- Update Codex OAuth presets to the gpt-5.6 family (haiku -> gpt-5.6-luna)
and bump the custom Codex template default model
- Inject CLAUDE_CODE_MAX_CONTEXT_TOKENS / CLAUDE_CODE_AUTO_COMPACT_WINDOW
(372000, the ChatGPT Codex catalog window with a ~353K effective budget,
openai/codex#31860) into effective live settings so Claude Code stops
assuming a 200K window and compacts before the upstream rejects the prompt
- Gate the injected defaults on every configured model being gpt-5.6*:
gpt-5.5's upstream catalog oscillates between 272K and 372K and must not
inherit them; explicit user values always win
- Strip the injected values on switch-away backfill (mirror-inverse of the
injection conditions) so program defaults never harden into per-provider
explicit values, and keep both keys out of the shared common-config snippet
Do not derive unauthenticated health-check targets from runtime adapter defaults for official providers. Batch checks now skip official entries, and direct base-URL resolution fails explicitly instead of probing first-party endpoints without credentials.
Cover the Codex official provider in the stream-check regression tests so future adapter changes cannot silently reintroduce these probes.
Allow the built-in Codex official provider to participate in takeover mode while preserving Codex's native OAuth or API-key credentials instead of persisting them into provider records.
Project official routing into a dedicated TOML provider, normalize inline tables, clean stale managed placeholders, and fail closed when the live configuration cannot be transformed safely.
Validate forwarded authorization, make official 401/403 responses non-retryable, avoid circuit-breaker pollution, and share the first-party ChatGPT endpoint across the Codex and Claude adapters.
Parse and retain Anthropic's ephemeral 5-minute and 1-hour cache-creation token buckets while preserving the existing aggregate cache-write metric for compatibility.
Price 1-hour writes at the documented premium relative to the configured 5-minute write rate, clamp inconsistent provider details safely, and include TTL buckets in usage diagnostics.
Persist 1-hour cache-write tokens with schema version 14 so zero-cost backfills and later pricing updates retain the original TTL semantics. Keep session import paths compatible through zero-valued detail fields.
Parse cache_write_tokens from OpenAI usage details and preserve cache creation data across Chat, Responses, and Anthropic conversion paths.
Add explicit input-token semantics to request logs and rollups so legacy rows subtract cache reads only while new total-inclusive rows subtract both cache reads and writes. Migrate v12 databases, normalize rollups to fresh input, and cover historical backfill behavior with regression tests.
Switching the Claude provider from a third-party (DeepSeek/MiMo/...) to a
managed Codex provider wrote both ANTHROPIC_API_KEY and
ANTHROPIC_AUTH_TOKEN as PROXY_MANAGED into ~/.claude/settings.json,
making Claude Code warn "Both ANTHROPIC_AUTH_TOKEN and ANTHROPIC_API_KEY
set - auth may not work as expected" on every run.
The double key is an accretion artifact: the ManagedAccount takeover
policy originally injected only ANTHROPIC_API_KEY (Copilot, #1049), and
the Codex fix that preserved ANTHROPIC_AUTH_TOKEN to avoid the login
prompt (#3784 / PR #3789) added the token on top without removing the
API key insert.
Inject exactly one placeholder instead: AUTH_TOKEN for Codex-managed
providers (keeps the #3784 fix), API_KEY for Copilot (keeps the #1049
behavior). Local-proxy auth is unaffected - the forwarder resolves the
PROXY_MANAGED placeholder from either header.
Updates the three Codex takeover tests to assert API_KEY is absent and
adds a regression test for the exact third-party -> Codex switch from
the report.
Fixes#4919
* Add test for Fable model env key exclusion
Add regression test to ensure Fable model env keys are excluded from common config.
* Add Fable model to CLAUDE_MODEL_OVERRIDE_ENV_KEYS
* Add files via upload
* Delete .github/workflows/build-windows-unsigned.yml
* feat(codex): support native Anthropic Messages protocol as upstream
Allow gateways that only expose the native Anthropic Messages protocol
(/v1/messages) to be used by Codex: the local proxy performs bidirectional
request/response/streaming conversion between Responses and Anthropic.
Backend:
- Add two conversion modules: transform_codex_anthropic / streaming_codex_anthropic
- codex.rs: add routing detection and auth: ANTHROPIC_AUTH_TOKEN→Bearer (default),
ANTHROPIC_API_KEY→x-api-key, mutually exclusive
- handlers.rs: add handle_codex_anthropic_to_responses_transform
- forwarder.rs: support optional Claude Code client fingerprint impersonation
(User-Agent / anthropic-beta / x-app / system prompt first-line injection)
and /responses→/v1/messages rewriting
- codex_config: the anthropic format reuses the NativeResponses profile to strip
custom tools
- ProviderMeta: add impersonateClaudeCode
Frontend:
- CodexApiFormat: add "anthropic"; the form adds auth field selection and an
impersonation toggle
- Add en/ja/zh/zh-TW copy
Robustness:
- Downgrade when tool history / forced tool_choice conflicts with extended
thinking, avoiding upstream 400s
- Emit cache_creation_input_tokens in usage and use saturating_add to guard
against overflow
- Append a unique suffix to non-streaming/streaming output-item ids to avoid
multi-segment text/thinking overwriting each other
* fix(codex): harden Anthropic bridge against empty text blocks and truncated streams
- Drop empty/whitespace-only text content blocks when rebuilding Anthropic
messages from Responses history; Anthropic 400s on empty text blocks (e.g. an
empty assistant text emitted alongside a tool_use), which broke follow-up and
tool-result requests. Also drop messages left without content.
