Get started

v2026.8.1: Models and Providers

Chat and the Models page now open from the catalog OpenClaw already has, so choosing a model no longer waits for a full provider scan. Live discovery runs only when you open a model screen or explicitly refresh it, keeps the last useful list when a lookup fails, and /model can change only the current conversation or deliberately update one agent or the shared default.

Once a model is selected, OpenClaw keeps the request inside the intended provider and authorized account order, preserves the real order and outcome of streamed replies and tool calls, carries reasoning and context settings with the model and runtime, and reports plan windows, token use, context pressure, and estimated cost more clearly.

Finding and choosing models

Chat and the Models page now start from the catalog OpenClaw already has, with supported providers looking for newer chat and text models when you open a picker or request a refresh. If that lookup fails, the built-in entries and last working list remain available.

A /model change can stay with the current conversation or deliberately apply to one agent or the shared default, with persistent changes requiring the right authority. Aliases and fallback keep the provider and account attached to the selected model.

The list shows models OpenClaw can identify, while the provider, account, region, endpoint, plan, limits, and pricing determine which ones you can use.

Sources and complete change list

Improvements

Migrate legacy Codex model routes into OpenAI 3c8269c
Show provider-configured models on Android #101912
Add Grok 4.5 to the bundled xAI provider #102316
Refresh Qwen, Cohere, and Mistral model catalogs #102489
Add current Cohere Command models #102563
Add bundled Meta Model API support for Muse Spark 1.1 #102873
Publish Meta under its canonical provider package #103070
Derive utility models from the primary provider #103769
Move active defaults and examples to GPT-5.6 Sol and Luna #104452
Reduce repeated agent turn setup latency #105220
Add official Baseten Model API support #108708
Add Kimi K3 support across Moonshot and Kimi Code #109202
Add provider-grouped models and per-session controls to macOS chat #110339
Discover current chat models from live provider catalogs #112412
Make provider catalogs authoritative for model compatibility #112542
Add end-to-end Claude Opus 5 support #113391
Complete Claude Opus 5 support across Anthropic routes #113392
Curate model pickers around current-generation choices #113594
Complete Claude Opus 5 support across Anthropic routes #113633
Refresh model catalogs without an OpenClaw release #113660
Refresh provider catalogs to current model lineups #113681
Start warm agent turns faster #113817
Add first-class Kimi K3 support #113909
Consolidate model settings on one Models page #115108
Follow xAI's OAuth default model automatically #115617
Clarify Model Setup and remove duplicate active routes #116086
Add session-only model selection and preserve fallback state #119325
Add Meta Muse Spark 1.2 models #120373
Add first-class Grok 4.6 catalog and reasoning support #122762
Add GLM 5.3 support for Z.AI Coding Plan #123523
Add first-class support for existing llama-server endpoints #125781
Unify managed and existing llama.cpp servers #126434
feat(models): add configurable model selection scopes #127813
feat(models): refresh provider models and live discovery #131294
Add first-class Alibaba Qwen Token Plan support #94419
Update the bundled Codex app-server to 0.144.1 #103156
Reduce warm Gateway model startup latency #105552
Refresh Z.AI, Kimi, Moonshot, and xAI catalog guidance #109666
Show bundled provider icons during model setup #112169
Speed up provider-filtered model listing #112752
Refresh OpenCode Zen models and pricing #113858
Add the Kimi K3 256K model option #113918
Keep code-mode tiers consistent across provider catalogs #115183
Remember the last-used model for future sessions #115717
Show full model names alongside aliases #115789
Clarify local model setup actions #116953
Reduce memory use for the default model list #117323
Clarify llama.cpp local model setup #117947
Speed up verified model setup #118352
Remove full model-catalog builds from turn startup #120834
Explain the Utility model setting in Models #123277
Reuse the WebUI model catalog across pages #124794
Avoid repeated plugin scans during model lookups #125090
perf(tui): prepare searchable picker rows once #127620
fix(ui): add brand icons for all provider auth choices #127668
perf(cli): lazy-load infer command domains #128150
feat(zai): add GLM-5.3 Flash model support #130309
Add Step 3.7 Flash to StepFun model catalogs #88082
Reuse prepared plugin metadata in static model catalogs #117288
Show model tables without waiting on promotions #117322
Avoid full Ollama plugin loading during model listing #117465
perf(models): reuse projected list rows #128248

