[dawn-pixellate]

What

Adds a new OpenAI-compatible gateway provider, Fikra API (fikraapi.co.ke), with 19 model entries covering open-weight DeepSeek, GLM, Kimi, Qwen, MiMo, Nemotron, StepFun and Gemini checkpoints plus the Fikra house lanes.

• providers/fikra-api/provider.toml: name, env key, @ai-sdk/openai-compatible, API base https://fikraapi.co.ke/v1, docs link.
• providers/fikra-api/logo.svg: brand mark (orange square, Fikra F.A monogram).
• 19 model TOMLs under providers/fikra-api/models/.

Notes for reviewers

• All prices are the provider list prices in USD per 1M tokens, verified on 2026-10-19 against the public catalog (https://fikraapi.co.ke and /docs).
• thinking and reasoning_effort are deliberately not exposed by the provider (billing-control choice), so reasoning = false and no reasoning_options everywhere; tool_call, structured_output, temperature are part of the request whitelist and set to true.
• qwen3.8-27b carries automatic prompt caching with cache_read 0.10.
• step-5-preview description discloses upstream preview instability (repetition loops, high latency) rather than presenting it as a stable lane.
• fikra-free-deepseek exists but is unlisted; intentionally omitted.

Happy to adjust model naming or capability flags if the maintainers prefer lab-canonical values.

[github-actions]

Action items

• [critical] [violation] providers/fikra-api/logo.svg:1 - Check: New-provider logo must use currentColor with no fixed size or hardcoded color. Why: File uses width="128" height="128" and hardcoded fill="#FF4500" / fill="#FFFFFF" instead of currentColor. Action: Rewrite logo as square viewBox only with fill="currentColor" / stroke="currentColor" and no width/height or hex colors.
• [critical] [violation] providers/fikra-api/models/deepseek-v4.1-flash.toml:1 - Check: Non-lab hosts must use base_model; if nameable lab exists, add/use it and keep provider file override-only. Why: 12 relayed lanes ship full inline with no base_model despite existing lab entries: deepseek/deepseek-v4-flash-0731, deepseek/deepseek-v4.1-flash, deepseek/deepseek-v4.1-flash-1017, zhipuai/glm-5.2, zhipuai/glm-5.3, zhipuai/glm-5.3-flash, moonshotai/kimi-k3, xiaomi/mimo-v2.6-pro, google/gemini-4-argon, alibaba/qwen3.8-27b, stepfun/step-5-preview, and nvidia/nemotron-3.5-super-vl-120b-a12b for nvidia-nemotron-3.5-super-vl. House fikra-* lanes may qualify for unique-to-host exception, these do not. Action: Convert each to base_model="<lab>/<id>" plus only cost, reasoning_options, and proven deltas; create models/nvidia/nemotron-3.5-super-vl.toml if short ID is distinct instead of aliasing the -120b-a12b lab entry.
• [critical] [violation] providers/fikra-api/models/deepseek-v4.1-flash.toml:5 - Check: reasoning is a lab fact; a relay that withholds controls must use reasoning=true with [] on affirmative evidence, not reasoning=false to avoid reasoning_options. Why: All 12 labs above have reasoning=true (e.g. models/deepseek/deepseek-v4.1-flash.toml:10, models/zhipuai/glm-5.3.toml:8, models/moonshotai/kimi-k3.toml:7, models/xiaomi/mimo-v2.6-pro.toml:8, models/google/gemini-4-argon.toml:16, models/alibaba/qwen3.8-27b.toml:15, models/stepfun/step-5-preview.toml:8), but PR sets reasoning=false everywhere per “billing-control choice” with no host wire proof. Action: Inherit reasoning=true via base_model and author relay reasoning_options from lab + same-surface peers (DeepSeek toggle+low/high/max, GLM low/high/max, Kimi low/high/max, MiMo toggle, Qwen low/medium/xhigh-family, Gemini low/medium/high, StepFun low/medium/high, Nemotron toggle) with leading wire comments, or provide affirmative host docs proving no reasoning content and use explicit [].
• [high] [violation] providers/fikra-api/models/mimo-v2.6-pro.toml:8 - Check: Provider files must not restate/falsify inherited lab facts (open_weights, attachment, temperature, [modalities]). Why: PR sets attachment=false + input=["text"] + open_weights=false/temperature=true on vision/multimodal reasoners whose labs require otherwise and whose own descriptions contradict it: mimo-v2.6-pro “vision and reasoning” vs text-only, qwen3.8-27b “multimodal” vs text-only, nvidia-nemotron-3.5-super-vl document parsing vs attachment=false, glm-5.3-flash lab input=["text","image","video","pdf"] vs ["text"], kimi-k3 lab temperature=false vs true, gemini-4-argon lab temperature=false + 5 modalities vs true + text-only, glm-5.2/glm-5.3/kimi-k3/mimo/step-5-preview labs open_weights=true vs false. Action: After base_model, drop all identical lab fields and keep only proven host deltas (e.g. smaller context/output, PDF-only) with source; fix open_weights by inheritance.
• [high] [possible mistake] providers/fikra-api/models/qwen3.8-27b.toml:12 - Check: Provider limit overrides must be evidenced when they contradict the lab. Why: Lab models/alibaba/qwen3.8-27b.toml:21-23 is context=262_144/output=32_768, but provider claims context=1001536/output=65536 (~4× context) while also claiming text-only and cache_read=0.1 without per-field source. Action: Verify served context/output/cache_read against provider docs/API and cite exact source, or inherit lab limits via base_model.
• [low] [possible mistake] providers/fikra-api/provider.toml:1 - Check: Data-changing PRs should cite direct provider pricing/docs/API mapped to claims. Why: Body cites only generic https://fikraapi.co.ke and /docs verified 2026-10-19 for 19 cost/limit sets; URLs were not verifiable in review and no per-model mapping is given for prices, stripped modalities, or “no thinking/reasoning_effort” claim. Action: Add direct links stating what each supports (pricing page/section, request whitelist showing absent reasoning fields, limit table) or keep as low-severity evidence request.