- Do not report a truncated Anthropic stream as completed: when the SSE
connection ends before message_stop with no stop_reason, emit an incomplete
response if partial output exists, or a failed (stream_truncated) response
otherwise, mirroring the chat converter's EOF handling.
* fix(codex): resolve Anthropic-bridge review findings
Blocking:
1. Defer stripping the [1m] long-context marker until after catalog matching and model write-back, re-stripping it on the final Codex→Anthropic body and setting a flag to emit the context-1m-2025-08-07 beta header, so the marker is no longer lost or overridden by the default model.
2. Gate Anthropic thinking on the trailing turn only (via trailing_turn_allows_thinking) instead of scanning full history, so a Codex session that resends history each round no longer permanently loses thinking after the first tool call.
3. Inject 5m ephemeral cache_control on the Codex→Anthropic body by reusing cache_injector (handling the system string→array conversion), so system/tools/history are cached instead of re-sent at full price every round.
4. Add a shared base_url_is_full_endpoint helper (normalizing whitespace/query/fragment/trailing slash) used by both the Anthropic and Chat paths, so a base URL already ending in /v1/messages is treated as a full endpoint instead of double-appending to /v1/messages/v1/messages.
5. Align the catalog tool-profile predicate with the routing predicate so resolve_codex_catalog_tool_profile returns the Anthropic profile whenever the request converts to Anthropic, preventing freeform tools like apply_patch from being silently filtered by a ProxyChat catalog.
6. Explicitly disable native web_search for the Anthropic profile (including the no-catalog branch) via set_codex_native_web_search_field, so Codex no longer treats it as available while the transform silently drops it.
7. Only forward tool_choice when tools survive filtering, dropping it otherwise, to avoid a non-retryable 400 from upstream when tool_choice is sent with no tools.
8. Lower the fallback default to max_tokens=8192 (only when max_output_tokens is omitted) and clamp the thinking budget to max_tokens/2, disabling thinking below the 1024 floor, to avoid hard 400s on low-output-ceiling models/gateways.
Minor:
9. Centralize the Codex/OpenAI fingerprint-header denylist in is_codex_client_fingerprint_header so impersonating Claude Code uniformly drops originator/session_id/conversation_id/chatgpt-account-id/x-client-request-id/openai-* and the x-stainless-*/x-codex-* prefixes.
10. Retain content_block_start.input as start_input and fall back to it at block close when no input_json_delta arrived, so a gateway that carries the full tool input on the start event no longer yields empty tool arguments.
11. Extract a shared codex_responses_sse module as the single Responses SSE envelope builder that both the chat and anthropic streaming emitters delegate to, with byte-for-byte-unchanged wire output, so future event-format fixes touch one place instead of two.
* fix(codex): add per-provider max_output_tokens override for Codex→Anthropic path
Codex does not forward model_max_output_tokens in the request body,
causing the proxy to fall back to a conservative 8192 default. This
truncates long or thinking-heavy responses (stop_reason=max_tokens).
- Add maxOutputTokens field to ProviderMeta (Rust + TypeScript)
- Inject the value into the Anthropic request body before transform,
taking precedence over request-supplied and default values
- Add numeric input in Codex form fields (Anthropic format only)
- Add i18n entries for label, placeholder, and hint (en/zh/zh-TW/ja)
- Include roundtrip and omission unit tests for the new field
* fix(codex): harden Anthropic protocol bridge
* fix(codex): address Anthropic bridge review
* fix(codex): preserve flattened Anthropic inputs
* fix(codex): harden Anthropic recovery paths
* fix(ci): satisfy Rust clippy
---------
Co-authored-by: Jason <farion1231@gmail.com>
updateTomlCommonConfigSnippet re-serialized the whole document through
smol-toml (parse -> deepMerge -> stringify): comments dropped, keys
reordered, and empty parent table headers synthesized -- the
long-standing "config.toml keeps getting reordered" symptom (audit C5,
introduced in 083e48bf).
Replace it with a backend command backed by the same
merge_toml_table_like / remove_toml_table_like used when writing live
configs, so the form preview and the live write share one merge
semantic and user formatting survives edit-time merges. The frontend
sync helper is deleted outright to keep the pattern from coming back.
Making the form operations async exposes them to interleaving, so a
result is discarded unless it is still current when it lands:
- a per-hook sequence number (last operation wins) covers rapid
toggle/save races where an earlier merge resolves after a later
removal and would flip the switch back;
- a config-baseline check covers the user hand-editing the TOML while
a merge is in flight -- the stale result must not clobber their edit.
The checkbox state self-heals via the existing inference effect.