Bug fixes

Refresh xAI models, reasoning, and server-tool defaults #103260
Align OpenAI model availability with its effective route #104685
Harden OpenAI Codex discovery, opt-out, and quota recovery #108683
Refresh bundled provider catalogs and remove retired models #109410
Separate model aliases from override allowlists #110888
Apply current request shaping to future Claude generations #114205
Unify dynamic model catalog selection #114284
Preserve selected fallback providers across alias resolution #114360
Keep model lists consistent across CLI and Gateway #114393
Keep agent model catalogs and switching reliable #114760
Improve gateway-node and local-model reliability #115185
Prevent multi-agent gateway startup stalls #116261
Stop cloud fallback after stale Gateway lifecycle aborts #117168
Keep large-roster chat startup off model catalog discovery #119742
Fall back when a model returns no usable answer #120148
Keep new-chat inference available across routes #120712
Unify agent failover error classification #121341
Keep model catalog discovery from stalling the Gateway #122350
Keep session model selection separate from response aliases and fallbacks #122910
Restore model lists when chat still works #123208
Keep Codex-routed turns working with request overrides #123258
Prevent multi-agent Gateway startup CPU spikes #124334
Let direct inference commands select an agent #125143
Honor per-agent model metadata everywhere #126194
fix(opencode): discover Zen and Go models on demand #129831
fix(models): keep native and alias-backed catalogs consistent #131677
fix: allow off-catalog chat models under unrestricted policy #132175
Keep collected follow-ups on their model routes 1647027
Keep model fallbacks working after cleanup takeovers #100118
Add Claude Fable 5 to the Claude CLI model catalog #101453
Expose normal model fallback transitions in diagnostics #102051
Apply hot-reloaded agent models without restarting the gateway #102305
Match Azure deployment map IDs across casing differences #103005
Keep replace-mode model listings limited to configured providers #103103
Restore current OpenCode Zen models in offline discovery #103197
Support managed ClawRouter gateway configuration #103299
Remove deprecated OpenCode Go MiMo aliases #103329
Preserve Grok behavior through the x-ai alias #103340
Allow the Codex runtime with codex/* models #103775
Surface utility-model narration failures in logs and model status #104770
Bound Amazon Bedrock model discovery requests #104841
Explain account-restricted model failures correctly #104878
Keep locked sessions on their stored model route #105335
Keep native thinking and fast controls on Codex routes #107588
Preserve newer model and auth choices during fallback recovery #108137
Bound Amazon Bedrock control-plane requests #108756
Bound stalled Z.AI endpoint probe responses #109026
Release Hugging Face connections after failed model discovery #109464
Preserve pinned models during temporary catalog failures #110364
Add OpenCode Zen GPT-5.6 model discovery #110405
Report unavailable Codex runtimes in model status #110447
Recover network error codes for model failover #110485
Report accurate OpenRouter limits in model scans #110855
Discover Vercel AI Gateway models through proxies #111209
Hide image-output models from the Kilocode chat catalog #111239
Discover Chutes models through eligible HTTP proxies #111266
Reuse model discovery across configless embedded runs #111423
Reject unknown text models before saving config #111571
Give cold model-runtime builds a longer startup budget #111983
Keep model catalogs coherent during config reloads #112331
Honor explicit subagent model requests #112393
Detect Claude CLI routes pinned to older models #113424
Hide live-discovered Claude models until request contracts support them #113757
Fix Anthropic live discovery for Claude subscription users #113906
Keep concurrent model status JSON clean #113953
Reduce reply startup delay with configured model policies #114117
fix(agents): retry live model switches outside fallback chain #114214
Show refreshed OpenAI models in provider-filtered lists #114265
Fall back on raw provider request errors #114457
Stabilize model metadata and config recovery under load #114679
Repair malformed generated model catalogs before startup #114697
Recognize local model endpoints and cap fallback output #114710
Keep Infer provider inspection and selection working #115054
Stabilize infer model selection, provider status, and shared-state access #115170
Show correct plugin-catalog metadata in configured model lists #115190
Canonicalize model references at their input owners #115965
Prevent large model configurations from delaying Gateway startup #116553
Show local provider identity after model setup #116655
Keep healthy fallback models available after a missing-model error #116768
Restore runtime-discovered models in browse views #116857
Reuse plugin metadata when building large model catalogs #117131
Keep alternate runtimes available after Codex preflight failures #117256
Use configured fallbacks after replay-safe silent failures #117356
fix(tui): honor restrictive agent model policies #117422
Restore Cerebras's documented static model catalog #117584
Hide retired NVIDIA models while preserving live discovery #117620
Prevent model discovery from hanging past its timeout #117825
Preserve model selection while the gateway is draining #118101
Preserve newer plugin metadata during models status #119183
Honor configured models for ACP agents #120046
Refresh OpenCode Zen and Go model catalogs #121030
Make exact dated model pins executable after Gateway startup #121119
Migrate retired OpenCode free-model references #121125
Show accurate defaults while a new session loads #121133
Deliver replies after CLI-to-embedded model fallback #121897
Show the resolved default model and its thinking levels #122907
Show resolved model identities in agents_list #123044
Keep scoped model checks from building the full catalog #123473
Stop Models page agent-roster retry storms #123483
Prevent truncated verified llama.cpp downloads #123738
Scope inference provider lists to a selected agent #123884
Let older llama.cpp providers load for upgrade repair #124041
Keep session status accurate through model fallback #124436
Preserve plugin completion agent selection on multi-agent installs #124954
Reject unknown providers when selecting default models #125558
Keep in-flight model requests on one plugin generation #125569
fix(providers): keep discovered image support for configured models #125670
Keep inherited parent model pins ahead of channel defaults #125880
Reject unusable explicit subagent model requests #125931
Preserve catalog capabilities for partial provider model configuration #126068
Scope local inference preparation to the selected model #126196
Let multi-agent infer commands select their owning agent #126234
Preserve explicit Codex runtime policy for embedded attempts #126259
fix(llm-task): trust explicit provider for slash-containing model ids #126518
Preserve Codex selection when model metadata comes from the catalog #126531
refactor(providers): return prepared dynamic models directly #126574
fix(codex): run newly available models through the existing Codex account #127322
Let Codex use newly available account models before catalog refresh #127394
fix(models): scope status alias resolution to the selected agent #127631
fix(models): stop /models from freezing the gateway while it rescans plugins #127847
Skip inherited automatic fallback overrides in child sessions #128521
perf(gateway): stop first models.authStatus call from reloading all plugins from source #128551
fix(workers): resolve catalog models on dispatched turns #128553
fix: /models stops working after an unrelated config edit and never recovers #128608
fix(tui): restore session setting reset directives #128731
fix(codex): keep /codex model scoped to the native binding #128751
fix(vercel): exclude non-language gateway models from chat catalog #129406
fix: agent-specific model selection repeatedly rescans every plugin #129510
fix(huggingface): honor available provider capabilities #129797
fix(cli): preserve human output for flag-like model option values #129911
fix(models): model browsing fails after automatic plugin activation #130481
fix(agents): preserve Codex for reasoning capability metadata #130752
fix(models): keep OpenCode fallback aligned with catalog lifecycle #130785
fix(models): keep model listings consistent after manifest changes #131518
fix(config): accept supported model pins outside curated catalogs #131562
fix(codex): keep native overrides effective at launch and turn start #131745
fix(agents): preserve native runtime auth across catalog refresh #131931
fix(agents): retain prepared catalog ownership through refresh #132244
fix(anthropic): defer Claude CLI capability probe until execution #132324
fix(fireworks): prevent unsupported images reaching GLM 5.2 Fast #132362
perf(plugins): consolidate provider discovery planning #133097
Recognize code-only provider errors for model failover #95729
Fix Codex ACP spawns with non-OpenAI fleet defaults #95852
Keep generated multimodal models in the registry #97858
Refresh NVIDIA featured models without breaking older configs #99242
Align Codex model discovery with client version 0.144.5 cfaa30a
Keep truncated fallback reasons valid Unicode #102631
Keep model tables aligned with wide Unicode names #102819
Align Meta validation and guidance with Muse Spark 1.1 #103163
Preserve Unicode in bounded OpenRouter Fusion model IDs #104433
Preserve Codex mock routing with model-specific image endpoints #105444
Repair legacy Codex model references in per-agent model maps #108369
Allow clearing agent model overrides #108734
Honor OpenRouter scan timeout while reading the catalog #108972
Release stalled GitHub Copilot catalog error streams #109071
fix(plugin-sdk): contextualize malformed provider catalogs #109754
Handle malformed provider catalog pagination URLs safely #109986
Make status reflect pending and completed live model switches #110474
Discover Hugging Face models through required HTTP proxies #110924
Sync OpenAI Codex OAuth catalog client version #110936
Reject malformed UTF-8 during self-hosted model discovery #111113
Nested model-policy wildcards respect namespace boundaries #111350
Let DeepInfra model discovery use trusted HTTP proxies #111428
Reject corrupted UTF-8 in live provider catalogs #111766
Keep new chats working through concurrent config reloads #112026
Report model switches before automatic thinking remaps #112352
Preserve local-provider errors during cron preflight #113409
Align Codex OAuth model discovery with the managed client version #113615
Prevent Meta catalog loading from rejecting unsupported inputs #113758
Keep empty model lists machine-readable #113897
Accept harmless separator padding in model policy references #114116
Make refreshed model catalogs usable from the CLI #114244
Recognize all IPv4 loopback model endpoints #114766
Reject incomplete model overrides in infer commands #115225
Align configured model rows with catalog availability #115239
Preserve valid LM Studio models in mixed catalogs #115336
fix(status): recent session listing resolves the wrong agent's model #115984
Reject malformed hosted model catalogs before saving #117744
Preserve explicit provider and model references in alias guidance #117869
Report accurate embedded-agent startup timing stages #117879
fix(config): ModelCompatSchema drops eight keys ModelCompatConfig declares, and the drift contract cannot see it #118674
Apply the global fast-mode default to the implicit agent #120697
fix(acpx): strip amazon-bedrock provider prefix from Claude ACP model refs #121105
Preserve concrete provider failure reasons through failover #121657
Propagate provider catalog failures instead of mistaking them for timeouts #124288
Preserve complete model fallback traces #126048
Stop stale provider tests after switching agents #126750
fix(models): --agent is silently ignored by models aliases and models scan #126864
fix(cli): return JSON for model list failures #127726
fix(agents): resolve pdf tool models through the canonical resolver #127952
fix(agents): warn when configured image model entries fail to resolve #127971
fix(cli): resolve alias targets from the write snapshot #128223
fix(cli): render models refresh JSON failures #128981
fix(cli): keep models --plain stdout clean of startup diagnostics #129037
fix(models): normalize mixed aliases consistently #129678
fix(agents): prevent selected-agent alias stalls #129700
fix(reply): keep fast abort off provider runtime #129704
fix(tui): dismiss 'loading models...' notice once the model picker opens #129770
fix(models): avoid retired automatic auth-probe models #129798
fix(cli): preserve populated model output on stdout in plain mode #129871
fix(ui): reset /model default through server directive #129895
refactor(agents): resolve runtime models through prepared owner #130007
fix(bedrock): reject invalid UTF-8 model discovery #130058
fix(agents): read the published model catalog owner from session_status #130277
refactor(models): preserve catalog metadata through one normalization path #130381
fix(cli): reject --agent for global model refresh #130448
fix(cli): reject empty agent scopes for global model commands #130739
fix: model picker refresh discovers configured agent catalog #131151
fix: cold inventories incorrectly quarantine provider-supported tools #131601
fix(config): preserve authored model rows during sparse patches #132080
fix(codex): explicitly selected hidden models fail bounded turns #132189
fix(sessions): honor per-agent model policies and aliases #132205

