fix(providers): scope think to Ollama endpoints

[AXEG0]

What does this PR do?

Restricts Ollama-only disabled-reasoning controls to endpoints positively identified as Ollama.

When reasoning is disabled, a verified Ollama endpoint receives both extra_body.think=false and top-level reasoning_effort="none". Generic OpenAI-compatible custom relays receive neither field, while explicit enabled reasoning effort remains available for custom reasoning APIs. Ollama detection uses the existing /api/tags probe instead of URL substrings or port numbers.

Related Issue

Fixes #11237.

Related/overlapping work:

• fix: only send think=false to Ollama, not all custom providers #11520 targets the same unsupported-think symptom but identifies Ollama from URL text/port. This PR uses positive /api/tags detection to avoid misclassifying arbitrary relays.
• fix: gate Ollama think/reasoning_effort on model's declared capabilities #63315 gates enabled reasoning controls on Ollama model capabilities. This PR instead scopes the disabled-control path and protects non-Ollama custom relays.
• fix(ollama): emit top-level reasoning_effort=none on /v1/chat/completions (#25758) #64608 added the current top-level disabled reasoning_effort behavior; this PR scopes that field under the same verified-Ollama predicate as think.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• plugins/model-providers/custom/__init__.py: identify Ollama through detect_local_server_type(base_url, api_key); scope both disabled controls to verified Ollama endpoints.
• agent/chat_completion_helpers.py and agent/transports/chat_completions.py: thread the API key through provider-profile and retry/fallback request construction.
• agent/auxiliary_client.py: keep auxiliary request construction in parity with the main transport path.
• Tests under tests/agent/, tests/plugins/model_providers/, and tests/providers/: cover Ollama, misleading URLs/ports, generic relays, API-key propagation, auxiliary requests, and retry/transport parity.

How to Test

1. Run scripts/run_tests.sh tests/plugins/model_providers/test_custom_profile.py tests/providers/test_transport_parity.py tests/agent/transports/test_chat_completions.py tests/agent/test_auxiliary_client.py tests/agent/test_probe_cache_followups.py -q.
2. Verify a positively identified Ollama endpoint with reasoning disabled receives think=false and reasoning_effort="none".
3. Verify generic relays—including URLs containing ollama or using port 11434 without a valid /api/tags response—receive neither disabled field.

Focused result: 240 passed.

Additional checks: ruff check ., git diff --check, and scripts/check-windows-footguns.py --all pass.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass (focused affected suites pass; the full GitHub Actions matrix is awaiting maintainer approval)
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: Linux

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — N/A; no user-facing interface or behavior documentation changed
• I've updated cli-config.yaml.example if I added/changed config keys — N/A; no configuration keys changed
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — N/A; no architecture or workflow changed
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — platform-independent Python change; Windows footgun check passes
• I've updated tool descriptions/schemas if I changed tool behavior — N/A; no tool schema changed

Screenshots / Logs

N/A — regression coverage and command results are listed above.

AI Assistance

OpenAI — GPT-5 via Codex CLI (implementation assistance, test execution, and PR-description drafting).

[tonydwb]

Code Review Summary

Verdict: Approved

This PR scopes the think feature to Ollama endpoints.

Looks Good

• Targeted fix ensuring think behavior is Ollama-specific.
• Well-scoped: 6 files, minimal additions.

No Issues Found

---

Reviewed by Hermes Agent

[teknium1]

Thanks for targeting a real custom-provider compatibility issue. Current main still sends disabled-reasoning controls to every custom endpoint (plugins/model-providers/custom/__init__.py:57-65).

Problems

• Mainline commit 8662254ab2fe added top-level reasoning_effort="none" beside think=False. This PR predates it and only gates think (plugins/model-providers/custom/__init__.py:65-67 on d9446975), so a salvage must scope both fields or generic relays can still receive an unsupported disabled-reasoning control.
• "ollama" in base_url and ":11434" in base_url are not positive endpoint identification (plugins/model-providers/custom/__init__.py:56-60). Current detect_local_server_type() verifies Ollama through /api/tags (agent/model_metadata.py:683-732).

Suggested changes

• Preserve the current Ollama /v1 fix while gating both disabled-case fields, with a transport-level generic-relay test asserting neither is sent.
• Use an endpoint test that does not classify an arbitrary matching relay URL as Ollama.

Automated hermes-sweeper review.

