Commit History
-
feat: theme dashboard charts and expose admin UI on HF Spaces
shinmentakezo078bdd0d7
-
feat: track local-path usage and enrich analytics dashboard
shinmentakezo07- Record server-tool and local-optimization requests in the usage ledger - Relay cache read/creation tokens into Responses usage details - Add cache/reasoning KPIs, request-log pagination, heatmap metric toggle, latency tooltips, and a sticky sidebar in the Admin UI - Rebuild admin static assets for the updated dashboard - Add usage-tracking and cache-token relay tests - Bump version to 4.28.0c860470 -
feat: add usage tracking with live throughput rates and analytics dashboard
shinmentakezo07 +1: ClaudeTrack per-request token usage (input/output/cache/reasoning) plus full prompts (capped at 64 KB) in a JSON ring buffer at ~/.fcc/usage.json, captured through a transparent UsageTrackingStream over the executor stream. Expose GET /admin/api/usage and a new Usage page in the admin UI with KPI cards, token mix, volume area chart, top endpoints, routing sankey, and hourly heatmap, auto-refreshing every 10 seconds. Throughput rates use a trailing 60-second live window on both backend and frontend so they track real-time traffic and decay to zero when idle; tokens/second counts output tokens only, tokens/minute counts input plus output. Bump 4.25.0 -> 4.27.0. Co-Authored-By: Claude <noreply@anthropic.com>e8f2acc -
feat: surface saved endpoint models and revamp the admin UI
shinmentakezo07Model ids saved per OpenAI-compatible endpoint on the admin page are now included in the admin model options and provider status, so numbered routes (openai_compatible_1/<model>) stay listed even when the endpoint's live /models route is unreachable. - settings: OpenAI-compatible instances persist an optional models list - status: endpoint cards report saved model ids alongside live ones - admin API: model options merge configured, discovered, and stored ids - tests: stored endpoint models appear in options/status; config round-trip Admin UI overhaul: - dark mode toggle (system-aware, persisted) with a light/dark token pass - responsive sidebar: off-canvas drawer with backdrop/Escape on mobile - sticky blurred header (view icon, title, description, version) and a sticky action bar with status icons and a dirty-count badge - config sections become cards with per-section icons and a model refresh action; fields gain source badges and description hints - provider/connected-account c…ee2cd20 -
fix: route numbered OpenAI-compatible instance model refs
shinmentakezo07Chat requests to a numbered endpoint model (MODEL=openai_compatible_1/<id>) failed with 400 "Unknown provider_type: 'openai_compatible_1'" because ModelRouter._validate_provider_id only accepted static catalog entries and never saw the dynamic numbered instance ids before the provider factory could resolve them. - ModelRouter now accepts the openai_compatible_<N> numbering scheme at route time; whether a specific number exists is still validated by the provider factory, which reports the configured instance ids - the instance id pattern is shared (config/provider_catalog) between the router and the runtime factory instead of a private duplicate regex - tests: numbered instance model refs route; unknown numbers defer to the factory; unknown provider ids still raise the typed error - version 4.23.1 -> 4.23.2af75bef -
fix: cache-bust admin UI and show the running version
shinmentakezo07The admin UI required a hard refresh after every deploy: the bundle was emitted with fixed filenames (assets/admin.js, assets/admin.css) even though the server already sent Cache-Control: no-store. Any cache that ignored or reused those URLs (browser memory/disk cache, back-forward cache, or an edge proxy in front of the deployment) kept serving the old bytes, since the URL never changed between deploys. - vite emits content-hashed filenames (admin-[hash].js/.css), so each build references new URLs and stale caches become inert by construction; index.html stays no-store and always fresh - admin config API now returns the installed package version (api -> core import, respecting the config package's dependency policy) and the header shows a vX.Y.Z badge, so a deployed update is visible without digging - cache-policy contract test now covers every built asset instead of the fixed names; version presence asserted in the manifest test - version 4.23.0 -> 4.23.11c54cf1 -
fix: make model combobox editable and searchable in Model Routing
shinmentakezo07The model field combobox looked searchable but was broken: the input was controlled by the committed value so keystrokes were reverted, typed text could never be committed (the empty state promised a custom slug that was impossible to enter), Enter only worked after arrow selection, and blur discarded the draft. Rewrite it as a true editable+searchable combobox: - the input shows the draft text while editing, so model ids can be typed or edited in place (WYSIWYG: the text in the box is the value) - filtering over the model list now reacts to visible typing - Enter commits the highlighted option, or the typed text as a custom model id when nothing is highlighted - blur commits the draft (exact option match selects the canonical option; anything else becomes a custom id); Escape reverts - stale filter/query and arrow index state reset on open/query change - version 4.22.1 -> 4.23.0c03ea68 -
feat: fetch models from OpenAI-compatible endpoints in admin UI
shinmentakezo07Add a Fetch models button to each numbered endpoint card on the OpenAI-Compatible page. It calls the endpoint's /models route through the existing provider test endpoint (GET {base}/models via the OpenAI client), lists the returned model ids on the card, adds them as provider/model options for the Model Config dropdowns, and offers "Use as default" to set the numbered route as the default model. - frontend: per-endpoint fetch/clear/use-model state, model list panel on each card, error surfacing through the action bar - tests: numbered instance list_model_infos() fetches model ids; contract status test accounts for numbered instances replacing the legacy openai_compatible entry when endpoints are configured - docs: README fetch-models notes - version 4.21.0 -> 4.22.0969cedb -