Documentation

Clarify OpenAI and Codex Fast mode precedence #120682
docs: correct numeric model selection guidance #132246
Document rolling Anthropic Opus aliases #113413
Repair OpenRouter copy-paste configuration examples #118747
Clarify model switching and cache trade-offs #124889
docs: correct chat model selection examples #132218

Provider accounts and sign-in

OpenClaw now keeps model requests inside the provider accounts and credential order you configured. If one account hits an authentication or quota cooldown, the next authorized account for that provider can take over without changing the selected provider or model, and the saved preference resumes when it recovers. Environment keys remain available when no explicit account list is configured.

OAuth registration stays with the setup conversation where it began, failed credential writes surface as failures, and a successful login applies only to the model route that authenticated. Tenant and custom-endpoint credentials stay on the account and origin they were configured for.

New GitHub Copilot device logins place the token in OpenClaw's protected local secret store by default and keep a reference in the auth profile. The store depends on state-directory permissions rather than encryption at rest, existing inline profiles are not migrated, and operators can still choose the prior plaintext mode explicitly.

Sources and complete change list

Improvements

Support loopback Codex proxies in the OpenAI provider #114555
Add GitHub Enterprise data-residency Copilot support #99221
Remove individual model auth profiles from the CLI #114407
Reuse SQLite readers for faster auth-profile loading #121358
Clarify auth profile order output #112203
Refresh macOS local-model and provider sign-in translations #113493
Retire legacy provider IDs and secret markers at runtime #115655

Bug fixes

Avoid false 12-hour OpenAI cooldowns after token expiry #102479
Restore native xAI routing for the x-ai provider alias #103488
Remove raw synthetic credentials from model status JSON #104734
Block unsupported GitHub Copilot OAuth domains #105584
Restore GitHub Copilot turns and account-scoped context limits #106198
Restore service-account authentication for Anthropic Vertex #108350
Keep OAuth providers isolated to their owning session #108680
Surface credential saves and harden agent sessions and tools #109622
Accept fine-grained GitHub tokens for Copilot #114282
Honor provider auth pins for ambient credentials #114944
Skip provider-auth scans on unrelated reloads #117449
Preserve compatible auth profiles when switching models #117550
Prevent undeclared environment credentials from taking over failover #118458
Keep OpenRouter custom-proxy credentials on the configured origin #118773
Stop retrying expired CLI auth profiles after fallback #119843
Refresh chat model availability after runtime auth succeeds #121193
Scope quota failures to individual auth profiles #121278
Make Doctor converge on missing OAuth sidecars #123093
Honor configured GitHub Copilot identity across model requests #127965
fix(github-copilot): device login no longer leaves plaintext tokens #133033
Refresh model availability after OAuth renewal #102289
Redact GitHub Copilot OAuth failure details #102953
Preserve OAuth diagnostics when provider messages contain braces #103106
Show actionable re-auth guidance for logged-out Claude CLI #103829
Keep Codex account identity when channel login switches profiles #104706
Keep unresolved API-key providers visible #106754
Stop false Claude CLI expiry warnings #108205
Redact credentials from GitHub Copilot embedding errors #109177
Ignore blank Google Vertex project and location values #109515
Improve provider login, Bedrock errors, and TUI rendering #109740
Keep Telegram responsive during Codex device login #110580
Ignore blank AWS region overrides for Bedrock #110676
Use canonical auth-profile ordering for background credential discovery #110678
Release stalled GitHub Copilot OAuth error responses promptly #112268
Use selected Claude profiles for CLI-backed agent turns #112458
Prevent Gateway crashes after Claude CLI auth-profile turns #112550
Retry transient OpenAI device-code polling failures #114094
Refuse mismatched Codex native-home billing routes #114719
Continue model fallback after Google invalid-key errors #114887
Keep Claude CLI logins refreshable instead of forwarding stale tokens #115134
Stop model fallback on rejected SSH sandbox secrets #115198
Isolate coding-agent Codex authentication #115611
Restore Copilot stored-profile authentication in external harnesses #116555
Guide recovery when agent-scoped Codex credentials were not imported #116807
Keep model-provider changes scoped and visible #116829
Recover from unavailable automatic auth-profile overrides #117846
Honor Google Cloud SDK credentials and Vertex billing projects #118745
Recognize configured credentials for media providers #118761
Keep transient timeouts from disabling inline API keys #119147
Preserve inherited auth profile order for secondary agents #119505
fix(ollama): redact reflected credentials from transport errors #119537
Apply auth-order changes without restarting Gateway #119946
Keep models available after Gateway restart #120977
Preserve rate-limit retries when no fallback model is configured #121294
Count only successful rate-limit profile rotations #121325
fix: inherited auth-profile cooldown writes target the child store and silently no-op #122604
Recover OpenAI sessions with a backup auth profile #123791
Keep auth-profile rotation on the current provider route #123894
fix(auth): reset only the profile replaced by a successful login #124014
Deprioritize OAuth profiles after refresh failures #125507
Restore explicit Bash environment imports without changing exec PATH probes #125624
Report the auth database that actually holds an agent's profiles #126918
Show the persisted auth database path #126955
fix(agents): surface provider auth guidance for terminal failures #128375
fix(auth): recover runtime profiles after publication failure #129234
fix(auth): inherit the shared auth store once it owns credentials #130264
fix(tui): reload auth store ownership after login #130349
fix(agents): publish SecretRef Codex API-key route success #130359
fix(models): verify native CLI auth in detailed lists #130749
fix(github-copilot): restore usage for domain-aware OAuth profiles #131137
Report Claude OAuth expiry with recovery steps in channel chats #131345
fix(gateway/ui): record model auth unavailability reasons; gate composers only on actionable auth failures #131697
fix(auth): keep working OpenAI models available after catalog refresh #131803
fix(ui): chat stays blocked after model credentials recover #132179
fix(process): make credential descriptors reopenable #132573
fix(agents): separate auth profiles from runtime model ids #133175
Cool down inline provider API keys after billing failures #88709
Keep Azure Responses models on Azure credentials #93833
Repair auth status and stale credential orders #98105
Materialize auth-profile secret refs for standalone local agents d4ae2bb
Bound Chutes OAuth requests with a 30-second deadline #102026
Bound Google Vertex credential refreshes to 30 seconds #102050
Keep Microsoft Foundry Azure CLI errors readable #102604
Keep Microsoft Foundry connection errors Unicode-safe #102605
Ignore malformed Microsoft Foundry endpoints #104796
Show re-auth hints for inline OAuth errors #105085
Reject malformed fallback-skip TTL values #107234
Bound stalled OpenAI device-code login requests #109494
Preserve Unicode in Microsoft Foundry Azure CLI login output #109499
fix(microsoft-foundry): bound az subprocess lifetime on exec timeout #110434
Keep OAuth refresh failures from using substitute credentials #112957
Continue model fallback after Google invalid-key errors #114825
fix(agents): separate retry count from API key position (#102938) #115507
Make Telegram Codex login codes tap-to-copy #116608
Use official Anthropic Vertex multi-region endpoints #116757
Stop no-op OAuth migration prompts and show failure causes #123164
Preserve Codex QA login identity across restarts #126777
Point missing API-key errors to the auth command #127011
fix(agents): stored auth profile order is ignored when resolving CLI runtime aliases #129165
fix(ollama): protect local credentials on all loopback addresses #129168
fix(auth): reject retired profiles from cold external providers #129580
fix(ui): clear stale provider API-key drafts when switching agents #129676
fix(github-copilot): configured account order is ignored #129978
fix(ai): require explicit endpoints for compatible providers #130453
fix(auth): preserve credential-source fact through model attempts #131220
fix(auth): keep SecretRef models available with explicit auth order #131255
Deduplicate Bedrock Mantle IAM failure diagnostics per region #81208