[aadamowski]

Confirming that this breaks Groq endpoints (https://console.groq.com/home):

⚠️   API call failed (attempt 1/3): BadRequestError [HTTP 400]
   🔌 Provider: custom  Model: llama-3.3-70b-versatile
   🌐 Endpoint: https://api.groq.com/openai/v1
   📝 Error: HTTP 400: property 'think' is unsupported
   📋 Details: {'message': "property 'think' is unsupported", 'type': 'invalid_request_error'}

I manually applied the fix from commit 4e1a167701427ef173565f7bd44c0b367def6554 onto my local Hermes install and it resolved the issue.

Please merge the contributed fix!

[AXEG0]

@teknium1 Thanks for the review. Both points are addressed, and the branch is rebased on current main (62a0253):

• Ollama detection is now a positive identification via detect_local_server_type() instead of the URL substring match. A non-Ollama endpoint whose URL merely contains "ollama", or a non-Ollama service on :11434, is no longer classified as Ollama.
• The disabled-branch reasoning_effort="none" is scoped under the same Ollama predicate as think, so non-Ollama custom endpoints (Groq included) receive neither field.
• The rebase also fixes an undefined api_key reference in _build_call_kwargs() that a broad exception handler had been masking; the provider profile now builds correctly on the retry/fallback paths.

Targeted suites pass locally (test_chat_completions, test_custom_profile, test_transport_parity). @aadamowski thanks for confirming the fix on Groq. Ready for another look.

[GottZ]

This was generated by AI during triage.

Summary

Thirty-seven PRs address or reference this broad Ollama/provider-reasoning complex. Their diffs cover Ollama Cloud registration/catalog handling, surrogate and truncation recovery, Windows-adjacent work, reasoning propagation and capability scoping, and—most directly for the target—preventing Ollama-only disable controls from reaching heterogeneous custom endpoints.

