Add Fikra API provider: OpenAI-compatible gateway for open-weight models

providers/fikra-api/models/deepseek-v4-flash-0731.toml

@@ -0,0 +1,24 @@
+# Relay of the lab entry behind Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Image content rejected (HTTP 400) live (2026-10-20).
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "deepseek/deepseek-v4-flash-0731"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 1048576
+output = 131072
+
+[cost.text.input]
+price = 0.045
+cache = { read = 0.025 }
+
+[cost.text.output]
+price = 0.17

providers/fikra-api/models/deepseek-v4.1-flash-1017.toml

@@ -0,0 +1,24 @@
+# Relay of the lab entry behind Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Image request accepted (200) live (2026-10-20) but returned empty content; attachment inherits the lab.
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "deepseek/deepseek-v4.1-flash-1017"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 1048576
+output = 393216
+
+[cost.text.input]
+price = 0.05
+cache = { read = 0.025 }
+
+[cost.text.output]
+price = 0.11

providers/fikra-api/models/deepseek-v4.1-flash.toml

@@ -0,0 +1,24 @@
+# Relay of the lab entry behind Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Image input verified live (2026-10-20): image_url content accepted and answered.
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "deepseek/deepseek-v4.1-flash"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 1048576
+output = 384000
+
+[cost.text.input]
+price = 0.05
+cache = { read = 0.025 }
+
+[cost.text.output]
+price = 0.11

providers/fikra-api/models/fikra-auto-pro.toml

@@ -0,0 +1,31 @@
+# Host lane on Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Router lane with no single nameable lab checkpoint, so it stays standalone; reasoning content is not exposed
+# on the wire and `reasoning_effort` is accepted silently (live probe 2026-10-20, data in the PR thread).
+# Image input verified live (2026-10-20) with a minimum image size enforced upstream.
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+name = "Fikra Auto Pro"
+description = "Auto-routed premium lane for long-context agent and coding runs on a 1M context"
+open_weights = false
+release_date = "2026-09"
+last_updated = "2026-10-20"
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[capabilities]
+tools = { supported = true }
+reasoning = { supported = false }
+structured_output = true
+temperature = true
+
+[limit]
+context = 1048576
+output = 256000
+
+[cost.text.input]
+price = 6.0
+
+[cost.text.output]
+price = 12.0

providers/fikra-api/models/fikra-auto.toml

@@ -0,0 +1,31 @@
+# Host lane on Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Router lane with no single nameable lab checkpoint, so it stays standalone; reasoning content is not exposed
+# on the wire and `reasoning_effort` is accepted silently (live probe 2026-10-20, data in the PR thread).
+# Image input verified live (2026-10-20): image_url content accepted and answered.
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+name = "Fikra Auto"
+description = "Auto lane: routes each request to the cheapest fitting lane on Fikra's OpenAI-compatible /v1"
+open_weights = false
+release_date = "2026-09"
+last_updated = "2026-10-20"
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[capabilities]
+tools = { supported = true }
+reasoning = { supported = false }
+structured_output = true
+temperature = true
+
+[limit]
+context = 262144
+output = 65536
+
+[cost.text.input]
+price = 0.045
+
+[cost.text.output]
+price = 0.35

providers/fikra-api/models/fikra-code-35b.toml

@@ -0,0 +1,23 @@
+# Host lane on Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Upstream identity (live probe 2026-10-20): an image-input request is rejected by the host with the verbatim error "model 'qwen3.6-35b-a3b' does not support image input".
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "alibaba/qwen3.6-35b-a3b"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 262144
+output = 65536
+
+[cost.text.input]
+price = 0.045
+
+[cost.text.output]
+price = 0.33

providers/fikra-api/models/fikra-pro-120b.toml

@@ -0,0 +1,23 @@
+# Host lane on Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Upstream identity (live probe 2026-10-20): an image-input request is rejected by the host with the verbatim error "Model openai/gpt-oss-120b does not accept image input".
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "openai/gpt-oss-120b"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 131072
+output = 65536
+
+[cost.text.input]
+price = 0.3
+
+[cost.text.output]
+price = 0.3