Documentation

docs(start): correct headless credential handoff to the SQLite auth store #131024
Add PowerShell commands to the minimal AI chat example #108567
Fix the Chutes API-key setup link #113545

Streaming Replies and Tool Calls

OpenClaw now distinguishes a complete streamed reply from one that stopped partway through. On supported Responses paths, reasoning, text, tool output, and available usage remain in order; if a stream fails, valid work that already arrived is preserved, and OpenClaw retries or moves to an authorized fallback only when replay is safe. With no safe recovery path, the partial reply remains attached to the error.

Tool calls wait for a complete name and arguments before they can run, including large streamed arguments and calls whose provider item IDs change along the way. Unmanaged native OpenAI Responses conversations on the official endpoint with storage enabled can continue by sending only new input after the first turn, then recover with full history if that upstream continuation state expires.

Sources and complete change list

Improvements

Add structured Codex questions and native goal controls #109724
Relay Claude CLI native tool requests through Gateway approvals #112918
Faster OpenAI agent turns through reusable WebSockets #121687
Continue stateful native OpenAI SSE turns #122194
feat(codex): upgrade main to app-server 0.149.1 #128370
Enable capability-gated tools on CLI-backed models #102835
Prewarm Codex app server during turn setup #114551
Reduce Codex coding-runtime and parity overhead #114574
Stream native Ollama tool-call lifecycle events #116809
perf(copilot): avoid duplicating BYOK request bodies #131145
Update the managed Codex app-server to 0.146.1 b269e65
Update the bundled Codex CLI to 0.144.4 #107208