Regression tests pin both by resolving a suspended merge after a newer
operation / an external edit, plus backend tests locking comment and
key-order preservation, scalar override, and value-matched removal.
import_mcp_from_apps swallowed every importer error with unwrap_or(0),
so a corrupt config.toml surfaced as "imported 0 servers" with no hint
that anything went wrong.
Move the aggregation into McpService::import_from_all_apps: each app
imports best-effort (one bad file doesn't block the rest), and failures
are collected into a single error naming the failing apps alongside the
count that did import. The frontend now refreshes the server list on
settle rather than success, since a partial failure still means new
servers were persisted.
sync_all_enabled iterated AppType::all() with `?`, so one app's corrupt
live file (e.g. a broken ~/.claude.json, which passes the existence
gate but fails to parse) blocked every app behind it in the iteration
order -- and bubbled the error into whatever operation triggered the
sync:
- switch/save had just rewritten only the target app's live file, yet a
broken unrelated file failed the whole operation after DB and live
were already updated, reporting a false "switch failed" to the user.
Both now project only the target app (sync_enabled_for_app) and
degrade projection failure to a warning: the primary operation has
already taken effect, and the projection self-heals on the next
switch or MCP toggle.
- sync_current_provider_for_app_to_live syncs a single app; it now
projects that app only, keeping failures (which can only concern the
target app) as errors.
- sync_current_to_live (config import / cloud-sync restore) keeps the
all-apps sweep, but sync_all_enabled is now best-effort: it projects
every app, collects failures, and reports them aggregated. Its error
is held until after skill sync so an MCP failure no longer skips it.
Claude already re-extracts the live config into the shared common-config
snippet right before switching away, so shared tweaks (plugins,
preferences) made in live propagate to all opted-in providers. Codex was
gated out because its TOML pipeline leaked provider-specific and
injected content into the snippet.
With the extractor now stripping all injection artifacts and routing
fields, and backfill stripping the MCP projection, open the gate to
Claude + Codex. The autosync-before-strip ordering also self-heals
stale snippet values previously baked into provider snapshots: the
re-extracted snippet matches the live values, so the value-match strip
removes them on the same switch.
End-to-end tests cover: new shared keys captured, deletions synced,
secrets/injected artifacts never entering the snippet, and provider A's
key not leaking into provider B's live.
Toggling unify_codex_session_history rewrites the current official
provider's live config.toml in full (intended design), which drops the
[mcp_servers] projection -- and nothing put it back, so enabled MCP
servers silently vanished until the next provider switch (#C2).
Re-project after the rewrite, with two deliberate choices:
- Project Codex only (new McpService::sync_enabled_for_app) instead of
sync_all_enabled: the all-apps sync short-circuits in AppType::all()
order, so a corrupt ~/.claude.json would error before Codex is ever
reached and the freshly wiped [mcp_servers] would stay missing. Only
Codex's live file was rewritten here, so only Codex needs
re-projection.
- Degrade projection failure to a warning: by this point the live file
already carries the new bucket state, so the toggle has taken effect.
Propagating the error would make save_settings roll back the setting,
creating the exact "setting=old, live=new bucket" session split the
rollback exists to prevent. The projection self-heals on the next
switch or any MCP toggle.
extract_codex_common_config kept several fields in the shared snippet
that must never cross providers:
- [mcp_servers] and the legacy [mcp.servers] form: owned by the DB
mcp_servers table; once in the snippet they get merged into every
opted-in provider and no sync path can ever clean them up.
- top-level experimental_bearer_token: normally lives inside
[model_providers.<id>] (stripped with the whole table), but three
fallbacks write it at top level -- leaking the API key into the
shared snippet.
- model_catalog_json: per-provider catalog projection pointer.
- web_search, only when it equals the injected "disabled" sentinel;
a user-set value remains a shareable preference.
- top-level wire_api: same provider-routing semantics as top-level
base_url (the fallback target when no model_provider is set); leaking
it would rewrite the next provider's protocol selection.
This makes the extractor safe as the source for switch-time
common-config autosync.
MCP servers are owned by the DB mcp_servers table; the [mcp_servers]
section in live config.toml is only a projection re-synced after every
live write. When switching away baked that projection into the stored
provider snapshot, servers deleted in the app were resurrected the next
time that provider was activated -- per-entry reconcile only knows rows
that still exist in the DB, so it can never clean up such orphans.
Strip [mcp_servers] (and the legacy [mcp.servers] form) from the live
settings during switch-away backfill. Previously polluted snapshots
self-heal the next time the user switches away from them.
* fix: sync openclaw live provider updates
* fix: sync hermes live provider updates
* fix test: hermes live import stores models as array after denormalize
The test import_hermes_providers_from_live_updates_existing_provider_from_live
seeded models as a dict, but import_hermes_providers_from_live reads via
get_providers() which calls denormalize_provider_models_for_read(), converting
models from YAML dict to UI-friendly array. The test assertion accessed models
as dict -> Null -> assertion failure.
Fix: access models as array by index, matching the actual storage format after
live import. Also verify the id field is preserved in the denormalized output.