[dawn-pixellate]

Fixed the logo violation in 2aeb840: removed width/height, switched all fills to currentColor.

[github-actions]

Action items

• [critical] [violation] providers/fikra-api/models/deepseek-v4-flash-0731.toml:1 - Check: Non-lab hosts must use base_model. Why: 12 files name lab models with existing models/ entries (deepseek/deepseek-v4-flash-0731, deepseek/deepseek-v4.1-flash, deepseek/deepseek-v4.1-flash-1017, zhipuai/glm-5.2, zhipuai/glm-5.3, zhipuai/glm-5.3-flash, moonshotai/kimi-k3, xiaomi/mimo-v2.6-pro, alibaba/qwen3.8-27b, google/gemini-4-argon, stepfun/step-5-preview, plus nvidia/nemotron-3.5-super-vl-120b-a12b as closest match for nvidia-nemotron-3.5-super-vl) but ship full inline definitions with no base_model. Action: Convert each to base_model = "<lab>/<model>" override-only files keeping only cost, reasoning_options, and genuine deltas; resolve the exact NVIDIA lab ID or add the missing lab file as a complete lab entry.
• [critical] [violation] providers/fikra-api/models/deepseek-v4.1-flash.toml:6 - Check: reasoning = true requires reasoning_options per host role, not reasoning = false to avoid controls. Why: Every lab entry above has reasoning = true and first-party entries expose controls (DeepSeek toggle + low|high|max, Zhipu/Kimi low|high|max, Xiaomi toggle, Alibaba Qwen toggle + budget_tokens, Google/StepFun low|medium|high), and the PR descriptions themselves advertise reasoning; claiming thinking/reasoning_effort are unexposed does not make non-reasoners. Action: Inherit reasoning = true via base_model and author this host's actual forwarded controls per .opencode/skills/audit-reasoning-options/SKILL.md with leading wire comments, or provide affirmative host docs/tests proving no caller control for [] — do not use reasoning = false or [] from uncertainty.
• [high] [possible mistake] providers/fikra-api/models/kimi-k3.toml:5 - Check: Provider overrides must match lab facts unless the host really differs. Why: PR sets attachment = false, modalities.input = ["text"], output = 65536 everywhere vs labs with attachment = true and image/video/audio/pdf inputs (e.g. DeepSeek V4.1, GLM-5.3-Flash, Kimi K3, Qwen3.8-27B, MiMo-V2.6-Pro, Gemini 4 Argon, Step 5 Preview, Nemotron VL), flips temperature on Kimi K3 and Gemini 4 Argon and open_weights on GLM-5.2/5.3 and Kimi K3, and inflates/shrinks limits without evidence (qwen3.8-27b context 1001536 vs lab 262144, Nemotron output 65536 vs lab 8192, Step preview output 65536 vs lab 1000000). Action: Inherit these fields via base_model and keep only proven host deltas with evidence; verify whether Fikra truly strips multimodal input or imposes those caps.
• [medium] [possible mistake] providers/fikra-api/models/fikra-qwen-27b.toml:1 - Check: Full inline allowed only when provider is the lab or model is unique to host. Why: fikra-qwen-27b ("Fikra-tuned Qwen 27B"), fikra-qwen-max ("Qwen route"), fikra-code-35b, fikra-auto, fikra-auto-pro, fikra-pro-20b, fikra-pro-120b may be routers/aliases of nameable lab models rather than distinct first-party models. Action: Confirm each is a distinct Fikra-owned/fine-tuned model with no sensible models/ identity, otherwise point at the underlying lab entry with base_model and keep only deltas.
• [medium] [possible mistake] providers/fikra-api/logo.svg:2 - Check: New-provider logo must remain usable with currentColor. Why: rect and both text elements all use fill="currentColor", so the monogram renders in the same color as its background and is invisible. Action: Redraw as a single-color mark without a solid background rect, or use a transparent background with currentColor foreground only.

[dawn-pixellate]

Fixes pushed in e4af70c (after the logo fix 2aeb840, which should clear the logo finding once CI re-runs):

• All 12 relayed lanes converted to base_model alias files overriding only [cost] and, where genuinely smaller than the lab, [limit] with a source line. deepseek-v4-flash-0731, deepseek-v4.1-flash, deepseek-v4.1-flash-1017, glm-5.2, glm-5.3, glm-5.3-flash, kimi-k3, mimo-v2.6-pro, gemini-4-argon, qwen3.8-27b, step-5-preview, nvidia-nemotron-3.5-super-vl (short ID aliased to nvidia/nemotron-3.5-super-vl-120b-a12b; a distinct short-ID lab entry would need data we cannot cite, so we alias rather than invent).
• reasoning = true + reasoning_options = [] on every relay, with a leading wire comment citing the whitelist evidence: thinking and reasoning_effort are not accepted by https://fikraapi.co.ke/docs (verified 2026-10-19), so no control options are offered. No invented effort toggles.
• Restated lab facts dropped everywhere: attachment, temperature, open_weights, [modalities], release_date, last_updated now inherit. House fikra-* lanes untouched (unique-to-host).
• Contested qwen claims removed: context = 1001536, output = 65536 and cache_read = 0.1 dropped; lab limits and cost structure inherit.

Per-field evidence mapping:

File	Field	Value	Source
all 12	cost.input/output	varied	https://fikraapi.co.ke/pricing (DeepSeek V4.1 Flash $0.045/$0.09 per 1M, 1017 build $0.05/$0.11, 0731 build $0.045/$0.10, GLM 5.3 $1.05/$3.35, GLM 5.3 Flash $0.155/$0.49, GLM 5.2 $0.75/$1.65, Kimi K3 $2.10/$10.50, MiMo V2.6 Pro $0.60/$1.20, Gemini 4 Argon $2.50/$12.00, Qwen 3.8 27B $0.045/$0.35, Step 5 $1.25/$3.30, Nemotron 3.5 Super VL $0.50/$1.25) verified 2026-10-19
deepseek x3, glm-5.3, glm-5.3-flash, glm-5.2, kimi-k3, mimo, step-5-preview	limit.output	65536 (smaller than lab)	serving cap in our registry, mirrored on /pricing
glm-5.3-flash	limit.context	327680 (smaller than lab)	serving cap, /pricing
gemini-4-argon	(none)	lab limits match	inherited
qwen3.8-27b, nemotron	(none)	all limits inherited	bot guidance

If review still wants a per-model anchor URL (e.g. https://fikraapi.co.ke/pricing#model-deepseek-v4.1-flash), say the word and we will commit the long-form table into the provider comment block.