Related pull requests

• feat(ollama): pass think=false to custom providers when reasoning_effort is none #3197 [closed] partial — (+51/-0) — n/a: Adds think=false for disabled reasoning on every custom provider; closed because its implementation was cherry-picked into merged feat: Ollama improvements — Cloud provider, GLM continuation, think=false, surrogate sanitization, /v1 hint #10782, although later evidence showed Ollama /v1/chat/completions also needs top-level reasoning_effort=none.
• fix(agent): sanitize surrogate characters from Ollama model output #5176 [closed] best fix — (+21/-2) — n/a: Sanitizes surrogate characters before requests, before assistant-message storage, and in the actual retry payload; closed in favor of earlier duplicate fix(agent): sanitize surrogate characters from API responses and before API calls #5074.
• feat: add Ollama Cloud as built-in provider #6038 [closed] fixes — (+596/-50) — n/a: Adds Ollama Cloud as a first-class provider with credentials, runtime routing, auxiliary defaults, dynamic discovery, caching, and tests; closed because it was cherry-picked into merged feat: Ollama improvements — Cloud provider, GLM continuation, think=false, surrogate sanitization, /v1 hint #10782.
• fix(agent): continue Ollama GLM replies after stop misreports #10740 [closed] best fix — (+172/-0) — n/a: Detects narrowly scoped Ollama/GLM stop misreports after tool turns and reuses the existing continuation path; closed because it was cherry-picked into merged feat: Ollama improvements — Cloud provider, GLM continuation, think=false, surrogate sanitization, /v1 hint #10782.
• feat: Ollama improvements — Cloud provider, GLM continuation, think=false, surrogate sanitization, /v1 hint #10782 [merged] best fix — (+822/-37) — n/a: Merged reference implementation consolidating Ollama Cloud registration, GLM continuation, think=false, surrogate sanitization, and local /v1 guidance, while its generic-custom disable control was later refined by fix(ollama): emit top-level reasoning_effort=none on /v1/chat/completions (#25758) #64608 and target fix(providers): scope think to Ollama endpoints #64286.
• fix(agent): stop sending think=false to remote custom providers #11296 duplicate — (+22/-7) — n/a: Gates think=false with local/Ollama heuristics in the obsolete run_agent.py builder; keep-open review says salvage the predicate into the live CustomProfile path.
• fix(agent): restrict think=False to Ollama/local endpoints only #11362 partial — (+10/-2) — n/a: Restricts think=false using URL, port, or locality checks in the obsolete builder; the keep-open verdict requires positive Ollama identification in CustomProfile and local non-Ollama coverage.
• fix: only send think=False to Ollama endpoints (#11237) #11395 [closed] partial — (+34/-7) — n/a: Adds URL/port Ollama detection and negative Mistral/Fireworks tests in the old builder; closed voluntarily to reduce duplicate volume.
• fix: only send think=false to Ollama, not all custom providers #11520 partial — (+8/-1) — n/a: Uses ollama_num_ctx, URL text, or port 11434 to scope think=false; the high-salvage keep-open verdict requires moving that guard to the live profile path with positive and negative tests.
• fix: restrict think=False to Ollama endpoints only #12488 [closed] partial — (+3/-1) — n/a: Adds a URL/port gate around think=false in stale run_agent.py code without tests; closed after being identified as a duplicate of fix(agent): restrict think=False to Ollama/local endpoints only #11362.
• fix(agent): scope think=False to Ollama endpoints only (#11237) #12914 [closed] partial — (+292/-3) — n/a: Adds _is_ollama() heuristics plus broad old-path tests for Ollama versus cloud and local non-Ollama endpoints; closed after its author concluded the then-current main no longer used that request path.
• fix(models): strip :cloud/-cloud suffix from models.dev Ollama Cloud IDs #16192 [closed] fixes — (+114/-3) — n/a: Normalizes models.dev :cloud/-cloud IDs before deduplication and adds focused tests; closed after its commit was salvaged into merged fix(models): strip :cloud/-cloud suffix from models.dev Ollama Cloud IDs #19881.
• fix(cli): implement self-healing sanitization for Ollama Cloud model IDs #16196 [closed] fixes — (+261/-9) — n/a: Extends suffix normalization to repair legacy caches and centralize deduplication; closed as superseded by merged fix(models): strip :cloud/-cloud suffix from models.dev Ollama Cloud IDs #19881.
• fix(windows): improve path normalization, shell execution, and encoding #16212 [closed] best fix — (+261/-9) — n/a: Despite its Windows title and recorded issue association, the visible diff implements Ollama Cloud suffix/cache sanitization rather than Windows path, shell, or encoding changes; closed under the contributor-documented consolidated Windows-support decision.
• fix(models): strip :cloud/-cloud suffix from models.dev Ollama Cloud IDs #19881 [merged] best fix — (+114/-3) — n/a: Merged salvage of fix(models): strip :cloud/-cloud suffix from models.dev Ollama Cloud IDs #16192 and reference implementation for hosted Ollama Cloud suffix stripping and duplicate suppression, now present as commit eadf34633e038c595dcf615845b992a45443e380.
• fix(models): prioritize Ollama Cloud models #25174 related — (+53/-1) — n/a: Prioritizes Qwen 3.5 and Gemma 4 in newly merged Ollama Cloud results; keep-open review requires applying the ordering to fresh and stale cache returns too.
• fix(agent): propagate reasoning_effort to Ollama + curator fork #25866 [closed] partial — (+23/-3) — n/a: Adds top-level reasoning_effort=none and propagates reasoning config into a curator fork; closed by its author, while fix(ollama): emit top-level reasoning_effort=none on /v1/chat/complet… #29820/fix(ollama): emit top-level reasoning_effort=none on /v1/chat/completions (#25758) #64608 cover the live Ollama disable path.
• fix(ollama): emit top-level reasoning_effort=none on /v1/chat/complet… #29820 [closed] best fix — (+170/-5) — n/a: Recorded best fix for [Bug]: agent.reasoning_effort: none silently ignored on Ollama — main agent stuck in medium mode, bg-review fork can spiral (up to 65k tokens / 28 min) #25758, adding top-level reasoning_effort=none plus background-review inheritance; closed after its provider half was salvaged with attribution into merged fix(ollama): emit top-level reasoning_effort=none on /v1/chat/completions (#25758) #64608 and the fork half was already on main.
• fix(ollama): add nemotron-3-ultra to model selector #42135 related — (+32/-0) — n/a: Injects nemotron-3-ultra beside existing Nemotron catalog entries; keep-open review requires cache-path coverage and evidence justifying an exception to dynamic discovery.