feat: numbered OpenAI-compatible endpoints admin page
shinmentakezo07Add a dedicated Admin UI page for OpenAI-compatible endpoints. Each endpoint becomes a numbered provider (openai_compatible_1/<model>, openai_compatible_2/<model>, ...) stored as a JSON array in OPENAI_COMPATIBLE_INSTANCES, with per-endpoint base URL, comma-separated key pool, and proxy. Endpoint 1 also answers the legacy openai_compatible/ route; the single-endpoint env vars remain the fallback. - settings: OpenAICompatibleInstance model, JSON parse/validate, numbered route validation across model fields - runtime: instance resolution in create_provider, provider config builder, model-list discovery includes numbered ids - admin: manifest field + status entries (numbered instances replace the single-endpoint entry when configured) - frontend: OpenAI-Compatble view with add/remove/import/test per endpoint - tests: settings parsing, route validation, factory resolution, discovery, manifest, admin status - version 4.20.0 -> 4.21.0f78c884 -
feat: add generic OpenAI-compatible endpoint provider
shinmentakezo07Add an openai_compatible provider so users can point at any OpenAI-compatible Chat Completions server with a custom base URL (server root or /v1) and a comma-separated API key pool. The profile sends no reasoning params and parses /think tags; keys rotate round-robin with failover on transient errors like every other provider. Admin manifest gains key/base URL/proxy fields, smoke override, and helpful descriptions. Bump to 4.20.0.8618eca -
feat: admin dedicated pages, pool key management, keypool quotas (#1350)
shinmentakezo07- Admin UI becomes a Vite build under frontend/ with hash-routed dedicated pages (#/providers, #/model-config, #/messaging); the providers view renders only on its page. - Pool fields support reveal/edit/remove of individual keys via a new loopback-only GET /admin/api/pools/{field_key} endpoint; apply resolves tokens mixed with raw edits against stored keys. - admin assets now serve Cache-Control: no-store, fixing stale-bundle breakage after the frontend migration. - ApiKeyPool gains per_key_quota rotation plus transient HTTP 429/5xx/overload failover rotation with mark_failure. - CI runs the frontend build and verifies committed admin assets.9b71fd0 -
feat: add key management and provider keypool (# new feature)
shinmentakezo07 +1: Claude- Add src/free_claude_code/config/admin/keys.py - Add src/free_claude_code/providers/keypool.py - Add corresponding test files - Update admin static assets and config files - Bump version to 4.18.0 Co-Authored-By: Claude <noreply@anthropic.com>2c96ecb
-
Use codebase .env for server and admin config
shinmentakezo07Settings and the Admin UI now load and persist config from the repo-local .env instead of the user-managed ~/.fcc/.env, so a codebase .env is always authoritative. Admin writes target the repo .env, and the managed env source label is removed from the Admin UI.3cc3731 -
Make OpenCode Zen mappings explicit (#1342)
Unverified signature## Problem OpenCode Zen uses the ambiguous `opencode/` identifier while OpenCode Go uses `opencode_go/`. Renaming Zen without migrating saved configuration would break existing installations. ## Changes | Before | After | | --- | --- | | OpenCode Zen used `opencode/` in model mappings and `OPENCODE` in diagnostics. | OpenCode Zen uses `opencode_zen/` and `OPENCODE_ZEN` throughout FCC. | | Existing FCC-owned environment files retained legacy Zen model references. | Server startup rewrites legacy Zen model references before settings validation. | | Zen used `OPENCODE_PROXY` and `FCC_SMOKE_MODEL_OPENCODE`. | Zen uses `OPENCODE_ZEN_PROXY` and `FCC_SMOKE_MODEL_OPENCODE_ZEN`. | | The package version was `4.16.7`. | The package version is `4.16.8`. | <!-- greptile_comment --> <details open><summary><h3>Greptile Summary</h3></summary> This change renames the OpenCode Zen provider and related environment settings, adds migration for legacy FCC-owned dotenv configuration on the server path…d837357
-
Preserve empty reasoning state across native tool turns (#1339)
Unverified signature## Problem Claude tool-only assistant turns omitted `reasoning_content` when a provider used native reasoning replay. DeepSeek V4 requires that field to remain present—even as an empty string—so OpenCode Zen and Go continuations could fail despite valid Claude history. Fixes #1336 Fixes #1338 ## Changes | Before | After | | --- | --- | | Native tool turns without thinking omitted `reasoning_content`. | Native `reasoning_content` tool turns send `""` when no replayable reasoning exists. | | OpenCode coverage injected an internal empty field before conversion. | Zen and Go coverage starts from real tool-only Anthropic history and asserts the final SDK wire JSON. | | DeepSeek owned a private wire-capture helper. | Provider tests share one OpenAI Chat wire-capture helper. | | The behavior was only simulated. | The exact tool-only continuation passed against the real DeepSeek V4 Flash endpoint. | <!-- greptile_comment --> <details open><summary><h3>Greptile Summary</h3></summary> Thi…81a9761
-
Restore OpenCode reasoning replay for tool turns (#1328)