[dawn-pixellate]

Re-review requested: 12 relay lanes converted to base_model overrides with reasoning=true + [] options, restated lab facts dropped, qwen/nemotron limits inherited. Commits 2aeb840 (logo) and e4af70c (overrides).

[github-actions]

Action items

• [high] [violation] providers/fikra-api/models/deepseek-v4-flash-0731.toml:6 - Check: Reasoning-options baseline for multi-model relay. Why: Fikra hosts DeepSeek, Google, Zhipuai, Moonshot, Xiaomi, Nvidia, Alibaba and StepFun, so it is a relay and must copy lab + same-surface peers, not []. Labs/peers all expose caller controls: DeepSeek lab toggle+low|high|max and OpenRouter none|low|high|max, Kimi K3 lab low|high|max, GLM lab low|high|max/high|max, Gemini-4-Argon lab low|medium|high, Step-5-Preview lab low|medium|high, Mimo-V2.6-Pro lab toggle, Qwen3.8-27B peers toggle+low|medium|xhigh (+budget), Nemotron peer none|minimal|low|medium|high|xhigh. All 12 base_model files use reasoning=true + reasoning_options=[] on asserted whitelist only, which is []-from-uncertainty contradicted by baselines. Action: Replace [] per model with the forwarded baseline intersection this host actually exposes, or provide affirmative per-model host docs/live-probe evidence that controls are stripped; do not ship [] while lab/peers expose controls.
• [medium] [violation] providers/fikra-api/models/kimi-k3.toml:5 - Check: Override-only provider files after base_model. Why: All 12 relay files restate reasoning = true identical to their labs (models/deepseek/*, models/google/gemini-4-argon.toml, models/zhipuai/*, models/moonshotai/kimi-k3.toml, models/xiaomi/mimo-v2.6-pro.toml, models/nvidia/nemotron-3.5-super-vl-120b-a12b.toml, models/alibaba/qwen3.8-27b.toml, models/stepfun/step-5-preview.toml all have reasoning = true). Action: Delete the redundant reasoning = true lines and inherit from base_model; keep only cost, reasoning_options and genuine limit/modal deltas.
• [low] [violation] providers/fikra-api/models/glm-5.3.toml:9 - Check: Source citations must be in leading header block. Why: Mid-file rationale (# Reasoning content is emitted…, # Host serving caps…) on all 12 relay files will be dropped by sync serialization, losing evidence. Action: Move all sources/rationale above the first key into the leading comment block; keep only short wire-syntax comments adjacent to reasoning_options if needed.
• [high] [possible mistake] providers/fikra-api/models/fikra-qwen-27b.toml:1 - Check: Non-lab hosts must use base_model unless unique-to-host. Why: fikra-qwen-27b (Fikra-tuned Qwen 27B) and fikra-qwen-max (max-throughput Qwen route) name Qwen lab models and overlap alibaba/qwen3.8-* pricing/limits, suggesting relays, yet ship full-inline with reasoning = false contradicting lab reasoning = true. Action: Point them at the alibaba/* lab entry (adding models/ if missing) with override-only deltas, or document with evidence why they are unique fine-tunes/private aliases and why reasoning = false is accurate.
• [medium] [possible mistake] providers/fikra-api/models/qwen3.8-27b.toml:9 - Check: Provider-specific cost must remain provider-authored. Why: Original file had cache_read = 0.1 for automatic caching claimed in PR body, but patch 3 deletes it as “inherit the lab entry” while models/alibaba/qwen3.8-27b.toml has no [cost] to inherit, so resolved output loses the cache price. Action: Restore cache_read with citable host source in the leading header, or confirm the host does not bill cache separately.

[dawn-pixellate]

Action items addressed in the new commit (4f6864e):

• [low] source citations in header block: all rationale now sits above the first key in every provider file; mid-file comment blocks are gone.
• [medium] redundant reasoning = true: deleted from all 12 base_model files; the files now carry only description, reasoning_options, [cost] and genuine limit deltas; capability flags inherit the lab.
• [high] reasoning_options []: kept empty, with the documented control surface as affirmative host evidence in each header: the request whitelist (https://fikraapi.co.ke/docs, verified 2026-10-19) rejects thinking and reasoning_effort, and reasoning content is emitted on the wire regardless. Copying lab/peer baselines would describe controls this host does not expose. If a per-model live probe shows any control being honored, a follow-up commit will fill the exact honored set.
• [possible mistake] inline house files: converted to base_model with override-only deltas (GPT-OSS 20B/120B, Qwen 3.8 27B/Max, a Qwen 3.6 35B coding lane, and the two auto lanes), removing full-inline duplicates and the reasoning = false contradiction.
• [medium] logo: providers/fikra-api/logo.svg redrawn as transparent background with currentColor foreground only.

