After studying clawdbot (OpenClaw) and OpenCode implementations:
## Beta headers
- Add interleaved-thinking-2025-05-14 and fine-grained-tool-streaming-2025-05-14
as common betas (sent with ALL auth types, not just OAuth)
- OAuth tokens additionally get oauth-2025-04-20
- API keys now also get the common betas (previously got none)
## Vision/image support
- Add _convert_vision_content() to convert OpenAI multimodal format
(image_url blocks) to Anthropic format (image blocks with base64/url source)
- Handles both data: URIs (base64) and regular URLs
## Role alternation enforcement
- Anthropic strictly rejects consecutive same-role messages (400 error)
- Add post-processing step that merges consecutive user/assistant messages
- Handles string, list, and mixed content types during merge
## Tool choice support
- Add tool_choice parameter to build_anthropic_kwargs()
- Maps OpenAI values: auto→auto, required→any, none→omit, name→tool
## Cache metrics tracking
- Anthropic uses cache_read_input_tokens / cache_creation_input_tokens
(different from OpenRouter's prompt_tokens_details.cached_tokens)
- Add api_mode-aware branch in run_agent.py cache stats logging
## Credential refresh on 401
- On 401 error during anthropic_messages mode, re-read credentials
via resolve_anthropic_token() (picks up refreshed Claude Code tokens)
- Rebuild client if new token differs from current one
- Follows same pattern as Codex/Nous 401 refresh handlers
## Tests
- 44 adapter tests (8 new: vision conversion, role alternation, tool choice)
- Updated beta header tests to verify new structure
- Full suite: 3198 passed, 0 regressions
* fix: /reasoning command output ordering, display, and inline think extraction
Three issues with the /reasoning command:
1. Output interleaving: The command echo used print() while feedback
used _cprint(), causing them to render out-of-order under
prompt_toolkit's patch_stdout. Changed echo to use _cprint() so
all output renders through the same path in correct order.
2. Reasoning display not working: /reasoning show toggled a flag
but reasoning never appeared for models that embed thinking in
inline <think> blocks rather than structured API fields. Added
fallback extraction in _build_assistant_message to capture
<think> block content as reasoning when no structured reasoning
fields (reasoning, reasoning_content, reasoning_details) are
present. This feeds into both the reasoning callback (during
tool loops) and the post-response reasoning box display.
3. Feedback clarity: Added checkmarks to confirm actions, persisted
show/hide to config (was session-only before), and aligned the
status display for readability.
Tests: 7 new tests for inline think block extraction (41 total).
* feat: add /reasoning command to gateway (Telegram/Discord/etc)
The /reasoning command only existed in the CLI — messaging platforms
had no way to view or change reasoning settings. This adds:
1. /reasoning command handler in the gateway:
- No args: shows current effort level and display state
- /reasoning <level>: sets reasoning effort (none/low/medium/high/xhigh)
- /reasoning show|hide: toggles reasoning display in responses
- All changes saved to config.yaml immediately
2. Reasoning display in gateway responses:
- When show_reasoning is enabled, prepends a 'Reasoning' block
with the model's last_reasoning content before the response
- Collapses long reasoning (>15 lines) to keep messages readable
- Uses last_reasoning from run_conversation result dict
3. Plumbing:
- Added _show_reasoning attribute loaded from config at startup
- Propagated last_reasoning through _run_agent return dict
- Added /reasoning to help text and known_commands set
- Uses getattr for _show_reasoning to handle test stubs
* fix: improve Kimi model selection — auto-detect endpoint, add missing models
Kimi Coding Plan setup:
- New dedicated _model_flow_kimi() replaces the generic API-key flow
for kimi-coding. Removes the confusing 'Base URL' prompt entirely —
the endpoint is auto-detected from the API key prefix:
sk-kimi-* → api.kimi.com/coding/v1 (Kimi Coding Plan)
other → api.moonshot.ai/v1 (legacy Moonshot)
- Shows appropriate models for each endpoint:
Coding Plan: kimi-for-coding, kimi-k2.5, kimi-k2-thinking, kimi-k2-thinking-turbo
Moonshot: full model catalog
- Clears any stale KIMI_BASE_URL override so runtime auto-detection
via _resolve_kimi_base_url() works correctly.
Model catalog updates:
- Added kimi-for-coding (primary Coding Plan model) and kimi-k2-thinking-turbo
to models.py, main.py _PROVIDER_MODELS, and model_metadata.py context windows.
