@linxiraos/pi-catalog 1.4.4 → 1.4.6

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (79) hide show
  1. package/CHANGELOG.md +8 -0
  2. package/dist/types/compat/anthropic.d.ts +20 -0
  3. package/dist/types/compat/behavior.d.ts +2 -0
  4. package/dist/types/compat/catalog-policy.d.ts +2 -2
  5. package/dist/types/compat/collapse.d.ts +10 -0
  6. package/dist/types/compat/factory-droid.d.ts +2 -0
  7. package/dist/types/compat/resolve.d.ts +2 -2
  8. package/dist/types/compat/types.d.ts +2 -0
  9. package/dist/types/discovery/antigravity.d.ts +7 -2
  10. package/dist/types/discovery/devin.d.ts +1 -1
  11. package/dist/types/provider-models/google.d.ts +2 -0
  12. package/dist/types/provider-models/openai-compat.d.ts +16 -3
  13. package/dist/types/types.d.ts +58 -3
  14. package/dist/types/wire/devin.d.ts +2 -2
  15. package/dist/types/wire/gemini-headers.d.ts +41 -22
  16. package/dist/types/wire/xiaomi-token-plan.d.ts +2 -0
  17. package/package.json +4 -4
  18. package/src/build.ts +47 -1
  19. package/src/compat/anthropic.d.ts +20 -0
  20. package/src/compat/anthropic.ts +47 -0
  21. package/src/compat/axes.ts +10 -3
  22. package/src/compat/behavior.d.ts +2 -0
  23. package/src/compat/behavior.ts +5 -0
  24. package/src/compat/catalog-policy.ts +10 -8
  25. package/src/compat/collapse.d.ts +10 -0
  26. package/src/compat/collapse.ts +63 -2
  27. package/src/compat/factory-droid.d.ts +2 -0
  28. package/src/compat/factory-droid.ts +3 -0
  29. package/src/compat/resolve.d.ts +2 -2
  30. package/src/compat/resolve.ts +7 -5
  31. package/src/compat/rules/README.md +13 -5
  32. package/src/compat/rules/auth/google-antigravity.kdl +2 -1
  33. package/src/compat/rules/auth/llama.cpp.kdl +1 -1
  34. package/src/compat/rules/auth/xiaomi-token-plan-ams.kdl +1 -1
  35. package/src/compat/rules/auth/xiaomi-token-plan-cn.kdl +1 -1
  36. package/src/compat/rules/auth/xiaomi-token-plan-sgp.kdl +1 -1
  37. package/src/compat/rules/classes/anthropic.kdl +6 -1
  38. package/src/compat/rules/classes/minimax.kdl +10 -0
  39. package/src/compat/rules/providers/anthropic.kdl +15 -1
  40. package/src/compat/rules/providers/deepseek.kdl +13 -10
  41. package/src/compat/rules/providers/factory-droid.kdl +48 -0
  42. package/src/compat/rules/providers/google-antigravity.kdl +7 -0
  43. package/src/compat/rules/providers/kimi-code.kdl +20 -0
  44. package/src/compat/rules/providers/moonshot.kdl +25 -0
  45. package/src/compat/rules/providers/muse-code.kdl +1 -1
  46. package/src/compat/rules/providers/openai-codex.kdl +17 -4
  47. package/src/compat/rules/providers/openai.kdl +23 -2
  48. package/src/compat/rules/providers/opencode-go.kdl +38 -3
  49. package/src/compat/rules/providers/opencode-zen.kdl +7 -0
  50. package/src/compat/rules/providers/xiaomi-token-plan-ams.kdl +1 -1
  51. package/src/compat/rules/providers/xiaomi-token-plan-cn.kdl +1 -1
  52. package/src/compat/rules/providers/xiaomi-token-plan-sgp.kdl +1 -1
  53. package/src/compat/rules/providers/xiaomi.kdl +1 -1
  54. package/src/compat/rules/runtime/behavior.kdl +30 -14
  55. package/src/compat/rules/taxonomy/kimi.kdl +7 -0
  56. package/src/compat/rules.json +1 -1
  57. package/src/compat/types.d.ts +2 -0
  58. package/src/compat/types.ts +2 -0
  59. package/src/discovery/antigravity.d.ts +7 -2