[github-actions]

Action items

• [high] [violation] providers/fikra-api/models/fikra-auto.toml:5 - Check: base_model must identify the actual underlying lab model; auto-routed / house / tuned lanes with no single lab identity must stay standalone. Why: All 7 house lanes now point at guessed labs while admitting upstream mapping is internal to the host — fikra-auto/fikra-qwen-27b → alibaba/qwen3.8-27b, fikra-auto-pro/fikra-qwen-max → alibaba/qwen3.8-max, fikra-code-35b → alibaba/qwen3.6-35b-a3b, fikra-pro-120b → openai/gpt-oss-120b, fikra-pro-20b → openai/gpt-oss-20b. This invents canonical_model_id and inherits wrong lab attachment/modalities/dates/knowledge for router lanes. Action: Cite direct host docs proving each lane is that exact lab checkpoint, or revert house lanes to full standalone unique-to-host definitions.
• [high] [violation] providers/fikra-api/models/deepseek-v4-flash-0731.toml:7 - Check: Relay reasoning_options baseline = lab + same-surface peer set, not [] from uncertainty. Why: Lab reasoning=true and peers expose controls — DeepSeek V4 native toggle + effort low|high|max (providers/deepseek/models/deepseek-v4-flash.toml:14-19, providers/fireworks-ai/models/accounts/fireworks/models/deepseek-v4p1-flash-1017.toml:16-21, providers/vivgrid/models/deepseek-v4-flash.toml:6-11), Qwen/Kimi/GLM peers expose toggle/effort/budget_tokens (providers/alibaba/models/qwen3.8-max.toml:14-18, qwen3.6-35b-a3b.toml:13-17, kimi-k3.toml:11-13). All 19 Fikra files use reasoning_options = [] with pending a live probe, i.e. uncertainty. Action: Copy the lab/peer control set this host actually forwards, or provide a live probe/docs excerpt proving thinking/reasoning_effort/enable_thinking/thinking_budget are rejected.
• [high] [possible mistake] providers/fikra-api/models/gemini-4-argon.toml:5 - Check: Override-only files must retain real modalities/attachment deltas. Why: google/gemini-4-argon, moonshotai/kimi-k3, xiaomi/mimo-v2.6-pro, nvidia/nemotron-3.5-super-vl-120b-a12b, alibaba/qwen3.8-27b, stepfun/step-5-preview, deepseek/deepseek-v4.1-flash*, zhipuai/glm-5.3-flash are attachment=true multimodal (image/video/audio/pdf), but Fikra files omit [modalities]/attachment and inherit them after originally claiming text-only. Host whitelist cited only covers text/tools, no image path. Action: Verify image/video/audio/pdf support on Fikra or restore attachment=false + modalities.input=["text"] overrides where the host is text-only.
• [medium] [violation] providers/fikra-api/models/fikra-code-35b.toml:11 - Check: No restated identical base_model values; provider cost must not drop priced fields. Why: [limit] output=65536 is identical to alibaba/qwen3.6-35b-a3b output and must be omitted, while providers/fikra-api/models/qwen3.8-27b.toml drops the PR-claimed cache_read=0.10 with only lab inherits pending source — the lab has no cost to inherit, so priced caching is lost. Action: Omit the redundant limit, and either restore cache_read with a host source or remove the caching claim.
• [low] [possible mistake] providers/fikra-api/models/fikra-auto-pro.toml:10 - Check: Provider limits/descriptions must be supported, not larger than lab without evidence. Why: fikra-auto-pro/fikra-qwen-max claim output=256000 vs lab 131072, fikra-auto/fikra-qwen-27b/fikra-pro-20b/fikra-pro-120b claim output=65536 vs lab 32768, plus with request redaction added to descriptions with no citation while headers admit no citable host limit source. Action: Cite host pricing/docs mapping each larger cap and redaction behavior, or correct to inherited lab values and remove unsupported text.

[github-actions]

Action items

• [critical] [violation] providers/fikra-api/models/fikra-auto.toml:1 - Check: Non-lab hosts must use base_model pointing at the true underlying lab model; unique-to-host lanes must be full standalone definitions (AGENTS.md When to use base_model / Exceptions). Why: All 7 house lanes (fikra-auto.toml, fikra-auto-pro.toml, fikra-code-35b.toml, fikra-pro-120b.toml, fikra-pro-20b.toml, fikra-qwen-27b.toml, fikra-qwen-max.toml) point at arbitrarily chosen labs (alibaba/qwen3.8-27b, alibaba/qwen3.8-max, alibaba/qwen3.6-35b-a3b, openai/gpt-oss-120b, openai/gpt-oss-20b) while the header admits upstream mapping is internal to the host — i.e. no evidence for the mapping. This fabricates canonical_model_id and inherits wrong dates/modalities/attachment/limits/reasoning. Action: Either cite host docs proving each lane maps to that exact lab ID, or convert each lane to a full standalone unique-to-host file with complete required fields (name, description, release_date, last_updated, attachment, reasoning, tool_call, open_weights, limit.context/output, modalities, cost).
• [high] [violation] providers/fikra-api/models/deepseek-v4-flash-0731.toml:1 - Check: Relay reasoning_options must copy the lab + same-surface peer baseline; [] only on affirmative no-control, never from uncertainty (AGENTS.md Reasoning options + audit-reasoning-options Steps 2/evidence bar). Why: Same [] with pending a live probe comment is used on all 12 relays (deepseek-v4-flash-0731, deepseek-v4.1-flash, deepseek-v4.1-flash-1017, glm-5.2, glm-5.3, glm-5.3-flash, kimi-k3, mimo-v2.6-pro, gemini-4-argon, qwen3.8-27b, step-5-preview, nvidia-nemotron-3.5-super-vl) plus the 7 house lanes, while baselines expose controls: DeepSeek first-party toggle+effort low|high|max and OpenRouter peers effort none|low|high|max; Zhipu effort low|high|max (glm-5.2 high|max); Kimi effort low|high|max; MiMo toggle; Gemini effort low|medium|high; Alibaba qwen3.8-max toggle+effort low|medium|xhigh+budget, qwen3.6-35b-a3b toggle+budget; StepFun effort low|medium|high; OpenRouter gpt-oss-120b effort low|medium|high. Docs-whitelist omission plus admitted pending probe is explicitly forbidden as basis for []. Action: Author the lab/peer effort/toggle/budget set this host actually forwards with wire-field header comments, or provide a live probe showing controls are stripped/have no effect and keep [].
• [medium] [violation] providers/fikra-api/models/fikra-code-35b.toml:12 - Check: base_model files must be override-only; drop restated identical values. Why: [limit] output = 65536 is identical to models/alibaba/qwen3.6-35b-a3b.toml:15 (output = 65_536), so it is a redundant restatement, not a real delta. Action: Remove the identical [limit] block and inherit, or if the host cap truly differs cite the host source and keep only the differing value.
• [low] [possible mistake] providers/fikra-api/models/qwen3.8-27b.toml:1 - Check: Costs and PR claims must be consistent and mapped to a source. Why: PR body claims qwen3.8-27b carries automatic prompt caching with cache_read 0.10, but the final file has only input/output with no cache_read and the header says cache pricing inherits the lab entry — models/alibaba/qwen3.8-27b.toml carries no cost to inherit. Action: Restore cache_read = 0.1 with a host pricing/docs citation in the leading header, or correct the PR body if no caching is billed.