Unverified signature## Problem OpenCode replayed reasoning tool turns as visible `<think>` tags instead of native `reasoning_content`. DeepSeek thinking models rejected the following tool-result request with HTTP 400. Fixes #1327. ## Changes | Before | After | | --- | --- | | OpenCode Zen and Go replayed prior reasoning through `<think>` tags. | OpenCode Zen and Go replay prior reasoning through native `reasoning_content`. | | Tool-follow-up coverage did not enforce OpenCode's native replay contract. | Zen and Go coverage verifies non-empty tool reasoning and explicit empty reasoning state. | | The package version was `4.16.5`. | The package version is `4.16.6`. | <!-- greptile_comment --> <details open><summary><h3>Greptile Summary</h3></summary> This change restores native reasoning replay for the OpenCode and OpenCode Go provider profiles and updates the package version to 4.16.6. The reasoning replay path was exercised for both profiles with empty reasoning and assistant thinking followed by a …972c110 -
Restore OpenAI subscription generation compatibility (#1326)
Unverified signature## Problem OpenAI subscription generations failed because FCC sent public Responses fields that the private Codex backend rejects. Valid backend SSE without a `Content-Type` header was then misclassified as non-streaming. Fixes #1320. ## Changes | Before | After | | --- | --- | | Private Codex requests forwarded `max_output_tokens` and `metadata`. | Private Codex requests omit those unsupported public fields while leaving the neutral Responses contract unchanged. | | Successful streams required `text/event-stream`. | Successful streams may omit `Content-Type`; explicitly declared non-SSE responses retain the bounded failure path. | | Non-SSE HTTP 2xx diagnostics depended on the header gate. | Non-SSE HTTP 2xx bodies remain retryable failures with bounded, redacted diagnostics. | | Compatibility coverage assumed canonical response headers. | Provider and API tests cover private projection, immutable input, headerless SSE success, explicit non-SSE failures, bounded diagnostics, and du…f412f5f
-
Restore Claude Code client controls for OpenAI Responses (#1311)
Unverified signature## Problem Claude Code sends effort and keep-all context controls that the connected OpenAI provider rejected before upstream I/O. The Responses converter revalidated application-owned effort and could not distinguish inert context edits from active unsupported behavior. Fixes #1310. ## Changes | Before | After | | --- | --- | | Every non-empty `output_config` failed OpenAI Responses preflight. | Application-consumed `output_config.effort` reaches one policy-derived OpenAI `reasoning` object. | | Every non-empty `context_management` failed OpenAI Responses preflight. | Empty edits and exact keep-all thinking edits are omitted as semantic no-ops. | | Structured output, active edits, and unknown controls shared coarse top-level diagnostics. | Unrepresentable controls still fail before I/O with precise nested output paths. | | The current Claude request shape had no end-to-end regression contract. | Core, real-provider, and `/v1/messages` tests lock conversion, immutability, wire paylo…1b6cc04 -
Keep tool names portable across OpenAI transports (#1309)
Unverified signature## Problem OpenAI-compatible upstreams receive Claude tool names unchanged, so names outside the portable OpenAI function grammar can be rejected before inference. Fixes #1307. ## Changes | Before | After | | --- | --- | | Long or non-portable tool names were forwarded verbatim to Chat Completions and upstream Responses APIs. | A shared deterministic codec emits portable aliases while leaving valid names unchanged. | | Upstream aliases could leak into tool validation, recovery, or client-visible SSE. | Structured, heuristic, recovered, and Responses tool calls restore the exact client name before downstream meaning. | | NIM function-name compatibility would require a provider-specific workaround. | The protocol boundary handles every OpenAI transport and composes with NIM's separate argument aliases. | <!-- greptile_comment --> <details open><summary><h3>Greptile Summary</h3></summary> This change makes tool names portable across OpenAI-compatible Chat and Responses transports by …522b31e
-
Add Codex App model catalog support (#1297)
Unverified signature## Problem The Codex App cannot discover FCC provider/model slugs because it loads a local model catalog at startup instead of querying FCC's `/v1/models` endpoint. Users can configure FCC as a Responses provider, but its native model picker remains incomplete. Fixes #1260. ## Changes | Before | After | | --- | --- | | FCC generated the Codex catalog only when `fcc-codex` launched. | FCC synchronizes the catalog from its canonical model lifecycle for independently launched Codex clients. | | Startup could leave a configured Codex App without a catalog file. | Startup creates a missing catalog after routed-provider warming and preserves an existing last-known-good catalog. | | Every catalog publication replaced the file. | Identical content is skipped and changed content is replaced atomically. | | Catalog filesystem failures could escape a direct writer call. | Runtime publication is fail-open and logs only the exception type. | | Codex App setup was undocumented. | The README docum…49fa921
-
Add Azure OpenAI provider (#1284)
Unverified signature## Problem Azure OpenAI deployments cannot be routed through FCC even though Azure exposes an OpenAI-compatible v1 Chat Completions endpoint. Fixes #1268. ## Changes | Before | After | | --- | --- | | Azure resource credentials and deployment routes had no FCC owner. | `azure_openai` requires an API key and exact `/openai/v1/` resource URL, then routes `azure_openai/<deployment-name>` through the shared OpenAI Chat adapter. | | A missing resource URL surfaced as an internal provider-construction failure. | Admin and runtime readiness distinguish a missing API key from a missing resource URL. | | OpenAI-compatible model discovery assumed every returned model ID was routable. | Azure validates its model endpoint without advertising underlying model IDs as custom deployment names. | | Provider documentation and smoke configuration omitted Azure OpenAI. | Setup, deployment-name guidance, proxy configuration, and opt-in smoke coverage include Azure OpenAI. | <!-- greptile_comment --> <…f77ae26