  60. package/src/discovery/antigravity.ts +41 -6
  61. package/src/discovery/cursor.ts +0 -6
  62. package/src/discovery/devin.d.ts +1 -1
  63. package/src/discovery/devin.ts +9 -3
  64. package/src/discovery/factory-droid.ts +38 -6
  65. package/src/model-cache.ts +4 -2
  66. package/src/models.json +1 -1
  67. package/src/provider-models/cache-provider-id.ts +5 -2
  68. package/src/provider-models/google.d.ts +2 -0
  69. package/src/provider-models/google.ts +13 -1
  70. package/src/provider-models/openai-compat.d.ts +16 -3
  71. package/src/provider-models/openai-compat.ts +127 -38
  72. package/src/types.d.ts +58 -3
  73. package/src/types.ts +63 -1
  74. package/src/wire/devin.d.ts +2 -2
  75. package/src/wire/devin.ts +2 -2
  76. package/src/wire/gemini-headers.d.ts +41 -22
  77. package/src/wire/gemini-headers.ts +68 -24
  78. package/src/wire/xiaomi-token-plan.d.ts +2 -0
  79. package/src/wire/xiaomi-token-plan.ts +4 -0
@@ -44,11 +44,31 @@ provider "kimi-code" {
44
44
  // Moonshot list price only when discovery reports $0, so a future priced
45
45
  // envelope keeps its real rates. `kimi-for-coding` is K2.8 Preview — no
46
46
  // public API SKU exists yet, so it uses the K2.7 Code line rate.
47
+ // K3 bills cache writes at the default 5m TTL rate (omp never requests
48
+ // 1h); K2.x SKUs have no cache-write charge.
49
+ // https://platform.kimi.ai/docs/guide/context-caching
47
50
  models "k3" "k3-256k" {
48
51
  cost-fallback {
49
52
  input 3.0
50
53
  output 15.0
51
54
  cache-read 0.3
55
+ cache-write 3.0
56
+ }
57
+ }
58
+ models "k3.1-pro" {
59
+ cost-fallback {
60
+ input 2.0
61
+ output 10.0
62
+ cache-read 0.2
63
+ cache-write 2.0
64
+ }
65
+ }
66
+ models "k3.1-pro-highspeed" {
67
+ cost-fallback {
68
+ input 4.0
69
+ output 20.0
70
+ cache-read 0.2
71
+ cache-write 4.0
52
72
  }
53
73
  }
54
74
  models "kimi-for-coding" {
@@ -14,6 +14,12 @@ provider "moonshot" {
14
14
  thinking-format "zai"
15
15
  family "k3" {
16
16
  thinking-format "openai"
17
+ // K3 and K3.1 Pro default `max_completion_tokens` to 131072 and
18
+ // accept up to the whole 1M window; the shared catalog reports that
19
+ // ceiling as the output limit, leaving no room for input. Pin the
20
+ // documented default, as `model-limits` does for /models rows.
21
+ // https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart
22
+ max-tokens-ceiling 131072
17
23
  }
18
24
  thinking-mode "effort"
19
25
  }
@@ -44,6 +50,25 @@ provider "moonshot" {
44
50
  cache-write 3.0
45
51
  }
46
52
  }
53
+ // Official K3.1 Pro rates. The discovery mapper seeds K3 list pricing on
54
+ // every $0 K3-family row, so these must patch rather than fall back.
55
+ // https://platform.kimi.ai/docs/pricing/chat
56
+ models "kimi-k3.1-pro" {
57
+ cost-patch {
58
+ input 2.0
59
+ output 10.0
60
+ cache-read 0.2
61
+ cache-write 2.0
62
+ }
63
+ }
64
+ models "kimi-k3.1-pro-highspeed" {
65
+ cost-patch {
66
+ input 4.0
67
+ output 20.0
68
+ cache-read 0.2
69
+ cache-write 4.0
70
+ }
71
+ }
47
72
  // The /models roster carries no modality metadata for the named vision/vl
48
73
  // SKUs; the bounded id token is the census signal.
49
74
  models token="vision" {
@@ -71,7 +71,7 @@ provider "muse-code" {
71
71
  // Spark 1.3 Contributor warmed siblings hit 10/10 vs 0/10 with per-child keys (2026-10-12).