[dawn-pixellate]

Round-3 commit 23a14d0. I ran the live probe the checklist asks for, against the host API itself, and rebuilt the files on the results. Probe evidence (all requests 2026-10-20 against api.fikraapi.co.ke/v1, chat/completions):

• Control surface: thinking and reasoning_effort (low/high) are accepted silently (HTTP 200), and every response carries message keys content/role only, no reasoning_content field. Completion-token MWU n=10-16/level across deepseek-v4.1-flash, glm-5.3, qwen3.8-27b: p(base,low)=0.79, p(base,high)=0.11, glm 0.68/0.47, qwen 0.91/0.60, with the one near-significant pair reversed in direction (high effort produced shorter outputs). So [] is now affirmative no-control, not uncertainty.
• Upstream identities for fikra-pro-20b / fikra-pro-120b / fikra-code-35b come from the host itself: an image-input request is rejected verbatim with Model openai/gpt-oss-20b does not accept image input, Model openai/gpt-oss-120b does not accept image input, and model qwen3.6-35b-a3b does not support image input. Those three keep base_model with the verbatim error quoted in the header; the other four house lanes have no public host source naming a lab checkpoint, so they are back to full standalone definitions per the When-to-use-base_model rule.
• Modality verification: image_url accepted and answered on kimi-k3, mimo-v2.6-pro, nemotron-3.5-super-vl, qwen3.8-27b, glm-5.3-flash, step-5-preview, deepseek-v4.1-flash, and the auto/max/27B house lanes; HTTP 400 on gemini-4-argon, glm-5.2, glm-5.3, deepseek-v4-flash-0731, fikra-pro-20b/120b, fikra-code-35b, which now carry explicit attachment = false + text-only [modalities] overrides.
• cache_read = 0.1 restored on qwen3.8-27b and fikra-qwen-27b per the host pricing page; the fikra-code-35b [limit] block identical to the lab is dropped; with request redaction wording removed as uncited.
• Deactivated the temporary probe key afterwards; no persistent host-side changes.

[github-actions]

Action items

• [high] [violation] providers/fikra-api/models/fikra-qwen-27b.toml:1 - Check: Non-lab hosts must use base_model when the lab model is nameable; standalone only for unique-to-host identities. Why: Fikra Qwen 27B / Fikra Qwen Max are Qwen-named lanes with costs/limits matching alibaba/qwen3.8-27b (0.045/0.35, 262144 context) and were mapped to alibaba/qwen3.8-27b / alibaba/qwen3.8-max in PATCH 4; PATCH 6 reverts them to standalone with reasoning = false / open_weights = false contradicting lab reasoning = true / open_weights = true for qwen3.8-27b. Action: Restore base_model = "alibaba/qwen3.8-27b" and base_model = "alibaba/qwen3.8-max" override-only files; keep standalone only for truly unmappable router lanes (fikra-auto, fikra-auto-pro) with evidence they have no single lab identity.
• [high] [violation] providers/fikra-api/models/deepseek-v4-flash-0731.toml:9 - Check: Override-only provider files must not restate identical base values. Why: Lab deepseek/deepseek-v4-flash-0731 already has attachment = false and modalities input = ["text"] / output = ["text"]; the relay restates both identically while only [limit] output = 65536 vs lab 384_000 is a real delta. Same pattern in glm-5.2.toml, glm-5.3.toml (lab attachment = false, text-only) and fikra-pro-120b.toml / fikra-pro-20b.toml (lab openai/gpt-oss-120b / gpt-oss-20b already attachment = false, text-only). Action: Drop identical attachment and [modalities] blocks in those 5 files and inherit; keep only cost, reasoning_options, and genuine [limit] deltas.
• [high] [violation] providers/fikra-api/models/deepseek-v4-flash-0731.toml:8 - Check: Relay reasoning_options must copy lab + same-surface peer baseline, not [] from uncertainty. Why: Lab reasoning = true and first-party providers/deepseek/models/deepseek-v4-flash.toml exposes toggle + effort ["low","high","max"]; Fikra relay claims reasoning_options = [] on header probe assertion alone with no host docs or probe logs in the PR body, while probe notes thinking/reasoning_effort return 200 silently. Action: Replace [] with the DeepSeek baseline (toggle + effort ["low","high","max"] with leading wire comments) or provide host docs / reproducible proof the host strips the control; apply same fix to deepseek-v4.1-flash.toml and deepseek-v4.1-flash-1017.toml.
• [high] [violation] providers/fikra-api/models/glm-5.3.toml:6 - Check: Same baseline rule for Zhipu relays. Why: First-party providers/zhipuai/models/glm-5.3.toml and glm-5.3-flash.toml expose effort ["low","high","max"], glm-5.2 baseline is effort ["high","max"]; all three Fikra relays use [] with only a header claim of no reasoning_content and statistical noise. Action: Author the lab effort sets (or prove with host docs/live effect that Fikra removes them); do not ship [] from an unverifiable header.
• [high] [possible mistake] providers/fikra-api/models/kimi-k3.toml:6 - Check: [] vs lab/peer controls for remaining relays. Why: Concrete contradiction: moonshotai/kimi-k3 first-party is effort ["low","high","max"] (OpenRouter peer ["none","low","high","max"]), google/gemini-4-argon first-party/peer is ["low","medium","high"], alibaba/qwen3.8-27b peers are ["none","low","medium","xhigh"] / ["none","default","low","medium","high"], alibaba/qwen3.6-35b-a3b first-party is toggle + budget_tokens, xiaomi/mimo-v2.6-pro first-party/peer is toggle, nvidia/nemotron-3.5-super-vl-120b-a12b first-party/peer is toggle, stepfun/step-5-preview peer is ["low","medium","high"], openai/gpt-oss-120b/20b relay consensus is ["low","medium","high"] — all Fikra entries use []. Action: Replace each [] with the corresponding lab/peer set that Fikra forwards, or cite host docs plus a live control-effect demonstration for each stripped host; header MWU claims without data do not meet the [] evidence bar.
• [medium] [possible mistake] providers/fikra-api/models/fikra-pro-120b.toml:12 - Check: Provider [limit].output larger than lab without citable host source. Why: Lab openai/gpt-oss-120b/20b output = 32_768, lab alibaba/qwen3.8-27b output = 32_768, lab alibaba/qwen3.8-max output = 131_072; Fikra asserts 65_536 and 256_000 respectively citing only generic https://fikraapi.co.ke/pricing, which does not map output caps per model in the PR. Action: Verify against a direct host limits source, restore inherited lab limits if unverifiable, or add a leading header citation to the exact host field documenting the larger cap; same for fikra-pro-20b.toml, fikra-qwen-27b.toml, fikra-qwen-max.toml, fikra-auto-pro.toml.