Bug fixes

Accept numeric and boolean enums in Gemini tool schemas #104567
Harden Codex aborts, dynamic-tool schemas, and result middleware #108105
Improve Codex full-auto execution, failure handling, and quota status #108626
Preserve tool arguments when Responses item IDs rotate #108630
Preserve tool-result text when media has no inline payload #108697
Restore ClawRouter Perplexity tools and model refreshes #108758
Keep tool-call preambles out of Chat Completions replies #109057
Move ACPX Codex sessions to the maintained adapter #109075
Harden OpenAI-compatible provider reliability and errors #109556
Harden Responses streaming across OpenAI, Azure, and ChatGPT #109615
Prevent repeated tool-call IDs from poisoning sessions #110518
Enforce bounded CLI input and provider response deadlines #110627
Unify OpenAI Responses stream processing #114263
Restore Codex patch execution and fast-mode priority #114907
Stop paired local inference after caller cancellation #115624
Preserve complete streamed Chat Completions responses #115652
Keep stressed Codex turns isolated and visible #115893
Make Codex and OpenAI turns finish truthfully #116012
Preserve complete terminal output in Responses streams #116836
Repair llama.cpp tool, reasoning, and context lifecycles #116903
Recover authoritative OpenAI Responses output #116910
Restore reliable Gemini and Vertex streaming and usage accounting #117042
Preserve fenced tool-call examples in replies #117089
Preserve structured OpenAI Chat responses #117136
Preserve Gemini signatures and provider failures #117193
Restore OAuth-backed tools for CLI agents #119166
Preserve the final streamed text when a run fails #119556
Flush trailing chat text after quiet streaming gaps #119566
Restore Venice Gemini tool-call continuations #119783
Preserve results from quiet long-running Codex tools #119835
Surface Claude CLI no-response failures #121589
Deliver partial replies when models reach their output limit #123546
Reuse Codex app servers after managed desktop fallback #124885
Reap Codex app-server descendants during retirement #126285
Surface Codex input prompts across native and ACP runs #126387
Reject malformed streamed tool calls #126391
fix(agents): retry incomplete model streams with partial output #127338
fix(gateway): fail streaming responses when agent runs fail #127662
fix(gateway): restore HTTP subagent completion announces [AI-assisted] #128068
perf(ai): keep streaming responsive while large tool call arguments assemble #128166
fix(agents): apply tool policy to Anthropic native calls #128805
fix(agents): quiet model turns are aborted by a false "no response from model" idle timeout (#117884) 0707fcb
fix(openai): reconcile rotated terminal message IDs (#122560) 5acfc2d
fix(openai-responses): reconcile terminal tool calls (#108461) 8c1a55d
Retry transient network failures consistently #101496
Show OpenAI refusal text instead of blank replies #102344
Restore image tool results for prefixed Gemini 2 models #102382
Update the Codex harness for app-server 0.143.0 #102444
Recognize zero-argument XML tool calls #103220
Preserve streamed answer order after tool-call repair #103585
Align the Meta provider with the official Model API #103680
Stop canceled local-provider startups from leaking sidecars #104070
Bound and unify serialized tool-call repair #104100
Preserve Unicode across streamed block-chunk boundaries #104441
Fix GPT-5.6 tool calls and catalog compatibility #104555
Preserve explicit abort outcomes for Codex turns with unfinished tools #104955
Clean up abort listeners after ChatGPT retry waits #105519
Preserve newer ChatGPT Responses WebSocket connections #105634
fix(deepinfra): apply request policy to video generation requests #105887
Update the managed Codex app-server to 0.144.3 #106098
Retry transient provider-wrapped 5xx errors #106850
Report empty proxy responses clearly #107325
Recover large Codex app-server messages #107369
Preserve Codex assistant message ordering during image saves #107680
Terminate Anthropic streams on circular provider errors #107800
Correct Codex tool-result failure classification #107807
Split embedded runner orchestration and harden timeout recovery #107900
Return visible text from constrained Z.AI GLM completions #108612
Recover Codex tool-terminal turns and align auth routing #108966
fix(agents): idle-timeout stalled provider response body reads #109067
Prevent broken Claude CLI output pipes from crashing node hosts #109794
Stop cancelled replies from retrying after transient provider errors #109805
Restore local GGUF asset resolution #110233
Reject invalid HTTP status prefixes in assistant errors #110554
Keep repeated Responses tool calls distinct #110956
Release guarded response stream reader locks #111259
ChatGPT Responses honor the server's retry delay fallback #111353
Fail fast on invalid provider TLS certificates #111818
Bound Ollama requests during stalled DNS preflight #111835
Suppress partial assistant output after agent errors or aborts #113012
Keep streamed OpenAI-compatible requests running after handler return #113514
Return 404 for disabled OpenAI-compatible API routes #113609
Close failed Responses streams without false completion #113629
Preserve OpenAI Responses cancellations #113801
Repair provider-polluted tool names safely #113814
Preserve Qwen Token Plan behavior for direct model references #113976
Unify OpenAI-compatible request behavior #114236
Honor per-turn OpenAI timeouts and retry limits #115017
Enable automatic code mode for Kimi Code K3 #115022
Preserve provider timeouts across plugin model rebuilds #115102
Preserve partial output from incomplete OpenAI Responses streams #115132
Show Codex startup diagnostics on initialize timeout #115161
Report embedded agent deadlines as timeouts #115319
Enforce one timeout for paired-node Ollama inference #115558
Accept the OpenAI SDK plain-text Responses format #115594
Restore tool calling for llama.cpp-compatible providers #115598
Keep progressing remote Ollama streams alive #115648
Strip unsupported keywords from every nested tool schema #115741
Keep OpenAI Responses replacement streams consistent #115799
fix(gateway): preserve OpenAI-compatible response contracts #116354
Let one-shot Claude CLI commands exit cleanly #116577
fix(runtime): prevent workspace, provider, and streaming lifecycle leaks #116650
Execute llama.cpp plaintext tool calls #116736
Honor Ollama stop sequences and asynchronous payload hooks #116737
Reject truncated Bedrock streams and preserve audio results #116743
Surface blocked Gemini prompts with usage details #116756
Preserve Google stream errors, usage, and tool signatures #116822
Preserve OpenAI Chat tool-call lifecycles #116922
fix(google): preserve stream terminals and provider tool-call identities #116926
Show clean retry guidance for malformed Codex streams #116966
Preserve fenced tool-marker examples in replies #116971
Repair Bedrock stream completion and OpenRouter model handling #116998
Enforce streamed OpenAI tool-argument limits consistently #117055
Clear sticky ChatGPT SSE fallback when sessions end #117123
Report provider HTTP status in diagnostic timelines #117403
Close unread Ollama error responses cleanly #117473
Preserve OpenAI streaming terminal delivery during Gateway shutdown #117505
Normalize Claude tool IDs for GitHub Copilot #117789
Preserve WebSocket transport for Codex OAuth sessions #117832
fix(agents): close proxy SSE bodies on terminal and errors #117887
Keep explicit llama.cpp HTTP routes on their configured transport #118293
Return the correct final agent failure without extra delay #118331
Report failed OpenAI-compatible chat streams as API errors #118428
Recover Gemini streams stalled before the first response body #118511
Keep llama.cpp auxiliary completions on native transport #119082
Let Codex-backed agents react in current conversations #119649
Detect truncated direct Anthropic streams without breaking proxy responses #120030
fix(ollama): reject invalid UTF-8 in streaming NDJSON responses #120240
Fail malformed LLM streams instead of hanging turns #120508
Support Codex app-server 0.147.0 #120594
fix: agent.wait can return error before the same run later ends successfully #120606
Honor explicit OpenAI service tiers over Fast mode #120696
Reject leaked Claude tool protocol output #120734
fix(agents): replace stale timeout payloads precisely #122036
Prevent replay after observed CLI timeout activity #122153
Speed up streamed assistant output #122206
Hide HTML error pages that follow HTTP reason phrases #122290
Restore cached OpenAI WebSocket continuation #122483
fix(tool-call-repair): stop re-parsing the whole response per candidate delta #122513
fix(ai): preserve structured error code through Responses failed-event normalization #122579
Honor the selected transport for embedded OpenAI turns #122946
Finish agent responses after post-tool provider overloads #123005
Retry structured transient network failures #123042
Preserve OpenAI terminal errors and avoid pointless failover #123151
Honor provider timeouts during stuck-session recovery #123877
Return failures for failed OpenAI-compatible HTTP runs #124039
Use the bundled Codex app-server for managed turns #124124
fix(agents): continue uncached when Google prompt-cache responses fail #124479
Preserve OpenAI WebSocket failure details and retry behavior #124591
Keep valid Codex tools when one tool name is unsupported #124932
Prevent duplicate audit starts for Codex dynamic tools #124976
fix(cli-runner): drop stock watchdog defaults that disable resume promotion (#125045) #125085
Deliver only the final answer after resumed reasoning #125149
Propagate local-model stream policy to diagnostic recovery (#125147) #125388
Report Ollama HTTP responses to diagnostics #125834
Stop Ollama response handling after cancellation #125961
Record provider request acceptance consistently #126028
fix(agents): stop fallback tasks hanging after primary model failures #126566
refactor(process): supervise Claude node invocations #127475
fix(codex): accept bounded upstream prompt provenance #127730
fix(mistral): retain provider-returned streaming model as responseModel #128066
fix(ai): preserve tool-result boundary whitespace for Mistral and Ollama #128266
fix(ai): keep SSE cancellation from blocking outcomes #128488
fix(agents): reject hidden output across model verification and system turns #128570
fix(cli): normalize echoed binary tool outputs #128699
fix(anthropic): show Claude questions as interactive prompts #128729
fix(agents): stop double-sanitizing openai-responses tool-call ids #128764
fix(deepseek): doubled-bar DSML tool calls are delivered as text and never executed #128882
fix: prevent custom Ollama cloud models from stalling indefinitely #129051
fix(mistral): report provider failures instead of silent success #129256
fix(ai): enforce Anthropic stream completion by provider ownership #129291
fix(codex): back off between app-server startup retries #129505
fix(gateway): report failed chat streams as errors #129787
fix(ai): report interrupted local-model streams instead of empty success #129794
fix(inbound-meta): avoid Anthropic billing classification #130298
fix(ai): replace streamed function name on same-id continuation #130537
fix: release stalled captured provider responses promptly #130804
fix(copilot): keep Gemini working without live discovery #131143
fix(copilot): own BYOK response stream lifecycles #131546
fix(agents): show Claude CLI replies while they stream #132002
fix: prevent first Grok tool requests from stalling #132475
fix: prevent cold image analysis from timing out during setup #133045
fix: prevent duplicate commentary after tool handoffs #133247
fix(gateway): report timed-out HTTP agent runs as failures #133275
Bound shared Codex app-server startup waits #89442
Stop fallback models from re-running completed agent work #92011
Make Claude CLI max-turn failures actionable and safe to retry #94130
Retry transient provider failures after earlier tool use #97966
Preserve complete streamed tool calls without finish_reason #98124
Warn without interrupting Codex projection when protocols drift #99287
Drop tool-result images for text-only Anthropic models #102329
Keep truncated provider errors valid Unicode #102496
Preserve emoji when Codex tool output is truncated #102522
Keep long Mistral errors Unicode-safe #102539
Keep Codex supervisor error messages valid around emoji #102590
Keep Ollama malformed-stream warnings Unicode-safe #102593
Keep truncated Codex dynamic-tool output valid Unicode #102622
Preserve complete characters in truncated Codex tool output #104116
Honor explicitly configured OpenAI Completions transport #105539
Sanitize ClawRouter attribution headers safely #106454
Preserve emoji boundaries in diagnostic timeline attributes #106986
Use the Anthropic SDK's SSE parser #107460
Bound oversized Ollama stream records #107473
Preserve Unicode in Codex app-server failure messages #108335
Restore llama.cpp compatibility for the cron tool schema #108360
Preserve Unicode in local provider startup errors #108518
Preserve Unicode in Codex sandbox HTTP output #108618
Report invalid UTF-8 in provider JSON accurately #108849
Preserve Unicode in Codex app-server error tails #109013
Report provider finish errors accurately #109313
Keep Unicode intact in truncated provider errors #109786
Prevent stream cleanup errors from overriding cancellation #110427
Keep OpenAI tool-call ID truncation Unicode-safe #110471
Stop retrying non-retryable ChatGPT Responses errors #110655
Reject non-text Codex WebSocket response frames #111138
Release Ollama stream locks after early exits #111641
Show provider 503 failures in group and channel chats #113656
Preserve Gemini compatibility flags on discovered model variants #114831
Preserve emoji in streamed tool arguments #115536
Separate streamed Chat Completions content from the finish event #115718
Preserve completed Codex answers when tool bookkeeping is incomplete #115849
Prevent native Anthropic attempts from carrying refusal markers #116433
Release Amazon Bedrock streaming clients #117039
Restore /tools inventory for dynamically prepared models #119306
Keep ACP retry statuses scoped to one logical dispatch #120560
Accept TLS-certificate cloud-worker fallback events #121285
Preserve provider local-service environment overrides on Windows #122027
Unify CLI failover policy across execution lanes #122164
Preserve reply cleanup when backend cancellation throws #122252
Retry transient Undici socket failures #122535
Recover rejected OpenAI WebSocket continuations #122727
Accept headerless text files in Open Responses URLs #123431
Close unhealthy ACP runtimes before replacement #126494
fix(cli): report exact Claude stream failures #127169
fix(gateway): Responses API returns blank or truncated assistant output #128295
fix(ai): record provider-returned model in managed openai completions stream #128324
fix(gateway): streamed tool calls silently omit assistant commentary #128353
fix(gateway): retain completed status in streamed tool snapshots #128510
fix(cli): convert HEIC sequences in model runs #128578
chore(deps): refresh dependencies after seven-day cooldown #129187
fix(codex): suppress under-development features warning in chat #129493
fix(google): retain provider-returned response model #129606
fix(ai): local-model tool calls fail in standalone library streams #129664
fix(llama-cpp): restore tools for fallback-capable chat templates #129684
fix: proxied model streams drop all events when the server omits the space after data: #129905
fix(codex): keep unsupported service-tier warnings out of chat #131190
fix(codex): hide managed Code Mode compatibility warning #131943
fix(agents): neutralize Anthropic refusal marker replacement #132556
fix(gateway): send finalized text with required tool calls #133206