* fix: sync existing opencode providers from live config
* fix: preserve OpenCode provider display names
* docs: update comments and log messages to reflect new update behavior
- fix stale comment in lib.rs that said existing providers are skipped
- change log message from 'Imported' to 'Synced' since count now
includes both new imports and updates
Usage/quota queries frequently showed spurious "query failed" states that
manual refresh could not clear (#3820). Root cause: all transport-level
failures were folded into Ok(success:false), so react-query's retry never
fired and the failure body poisoned the cache as regular data.
Backend:
- balance/coding_plan/subscription services now return Err for send
failures and body-read failures (read body via bytes() before
serde_json::from_slice; reqwest's json() wraps read errors as Decode,
making error-kind checks on it dead code). Auth/4xx/parse errors stay
Ok(success:false) and surface immediately.
- Script path maps transient AppError keys (request_failed /
read_response_failed) to Err; Volcengine adds a Transient call variant.
- Command layer skips snapshot persistence, usage-cache-updated emit and
tray refresh on Err so the cache bridge cannot overwrite retained data.
- Expired-credential retry propagates transient errors instead of
rewriting them as "OAuth token has expired".
Frontend:
- resolveDisplayUsage generalized to subscription quotas; 5th param is now
an options object with a `rejected` flag: stale success data retained by
react-query across rejections is re-anchored to dataUpdatedAt and expires
through the same 10-minute keep-last-good window instead of being shown
indefinitely (a total outage used to mask longer than a single 5xx).
- useSubscriptionQuota / useCodexOauthQuota gain keep-last-good with
per-scope snapshot reset; rejected queries with no displayable value
synthesize a failure placeholder carrying the real error message so
footers keep rendering a retry entry point.
- HTTP 429 is classified transient (retry-later) alongside 5xx.
- UsageScriptModal surfaces string rejections via extractErrorMessage.
Tests: 6 backend behavior tests drive real HTTP against a local listener
to pin the Err/Ok channel semantics; frontend suite extends
keepLastGoodUsage coverage (405 passing).
Since switching projects automatically saves the current configuration
back to the project being left, the per-project "update from current"
button is redundant and even harmful: it allowed overwriting another
project's snapshot with the current state, breaking the invariant that
a project holds the state you last left it in.
- Remove the resnapshot button and confirm dialog from the manage
dialog; drop the now-unused scope prop and PROFILE_SCOPE_LABELS
- Add a footer Close button to the manage dialog (overlay click is
disabled app-wide, so it previously could only be closed via Esc)
- Update manageDescription in all four locales to describe the
autosave behavior; remove the two dead i18n keys
- Fix a shadowed `warnings` vec in ProfileService::apply that silently
dropped autosave-failure warnings before they reached the frontend
- Mention auto-capture on next switch in the "no snapshot for this
scope yet" warning and doc comments
The backend update command keeps its resnapshot parameter; it is now
only exercised by the pre-switch autosave path.
When switching to a different project, the current configuration of the
previously active project is automatically captured into that project's
snapshot before the target project's snapshot is applied. Scope is limited
to the active group (Claude/Codex); autosave failures are collected as
warnings and do not block the switch.
- Move autosave logic into ProfileService::apply for consistent behavior
across UI, tray, and tests
- Add integration test covering bidirectional autosave roundtrip
Refs: profile switch B方案
Projects are now global shared entities; Claude and Codex groups switch
independently via scoped current pointers and scoped payload slots.
- Remove scope column from profiles; keep current_profile_id_<scope>
- Use Option<Vec<String>> for mcp/skills to distinguish 'never captured'
from 'captured empty', preventing cross-side accidental disable
- Update/apply operations scoped to the active group via merge_scope_from
- Tray menu nests same shared list under Claude Code/Codex groups
- Add i18n for per-scope tooltips and 'not saved for this side' hint
Refs: profile P1 shared-entity redesign
Claude Desktop's only dimension managed by cc-switch is its provider
(MCP/Skills are hardcoded unsupported and prompts have no live file),
so snapshots capture just the current desktop provider while the empty
MCP/Skills sets and None prompt make apply a natural no-op for the
other dimensions - no per-dimension special casing needed.
- Add claude-desktop slot to PROFILE_APPS and PerApp (serde key uses
the hyphenated app id); old payloads without the key deserialize to
None and leave Claude Desktop untouched on apply
- Gate the tray Projects submenu by iterating PROFILE_APPS instead of
hardcoding Claude/Codex visibility
- Show the profile switcher on the Claude Desktop tab, mirror the
PerApp type, invalidate the claude-desktop providers cache on apply,
and extend the switcher tooltip in all four locales
- Extend the roundtrip integration test with the desktop provider
dimension; the desktop switch is cfg-gated to macOS/Windows because
desktop live writes error on Linux where CI runs cargo test
Add a profile feature that captures the current provider, MCP,
skills and prompt state for Claude Code and Codex as a named
snapshot, and re-applies it in one click from the header switcher
or the tray Projects submenu.