- Updated User-Agent from KimiCLI/1.0 to KimiCLI/1.3 (Kimi's coding
endpoint whitelists known coding agents via User-Agent sniffing).
- gateway/run.py: Take main's _resolve_gateway_model() helper
- hermes_cli/setup.py: Re-apply nous-api removal after merge brought
it back. Fix provider_idx offset (Custom is now index 3, not 4).
- tests/hermes_cli/test_setup.py: Fix custom setup test index (3→4)
Model selection now comes exclusively from config.yaml (set via
'hermes model' or 'hermes setup'). The LLM_MODEL env var is no longer
read or written anywhere in production code.
Why: env vars are per-process/per-user and would conflict in
multi-agent or multi-tenant setups. Config.yaml is file-based and
can be scoped per-user or eventually per-session.
Changes:
- cli.py: Read model from CLI_CONFIG only, not LLM_MODEL/OPENAI_MODEL
- hermes_cli/auth.py: _save_model_choice() no longer writes LLM_MODEL
to .env
- hermes_cli/setup.py: Remove 12 save_env_value('LLM_MODEL', ...)
calls from all provider setup flows
- gateway/run.py: Remove LLM_MODEL fallback (HERMES_MODEL still works
for gateway process runtime)
- cron/scheduler.py: Same
- agent/auxiliary_client.py: Remove LLM_MODEL from custom endpoint
model detection
Phase 2 of the provider router migration — route the main agent's
client construction and fallback activation through
resolve_provider_client() instead of duplicated ad-hoc logic.
run_agent.py:
- __init__: When no explicit api_key/base_url, use
resolve_provider_client(provider, raw_codex=True) for client
construction. Explicit creds (from CLI/gateway runtime provider)
still construct directly.
- _try_activate_fallback: Replace _resolve_fallback_credentials and
its duplicated _FALLBACK_API_KEY_PROVIDERS / _FALLBACK_OAUTH_PROVIDERS
dicts with a single resolve_provider_client() call. The router
handles all provider types (API-key, OAuth, Codex) centrally.
- Remove _resolve_fallback_credentials method and both fallback dicts.
agent/auxiliary_client.py:
- Add raw_codex parameter to resolve_provider_client(). When True,
returns the raw OpenAI client for Codex providers instead of wrapping
in CodexAuxiliaryClient. The main agent needs this for direct
responses.stream() access.
3251 passed, 2 pre-existing unrelated failures.
Add centralized call_llm() and async_call_llm() functions that own the
full LLM request lifecycle:
1. Resolve provider + model from task config or explicit args
2. Get or create a cached client for that provider
3. Format request args (max_tokens handling, provider extra_body)
4. Make the API call with max_tokens/max_completion_tokens retry
5. Return the response
Config: expanded auxiliary section with provider:model slots for all
tasks (compression, vision, web_extract, session_search, skills_hub,
mcp, flush_memories). Config version bumped to 7.
Migrated all auxiliary consumers:
- context_compressor.py: uses call_llm(task='compression')
- vision_tools.py: uses async_call_llm(task='vision')
- web_tools.py: uses async_call_llm(task='web_extract')
- session_search_tool.py: uses async_call_llm(task='session_search')
- browser_tool.py: uses call_llm(task='vision'/'web_extract')
- mcp_tool.py: uses call_llm(task='mcp')
- skills_guard.py: uses call_llm(provider='openrouter')
- run_agent.py flush_memories: uses call_llm(task='flush_memories')
Tests updated for context_compressor and MCP tool. Some test mocks
still need updating (15 remaining failures from mock pattern changes,
2 pre-existing).
Route all remaining ad-hoc auxiliary LLM call sites through
resolve_provider_client() so auth, headers, and API format (Chat
Completions vs Responses API) are handled consistently in one place.
Files changed:
- tools/openrouter_client.py: Replace manual AsyncOpenAI construction
with resolve_provider_client('openrouter', async_mode=True). The
shared client module now delegates entirely to the router.
- tools/skills_guard.py: Replace inline OpenAI client construction
(hardcoded OpenRouter base_url, manual api_key lookup, manual
headers) with resolve_provider_client('openrouter'). Remove unused
OPENROUTER_BASE_URL import.