• Disable qwen3 thinking on local Ollama so tool-calling works (fixes #6152) #42201 [closed] fixes — (+132/-0) — n/a: Adds local auto-disable plus explicit model.extra_body forwarding for remote Ollama; closed as implemented on main via commit 67df958dbe06bb10ca16b8686c76baf0de3bac03, though later fixes refine the custom-endpoint scope.
• fix: respect reasoning_config for custom providers and fix OpenRouter… #43613 [closed] partial — (+21/-1) — n/a: Makes legacy transport reasoning honor configuration and sends top-level effort to otherwise unsupported providers; closed as implemented through the live profile path by 67df958dbe06bb10ca16b8686c76baf0de3bac03.
• fix(transports): promote Ollama reasoning field to content when content is empty #46489 best fix — (+66/-3) — n/a: Recorded best fix for Fix: Ollama reasoning models return empty content — send reasoning_effort to disable thinking #46131, extracting Ollama reasoning and promoting it to visible content when content is blank; keep-open review requires preserving reasoning separation or proving a provider-scoped final-answer contract end to end.
• fix: respect reasoning_config effort for custom providers #49607 [closed] partial — (+85/-1) — n/a: Reads configured effort and adds a config-based custom-model reasoning opt-in in legacy paths; closed because its author found provider normalization prevented the approach from working end to end.
• feat(custom-provider): forward session_id as x-session-id HTTP header #52413 related — (+132/-2) — n/a: Adds x-session-id forwarding to the custom profile and tests profile/transport composition; keep-open review supports salvaging the header while dropping the stale custom:* lookup hunk.
• fix(ollama): preserve cloud model suffixes in picker #52613 partial — (+38/-18) — n/a: Preserves suffixed models.dev IDs while deduplicating by normalized keys; keep-open review says this reverses hosted-API normalization and must distinguish hosted versus local-proxy endpoints.
• fix(custom-provider): pass through reasoning_effort to extra_body for OpenAI-compatible backends #55280 [closed] partial — (+121/-2) — n/a: Emits enabled reasoning as extra_body.reasoning from CustomProfile; closed because merged fix(custom-provider): forward reasoning_effort at the live profile path (GLM-5.2/ARK) #58156 implemented the issue on the live path using top-level reasoning_effort instead.
• fix: pass through reasoning_effort for custom/vLLM/Ollama providers (#55276) #55428 [closed] partial — (+13/-1) — n/a: Emits think=True and effort while widening a legacy reasoning gate; closed because merged fix(custom-provider): forward reasoning_effort at the live profile path (GLM-5.2/ARK) #58156 implemented safer top-level effort without forcing Ollama-specific think=True.
• fix: forward reasoning_effort for custom providers (GLM-5.2 on ARK) #57601 [closed] partial — (+21/-2) — n/a: Adds max, removes hardcoded medium effort, and attempts custom forwarding in an unreachable legacy branch; closed after salvage into merged fix(custom-provider): forward reasoning_effort at the live profile path (GLM-5.2/ARK) #58156 relocated the fix to CustomProfile.
• fix(custom-provider): forward reasoning_effort at the live profile path (GLM-5.2/ARK) #58156 [merged] best fix — (+153/-11) — n/a: Merged reference implementation for enabled custom-provider effort on the live profile path, including max support and configured legacy fallback effort; later reports show unconditional emission is incompatible with some heterogeneous endpoints.
• fix(custom-profile): gate reasoning_effort on known reasoning-capable models (#59660) #59678 partial — (+187/-4) — n/a: Gates enabled effort through a model-name allowlist to avoid non-reasoning-model 400s; keep-open review requires delimiter-safe matching and a transport-level fallback regression.
• fix(ollama-cloud): preserve :cloud tags for local-daemon proxy in model catalog #59819 best fix — (+108/-44) — n/a: Recorded best fix for [Bug]: Ollama Cloud model picker shows local models instead of cloud models #52599, preserving or synthesizing cloud tags for local-daemon routing while retaining hosted IDs; keep-open review requires endpoint-aware caching before this salvage is safe.
• fix: gate Ollama think/reasoning_effort on model's declared capabilities #63315 best fix — (+260/-5) — n/a: Recorded best fix for Bug: CustomProfile (Ollama fallback) forwards reasoning_effort to non-reasoning models, causing HTTP 400 'does not support thinking' #59660, probing Ollama /api/show capabilities before sending reasoning controls; keep-open review requires distinguishing probe failure from a successful non-thinking declaration.
• fix(providers): scope think to Ollama endpoints #64286 best fix — (+245/-58) — n/a: Recorded best fix for [Bug]: think=False incorrectly sent to all provider=custom endpoints, not just Ollama #11237, Bug: CustomProfile sends reasoning_effort="none" to remote APIs, causing HTTP 400 #65233, and custom provider: unconditional top-level reasoning_effort is fatal on OpenAI-compatible proxies that reject unknown params #72649: positively identifies Ollama via the shared /api/tags probe, threads API keys through main, auxiliary, retry, and fallback paths, and scopes both disable fields while retaining explicit enabled effort.
• fix(ollama): emit top-level reasoning_effort=none on /v1/chat/completions (#25758) #64608 [merged] best fix — (+21/-5) — n/a: Merged salvage of fix(ollama): emit top-level reasoning_effort=none on /v1/chat/complet… #29820 that adds top-level reasoning_effort=none beside think=false, fixing Ollama /v1; it remains relevant as the reference implementation that target fix(providers): scope think to Ollama endpoints #64286 narrows to verified Ollama endpoints.
• fix(custom-profile): omit reasoning_effort="none" for remote endpoints #65234 [closed] duplicate — (+202/-20) — n/a: Uses endpoint locality plus optional configuration to omit disable fields remotely; closed after review found locality does not establish reasoning capability and requested final-wire coverage.
• fix(custom): move reasoning_effort to extra_body for custom providers #72656 best fix — (+19/-31) — n/a: Recorded best fix for custom provider: unconditional top-level reasoning_effort is fatal on OpenAI-compatible proxies that reject unknown params #72649 by omitting enabled reasoning_effort for heterogeneous custom providers after correcting the ineffective extra_body idea; keep-open review requires the existing profile tests to match that omission contract.
• fix(responses): preserve explicit reasoning none #75234 best fix — (+96/-1) — n/a: Recorded best fix for [Bug]: reasoning_effort: none silently falls back to the model default on OpenAI Responses #75227, preserving explicit none on known OpenAI/Codex GPT-5.6 Responses routes and rejecting unsupported or unknown relays; the keep-open review's endpoint-capability objection is addressed in the shown diff and follow-up tests.