- New profiles table (schema v12) with current marker in settings
- ProfileService orchestrates the four existing switch primitives
(provider first, then MCP diff, skills diff, prompt enable)
- Best-effort apply: dangling references become warnings, no rollback
- Header combobox switcher + snapshot-style manage dialog
- Tray Projects submenu shared with the UI apply/event pipeline
- i18n for zh/en/ja/zh-TW under the new profiles domain
- Integration tests covering roundtrip, dangling refs and clear
* feat: add Claude subagent takeover config
* feat: add Claude subagent model field
* i18n: add Claude subagent model labels
* fix(proxy): preserve configured subagent model mapping
* fix(providers): exclude subagent model from Claude common config
* style: format rust code
* fix(subscription): display Codex free-plan 30-day quota window (#3651)
Codex free accounts are metered on a rolling 30-day secondary window
(limit_window_seconds = 2,592,000) instead of the weekly window used by
paid plans. The backend already maps this correctly to the tier name
"30_day", but the frontend TIER_I18N_KEYS whitelist had no entry for it,
so SubscriptionQuotaView filtered the tier out. When that was the only
surviving tier (as happens for free accounts), the whole quota footer
rendered nothing — free accounts showed no remaining quota or reset time
while paid accounts worked.
- Add "30_day" -> subscription.thirtyDay to TIER_I18N_KEYS
- Add the thirtyDay label to all four locales (zh/en/ja/zh-TW)
- Add a unit test locking in window_seconds_to_tier_name mappings,
including the 30-day case
Fixes#3651
* fix(tray): render Codex free-plan 30-day window in tray menu
The tray keeps its own tier-name whitelist (TIER_LABEL_GROUPS) independent
of the frontend TIER_I18N_KEYS. Its month group only listed "monthly", so a
free Codex account whose only tier is "30_day" produced empty labeled parts
→ format_subscription_summary returned None → the tray showed no quota, even
though the footer now does. That breaks the invariant the existing
gemini_summary_lite_only_still_renders test guards: any tier visible in the
footer must not leave the tray blank.
- Add TIER_THIRTY_DAY ("30_day") constant so the string is single-sourced
across backend mapping, tray grouping, and the frontend whitelist
- Map 2_592_000s explicitly to it in window_seconds_to_tier_name
- Add TIER_THIRTY_DAY to the tray month ("m") group
- Add codex_summary_thirty_day_only_still_renders regression test
Follow-up to the review on #3651 / PR for the 30-day footer fix.
When switching away from a Claude provider that opted into common config, re-extract the shareable portion of its live settings.json and replace the stored snippet. This captures config the user added directly in the app (plugins/hooks/shared prefs) so it isn't lost on switch, and propagates deletions so a removed key isn't re-injected on the next switch.
Scoped to Claude providers with common_config_enabled, skipped when the snippet was explicitly cleared. All failures are non-fatal (warn only) and never block the switch.
Also harden extract_claude_common_config to strip ALL credential-like keys via pattern match (*_API_KEY / *_AUTH_TOKEN / secret / token / etc.), not just ANTHROPIC_API_KEY / ANTHROPIC_AUTH_TOKEN. Claude providers legitimately use OPENROUTER_API_KEY / GOOGLE_API_KEY (and may carry OpenAI/Gemini/AWS Bedrock/Vertex creds), which previously could leak into the shared snippet and be injected into other providers. This also fixes the pre-existing leak in the manual Extract and one-time auto-extract paths. Plural _TOKENS (e.g. MAX_OUTPUT_TOKENS) is preserved as shareable.
Tests: 4 integration (capture / delete-sync / opt-out / cleared) + 1 unit (credential stripping).
Volcengine Ark rejects the bare model name doubao-seed-2-1-pro with a 404
("model does not exist or you do not have access to it") even after the
model is activated; the API requires the full dated id
doubao-seed-2-1-pro-260628. Update the Doubao preset model id across all
apps (config default, generated catalog, and OpenClaw namespaced refs).
Pricing lookup only stripped 8-digit (YYYYMMDD) and ISO date suffixes, not
the 6-digit YYMMDD format Volcengine uses (-260628, -250615), so real
Doubao usage never matched the bare-name pricing seed and showed $0 cost.
Extend strip_model_date_suffix to also strip 6-digit YYMMDD (with
month/day validation to avoid eating non-date version suffixes), keeping
the bare-name pricing seed as the canonical identity. This also fixes
pricing for every other Volcengine Doubao model. Add unit and end-to-end
regression tests.
Codex providers can now run in two modes per provider:
- Proxy-Chat (route takeover): apiFormat=openai_chat, existing Responses<->Chat conversion. Unchanged.
- Native-Responses (direct): apiFormat=openai_responses, no proxy. cc-switch generates ~/.codex/cc-switch-model-catalog.json so Codex shows the custom models and tools work without the freeform apply_patch (type=custom) tool that native gateways like MiMo reject; editing falls back to shell_command.
Catalog generation is keyed on apiFormat (CodexCatalogToolProfile), decoupled from the takeover toggle, so native providers persist a catalog without enabling local route mapping.
Codex's catalog parser requires base_instructions on every entry; the native template carries a neutral default and per-vendor official text overrides it (MiMo, MiniMax). Synthesized catalogs for Qwen/Doubao/LongCat use the neutral default.
Existing native providers must be re-saved once to regenerate a valid catalog (no DB migration).
Query coding-plan / agent-plan usage for Volcengine Ark via the
control-plane OpenAPI (open.volcengineapi.com), which requires
account-level AccessKey signing rather than the inference API key.