- trajectory_compressor.py: Add _detect_provider() to map config
base_url to a provider name, then route through
resolve_provider_client. Falls back to raw construction for
unrecognized custom endpoints.
- mini_swe_runner.py: Route default case (no explicit api_key/base_url)
through resolve_provider_client('openrouter') with auto-detection
fallback. Preserves direct construction when explicit creds are
passed via CLI args.
- agent/auxiliary_client.py: Fix stale module docstring — vision auto
mode now correctly documents that Codex and custom endpoints are
tried (not skipped).
Three interconnected fixes for auxiliary client infrastructure:
1. CENTRALIZED PROVIDER ROUTER (auxiliary_client.py)
Add resolve_provider_client(provider, model, async_mode) — a single
entry point for creating properly configured clients. Given a provider
name and optional model, it handles auth lookup (env vars, OAuth
tokens, auth.json), base URL resolution, provider-specific headers,
and API format differences (Chat Completions vs Responses API for
Codex). All auxiliary consumers should route through this instead of
ad-hoc env var lookups.
Refactored get_text_auxiliary_client, get_async_text_auxiliary_client,
and get_vision_auxiliary_client to use the router internally.
2. FIX CODEX VISION BYPASS (vision_tools.py)
vision_tools.py was constructing a raw AsyncOpenAI client from the
sync vision client's api_key/base_url, completely bypassing the Codex
Responses API adapter. When the vision provider resolved to Codex,
the raw client would hit chatgpt.com/backend-api/codex with
chat.completions.create() which only supports the Responses API.
Fix: Added get_async_vision_auxiliary_client() which properly wraps
Codex into AsyncCodexAuxiliaryClient. vision_tools.py now uses this
instead of manual client construction.
3. FIX COMPRESSION FALLBACK + VISION ERROR HANDLING
- context_compressor.py: Removed _get_fallback_client() which blindly
looked for OPENAI_API_KEY + OPENAI_BASE_URL (fails for Codex OAuth,
API-key providers, users without OPENAI_BASE_URL set). Replaced
with fallback loop through resolve_provider_client() for each
known provider, with same-provider dedup.
- vision_tools.py: Added error detection for vision capability
failures. Returns clear message to the model when the configured
model doesn't support vision, instead of a generic error.
Addresses #886
Allow users to interact with Hermes by sending and receiving emails.
Uses IMAP polling for incoming messages and SMTP for replies with
proper threading (In-Reply-To, References headers).
Integrates with all 14 gateway extension points: config, adapter
factory, authorization, send_message tool, cron delivery, toolsets,
prompt hints, channel directory, setup wizard, status display, and
env example.
65 tests covering config, parsing, dispatch, threading, IMAP fetch,
SMTP send, attachments, and all integration points.
- Add agent/embeddings.py with Embedder protocol, FastEmbedEmbedder, OpenAIEmbedder
- Factory function get_embedder() reads provider from config.yaml embeddings section
- Lazy initialization — no startup impact, model loaded on first embed call
- cosine_similarity() and cosine_similarity_matrix() utility functions included
- Add fastembed as optional dependency in pyproject.toml
- 30 unit tests, all passing
Closes#675
The KawaiiSpinner animation would occasionally spam dozens of duplicate
lines instead of overwriting in-place with \r. This happened because
prompt_toolkit's StdoutProxy processes each flush() as a separate
run_in_terminal() call — when the write thread is slow (busy event loop
during long tool executions), each \r frame gets its own call, and the
terminal layout save/restore between calls breaks the \r overwrite
semantics.
Fix: rate-limit flush() calls to at most every 0.4s. Between flushes,
\r-frame writes accumulate in StdoutProxy's buffer. When flushed, they
concatenate into one string (e.g. \r frame1 \r frame2 \r frame3) and
are written in a single run_in_terminal() call where \r works correctly.
The spinner still animates (flush ~2.5x/sec) but each flush batches
~3 frames, guaranteeing the \r collapse always works. Most visible
with execute_code and terminal tools (3+ second executions).
Vision auto-mode previously only tried OpenRouter, Nous, and Codex
for multimodal — deliberately skipping custom endpoints with the
assumption they 'may not handle vision input.' This caused silent
failures for users running local multimodal models (Qwen-VL, LLaVA,
Pixtral, etc.) without any cloud API keys.