Documentation

Document llama.cpp schema compatibility for custom endpoints #117120

Reasoning and context limits

Reasoning settings now stay attached to the selected model and runtime when a conversation is restored or its authentication route is rebuilt. Native Codex supports Ultra for Sol and Terra and Max for Luna; embedded OpenClaw maps Ultra to the provider's highest supported effort and adds guidance for delegated work, which is a different behavior from native Codex Ultra.

Supported GPT and Claude routes expose larger context options. Normal GPT-5.5 and GPT-5.6 runs use a 272,000-token budget with an opt-in 922,000-token input window, and the Control UI can choose 200K or 1M for supported Claude 5 CLI conversations. These options remain limited to the routes, interfaces, and accounts that support them.

Compaction now uses the latest trustworthy context count rather than accumulated cache billing or duplicated history, preserves tool output across supported Responses checkpoints, and falls back to full history when a stored checkpoint is rejected. Context limits are now configured per model, and Doctor can migrate supported provider-level settings tied to explicit model entries.

Sources and complete change list

Improvements

Keep sent prompt history stable within a session #102610
Reduce recurring prompt and tool-instruction overhead #105095
Make million-token OpenAI context an explicit opt-in #112916
Consolidate model context budgets #124665
feat(agents): selectable Claude CLI context window (200K/1M) #127951
Add GPT-5.6 Ultra across OpenClaw and Codex runtimes #98021
Update the managed Codex app-server to 0.144.6 #110821
Reduce repeated per-turn conversation metadata tokens #113616
Stabilize Codex tool prompts and expose cache losses #115238
Add Gemini 3.7 Flash with compatible thinking levels #123366
Add session thinking level to the bundled Swift protocol client 0d215c1
Show the effective thinking level in /model summaries #111709
Preserve prompt-cache reuse in history-limited sessions #115281
Reduce duplicate ambient-room prompt content #119860