- Implement Volcengine Signature V4 (an AWS SigV4 variant: fixed
canonical header order, "HMAC-SHA256" algorithm without the AWS4
prefix, scope ending in "request", service "ark") in
services/coding_plan.rs
- Auto-detect the plan by probing GetAFPUsage (Agent Plan, AFP
five-hour/weekly/monthly quotas) first, falling back to
GetCodingPlanUsage (Coding Plan, session/weekly/monthly percentages);
a single credential covers whichever plan the account holds
- Parse the real response shape: the window label lives in the `Level`
field; guard ResetTimestamp <= 0 (session returns -1)
- Store account-level credentials on UsageScript (accessKeyId /
secretAccessKey), threaded through both the test command and the
TOKEN_PLAN auto-refresh path
- Add an independent AK/SK input block with a detailed hint pointing to
the Volcengine console -> API Access Keys
- Add the `monthly` tier label (frontend TIER_I18N_KEYS + tray
TIER_LABEL_GROUPS + subscription service)
- Sync zh / en / ja / zh-TW locales and the usage-query docs
* feat(usage): support importing model pricing from models.dev
Add an "Import from models.dev" button to the Add Pricing panel that
fetches https://models.dev/api.json, lists priced models sorted by
release date (newest 50 by default, full-text search across ~4800),
and bulk-imports the selected entries through the same
update_model_pricing command as manual entry.
- Normalize imported model IDs to match the backend's
clean_model_id_for_pricing rules (strip vendor prefix, lowercase,
truncate ':' suffix, map '@' to '-', drop the [1m] marker) so the
stored rows actually match cost-attribution lookups
- Dedupe selections that collapse to the same model_id and report
skipped duplicates in the success toast
- Invalidate usage queries on settled (not just success) so partial
import failures still refresh the pricing list
- Keep ESC inside the picker's search input from closing the dialog
and discarding the selection
- Add i18n keys for zh/en/zh-TW/ja and unit tests for the
normalization, price formatting and flattening logic
Fixes#4017
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(usage): match scoped cost backfill against raw model aliases
The scoped backfill selected zero-cost rows via exact SQL string match,
but log columns store raw model strings (route prefixes, :free variants,
date suffixes), so alias rows were skipped until the next full backfill
on startup. Filter rows in Rust with the same model_pricing_candidates
normalization used by the pricing lookup; pricing decision logic is
untouched. Pre-existing gap from schema v11, surfaced by bulk import.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* refactor(usage): restrict models.dev pricing import to a single model
Each update_model_pricing call triggers a backfill pass that loads
every zero-cost usage row before filtering by model, so bulk-importing
N entries scaled as selectedModels x allZeroCostLogs full scans.
Re-importing pricing is rare, so drop the batch path instead of
optimizing it: the picker is now single-select, one import runs exactly
one update_model_pricing call and one backfill pass. This also removes
the normalized-ID dedup logic and the useImportModelPricing hook in
favor of the existing useUpdateModelPricing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
- useUsageQuery: retry once + keep-last-good — show the last successful
result for up to 10min when a query fails transiently (network/timeout/
HTTP 5xx), so a single blip no longer flips the card to red. Deterministic
failures (auth, empty key, unknown provider, 4xx) surface immediately and
clear the snapshot so a stale quota can't resurface after credentials change.
- bump native balance/coding-plan/subscription request timeouts 10s -> 15s
for slow cross-border endpoints.
- coding_plan: return explicit errors ("API key is empty" / "Unknown coding
plan provider") instead of a blank failure, mirroring balance.
- add unit tests for keep-last-good and transient/deterministic classification.
Official providers (Claude/Codex/Gemini/Claude Desktop) use OAuth with an intentionally empty base_url and connect via the client's default endpoint, so cc-switch has no reliable reachability target. Probing a guessed endpoint either hits the wrong target or returns a meaningless green light. Hide the connectivity button for category === 'official'; reachability stays available for copilot/codex-oauth/third-party/custom providers, which is where the old real-request probe produced false negatives. Revert the official base_url fallback added earlier — resolve_base_url is back to extract-or-error.
The 1500ms degraded threshold was too strict; normal ~1s probe latencies showed as 'slow'. Restore the original 6000ms scale (default + config panel + per-provider range). Keep the reachability-appropriate 8s timeout / 1 retry.
The provider panel health check sent a real streaming model request, which many third-party providers block (401/403/WAF), causing false negatives while only stable official endpoints passed. Replace it with a lightweight reachability probe: GET the provider base_url and treat any HTTP response (200/4xx/5xx) as reachable; only DNS/connect/TLS/timeout count as failure. Latency is the probe's TTFB.
Backend (services/stream_check.rs): rewrite ~2200 -> ~350 lines, dropping real-request building, format conversion, auth and API-path resolution while keeping per-app base_url extraction. Defaults: 8s timeout, 1 retry, 1500ms degraded threshold.
Failover invariant: the reachability check must never reset the circuit breaker (reachable != usable; a 403 host is reachable but broken for real traffic). Remove the resetCircuitBreaker call from useStreamCheck; failover failure detection stays driven solely by real proxy traffic (forwarder/circuit_breaker untouched). useResetCircuitBreaker is kept dormant for a future manual-recovery entry.