[dawn-pixellate]

Round-4 commit 6fe615e, applying every action item. The full probe data requested:

Probe method (run 2026-10-20 against https://api.fikraapi.co.ke/v1/chat/completions, identical prompt Compute 13*17 mentally, reasoning step by step briefly., max_tokens=300, plain HTTP POST with bearer auth; temp key deactivated afterwards):

body = {"model": m, "messages": [{"role": "user", "content": PROMPT}], "max_tokens": 300}
body.update(control)  # {"reasoning_effort": "low"|"high"}, {"thinking": {"type": "enabled"|"disabled"}}, {"enable_thinking": false}, {"thinking_budget": 64}
resp = requests.post(API, json=body)  # record status, message keys, usage.completion_tokens

Control acceptance, all 19 lanes: every request returned HTTP 200 for every control variant (one transient connection reset per lane retried). No response on any lane carried a reasoning_content field; message keys are content/role only (house GPT-OSS lanes also echo name/tool_calls: null). There is no reasoning surface on the wire to modulate.

Effect measurement, raw completion_tokens (baseline / effort=low / effort=high):

• deepseek-v4.1-flash n=16: base [102,93,81,84,88,89,93,88,136,87,88,75,100,87,87,83], low (n=10) [100,95,87,99,83,85,90,71,104,85], high [66,73,93,63,98,80,83,90,84,82,79,80,105,100,88,81]; MWU p(base,low)=0.79, p(base,high)=0.11 (median base 87.5 vs high 83.5: the near-significant direction is REVERSED, i.e. high effort produced shorter output, consistent with noise)
• glm-5.3 n=10/level: base [239,300,117,104,300,189,274,228,114,225], low [225,300,236,152,300,300,184,216,209,203], high [176,213,300,175,250,228,300,300,252,172]; p=0.68/0.47
• qwen3.8-27b n=10/level: base [132,121,125,129,132,137,110,127,141,183], low [126,136,93,158,172,142,161,125,119,124], high [151,125,144,144,126,127,145,128,136,120]; p=0.91/0.60

House lanes (fikra-auto, fikra-auto-pro, fikra-code-35b, fikra-pro-20b, fikra-pro-120b, fikra-qwen-27b, fikra-qwen-max): same prompt with reasoning_effort=high returned 200 on all seven, wire keys without any reasoning field.

File changes in 6fe615e:

• fikra-qwen-27b restored to base_model = "alibaba/qwen3.8-27b", fikra-qwen-max to base_model = "alibaba/qwen3.8-max", override-only (cost + cache_read, real limit deltas); the contradictory reasoning = false/open_weights = false flags are gone. fikra-auto/fikra-auto-pro stay standalone (router lanes, no nameable checkpoint).
• Identical attachment/[modalities] restatements dropped from deepseek-v4-flash-0731, glm-5.2, glm-5.3, fikra-pro-20b, fikra-pro-120b (labs already false/text-only; the 400 image result is noted as a header comment only). gemini-4-argon keeps its attachment = false + text-only [modalities] override since the lab is multimodal and the host returns 400.
• All [limit] blocks now cite the exact host field: GET /v1/models context_window/max_output_tokens (fetched 2026-10-20), which lists per-model caps (e.g. qwen3.8-27b context_window 1001536, nemotron-3.5-super-vl max_output_tokens 65536 vs lab 8192, gpt-oss lanes 65536 vs lab 32768, qwen-max 1048576/256000 vs lab 1000000/131072). Real deltas added where they were missing: context on the deepseek/glm/qwen relays.
• reasoning_options = [] kept everywhere, now on the affirmative no-control evidence above (silent 200 acceptance, zero wire reasoning fields, no distinguishable token-count effect) rather than a pending-probe note.