Bug fixes

fix(agents): honor account-scoped history limits instead of silently ignoring them (#119326) 3b3372b
Prevent stale Apple chat thinking after model switches and reconnects #103767
Serialize chat model and reasoning changes #106534
Stop cumulative cache usage inflating session context #108323
Keep ACP sessions from failing on unsupported automatic thinking controls #108392
Correct Claude 1M limits for direct Anthropic sessions #108751
Prevent premature compaction in tool-heavy sessions #110297
Preserve context for policy-routed CLI sessions #115284
Enable custom providers to expose Max and Ultra reasoning #115690
Preserve Claude CLI conversation context across prompt updates #116292
Preserve Ollama thinking through reply maintenance #116963
Keep runtime context focused on the active user request #117364
Honor per-agent context limits in embedded runs #120343
Let Claude CLI sessions compact with subscription authentication #120496
Prevent premature compaction from inflated token counts #120497
Preserve full Responses history across compaction retries #120729
Align server and local compaction decisions #123397
Prevent premature compaction from double-counted context #124267
Let tool-heavy sessions survive runtime cutover #125181
Apply Claude CLI thinking and preserve warm sessions #125528
Preserve GPT-5.6 Max and Ultra selections through Codex #126492
Use real Ollama Cloud context windows and capabilities #126653
fix(acp): preserve configured thinking across session reuse #128169
fix: Responses sessions compact before reaching context limit #130993
fix: preserve ACP model switches and reply completion #132040
fix(compaction): preserve headroom for local chat follow-up turns #132190
Cap oversized Codex tool output before replay #99578
Default GPT-5.6 Sol to medium reasoning e9c6c15
Use GitHub Copilot’s actual prompt limits #103275
Let Claude CLI choose adaptive thinking effort #103815
Show model-specific thinking levels in Android chat #103893
Apply provider-level token limits to configured models #105807
Persist iOS chat thinking selections #106287
Preserve Fable 5 effort settings on Claude CLI routes #107636
fix(github-copilot): preserve catalog thinking efforts in requests #107834
Use Z.AI's documented thinking payload #108633
Honor thinking off for Kimi Code K3 #109335
Complete Claude transcript pages across short file reads #109431
Persist Codex automatic compaction counts #110587
Default missing Anthropic output-token limits #110617
Let Apple Watch chat inherit configured thinking defaults #111301
Honor provider-specific reasoning effort mappings #112632
Restore full verbosity and streamed reasoning controls in the TUI #113782
Read Anthropic token limits during live model discovery #113784
ACP turns continue when optional thinking values are rejected #113935
Keep Copilot tool transcripts complete after interruptions #114403
Prevent false idle timeouts during hidden reasoning #114997
Make auto-compaction cut points free real context #115309
Recover channel conversations from expired Claude CLI sessions #115513
Preserve native GPT limits for Microsoft Foundry deployments #115542
Use self-hosted providers' advertised context windows #115566
Restore tagged Claude CLI reasoning delivery #117264
Honor stored model overrides and cached tokens in /compact #117580
fix(bedrock): requests respect model output limit when reasoning is off #119155
Stop bodyless 400 errors from triggering false compaction #119596
fix(agent-core): preserve explicit reasoning off through Agent Core handoff #120020
Honor Claw Router model reasoning effort metadata #120631
Preserve long Responses sessions across cloud-worker handoff #120803
Enable native max thinking for supported Ollama Cloud models #121074
Preserve GPT-5 personality settings during doctor migration #121346
Preserve Claude CLI context across heartbeats #121509
Restore all GPT-5.6 Sol thinking levels #121865
Resolve configured model aliases during compaction #122422
Refresh Codex context totals after native compaction #123478
Stabilize prompt caching across discovery order #123543
Honor per-model context caps on Codex-routed runs #124735
Restore GPT-5.6 reasoning effort options in Codex #126182
fix(agents): prevent persona drift with generic slot IDs #128559
fix(agents): native runs ignore the selected 200K context window #128838
fix(models): preserve provider-defined reasoning levels across model catalogs #129623
fix(agents): keep active hidden model streams out of stalled recovery #130140
fix(ai): stabilize runtime context across Responses tool rounds #130166
fix(ai): carry the Responses system prompt via instructions, not input[0] #130206
fix(failover): oversized Groq requests stall the session with unusable rate-limit advice #130275
fix(copilot): prevent static GPT-5.4 mini thinking failures #131160
fix(anthropic): reload compacted Claude sessions #131288
fix(agents): show token limits for rejected model requests #131990
fix: prompt cache collapses in long sessions — aggregate tool-result truncation rewrites already-sent history #132017
fix: keep CJK compaction summaries within budget #132145
fix(codex): honor thinking off on supported Platform routes #132276
fix: keep session thinking controls aligned with their model owner #132389
fix: preserve model thinking policy and repair live validation #132431
fix(google): preserve Gemini native compaction ownership #132527
fix: keep summary checks after custom compactor timeouts #132536
fix: preserve compact focus in split-turn summaries #132564
fix(compaction): recover split-turn summaries without duplicate sections #132618
Honor configured context limits for Claude CLI sessions #93198
Enable maximum thinking for OpenCode Go DeepSeek V4 #99643
Prevent silent reasoning turns from bricking Anthropic sessions #99772
Keep Gemini thinking disabled when reasoning clamps to off #101832
Preserve session thinking levels during unrelated updates #102866
Match canonical reasoning efforts regardless of casing #102993
Keep Claude fallback summaries valid at emoji boundaries #104454
Honor Z.AI GLM-5.2 max thinking through the gateway #104778
Enforce safe output limits for known Mistral models #105040
fix(extra-params): preserve resolved cacheRetention against undefined own-property clobber #106069
Show the active model and thinking level in the TUI #108669
Preserve thinking levels in status for discovered Ollama models #108789
Keep Claude CLI Mythos 5 context at its fallback limit #110461
Recover from Z.AI token-limit errors #111744
TUI footer reflects the launch thinking override #112237
Preserve Ollama cloud model context windows #112430
Let native iOS Talk inherit the session thinking level #112901
Normalize standalone AI package transcripts by default #114843
fix(agents): resolve configured context tokens for provider-self-prefixed refs #116521
Reject invalid Chat Completions token caps #116563
Preserve thinking levels for live-discovered Ollama models #116584
Use Kilocode's serving-provider context limits #118868
Preserve Chutes serving context limits #119196
Retain images within tool-result context limits #120808
Make /think suggestions match the active model #123507
Preserve prompt-cache boundaries for long OpenAI session IDs #126380
fix: show context window switcher on new sessions #128621
fix(apple): surface rejected chat session settings #128737
fix(ai): recover input-only provider context overflows #129536
fix(lmstudio): preserve saved model reasoning and code mode #129675
fix(github-copilot): allow xhigh and max thinking levels on claude-opus-5 #129690
fix(tui): make provider-specific thinking labels selectable #129952
fix(deepinfra): preserve manifest reasoning during live discovery #129998
fix(ollama): stop replaying images to text-only models #132106
fix(compaction): restore fallback after safeguard provider failures #132449
fix: keep CLI context budgets stable across provider discovery #132544
Warn when prompt cache retention is invalid #94494

Documentation

Correct configuration and runtime documentation to match shipped behavior #121324
Fix the Anthropic prompt-caching documentation link #114028
Fix Mistral reasoning link and audit links inside docs components #114867
docs(codex): correct effort support and provenance in harness reference #130882

Usage, limits, and pricing

OpenClaw now separates subscription-plan information from estimated API cost. Chat can show plan windows, reset times, credits, and the account email attached to a snapshot, while completed iOS replies can show supported input, output, cache, cost, and context-pressure details. These figures are snapshots or estimates, not provider invoices, and in mixed API-key and subscription setups the account label identifies the plan snapshot rather than every run.