Now custom endpoints are tried as a last resort in auto mode. If the
model doesn't support vision, the API call fails gracefully — but
users with local vision models no longer need to manually set
auxiliary.vision.provider: main in config.yaml.
Reported by @Spadav and @kotyKD.
Skills can now declare fallback_for_toolsets, fallback_for_tools,
requires_toolsets, and requires_tools in their SKILL.md frontmatter.
The system prompt builder filters skills automatically based on which
tools are available in the current session.
- Add _read_skill_conditions() to parse conditional frontmatter fields
- Add _skill_should_show() to evaluate conditions against available tools
- Update build_skills_system_prompt() to accept and apply tool availability
- Pass valid_tool_names and available toolsets from run_agent.py
- Backward compatible: skills without conditions always show; calling
build_skills_system_prompt() with no args preserves existing behavior
Closes#539
New config option:
security:
redact_secrets: false # default: true
When set to false, API keys, tokens, and passwords are shown in
full in read_file, search_files, and terminal output. Useful for
debugging auth issues where you need to verify the actual key value.
Bridged to both CLI and gateway via HERMES_REDACT_SECRETS env var.
The check is in redact_sensitive_text() itself, so all call sites
(terminal, file tools, log formatter) respect it.
The summary message was always injected as 'user' role, which causes
consecutive user messages when the last preserved head message is also
'user'. Some APIs reject this (400 error), and it produces malformed
training data.
Fix: check the role of the last head message and pick the opposite role
for the summary — 'user' after assistant/tool, 'assistant' after user.
Based on PR #328 by johnh4098. Closes#328.
Complete Signal adapter using signal-cli daemon HTTP API.
Based on PR #268 by ibhagwan, rebuilt on current main with bug fixes.
Architecture:
- SSE streaming for inbound messages with exponential backoff (2s→60s)
- JSON-RPC 2.0 for outbound (send, typing, attachments, contacts)
- Health monitor detects stale SSE connections (120s threshold)
- Phone number redaction in all logs and global redact.py
Features:
- DM and group message support with separate access policies
- DM policies: pairing (default), allowlist, open
- Group policies: disabled (default), allowlist, open
- Attachment download with magic-byte type detection
- Typing indicators (8s refresh interval)
- 100MB attachment size limit, 8000 char message limit
- E.164 phone + UUID allowlist support
Integration:
- Platform.SIGNAL enum in gateway/config.py
- Signal in _is_user_authorized() allowlist maps (gateway/run.py)
- Adapter factory in _create_adapter() (gateway/run.py)
- user_id_alt/chat_id_alt fields in SessionSource for UUIDs
- send_message tool support via httpx JSON-RPC (not aiohttp)
- Interactive setup wizard in 'hermes gateway setup'
- Connectivity testing during setup (pings /api/v1/check)
- signal-cli detection and install guidance
Bug fixes from PR #268:
- Timestamp reads from envelope_data (not outer wrapper)
- Uses httpx consistently (not aiohttp in send_message tool)
- SIGNAL_DEBUG scoped to signal logger (not root)
- extract_images regex NOT modified (preserves group numbering)
- pairing.py NOT modified (no cross-platform side effects)
- No dual authorization (adapter defers to run.py for user auth)
- Wildcard uses set membership ('*' in set, not list equality)
- .zip default for PK magic bytes (not .docx)
No new Python dependencies — uses httpx (already core).
External requirement: signal-cli daemon (user-installed).
Tests: 30 new tests covering config, init, helpers, session source,
phone redaction, authorization, and send_message integration.
Co-authored-by: ibhagwan <ibhagwan@users.noreply.github.com>
The 'openai' provider was redundant — using OPENAI_BASE_URL +
OPENAI_API_KEY with provider: 'main' already covers direct OpenAI API.
Provider options are now: auto, openrouter, nous, codex, main.
- Removed _try_openai(), _OPENAI_AUX_MODEL, _OPENAI_BASE_URL
- Replaced openai tests with codex provider tests
- Updated all docs to remove 'openai' option and clarify 'main'
- 'main' description now explicitly mentions it works with OpenAI API,
local models, and any OpenAI-compatible endpoint
Tests: 2467 passed.
The Codex Responses API (chatgpt.com/backend-api/codex) supports
vision via gpt-5.3-codex. This was verified with real API calls
using image analysis.