-
fix: detect NIM 500 context overflow as context_window_exceeded (#1280)
Unverified signatureNVIDIA NIM can return HTTP 500 for context overflow during long-context sessions, but _provider_failure_override() only checked openai.BadRequestError (400). The 500 fell through to generic UPSTREAM/500 classification, causing the admission controller to wastefully retry the same oversized payload up to 5 times instead of returning the context_window_exceeded failure that triggers Claude Code's compaction. Extend the override to also handle openai.InternalServerError (500) with the same negative-max-tokens body pattern. Degraded function detection stays 400-only. Also fix a Python except syntax bug in native_tool_stream.py where except json.JSONDecodeError, ValueError: caught only JSONDecodeError and shadowed the ValueError builtin. <!-- greptile_comment --> <h3>Greptile Summary</h3> Updates NVIDIA NIM error handling and native tool argument coercion. - **[Files Changed]** `client.py` now recognizes NIM’s negative-`max_tokens` context-overflow signature in both HTTP 400 and 500 SD…ae9c841 -
Add ChatGPT OAuth as an OpenAI Codex provider (#1279)
Unverified signature## Problem FCC supports API-key and local providers, but customers cannot use models included with their ChatGPT subscription. Reusing Codex credentials would also couple FCC to another application's mutable authentication state. ## Changes | Before | After | | --- | --- | | OpenAI subscription access was unavailable. | Customers connect ChatGPT from the local Admin UI with browser PKCE or device code. | | Provider availability came only from static settings. | A process-owned connected account adds, refreshes, and evicts OpenAI models dynamically. | | FCC had no ChatGPT Codex transport. | `openai/<model-id>` uses FCC-owned renewable credentials and neutral Responses conversion. | | Retries and reasoning continuity had no subscription-provider coverage. | Commit-boundary retries prevent duplicate output while encrypted reasoning and tool IDs survive round trips. | <!-- greptile_comment --> <details open><summary><h3>Greptile Summary</h3></summary> This PR adds ChatGPT OAuth and a…9fb95ba
-
Make Kilo discovery and reasoning protocol-correct (#1274)
Unverified signature## Problem Kilo was registered as a plain OpenAI Chat profile even though its gateway uses [OpenRouter-format reasoning](https://github.com/Kilo-Org/kilo-pi-provider/blob/master/kilo.ts), identifies Responses-only models, and exposes output-modality metadata. That made incompatible models discoverable and could drop or duplicate structured reasoning across tool turns. The live stream helpers also relied on omitted-stream behavior instead of explicitly exercising SSE. ## Changes | Before | After | | --- | --- | | Kilo used generic ID-only discovery. | Kilo advertises tool-capable Chat Completions models and reports reasoning support from live metadata. | | Image-output and Responses-only models appeared in the Chat catalog. | Kilo filters those catalog entries while direct configured-model execution remains upstream-authoritative. | | Kilo read a legacy reasoning field and discarded structured details. | Kilo reads `reasoning`, emits plaintext once, and preserves opaque details acros…52c1e38 -
feat: add Kilo.ai provider (#1273)
tdogghockeyUnverified signature2d79060
-
Close interrupted tool turns before new user input (#1271)
Unverified signature## Problem The remaining interrupted-session failure in #1269 replayed completed tool output directly before a new user prompt. Strict OpenAI-compatible chat templates reject this `tool → user` sequence. ## Changes | Before | After | | --- | --- | | Shared conversion emitted a user turn immediately after completed tool output. | Shared conversion inserts one neutral whitespace-only assistant boundary before subsequent user input. | | Terminal tool results and already-completed assistant turns had no explicit regression distinction. | Boundaries are added only for `tool → user`; terminal tools and existing assistant completions remain unchanged. | | OpenRouter reasoning replay treated every converted assistant message as source history. | OpenRouter skips the structural boundary when restoring encrypted reasoning details. | | The interrupted resume shape had no permanent live product scenario. | An opt-in provider smoke exercises the exact resume transcript. | <!-- greptile_comment …5a80185 -
Preserve gateway model identity in Anthropic responses (#1270)
Unverified signature## Problem FCC replaced the stable gateway model selected by Claude Code with the provider-facing model when serializing Anthropic responses. Claude then persisted an ID absent from FCC's `/v1/models` catalog, so later sessions warned that the model could not be restored and fell back. This partially addresses #1269. ## Changes | Before | After | | --- | --- | | Provider streams, local optimizations, and local server-tool responses exposed the routed upstream model. | Every Anthropic response path receives and emits the original gateway model while upstream requests retain the provider model. | | Traces labeled the routed upstream ID as the gateway model. | Routing, execution, provider, and token-count traces preserve both identities accurately. | | No contract guarded model identity across routing and response serialization. | Tests cover executor handoff, upstream request bodies, Anthropic `message_start`, local optimizations, and server tools; the patch version is `4.12.10`. | <…82e8d6e
-
Contain MiniMax native tool markup across NIM streams (#1259)