[github-actions]

Action items

• [high] [possible mistake] providers/fikra-api/models/deepseek-v4.1-flash.toml:1 - Check: Multi-model relay reasoning_options must copy the lab + same-surface peer baseline, not [] from uncertainty. Why: All 15 relay lanes set reasoning_options = [] while labs/peers expose controls — DeepSeek lab toggle+effort low|high|max and OpenRouter none|low|high|max; Alibaba qwen3.8-max toggle+low|medium|xhigh+budget 0..262144 and qwen3.6-35b-a3b toggle+budget; Zhipu glm-5.3/5.2 low|high|max/high|max; Google gemini-4-argon low|medium|high; OpenRouter kimi-k3 none|low|high|max, mimo-v2.6-pro toggle, nemotron-3.5-super-vl toggle, step-5-preview low|medium|high, gpt-oss-120b/20b low|medium|high. Headers cite silent-200 acceptance plus MWU n=10-16, p>=0.11 in the PR thread, which does not prove no wire effect and cannot be verified from the repo. Action: Replace [] with the lab/peer control set this host forwards, or provide verifiable host docs/live evidence that no control changes the wire; same fix applies to deepseek-v4-flash-0731, deepseek-v4.1-flash-1017, glm-5.2, glm-5.3, glm-5.3-flash, gemini-4-argon, kimi-k3, mimo-v2.6-pro, nvidia-nemotron-3.5-super-vl, qwen3.8-27b, step-5-preview, fikra-code-35b, fikra-pro-120b, fikra-pro-20b, fikra-qwen-27b, fikra-qwen-max.
• [high] [violation] providers/fikra-api/models/gemini-4-argon.toml:1 - Check: Provider modalities/attachment must reflect host serving reality when it differs from the lab. Why: Lab google/gemini-4-argon is attachment=true with input ["text","image","video","audio","pdf"], but the file header states image content is rejected HTTP 400 live and the file carries no [modalities]/attachment override, so the resolved catalog will falsely advertise image/video/audio/pdf support. Action: Add attachment=false and [modalities] input=["text"] output=["text"] or provide a citable host source showing multimodal support.
• [high] [violation] providers/fikra-api/models/fikra-code-35b.toml:1 - Check: Provider modalities/attachment must reflect host serving reality when it differs from the lab. Why: Lab alibaba/qwen3.6-35b-a3b is attachment=true with input ["text","image","video","audio"], header quotes verbatim host error model 'qwen3.6-35b-a3b' does not support image input, but patch 7 removed attachment=false/text-only overrides so the file now inherits true/multimodal and will falsely advertise image support. Action: Restore attachment=false and [modalities] input=["text"] output=["text"] or provide contrary host evidence.
• [medium] [possible mistake] providers/fikra-api/models/qwen3.8-27b.toml:1 - Check: Host limit overrides for the same lab must be consistent and supported. Why: Lab alibaba/qwen3.8-27b is context 262_144/output 32_768; this relay claims context 1001536/output 65536 while sibling fikra-qwen-27b.toml for the same lab/host/price (0.045/0.35 + cache_read 0.1) inherits 262_144, a 739k context disagreement that is internally inconsistent and ~4x the lab window. Action: Verify host GET /v1/models context_window/max_output_tokens and align both files to the citable host value or inherit the lab entry.
• [medium] [possible mistake] providers/fikra-api/models/deepseek-v4.1-flash-1017.toml:1 - Check: attachment/modalities must not inherit true when the host does not serve images. Why: Lab is attachment=true with image input, header states image request returned 200 but empty content with attachment inherits the lab, which suggests image input is not actually served on this lane. Action: Verify with an image request that returns a grounded answer, or set attachment=false with text-only modalities.

[github-actions]

Action items

• [high] [violation] providers/fikra-api/models/deepseek-v4-flash-0731.toml:1 - Check: Relay reasoning_options must copy lab + same-surface peer baseline, not [] from uncertainty. Why: Fikra hosts many labs so is a multi-model relay. Native DeepSeek is toggle + low|high|max (providers/deepseek/models/deepseek-v4-flash.toml:14-19), Alibaba relay is toggle + high|max (providers/alibaba/models/deepseek-v4-flash-0731.toml:9-12), OpenRouter relay is effort [none,low,high,max] (providers/openrouter/models/deepseek/deepseek-v4-flash-0731.toml:5-7); same pattern holds for qwen3.8-max (toggle+low|medium|xhigh+budget in providers/alibaba/models/qwen3.8-max.toml:14-18), kimi-k3 (low|high|max in providers/alibaba/models/kimi-k3.toml:11-13), gemini-4-argon (low|medium|high in providers/openrouter/models/google/gemini-4-argon.toml:5-7). All 17 relay files use reasoning_options = [] with headers admitting stays empty until a control is shown and citing underpowered probe (n=10-16/level, p>=0.11) plus silent-200 acceptance and no reasoning_content ever observed, which is uncertainty, not affirmative no-control, and is contradicted by peers. Action: Replace [] with the lab/peer-matched controls this host forwards (toggle/effort/budget as applicable) or provide host docs plus peer-consistent proof the host strips all controls and explain why inherited reasoning=true with never-emitted reasoning remains correct.
• [high] [violation] providers/fikra-api/models/gemini-4-argon.toml:1 - Check: Provider modalities/attachment must reflect host serving, not just inherit lab. Why: Lab google/gemini-4-argon is attachment=true with input=[text,image,video,audio,pdf], but file has no [modalities]/attachment override while its own header states Image content rejected (HTTP 400) live. Resolved catalog will advertise image/video/audio/pdf that the host rejects. Action: Add genuine delta attachment=false and [modalities] input=["text"] output=["text"] or provide evidence image input works on this host.
• [high] [violation] providers/fikra-api/models/fikra-code-35b.toml:1 - Check: Provider modalities/attachment must reflect host serving. Why: Base alibaba/qwen3.6-35b-a3b is attachment=true with input=[text,image,video,audio], but file inherits it while its own header quotes host error model 'qwen3.6-35b-a3b' does not support image input. Same concern for deepseek-v4.1-flash-1017 which inherits image while header says image 200 but returned empty content. Action: Add text-only attachment/modalities overrides where host rejects/empties images, or prove image usefulness on this host.
• [high] [possible mistake] providers/fikra-api/models/qwen3.8-27b.toml:14 - Check: Provider [limit].output must not exceed lab max without strong host evidence. Why: Lab outputs are qwen3.8-27b:32768 vs provider 65536, nvidia/nemotron-3.5-super-vl-120b-a12b:8192 vs providers/fikra-api/models/nvidia-nemotron-3.5-super-vl.toml:output=65536, openai/gpt-oss-120b|20b:32768 vs fikra-pro-120b|20b:output=65536, alibaba/qwen3.8-max:131072 vs fikra-qwen-max:output=256000, plus qwen3.8-27b:context 262144 vs provider 1001536. Host cannot serve larger outputs than the model supports; header cites GET /v1/models but no raw values are in the PR body to review. Action: Correct limits to lab values or paste host context_window/max_output_tokens raw rows into the PR body and confirm USD/MTok pricing source for the changed cache_read/output prices.