Open the check to all providers: drop the official/copilot/codex-oauth/third-party gating and the 'sends a real request' confirm dialog. For official providers whose base_url is intentionally empty, fall back to the endpoint the client actually uses (Claude -> api.anthropic.com, Codex -> chatgpt.com/backend-api/codex, Gemini -> generativelanguage). Non-official providers with a missing base_url still error to avoid a false green light. Claude Desktop Official is native 1P mode (talks to claude.ai, cc-switch not in the request path, no reliable endpoint) so its button stays hidden.
Slim StreamCheckConfig and per-provider testConfig to timeout/threshold/retries (drop test model + prompt); sync zh/en/ja/zh-TW. Retain the now-unused anthropic_to_openai/anthropic_to_gemini transform utilities and their test suites.
Codex buckets resume history by the model_provider id recorded in each
session: official runs (no key, built-in "openai") and cc-switch
third-party runs (shared "custom") are mutually invisible in the resume
picker. Add an opt-in setting that runs official providers under the
shared "custom" id so future official sessions land in the same history
bucket as third-party ones. Forward-only by design: existing sessions
are not migrated.
When enabled, official live config.toml gets model_provider = "custom"
plus a [model_providers.custom] entry that mirrors the built-in openai
provider (requires_openai_auth routes auth to the ChatGPT login in
auth.json, name "OpenAI" keeps is_openai() feature gates, explicit
supports_websockets/wire_api restore built-in defaults). auth.json is
untouched.
Key invariants:
- Injection lives only in the live config: switch-away backfill strips
the exact injected shape, so stored provider configs stay clean and
turning the toggle off fully reverts on the next write.
- Toggle changes apply immediately via a takeover-aware reapply: when
the proxy owns the live config (backup/placeholder present), only the
live backup is updated, mirroring the provider-switch path.
- The takeover backup path runs the same injection so a takeover
release restores a config that still carries the unified routing.
- Injection refuses to activate a foreign [model_providers.custom]
table (e.g. stale entry with a third-party base_url) to avoid routing
ChatGPT OAuth traffic to an unknown backend.
The toggle lives under Settings → Codex App Enhancements; the
description warns that resuming old sessions across providers may fail
because encrypted_content reasoning only decrypts on the backend that
created it (upstream treats cross-provider resume as unsupported).
The provider/model filters only lived inside the request-log table, so
there was no way to see "how much did app X spend on source Y" across
the whole dashboard. Promote them to the top bar next to the app
filter, applying globally to the hero summary, trend chart, request
logs, and both stats tabs.
Backend: the five stats queries (summary, summary-by-app, trends,
provider stats, model stats) accept optional provider_name/model
filters, applied to both the detail and daily-rollup branches (the
rollup PK already carries provider_id/model/pricing_model). Sources
match by exact display name via provider_name_coalesce, so session
placeholder rows like "Claude (Session)" are selectable; models match
by effective pricing model (pricing_model falling back to model), the
same grouping key the model-stats tab uses. Request-log filtering
switches from LIKE to these exact semantics.
Frontend: two truncating dropdowns list only sources/models that have
data in the current range, with the model list cascading from the
selected source. Dynamic option values are prefix-encoded so a source
literally named "all" cannot collide with the sentinel, query keys
fall back to null instead of "all", and the option queries follow the
dashboard refresh interval (otherwise their default 30s polling drags
same-key stats queries along even with refresh disabled). The
request-log filter bar keeps only the log-specific status-code select.
Labels read "sources" rather than "providers" because direct-connect
session buckets sit alongside real providers. i18n updated across
zh/en/ja/zh-TW.
The Desktop gateway's proxy traffic is still recorded under its own
app_type for route-takeover billing audit (the request detail panel
shows the real value), but the dashboard now folds it into `claude`
for display. A standalone "Claude Desktop" bucket only ever showed a
partial number: Desktop chat usage never passes through the proxy and
has no scannable local source, while its Code-tab sessions are just the
embedded Claude Code runtime writing into the shared ~/.claude/projects
tree — so a separate bucket misled users into reading it as Desktop's
full usage.
Backend: new `folded_app_type_sql` helper wraps the app_type column in
every dashboard read path (10 filter sites + the by-app projection) so
`= 'claude'` also matches claude-desktop and GROUP BY merges the two,
without changing bound-param counts. Dedup matching and provider-limit
checks keep exact comparison; get_request_logs folds only the WHERE
filter and keeps the raw app_type in its row projection.
Frontend: drop claude-desktop from the dashboard AppType/KNOWN_APP_TYPES
filter list, the UsageHero theme, and the now-dead appFilter locale key
in all four languages (the managed-app apps.claude-desktop label stays).
Adds test_claude_desktop_folds_into_claude_for_display.
* fix(proxy): preserve Codex OAuth auth token on takeover
* style(proxy): format Codex OAuth takeover fix
* fix(proxy): unconditionally inject AUTH_TOKEN placeholder for codex takeover
The preserve-if-exists condition left #3784 unfixed on three paths:
hot-switch passes the provider's settings (presets carry no
ANTHROPIC_AUTH_TOKEN key), fresh installs never had the key, and live
configs already stripped by older releases stay stripped.