Unverified signature## Problem MiniMax models served through NVIDIA NIM can emit private tool-call markup in content or reasoning output, including requests without tool schemas. FCC only inspected tool-bearing content streams, so control tokens could reach clients; this partially addresses #1254. ## Changes | Before | After | | --- | --- | | Native normalization ran only when tool schemas were present. | Every NVIDIA NIM stream contains the MiniMax control namespace while preserving ordinary chunks unchanged. | | Only content deltas were inspected for native tool calls. | Content and reasoning deltas share the same framing and validation contract. | | Undeclared or malformed native calls could appear as assistant text. | Declared calls become structured tool calls, while invalid calls enter FCC's existing bounded recovery lifecycle without leaking markup. | | The normalizer could let raw-stream cleanup replace the established request outcome. | The normalizer delegates closure so the provider runner p…ab33496 -
Make the FCC logo background transparent (#1257)
Unverified signature## Problem The FCC mark is enclosed by a dark rounded-square tile. The tile prevents the symbol from blending naturally with README, desktop, tray, and menu-bar backgrounds. ## Changes | Before | After | | --- | --- | | The three branches and terminal sit on a fixed dark tile. | The branches and terminal sit on a transparent canvas. | | Packaged and README artwork include the outer background. | SVG, PNG, ICO, ICNS, and theme-aware README artwork retain only the FCC symbol. | | The package version is 4.12.7. | The package version is 4.12.8. |07abc6b -
Make the FCC terminal mark square (#1256)
Unverified signature## Problem The terminal frame in the FCC mark is taller than it is wide. This makes the convergence symbol end in a portrait rectangle instead of the intended square terminal. ## Changes | Before | After | | --- | --- | | The terminal frame measures 244 by 280 units. | The terminal frame measures 244 by 244 units and remains vertically centered. | | Packaged and README artwork use the portrait frame. | SVG, PNG, ICO, ICNS, and theme-aware README artwork use the same square frame. | | The package version is 4.12.6. | The package version is 4.12.7. |1a07ecc -
Add a native Free Claude Code app icon (#1252)
Unverified signature## Problem Free Claude Code's desktop surfaces had no shared visual identity. The tray used a generated blue `F`, Windows shortcuts inherited the generic executable icon, and the macOS app bundle declared no icon at all. ## Changes | Before | After | | --- | --- | | Desktop artwork was generated independently at runtime. | A deterministic SVG source and packaged PNG, ICO, and ICNS variants define one high-contrast orange FCC mark. | | The tray, Windows shortcuts, and macOS bundle displayed unrelated or generic artwork. | Every desktop surface now consumes the same packaged brand asset. | | Installers had no native icon handoff contract. | `fcc-desktop --export-icon` copies the wheel-owned format into installer-managed locations, with fail-fast verification and uninstall coverage through the existing managed paths. | | Icon packaging and platform wiring were implicit. | Regression tests validate native containers, byte-preserving export, Windows shortcut metadata, macOS bundle metada…1bfa806 -
Report provider permission denials as HTTP 403 (#1249)
Unverified signature## Problem Provider HTTP 403 responses were categorized as generic API failures or rewritten as authentication 401 errors. Users received misleading diagnostics for credential or model-access denials, as reported in #1242. ## Changes | Before | After | | --- | --- | | Typed and generic OpenAI 403 errors fell through to generic API failures. | Typed and generic OpenAI 403 errors become canonical permission failures. | | Raw HTTP 403 errors were rewritten as authentication failures with status 401. | Raw HTTP 403 errors retain permission semantics and status 403. | | Permission diagnostics used generic provider-failure wording. | Permission diagnostics advise checking credential permissions and model access while retaining redacted upstream detail. | | Access-denial wire behavior was implicit. | Messages and Responses return `permission_error` with `x-should-retry: false`. | | The package version was 4.12.2. | The package version is 4.12.3. | <!-- greptile_comment --> <details open>…7b4e84b -
Recover MiniMax-M3 native tool calls on NVIDIA NIM (#1246)
Unverified signature## Problem NVIDIA NIM can emit MiniMax-M3's native tool envelope in streamed content instead of OpenAI `tool_calls`. FCC forwarded that markup as assistant text, so Claude displayed control tokens and terminated the request without executing the tool. Fixes #1243 ## Changes | Before | After | | --- | --- | | NIM streamed MiniMax-M3 native tool markup as visible assistant text. | NIM recognizes the exact native signature on tool-bearing requests and converts complete blocks into OpenAI tool-call deltas. | | Native argument markup was untyped and unvalidated. | Recursive objects, arrays, scalars, aliases, and parallel calls are decoded and validated against the accepted request schemas. | | Malformed or incomplete native blocks escaped as successful HTTP 200 content. | Invalid native blocks stay behind the stream boundary and trigger bounded protocol retry or recovery. | | Midstream continuation removed tools and could only recover text. | Tool-protocol continuation retains the origi…a60e7bf
-
Preserve Claude feature flags in interactive clients (#1239)
Unverified signature## Problem `fcc-claude` and the documented IDE setups forced `DISABLE_TELEMETRY=1`. Claude Code disables feature-flag fetching with that variable, so rollout features such as fullscreen rendering and mouse support fell back to disabled defaults. Fixes #1238. ## Changes | Before | After | | --- | --- | | Shared proxy policy forced telemetry off for interactive and managed Claude sessions. | Interactive Claude sessions preserve the user's telemetry choice, while managed headless tasks own the opt-out. | | VS Code and JetBrains examples disabled Claude feature-flag fetching. | Both IDE examples allow Claude to fetch rollout feature flags. | | Environment contracts expected every Claude surface to disable telemetry. | Regression tests distinguish interactive inheritance from managed enforcement. | | Package metadata reported version 4.12.0. | Package metadata and the lockfile report version 4.12.1. | <!-- greptile_comment --> <details open><summary><h3>Greptile Summary</h3></summary> …3f60854