providers/fikra-api/models/fikra-pro-20b.toml

@@ -0,0 +1,23 @@
+# Host lane on Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Upstream identity (live probe 2026-10-20): an image-input request is rejected by the host with the verbatim error "Model openai/gpt-oss-20b does not accept image input".
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "openai/gpt-oss-20b"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 131072
+output = 65536
+
+[cost.text.input]
+price = 0.2
+
+[cost.text.output]
+price = 0.2

providers/fikra-api/models/fikra-qwen-27b.toml

@@ -0,0 +1,23 @@
+# Host lane on Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Image input verified live (2026-10-20): image_url content accepted and answered.
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "alibaba/qwen3.8-27b"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 262144
+output = 65536
+
+[cost.text.input]
+price = 0.045
+
+[cost.text.output]
+price = 0.35

providers/fikra-api/models/fikra-qwen-max.toml

@@ -0,0 +1,23 @@
+# Host lane on Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Image input verified live (2026-10-20) with a minimum image size enforced upstream.
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "alibaba/qwen3.8-max"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 1048576
+output = 256000
+
+[cost.text.input]
+price = 6.0
+
+[cost.text.output]
+price = 12.0

providers/fikra-api/models/gemini-4-argon.toml

@@ -0,0 +1,23 @@
+# Relay of the lab entry behind Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Image content rejected (HTTP 400) live (2026-10-20).
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "google/gemini-4-argon"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 1048576
+output = 65536
+
+[cost.text.input]
+price = 2.5
+
+[cost.text.output]
+price = 12.0

providers/fikra-api/models/glm-5.2.toml

@@ -0,0 +1,24 @@
+# Relay of the lab entry behind Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Image content rejected (HTTP 400) live (2026-10-20).
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "zhipuai/glm-5.2"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 262144
+output = 131072
+
+[cost.text.input]
+price = 0.75
+cache = { read = 0.55 }
+
+[cost.text.output]
+price = 1.65

providers/fikra-api/models/glm-5.3-flash.toml

@@ -0,0 +1,24 @@
+# Relay of the lab entry behind Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Image input verified live (2026-10-20): image_url content accepted and answered.
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "zhipuai/glm-5.3-flash"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 327680
+output = 131072
+
+[cost.text.input]
+price = 0.155
+cache = { read = 0.08 }
+
+[cost.text.output]
+price = 0.49

providers/fikra-api/models/glm-5.3.toml

@@ -0,0 +1,24 @@
+# Relay of the lab entry behind Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Image content rejected (HTTP 400) live (2026-10-20).
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "zhipuai/glm-5.3"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 327680
+output = 131072
+
+[cost.text.input]
+price = 1.05
+cache = { read = 0.55 }
+
+[cost.text.output]
+price = 3.35

providers/fikra-api/models/kimi-k3.toml

@@ -0,0 +1,24 @@
+# Relay of the lab entry behind Fikra's OpenAI-compatible /v1 (prices: https://fikraapi.co.ke/pricing, verified 2026-10-19).
+# Control surface (live probe 2026-10-20, raw counts and p-values in the PR thread): the documented request whitelist
+# (https://fikraapi.co.ke/docs, verified 2026-10-19) covers only tools/tool_choice/n/seed/stop/penalties/response_format;
+# `thinking` and `reasoning_effort` return 200 and are accepted silently, and no response on any lane carries a
+# `reasoning_content` field (message keys: content/role only). reasoning_options stays empty until a control is shown to
+# change the wire; effort levels are indistinguishable from sampling noise (MWU n=10-16/level, all p >= 0.11).
+# Image input verified live (2026-10-20): image_url content accepted and answered.
+# Limits: host GET /v1/models `context_window`/`max_output_tokens` (fetched 2026-10-20).
+schema = 2
+base_model = "moonshotai/kimi-k3"
+
+[capabilities]
+reasoning = { supported = true }
+
+[limit]
+context = 1113088
+output = 1048576
+
+[cost.text.input]
+price = 2.1
+cache = { read = 0.6 }
+
+[cost.text.output]
+price = 10.5