Changes to _CodexCompletionsAdapter:
- Added _convert_content_for_responses() to translate chat.completions
multimodal format to Responses API format:
- {type: 'text'} → {type: 'input_text'}
- {type: 'image_url', image_url: {url: '...'}} → {type: 'input_image', image_url: '...'}
- Fixed: removed 'stream' from resp_kwargs (responses.stream() handles it)
- Fixed: removed max_output_tokens and temperature (Codex endpoint rejects them)
Provider changes:
- Added 'codex' as explicit auxiliary provider option
- Vision auto-fallback now includes Codex (OpenRouter → Nous → Codex)
since gpt-5.3-codex supports multimodal input
- Updated docs with Codex OAuth examples
Tested with real Codex OAuth token + ~/.hermes/image2.png — confirmed
working end-to-end through the full adapter pipeline.
Tests: 2459 passed.
Users can now set provider: "openai" for auxiliary tasks (vision, web
extract, compression) to use OpenAI's API directly with their
OPENAI_API_KEY. This hits api.openai.com/v1 with gpt-4o-mini as the
default model — supports vision since GPT-4o handles image input.
Provider options are now: auto, openrouter, nous, openai, main.
Changes:
- agent/auxiliary_client.py: added _try_openai(), "openai" case in
_resolve_forced_provider(), updated auxiliary_max_tokens_param()
to use max_completion_tokens for OpenAI
- Updated docs: cli-config.yaml.example, AGENTS.md, and user-facing
configuration.md with Common Setups section showing OpenAI,
OpenRouter, and local model examples
- 3 new tests for OpenAI provider resolution
Tests: 2459 passed (was 2429).
Improvements on top of PR #606 (auxiliary model configuration):
1. Gateway bridge: Added auxiliary.* and compression.summary_provider
config bridging to gateway/run.py so config.yaml settings work from
messaging platforms (not just CLI). Matches the pattern in cli.py.
2. Vision auto-fallback safety: In auto mode, vision now only tries
OpenRouter + Nous Portal (known multimodal-capable providers).
Custom endpoints, Codex, and API-key providers are skipped to avoid
confusing errors from providers that don't support vision input.
Explicit provider override (AUXILIARY_VISION_PROVIDER=main) still
allows using any provider.
3. Comprehensive tests (46 new):
- _get_auxiliary_provider env var resolution (8 tests)
- _resolve_forced_provider with all provider types (8 tests)
- Per-task provider routing integration (4 tests)
- Vision auto-fallback safety (7 tests)
- Config bridging logic (11 tests)
- Gateway/CLI bridge parity (2 tests)
- Vision model override via env var (2 tests)
- DEFAULT_CONFIG shape validation (4 tests)
4. Docs: Added auxiliary_client.py to AGENTS.md project structure.
Updated module docstring with separate text/vision resolution chains.
Tests: 2429 passed (was 2383).
- Added support for auxiliary model overrides in the configuration, allowing users to specify providers and models for vision and web extraction tasks.
- Updated the CLI configuration example to include new auxiliary model settings.
- Enhanced the environment variable mapping in the CLI to accommodate auxiliary model configurations.
- Improved the resolution logic for auxiliary clients to support task-specific provider overrides.
- Updated relevant documentation and comments for clarity on the new features and their usage.
Skills can now declare runtime prerequisites (env vars, CLI binaries) via
YAML frontmatter. Skills with unmet prerequisites are excluded from the
system prompt so the agent never claims capabilities it can't deliver, and
skill_view() warns the agent about what's missing.
Three layers of defense:
- build_skills_system_prompt() filters out unavailable skills
- _find_all_skills() flags unmet prerequisites in metadata
- skill_view() returns prerequisites_warning with actionable details
Tagged 12 bundled skills that have hard runtime dependencies:
gif-search (TENOR_API_KEY), notion (NOTION_API_KEY), himalaya, imessage,
apple-notes, apple-reminders, openhue, duckduckgo-search, codebase-inspection,
blogwatcher, songsee, mcporter.
Closes#658Fixes#630
browser_vision now saves screenshots persistently to ~/.hermes/browser_screenshots/
and returns the screenshot_path in its JSON response. The model can include
MEDIA:<path> in its response to share screenshots as native photos.
