rollups are stored at day granularity
Rejected: Require exact 23:59:59 end timestamps | unreachable from the current picker UI
Rejected: Rebuild applied filters from the full draft state on select changes | silently commits unsaved text input
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep request-log text fields on explicit apply semantics even when select filters remain immediate
Tested: cargo test --manifest-path src-tauri/Cargo.toml usage_stats
Tested: pnpm exec vitest run tests/components/RequestLogTable.test.tsx
Tested: pnpm typecheck
Not-tested: Manual Tauri dashboard interaction
* refactor(usage): move range presets into date picker, single-row layout
- UsageDateRangePicker: add preset shortcuts (今天/1d/7d/14d/30d) inside
popover top; clicking a preset applies immediately and closes popover
- UsageDashboard: collapse to single row (app filters + refresh + picker);
remove standalone preset buttons and summary stats bar
- RequestLogTable: replace static Calendar badge with interactive
UsageDateRangePicker via onRangeChange prop; single filter row
* Keep usage pagination regression coverage aligned with the rendered UI
The new regression test was asserting a non-existent pagination label and page summary text, so it failed before it could verify the real page-reset behavior. This commit switches the assertions to the numbered pagination buttons that the component actually renders and validates the reset through the query hook arguments.
Constraint: RequestLogTable exposes numbered pagination buttons, not a "Next page" label or "2 / 6" summary text
Rejected: Add synthetic pagination labels solely for the test | would couple production markup to a test-only assumption
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Prefer pagination assertions that follow the rendered controls or hook inputs instead of invented summary text
Tested: pnpm vitest run tests/components/RequestLogTable.test.tsx; pnpm typecheck; pnpm test:unit
* refactor(usage): clean up dead code and polish date range picker
- Remove unused exports MAX_CUSTOM_USAGE_RANGE_SECONDS,
timestampToLocalDatetime, and localDatetimeToTimestamp from
usageRange.ts (replaced by the calendar picker)
- Deduplicate getPresetLabel from UsageDashboard and
UsageDateRangePicker into shared getUsageRangePresetLabel helper
- Add aria-label, aria-current and aria-pressed to calendar day
buttons so screen readers can disambiguate same-numbered days
across adjacent months
- Drop unused cacheReadShort and cacheWriteShort i18n keys (zh/en/ja);
the request log table renders R/W prefixes inline
- Align customRangeHint copy with the removed 30-day limit by
dropping "up to 30 days" wording (zh/en/ja)
* fix(usage): align rollup cutoff to local midnight to keep days complete
`rollup_and_prune` previously used `Utc::now() - retain_days * 86400`
as the cutoff. Because rollups are bucketed by *local* date and detail
rows below the cutoff are pruned, an unaligned cutoff left the youngest
rolled-up day half-rolled-up and half-pruned. Combined with the new
`compute_rollup_date_bounds` boundary trimming (which excludes any
rollup day not fully covered by the requested range), custom range
queries that touch that day silently under-count summary, trend,
provider, and model stats.
Fix the invariant at the source: snap the cutoff to the next local
midnight after `(now - retain_days)`. Every rollup row now reflects a
complete local day, so the boundary trimmer's all-or-nothing assumption
holds.
Includes unit tests for the cutoff math (typical case + already-on-
midnight case). DST gap is handled defensively by bumping forward by
an hour.
Addresses Codex P2 review finding on PR #2002.
---------
Co-authored-by: Jason <farion1231@gmail.com>
* chore(stream-check): update default health check models to latest
Replaces deprecated gpt-5.1-codex@low with gpt-5.4@low and switches
the Gemini default from gemini-3-pro-preview to gemini-3-flash-preview
to pick the lightest variant of the latest series for fast, low-cost
health checks.
https://claude.ai/code/session_01NGWLchcTP76rJHjiP5Ehte
* feat(stream-check): detect model-not-found errors with dedicated toast
Health check previously classified failures purely by HTTP status code,
which meant deprecated/invalid models showed up as a generic "Not found
(404)" error pointing users to check the Base URL — misleading when the
URL is fine and only the test model is wrong (e.g. gpt-5.1-codex after
it was retired).
Backend: add detect_error_category() that inspects 4xx response bodies
for model-not-found indicators (model_not_found, does not exist,
invalid model, not_found_error, etc.) and returns a "modelNotFound"
category. Thread the resolved test model through build_stream_check_result
so the failed result carries it in model_used. Add StreamCheckResult
.error_category field (serde-skipped when None).
Frontend: useStreamCheck branches on errorCategory === "modelNotFound"
before the HTTP-status fallback and renders a toast.error with the model
name and a description pointing to Model Test Config. Add i18n keys
(modelNotFound / modelNotFoundHint) for zh/en/ja.
Tests: unit-test detect_error_category against real OpenAI/Anthropic
error shapes, 5xx false-positive avoidance, and plain 401 auth errors.
https://claude.ai/code/session_01NGWLchcTP76rJHjiP5Ehte
* fix(stream-check): add missing error_category field in fallback
The error_category field was added to StreamCheckResult in this branch
but the fallback constructor in stream_check_all_providers was not
updated, which broke cargo build.
---------
Co-authored-by: Claude <noreply@anthropic.com>
Distinguish between "provider rejects probe" (yellow warning) and
"genuinely broken" (red error) in health check results.
Backend: add AppError::HttpStatus variant to carry structured HTTP
status codes, populate http_status on error results, classify codes
into short labels (e.g. "Auth rejected (401)"), and truncate overly
long response bodies.
Frontend: route 401/403/400/429/5xx to toast.warning with localized
hints explaining the error may not indicate actual unusability; route
404/402/connection errors to toast.error. Add i18n keys for all three
locales (zh/en/ja).
Also deduplicate check_once by reusing build_stream_check_result.
When a search query matches text beyond the collapse point, the message
automatically expands to show the highlighted match. Also adds
aria-expanded for accessibility.
Jason