-
Add a native FCC desktop launcher for Windows and macOS (#1225)
Unverified signature## Problem Free Claude Code currently has to remain attached to a terminal, so closing that window stops the proxy and users have no native way to reopen Admin or control the background server. The contributed Windows wrapper also would have introduced a second bundled server lifecycle instead of reusing FCC's cleanup and restart ownership. Fixes #1147. ## Changes | Before | After | | --- | --- | | Users keep `fcc-server` running in a terminal. | Windows and macOS users can launch a console-free FCC Desktop host from a desktop/application shortcut and control it from the tray or menu bar. | | A wrapper would need to spawn and terminate a child server. | The terminal and desktop paths share one in-process supervisor, one graceful runtime shutdown path, and an OS-held singleton lock. | | Installers manage only command entry points. | Windows installs desktop and Start-menu shortcuts; macOS installs a per-user app bundle and owned desktop link; uninstallers remove only those FCC artifa…db3606c -
Make ProviderModelInfo the sole model-catalog contract (#1222)
Unverified signature## Problem Provider model discovery consumes metadata, but providers and the cache still expose a parallel IDs-only contract. The duplicate contract adds adapters and lets tests bypass capability metadata. ## Changes | Before | After | | --- | --- | | `BaseProvider` exposed `list_model_ids()` plus a metadata adapter. | `BaseProvider` exposes only abstract `list_model_infos()` returning application-owned metadata. | | Ordinary providers parsed IDs and converted them later. | Ordinary providers parse OpenAI-compatible catalogs directly into `ProviderModelInfo` values. | | OpenRouter, Cloudflare, and GitHub Models maintained redundant IDs-only wrappers. | Provider-specific filters and capability metadata have one return path. | | Vertex returned paginated IDs for the base adapter to wrap. | Vertex returns metadata after completing the same paginated discovery flow. | | The runtime cache exposed test-only raw-ID write and prefixed-ID read helpers. | The runtime cache accepts and returns…7c0961b -
Make provider model discovery the sole catalog owner (#1221)
Unverified signature## Problem Startup model-list I/O was split between a non-enforcing configured-model validator and the real discovery path. Both queried providers and populated the same cache even though synchronous cache warm-up was the validator's only required effect. ## Changes | Before | After | | --- | --- | | Validation and discovery independently resolved providers, queried model lists, and cached results. | `ProviderModelDiscovery` solely owns model-list queries, failure reporting, and cache population. | | Startup ran configured-model validation and then launched discovery. | Startup synchronously warms referenced providers through discovery, then launches the existing missing-provider background pass. | | Absence from a provider catalog produced a non-enforcing missing-model warning. | Provider catalogs remain discovery metadata and provider execution remains authoritative. | | Configured model references retained environment-source metadata for validator diagnostics. | Configured model …1fe45a3 -
Remove obsolete init and server alias commands (#1220)
Unverified signature## Problem FCC still publishes `fcc-init` and the `free-claude-code` server alias from an obsolete installation flow. Configuration creation now belongs to the Admin UI, while `fcc-server` owns startup and legacy migration. ## Changes | Before | After | | --- | --- | | The package installs six console commands. | The package installs only `fcc-server`, `fcc-claude`, `fcc-codex`, and `fcc-pi`. | | `fcc-init` duplicates configuration creation outside the Admin UI. | The Admin UI exclusively creates managed configuration, and server startup retains legacy migration. | | Smoke contracts and architecture describe the retired commands. | Smoke contracts and architecture describe the supported command surface. | | Installer safety lists contain unexplained retired names. | Installer safety lists document why retired names remain detectable during updates and uninstall. | | The package version is `4.11.4`. | The package version is `4.11.5`. | <!-- greptile_comment --> <details open><summa…451ed03
-
Recover Claude sessions from provider context overflow (#1215)
Unverified signature## Problem NVIDIA NIM can report context exhaustion as a 400 `BadRequestError` saying its derived `max_tokens` is negative. FCC treated that as a generic invalid request, so Claude Code could not recognize the smaller upstream context window, compact the conversation, and replay the interrupted turn. LM Studio also encoded Claude's recovery phrase inside provider code instead of reporting a protocol-neutral failure. Fixes #1198. ## Changes - Add one protocol-neutral `context_window_exceeded` execution failure with a non-retryable 400 contract. - Narrowly classify only NVIDIA NIM's negative derived-`max_tokens` signature while preserving the complete redacted provider diagnostic and request ID. - Let the Anthropic serializer alone add Claude's `prompt is too long` compaction trigger; OpenAI Responses keeps a standard invalid-request envelope. - Migrate LM Studio's existing context preflight to the same neutral semantic and document the ownership boundary. - Add provider, protocol, API,…838ef01 -
Keep FCC-local traffic off outbound proxies (#1214)