Changes:
- browser_tool.py: Save screenshots persistently, return screenshot_path,
auto-cleanup files older than 24 hours, mkdir moved inside try/except
- telegram.py: Add send_image_file() — sends local images via bot.send_photo()
- discord.py: Add send_image_file() — sends local images via discord.File
- slack.py: Add send_image_file() — sends local images via files_upload_v2()
(WhatsApp already had send_image_file — no changes needed)
- prompt_builder.py: Updated Telegram hint to list image extensions,
added Discord and Slack MEDIA: platform hints
- browser.md: Document screenshot sharing and 24h cleanup
- send_file_integration_map.md: Updated to reflect send_image_file is now
implemented on Telegram/Discord/Slack
- test_send_image_file.py: 19 tests covering MEDIA: .png extraction,
send_image_file on all platforms, and screenshot cleanup
Partially addresses #466 (Phase 0: platform adapter gaps for send_image_file).
Kimi Code (platform.kimi.ai) issues API keys prefixed sk-kimi- that require:
1. A different base URL: api.kimi.com/coding/v1 (not api.moonshot.ai/v1)
2. A User-Agent header identifying a recognized coding agent
Without this fix, sk-kimi- keys fail with 401 (wrong endpoint) or 403
('only available for Coding Agents') errors.
Changes:
- Auto-detect sk-kimi- key prefix and route to api.kimi.com/coding/v1
- Send User-Agent: KimiCLI/1.0 header for Kimi Code endpoints
- Legacy Moonshot keys (api.moonshot.ai) continue to work unchanged
- KIMI_BASE_URL env var override still takes priority over auto-detection
- Updated .env.example with correct docs and all endpoint options
- Fixed doctor.py health check for Kimi Code keys
Reference: https://github.com/MoonshotAI/kimi-cli (platforms.py)
Updated the _generate_summary method to attempt summary generation using the auxiliary model first, with a fallback to the main model. If both attempts fail, the method now returns None instead of a placeholder, allowing the caller to handle missing summaries appropriately. This change enhances the robustness of context compression and improves logging for failure scenarios.
Reduces token usage and latency for most tasks by defaulting to
medium reasoning effort instead of xhigh. Users can still override
via config or CLI flag. Updates code, tests, example config, and docs.
Enhance message compression by adding a method to clean up orphaned tool-call and tool-result pairs. This ensures that the API receives well-formed messages, preventing errors related to mismatched IDs. The new functionality includes removing orphaned results and adding stub results for missing calls, improving overall message integrity during compression.
Add a 'platforms' field to SKILL.md frontmatter that restricts skills
to specific operating systems. Skills with platforms: [macos] only
appear in the system prompt, skills_list(), and slash commands on macOS.
Skills without the field load everywhere (backward compatible).
Implementation:
- skill_matches_platform() in tools/skills_tool.py — core filter
- Wired into all 3 discovery paths: prompt_builder.py, skills_tool.py,
skill_commands.py
- 28 new tests across 3 test files
New bundled Apple/macOS skills (all platforms: [macos]):
- imessage — Send/receive iMessages via imsg CLI
- apple-reminders — Manage Reminders via remindctl CLI
- apple-notes — Manage Notes via memo CLI
- findmy — Track devices/AirTags via AppleScript + screen capture
Docs updated: CONTRIBUTING.md, AGENTS.md, creating-skills.md,
skills.md (user guide)
These direct providers don't return cost in API responses and their
per-token pricing isn't readily available externally. Treat as local
models with zero cost so they appear in /insights without fake estimates.
When the user only has a z.ai/Kimi/MiniMax API key (no OpenRouter key),
auxiliary tasks (context compression, web summarization, session search)
now fall back to the configured direct provider instead of returning None.
Resolution chain: OpenRouter -> Nous -> Custom endpoint -> Codex OAuth
-> direct API-key providers -> None.
Uses cheap/fast models for auxiliary tasks:
- zai: glm-4.5-flash
- kimi-coding: kimi-k2-turbo-preview
- minimax/minimax-cn: MiniMax-M2.5-highspeed
Vision auxiliary intentionally NOT modified — vision needs multimodal
models (Gemini) that these providers don't serve.
Adds DEFAULT_CONTEXT_LENGTHS entries for kimi-k2.5 (262144), kimi-k2-thinking
(262144), kimi-k2-turbo-preview (262144), kimi-k2-0905-preview (131072),
MiniMax-M2.5/M2.5-highspeed/M2.1 (204800), and glm-4.5/4.5-flash (131072).
Avoids unnecessary 2M-token probe on first use with direct providers.