Duplicates

#3197 was absorbed by #10782 and later completed by #64608; #6038 and #10740 were also absorbed by #10782; #11296, #11362, #11395, #11520, #12488, and #12914 are old-path variants of the #11237 fix now best represented by #64286; #16192 was salvaged as #19881 and #16196 was superseded by it; #25866 and #29820 overlap, with #29820 salvaged as #64608; #55280, #55428, and #57601 overlap with merged #58156; #52613 and #59819 share suffix-key deduplication, but #59819 adds the endpoint-specific local-proxy behavior required by its recorded best-fix verdict; #65234 overlaps the disabled-field scope solved more robustly by #64286.

Suggested consolidation

Keep #64286 open with a salvage path: preserve its positive /api/tags Ollama identification, shared scoping of both think=false and reasoning_effort="none", API-key propagation, and generic-relay final-wire tests. This follows the visible keep-open review rather than deviating from it; close #11296, #11362, #11520, and other still-open old-path #11237 variants as duplicates of #64286, while keeping #63315, #59819, #72656, and #75234 separate for their distinct capability, endpoint-shaped catalog, enabled-effort, and Responses issues.

Complex graph

flowchart TD
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    I11237(["issue #11237 (open)"])
    I65233(["issue #65233 (open)"])
    I72649(["issue #72649 (open)"])
    subgraph Dup11296 ["PRs duplicating each other"]
        P11296["PR #11296 (open)"]
        P11362["PR #11362 (open)"]
        P11395["PR #11395 (closed)"]
        P11520["PR #11520 (open)"]
        P12488["PR #12488 (closed)"]
        P12914["PR #12914 (closed)"]
        P64286["PR #64286 (open)"]
        P65234["PR #65234 (closed)"]
        P72656["PR #72656 (open)"]
    end
    P64286 -->|best fix| I11237
    P64286 -->|best fix| I65233
    P64286 -->|best fix| I72649
    class I11237 open
    class I65233 open
    class I72649 open
    class P11296 open
    class P11362 open
    class P11395 closed
    class P11520 open
    class P12488 closed
    class P12914 closed
    class P64286 open
    class P65234 closed
    class P72656 open
    class P64286 best
    class P64286 best
    class P64286 best
    class P72656 best
    class P64286 target
    click I11237 "https://github.com/NousResearch/hermes-agent/issues/11237"
    click I65233 "https://github.com/NousResearch/hermes-agent/issues/65233"
    click I72649 "https://github.com/NousResearch/hermes-agent/issues/72649"
    click P11296 "https://github.com/NousResearch/hermes-agent/pull/11296"
    click P11362 "https://github.com/NousResearch/hermes-agent/pull/11362"
    click P11395 "https://github.com/NousResearch/hermes-agent/pull/11395"
    click P11520 "https://github.com/NousResearch/hermes-agent/pull/11520"
    click P12488 "https://github.com/NousResearch/hermes-agent/pull/12488"
    click P12914 "https://github.com/NousResearch/hermes-agent/pull/12914"
    click P64286 "https://github.com/NousResearch/hermes-agent/pull/64286"
    click P65234 "https://github.com/NousResearch/hermes-agent/pull/65234"
    click P72656 "https://github.com/NousResearch/hermes-agent/pull/72656"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 37 pull requests and 16 issues in this complex. Each diff was read against this issue; Assessment working set: 331 kB of PR diffs, 187 kB of issue/PR text, 71 kB of discussion (93 comments), 74 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.