Unverified signature## Problem FCC-local health and model-catalog requests inherited machine proxy settings, so an outbound proxy could return HTTP 502 even while `fcc-server` was healthy. Spawned agents inherited the same missing loopback bypass. Fixes #1199. ## Changes | Before | After | | --- | --- | | Launcher probes used the process-wide urllib proxy policy. | FCC-local probes use one proxy-disabled transport owner. | | Codex catalog loading could leave through an outbound proxy. | Codex catalog loading uses the same direct local transport. | | Claude, Codex, Pi, and managed messaging inherited incomplete bypass lists. | Spawned clients preserve outbound proxies while merging FCC and loopback hosts into `NO_PROXY` and `no_proxy`. | | Local and upstream proxy responsibilities overlapped. | CLI-local transport and provider-specific upstream proxies have explicit separate owners. | | Tests mocked launcher reachability without exercising proxy interception. | A real fake-proxy regression proves loopba…ef18dfc -
Give Google reasoning controls one owner (#1211)
Unverified signature## Problem Gemini requests with FCC reasoning could send both `reasoning_effort` and `extra_body.google.thinking_config`, which [Google documents as mutually exclusive](https://ai.google.dev/gemini-api/docs/openai#thinking). Google reasoning and thought-signature postprocessing had overlapping request ownership. Fixes #1206. ## Changes | Before | After | | --- | --- | | Shared Google quirks injected thought output independently of the profile encoder. | One provider-selected Google encoder owns every reasoning wire field. | | Gemini could send named effort beside a custom thinking config. | Gemini selects one channel, with exact budgets taking precedence over named effort. | | Vertex reasoning and thought-signature behavior shared one quirks module. | Vertex retains its budget mapping while thought signatures have a separate owner. | | Caller-native Google controls could collide with FCC controls. | Native controls are preserved only under provider-default reasoning; controlled coll…951f1d1 -
Coordinate provider recovery under concurrent load (#1205)
Unverified signature## Problem Concurrent transient failures could start independent retry, replay, continuation, and repair loops while holding provider concurrency slots. This multiplied upstream attempts and could delay or strand terminal errors under fan-out. ## Changes | Before | After | | --- | --- | | Retry paths owned separate attempt budgets. | One logical-execution session caps all upstream work at five attempts. | | Concurrent failures backed off independently. | One provider-owned recovery episode elects a single half-open probe while followers coalesce. | | Backoff occupied stream concurrency. | Concurrency is held only while an upstream operation or stream is active. | | Provider catalog calls and stream creation used separate admission paths. | Every upstream operation uses one provider-generation admission controller. | | Cancellation could leave recovery ownership or follower state unresolved. | Cancellation releases permits, transfers probe ownership, and unregisters waiting followers…3a9bf35
-
Add Google Vertex AI with renewable ADC (#1193)
Unverified signature## Problem | Before | After | | --- | --- | | FCC supported Google AI Studio API keys but could not route coding agents through a Google Cloud Vertex AI project. | `vertex/...` routes through Google's [documented OpenAI-compatible Chat Completions endpoint](https://cloud.google.com/vertex-ai/generative-ai/docs/start/openai), using the global endpoint by default or an explicitly configured region. | | A pasted Vertex access token would expire, while Application Default Credentials were not part of provider construction. | FCC loads [Application Default Credentials](https://cloud.google.com/docs/authentication/application-default-credentials), supplies a renewable credential callback to the OpenAI transport, coalesces concurrent refreshes, and returns typed authentication or transient failures. | | Vertex does not expose its model catalog through the compatible OpenAI `/models` route. | FCC translates its generic discovery operation to Google's paginated [publisher-model list API](https…cabd2d4 -
Add Amazon Bedrock Mantle support (#1192)
Unverified signature## Problem FCC cannot route coding-agent requests through Amazon Bedrock even though Bedrock Mantle exposes an OpenAI-compatible streaming Chat Completions API. Users currently need a separate compatibility layer, and FCC has no catalog, Admin UI, model-discovery, or smoke-test contract for Bedrock. Fixes #863. ## Changes | Before | After | | --- | --- | | Amazon Bedrock was absent from provider routing. | `bedrock/` uses the existing OpenAI Chat provider with AWS's [Bedrock Mantle endpoint](https://docs.aws.amazon.com/bedrock/latest/userguide/inference-chat-completions-mantle.html). | | Bedrock integration would have implied a new AWS-native transport. | The ordinary profile owns streaming, tools, retries, and `/models` discovery without boto3, SigV4, Converse, or Invoke machinery. | | FCC had no Bedrock authentication or regional endpoint configuration. | `AWS_BEARER_TOKEN_BEDROCK`, `BEDROCK_BASE_URL`, and `BEDROCK_PROXY` are available through environment and Admin UI configuratio…cc8e5d2
-
Add Kimi Code subscription support (#1183)
Unverified signature## Problem FCC's existing `kimi` provider targets Kimi Open Platform credits. Kimi Code subscription keys use a separate coding-agent endpoint, so subscribers cannot currently use their plan or select its K3 and coding models. Fixes #1161. ## Changes | Before | After | | --- | --- | | `KIMI_API_KEY` was the only Kimi contract and routed to the credit-based API platform. | `KIMI_API_KEY` remains unchanged, while `KIMI_CODE_API_KEY` routes the separate [Kimi Code subscription](https://www.kimi.com/code/docs/en/) through `https://api.kimi.com/coding/v1`. | | OpenAI-chat profiles could not declare an upstream client identity. | The Kimi Code profile sends an honest `free-claude-code` user agent as required by [Kimi's integration policy](https://www.kimi.com/code/docs/en/kimi-code/community-guidelines.html), without adding a specialized provider class. | | FCC's fallback output limit and generic token field would override Kimi's default when the client omitted a limit. | Explicit client …111b54a -