- Fold the bool parameter into the policy enum as
ManagedAccount { keep_auth_token } so every construction site
declares intent
- Decide via !is_github_copilot() within the managed branch so
URL-only codex providers (no provider_type meta) are covered,
matching the predicate family used for policy selection
- Inject the placeholder unconditionally instead of only when the
key pre-exists; Copilot behavior is unchanged (API_KEY only)
- Pin the previously uncovered cases with tests: codex without a
pre-existing key, URL-only codex, and Copilot removing a stale
AUTH_TOKEN
---------
Co-authored-by: codeasier <liuyekang@huawei.com>
Co-authored-by: Jason <farion1231@gmail.com>
Changing app_config_dir relocates the SQLite database, so a restart
triggered while proxy takeover is active used to strand the live
configs: the new instance reads a fresh DB with no live backups, the
first-run import then persisted the PROXY_MANAGED placeholder as the
`default` provider, and the no-backup recovery path wrote that
placeholder right back to the live files — leaving Claude/Codex/Gemini
pointed at a dead local proxy with no automatic way out.
Three orthogonal fixes, defense in depth:
- restart_app now awaits cleanup_before_exit() before app.restart().
Since #4069 the ExitRequested handler intentionally defers restart
requests to Tauri's default re-exec without custom cleanup, which is
correct for same-DB restarts but not for this command's dir-change
use case: only the old instance holds the backups needed to restore
the taken-over live files, so it must restore them while its event
loop is still alive.
- import_default_config refuses to import a live config that is under
proxy takeover (placeholder detected), instead of persisting it as
the current provider.
- restore_live_from_ssot_for_app validates that the current provider's
settings_config does not itself contain takeover placeholders before
writing it back; polluted SSOT now falls through to the placeholder
cleanup fallback.
Regression tests cover the import guard and the no-backup recovery
path (the latter fails before this change by writing PROXY_MANAGED
back to live).
Extract a shared `parse_custom_user_agent` helper in provider.rs returning
`Result<Option<HeaderValue>>`, and reuse it in the forwarder, stream check,
and model fetch paths so detection, forwarding, and model listing all apply
the same provider-level User-Agent. Previously only the forwarder honored it,
so stream check could fail (or model listing 403) on UA-gated upstreams that
the proxy itself handled fine.
- stream_check injects the provider's custom UA on the claude/codex paths and
still skips the GitHub Copilot fingerprint UA.
- model_fetch service + command and the model-fetch.ts wrapper thread an
optional UA through to GET /v1/models.
- runtime callers silently ignore invalid values via `.ok().flatten()`
(no save-time block, so deeplink imports stay lenient).
- proxy_request_logs: add pricing_model column recording the basis actually
used at write time (NULL = pre-v11 rows, '' = unpriced error rows)
- cost backfill recomputes strictly by the persisted basis; the
request_model fallback now only applies to placeholder models, so
real-but-unpriced takeover rows stay at zero cost until pricing is
added instead of being permanently frozen at the alias's price
- backfill_missing_usage_costs_for_model can locate rows by pricing_model
- usage_daily_rollups: rebuild with request_model + pricing_model in the
primary key so the alias-to-real-model mapping and the pricing basis
survive the 30-day prune; legacy rows migrate with ''
- rollup_and_prune backfills costs before pruning: prune is irreversible
and used to run before the startup backfill, permanently booking
then-unpriced rows as zero
- get_model_stats groups by the effective pricing model
(COALESCE(NULLIF(pricing_model,''), model)) so costs aggregate under
the model whose prices produced them; response-mode behavior unchanged
The Zhipu quota API returns two TOKENS_LIMIT entries whose identity was
inferred by sorting nextResetTime ascending (nearest = five_hour). In the
last hours of each weekly cycle the weekly window resets sooner than the
current 5-hour session window, so the two buckets were swapped exactly
when users check their weekly quota most.
Classify by the explicit unit field instead (3 = hour window -> five_hour,
6 = week window -> weekly_limit; same shape on bigmodel.cn and api.z.ai,
weekly observed with number 7 and 1 so only unit is matched), falling back
to the old reset-time heuristic when the field is missing.
The local session-log scanner dropped any assistant message that lacked
a stop_reason or had output_tokens==0. Claude Code Workflow / sub-agent
fan-out frequently produces messages that only wrote a message_start
snapshot (output=1, stop_reason=None) without a final block, yet their
input + cache_read + cache_creation tokens are already billed by
Anthropic (charged once the request is accepted). Dropping them
under-counted usage by ~4.1% overall, 92% concentrated in
workflow/subagent transcripts.
Replace the stop_reason/output gate with a billable-token check (any of
input/output/cache_read/cache_creation > 0). The per-message-id dedup
selection is unchanged, and request_id = "session:"+msg_id PRIMARY KEY
with INSERT OR IGNORE keeps each message single-inserted, so relaxing
the gate cannot double-count. Add a regression test covering a
stop_reason-less message with real cache cost plus an all-zero skip.
This is the parser-layer half of the Workflow under-counting fixed at
the collector layer in 8d332925.
Jason