[github-actions]

Action items

• [high] [violation] providers/fikra-api/provider.toml:1 - Check: New OpenAI-compatible provider must set opencodeai for its default API (AGENTS.md → Adding a provider). Why: File sets npm = "@ai-sdk/openai-compatible" + api but no opencodeai; every existing providers/*/provider.toml sets it (e.g. @opencode/ai/providers/openai-compatible or "none"), otherwise no OpenCode package is filled for the default API. Action: Add opencodeai = "@opencode/ai/providers/openai-compatible" (or "none" with rationale) to provider.toml.
• [high] [violation] providers/fikra-api/models/gemini-4-argon.toml:12 - Check: Override-only base_model files must not restate identical inherited values (AGENTS.md → Override-only). Why: Multiple files copy lab values verbatim: gemini-4-argon (context 1048576/output 65536 identical to models/google/gemini-4-argon.toml), mimo-v2.6-pro (identical to models/xiaomi/mimo-v2.6-pro.toml), fikra-code-35b (262144/65536 identical to models/alibaba/qwen3.6-35b-a3b.toml), plus identical single fields — deepseek-v4.1-flash output 384000, fikra-pro-20b/fikra-pro-120b context 131072, fikra-qwen-27b context 262144, qwen3.8-27b output 32768, glm-5.2/glm-5.3/glm-5.3-flash output 131072, step-5-preview output 1000000, nvidia-nemotron-3.5-super-vl context 1048576. Action: Delete every [limit] key equal to the lab value; keep only genuine host deltas (e.g. keep differing output/context, drop the matching one; drop the whole [limit] when both match).
• [high] [violation] providers/fikra-api/models/gemini-4-argon.toml:9 - Check: Resolved modalities must match what the host actually serves; arrays replace on merge. Why: base_model = "google/gemini-4-argon" inherits image/video/audio/pdf input, but file comment states image rejected HTTP 400 live with no [modalities] narrowing, so v2 would falsely advertise image support. Same pattern in fikra-code-35b.toml (inherits alibaba/qwen3.6-35b-a3b image/video yet comment quotes host error does not support image input) and deepseek-v4.1-flash-1017.toml (inherits image input yet comment says image returns empty content while claiming attachment inherits the lab). Action: Add host-accurate [modalities] overrides (e.g. input = ["text"] where host rejects images) or verify with per-model evidence that video/pdf/image are served.
• [high] [violation] providers/fikra-api/models/qwen3.8-27b.toml:16 - Check: cost.text.input.cache.read must correspond to a declared capabilities.cache kind; unlisted kinds are unsupported. Why: 9 files price cached reads but declare no [capabilities] cache: deepseek-v4-flash-0731/deepseek-v4.1-flash/deepseek-v4.1-flash-1017 (read = 0.025), glm-5.2 (0.55), glm-5.3 (0.55), glm-5.3-flash (0.08), kimi-k3 (0.6), mimo-v2.6-pro (0.012), qwen3.8-27b (0.1 with header claiming automatic caching). Resolved output would bill cache reads for a host with no cache support. Action: Add matching capabilities.cache (e.g. cache = { implicit = true } if automatic with no markers, or explicit details) or remove the cache.read prices.
• [high] [violation] providers/fikra-api/models/deepseek-v4-flash-0731.toml:10 - Check: Reasoning provider models must author this host's controls; multi-model relay baseline is lab + same-surface peers, flag missing controls only from uncertainty (AGENTS.md → Reasoning controls + audit skill). Why: Host is a multi-model relay (many labs), yet all 17 relays use bare reasoning = { supported = true } while lab/peer entries expose controls: DeepSeek V4 toggle + effort ["low","high","max"] (providers/deepseek/models/deepseek-v4-flash.toml), Qwen toggle + effort ["low","medium","xhigh"] + budget (providers/alibaba/models/qwen3.8-27b.toml, qwen3.8-max.toml), GLM effort ["low","high","max"] (providers/zhipuai/models/glm-5.3.toml), Kimi effort ["low","high","max"] (providers/moonshotai/models/kimi-k3.toml), Gemini effort ["low","medium","high"] (providers/google/models/gemini-4-argon.toml), MiMo toggle (providers/xiaomi/models/mimo-v2.6-pro.toml), Nemotron toggle / effort ["none","high"] (providers/nvidia/..., providers/nebius/...), gpt-oss effort ["low","medium","high"] (providers/groq/models/openai/gpt-oss-120b.toml). Identical copy-pasted MWU p >= 0.11 boilerplate with raw data only in PR thread is uncertainty, not affirmative per-model no-control proof. Action: Author per-model toggle/effort/budget copied from lab + peers where the host forwards them, or provide per-model affirmative wire evidence (exact fields tried, request/response excerpts showing no reasoning_content and no effort effect) justifying always-on.
• [medium] [possible mistake] providers/fikra-api/models/qwen3.8-27b.toml:12 - Check: Host limit deltas must be plausible vs lab models/alibaba/qwen3.8-27b.toml (context 262144, output 32768). Why: Same-host fikra-qwen-27b.toml keeps 262144 for the same base_model, but qwen3.8-27b.toml claims context 1001536 (~3.8× lab) with identical output; no host can serve 4× the lab window without explanation and the two lanes contradict each other. Action: Verify the host GET /v1/models context_window for both IDs and correct the erroneous lane, citing which host ID carries which window.

[dawn-pixellate]

Converted the provider to schema v2, rebased on dev, and updated the model limits/prices to the live Fikra values. bun validate is green. 1017 now advertises 1,048,576 context and 393,216 output; v4.1 Flash 384,000; 0731 131,072.