The Control UI adds a Profile page for lifetime activity recorded by OpenClaw, with Usage and Profile views grouped in the selected time zone, and plugin-initiated model calls now contribute to aggregate totals. OpenClaw cannot reconstruct activity it never recorded.

Supported failed or incomplete turns can retain the provider's token and cost data without being marked successful. Permanent authentication, model, media, and long-window quota failures stop retrying, while transient rate limits and retryable server errors keep their existing retry or authorized same-provider fallback behavior; Codex subscription runs do not silently switch to pay-as-you-go API keys.

Sources and complete change list

Improvements

Show subscription plan limits in the chat context popover #102784
Add a lifetime activity profile to Control UI settings #102842
Show token, cost, and context usage under iOS replies #103471
Show the account behind plan usage #106200
Focus Profile on identity and move analytics to Usage #115245
Label Codex usage windows with the account email #106500
Trace Claude Code CLI turns in model-call telemetry #108304
Source model pricing from the hosted catalog #114060
Speed up repeat Usage page loads with short-lived caches #115369
Label onboarding CLI auth as subscription or usage-billed #122140
Make cached provider usage return immediately #125520

Bug fixes

Wait for complete all-agent usage totals #103998
Stop retries for permanent provider errors and long quotas #105258
Include plugin LLM calls in usage diagnostics #107937
Make Codex quota failover model-aware and billing-safe #108254
Recover automatically from temporary provider overloads #109354
Harden provider caching, retries, and overflow recovery #109796
Avoid unused Claude CLI observability overhead and bound diagnostics #110246
Bound Control UI session-observer model calls #114052
Restore trajectory capture for Codex sessions #115220
Show the real Codex context window and simplify its popover #121491
restore Codex subscription usage in /status 7eeb109
Make provider usage load without waiting on slow providers e9620fb
Limit all-agent session usage scans #102013
Show periodic usage-limit errors in group chats #102643
Keep Usage and Profile day buckets correct across DST #103097
Correct DeepSeek V4 pricing and legacy alias metadata #103241
Recover Groq Llama turns rejected by TPM limits #104904
Honor Retry-After floors through one retry scheduler #105789
Keep Usage insights aligned with active filters #105792
Retry provider operations after HTTP 429 responses #106365
Stop provider usage reads from hanging after headers #107454
Report truthful Codex context freshness #107813
Count instant sessions in hourly usage metrics #108160
Honor the selected time zone in Usage session details #108579
Record usage and partial output for incomplete Responses turns #109904
Improve Ollama CJK usage estimates #110073
Avoid false Codex subscription-exhaustion messages when account limits remain available #110381
Honor Anthropic Retry-After cooldowns #111072
Keep Claude usage refreshes stable on malformed responses #111088
Report Chutes cache-read pricing correctly #111253
Price usage with each agent's model registry #111389
Count OpenRouter cache-write tokens correctly #111435
Show accurate quota and rate-limit alerts #112047
Reject malformed UTF-8 in provider usage responses #112081
Show current MiniMax quota windows correctly #115030
Restore CLI token usage for OpenAI-compatible providers #115564
Correct OpenRouter budgets and Bedrock cache accounting #116896
Correct cached-token and cost accounting for managed OpenAI chat #117043
Correct managed Anthropic cache accounting through shared streaming policy #117086
Report complete Responses token usage #117533
Keep multi-call token totals cumulative #117930
Classify exhausted xAI credits as billing failures #118615
Apply quota suspensions to the session's real owning agent #120262
Correct LongCat cache pricing and add provider branding #121369
Unify provider-error handling and prevent false rate-limit matches #121817
Correct GPT-5.5 fast-mode session costs #123152
Restore chat cost and usage empty states #123269
Keep provider usage tied to reloaded account config #124131
Restore Codex harness trajectory persistence #125941
Compact recovery ownership and model metadata #126751
fix(announce): report unknown token usage instead of zero on the stats line #126976
fix(usage): preserve agent ownership for matching session IDs #128803
fix: restore usage for prepared provider credentials #129220
fix(ui): render model provider controls progressively #130546
fix: include compaction and branch-summary tokens in run usage #131115
fix: reply status shows missing or stale context estimates #131554
fix(usage): retain reset history and unify accounting #131667
fix: usage dashboard loads slowly for large session histories #132064
fix(ui): show plan usage only for the session provider #132972
Keep selected-day usage totals free of undated historical records #89745
Retry transient bare 429 provider errors #99097
fix(agents): preserve native session accounting at finalization beb07f5
Preserve explicit zero usage in model-call telemetry #100115
Prevent Sonnet 5 model listings from crashing when cost data is missing #102186
Ignore malformed provider Retry-After dates #102987
Classify model-provider telemetry as client spans #104211
Prevent usage-cost audits from stalling #104823
Validate usage-footer fixed precision without breaking high precision #105297
Keep the Usage page working during cache refreshes #105611
Handle malformed Gemini quota responses safely #106420
Drop invalid provider usage reset timestamps #106421
fix(ui): usage page stays loading after cache rebuild #108270
Cap usage history requests before Date overflow #110552
Handle malformed Z.AI usage responses safely #110741
Handle malformed DeepSeek balance responses #110785
Prevent malformed Copilot usage payloads from breaking refreshes #110795
Reject corrupted ClawRouter usage responses #111183
Preserve throttling details in fallback notices #112702
Reject incomplete usage date ranges #113259
Preserve usage for failed OpenAI Responses #115175
Normalize partial model costs before catalog publication #116317
Preserve Claude thinking-token usage #116993
Correct Codex app-server token usage breakdowns #117013
fix(ios): show recent usage days first #117222
Honor configured pricing for Claude Sonnet 5 and Opus 5 #117275
Keep pending usage sessions visible in UTC mode #118004
ClawRouter shows a stable error for malformed usage responses #119580
Preserve cumulative CLI usage in turn diagnostics #120100
Keep failover messages aligned with failure reasons #121717
Include system-agent costs in all-agent usage totals #123271
Keep healthy provider usage visible after auth failures #124367
Keep provider usage tied to the refreshed account #125709
Preserve terminal facts in oversized Codex trajectories #126050
Preserve verified session context limits during catalog gaps #126422
Restore /usage cost in terminal sessions #126668
Refresh Control UI locale catalogs #126927
fix(ui): keep Usage results aligned with active filters #127720
fix(cli): reject malformed usage-cost --days instead of silently defaulting #127978
fix(ui): usage filters hide sessions when multiple providers are selected #129482
fix(kilocode): paid default router incorrectly reports zero cost #129546
fix(ui): usage copy actions silently hide clipboard failures #129562
fix(venice): preserve authoritative model pricing #129694
fix(usage): prevent model identity collisions #130307
fix(ui): Usage Shift-click selects the wrong session range #130375
fix(usage): preserve MiniMax reset alias order #132897

Documentation

Explain Z.AI rate-limit and overload responses #107940
Explain model usage events and supported usage APIs #118483

Was this useful?