Report output-limit truncation as incomplete Responses (#1180)
Unverified signature## Problem Responses streams discarded the canonical Anthropic `max_tokens` stop reason and emitted `response.completed`. Codex therefore treated truncated provider output as a successful end and could stop midway without explaining why. Fixes #1178. ## Changes | Before | After | | --- | --- | | The Responses boundary discarded `message_delta.stop_reason`. | The Responses assembler retains the canonical terminal stop reason. | | Output-limit streams ended with `response.completed`. | Output-limit streams end with `response.incomplete` and `incomplete_details.reason=max_output_tokens`. | | Truncated output had no dedicated Responses contract coverage. | Core and API tests cover partial-output and zero-visible-output truncation while preserving response ID and usage. | | The package version was `4.8.6`. | The patch release is `4.8.7`. | <!-- greptile_comment --> <details open><summary><h3>Greptile Summary</h3></summary> This PR reports Anthropic output-limit termination as an incom…3608de3
-
Prevent the admin UI from being served from cache (#1160)
Unverified signature## Problem Browsers could retain the local admin page, assets, or error responses, leaving users on stale UI state after an FCC update. ## Changes | Before | After | | --- | --- | | Admin responses had no complete cache policy, and route-level handling missed exception responses. | A dedicated admin response boundary sends `Cache-Control: no-store` for every `/admin` response, including 403, 404, validation, and unexpected errors. | | Admin API fetches used the browser's default cache mode. | Admin API fetches explicitly use `cache: "no-store"`. | | Cache behavior was untested. | Tests cover HTML, assets, JSON, failure responses, and path scoping. | <!-- greptile_comment --> <details open><summary><h3>Greptile Summary</h3></summary> This PR prevents browsers from retaining stale admin UI responses. The main changes are: - Adds `Cache-Control: no-store` to successful and error responses under `/admin`. - Sets admin API fetches to use the browser's `no-store` cache mode. - Adds te…98786ea -
Honor reasoning tiers on numeric-budget providers (#1150)
Unverified signature## Problem FCC preserved named reasoning effort but numeric-budget providers received no intensity unless a client supplied an exact token budget. NIM and llama.cpp therefore treated Low through Max like the same provider default. ## Changes | Before | After | | --- | --- | | Named effort had no FCC-owned numeric meaning. | FCC maps Low, Medium, High, X-High, and Max to 512, 1,024, 2,048, 4,096, and 8,192 tokens. | | NIM received only thinking booleans for named effort. | [NIM](https://docs.nvidia.com/nim/large-language-models/1.15.0/thinking-budget-control.html) receives thinking booleans plus the mapped `reasoning_budget`, retaining its existing retry without rejected budget control. | | llama.cpp forwarded only exact client budgets. | [llama.cpp](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md) receives the mapped `thinking_budget_tokens` value. | | Named, boolean, and provider-default adapters shared no explicit numeric contract. | Named adapters keep wo…4faabdf -
Keep inline system reminders cache-stable across providers (#1154)
Unverified signature## Problem Inline Anthropic system reminders were forwarded as mid-conversation OpenAI system roles. Compatible provider chat templates could reposition those roles, changing prior prompt tokens and causing periodic full cache misses. Fixes #1152. ## Changes | Before | After | | --- | --- | | Top-level and inline system content both used downstream system roles. | Only top-level system content uses the leading downstream system role. | | Inline system reminders could trigger provider-side prompt retemplating. | Inline system reminders keep their text and position as downstream user messages. | | Provider policies could reinterpret the shared role mapping. | Shared conversion owns one provider-independent role mapping. | | Cache-prefix coverage allowed mid-conversation system roles. | Cache-prefix coverage requires append-only user-encoded reminders. | | The package version was 4.8.0. | The package version is 4.8.1. | <!-- greptile_comment --> <details open><summary><h3>Greptile Su…fe9d430 -
Make reasoning policy provider-neutral and client-aware (#1148)
Unverified signature## Problem FCC reduced reasoning to global and route booleans, mixing client intent, configuration, provider wire capabilities, output visibility, and history replay. That discarded named client efforts, encouraged model-name checks, and made provider behavior inconsistent. ## Changes | Before | After | | --- | --- | | Admin exposed global and route thinking toggles. | Admin exposes **Off**, **From client**, **Low**, **Medium**, **High**, **X-High**, and **Max**; Fable, Opus, Sonnet, and Haiku also expose **Inherit**. | | Request intent was repeatedly reduced to a boolean across routing and providers. | The application boundary resolves one immutable `ReasoningPolicy` with independent control, named effort, and exact positive token budget. | | Provider adapters could infer reasoning behavior from upstream model names or versions. | Provider profiles translate only documented provider-wide wire capabilities; architecture and contributor rules prohibit model-specific reasoning branche…dc04eee