72
72
  // Meta documents shared-prefix prompt_cache_key routing; keep the opt-in on these exact Muse Code IDs.
73
73
  models "muse-spark-1.3" "muse-spark-1.3-contributor" "muse-wonder-1.0" "muse-wonder-1.0-contributor" "muse-wonder-full-1.0" "muse-wonder-full-1.0-contributor" {
74
- task-sibling-prompt-cache #true
74
+ shared-prompt-cache #true
75
75
  }
76
76
  }
77
77
  // Same capability-less `/v1/models` as Meta: unknown ids are opt-in.
@@ -140,6 +140,19 @@ provider "openai-codex" {
140
140
  ultrafast 8
141
141
  }
142
142
  }
143
+ // GPT-6.1 Terra lists at half GPT-6.1 Sol: $1/$5 with $0.05 cached input
144
+ // (OpenAI API docs, 2026-10-26). OpenAI documents 922K max input (1.05M
145
+ // total); the curated maximum applies only with extended context on, and a
146
+ // higher live maximum still wins. Codex ships no Ultrafast tier for it.
147
+ models "gpt-6.1-terra" "gpt-6.1-terra-wm" {
148
+ cost-patch {
149
+ input 1.0
150
+ output 5.0
151
+ cache-read 0.05
152
+ cache-write 0
153
+ }
154
+ max-context-window 922000
155
+ }
143
156
  models "gpt-6-luna" "gpt-6-luna-wm" {
144
157
  cost-patch {
145
158
  input 0.1
@@ -209,10 +222,10 @@ provider "openai-codex" {
209
222
  models "gpt-5.6-sol*" "gpt-5.6" {
210
223
  long-context-cost {
211
224
  input-threshold 272000
212
- input 10.0
213
- output 45.0
214
- cache-read 1.0
215
- cache-write 12.5
225
+ input 8.0
226
+ output 30.0
227
+ cache-read 0.8
228
+ cache-write 10.0
216
229
  }
217
230
  }
218
231
  models "gpt-5.6-terra*" {
@@ -152,8 +152,19 @@ provider "openai" {
152
152
  apply-patch-tool-type "freeform"
153
153
  }
154
154
  }
155
- // First-party GPT-5.6 whole-request long-context tiers (>272K input).
156
- models "daybreak-blue-latest" "gpt-5.6" "gpt-5.6-sol*" {
155
+ // First-party GPT-5.6 whole-request long-context tiers (>272K input): 2x
156
+ // input and cache rates, 1.5x output over the standard card.
157
+ models "gpt-5.6" "gpt-5.6-sol*" {
158
+ long-context-cost {
159
+ input-threshold 272000
160
+ input 8.0
161
+ output 30.0
162
+ cache-read 0.8
163
+ cache-write 10.0
164
+ }
165
+ }
166
+ // Daybreak Blue keeps the pre-2026-08-21 $5/$30 Sol base.
167
+ models "daybreak-blue-latest" {
157
168
  long-context-cost {
158
169
  input-threshold 272000
159
170
  input 10.0
@@ -193,6 +204,16 @@ provider "openai" {
193
204
  cache-write 25.0
194
205
  }
195
206
  }
207
+ // GPT-6 Sol, GPT-6 Luna, and GPT-6.1 Terra use the same >272K whole-request
208
+ // multipliers as Astra. Derive the tier from each API card so upstream price
209
+ // changes carry through.
210
+ models "gpt-6-sol" "gpt-6-luna" "gpt-6.1-terra" {
211
+ long-context-cost {
212
+ input-threshold 272000
213
+ multiplier 2.0
214
+ output-multiplier 1.5
215
+ }
216
+ }
196
217
  // API Ultrafast is 6x the standard card on both context tiers: GPT-6 Astra
197
218
  // $60/$6/$75/$300 and GPT-6.1 Sol $12/$0.60/$15/$60 per 1M short-context.
198
219
  models "gpt-6-astra*" "gpt-6.1-sol*" {
@@ -7,6 +7,34 @@ provider "opencode-go" {
7
7
  // Replaces the provider check that disabled forced tool choice on Responses.
8
8
  on-api "openai-responses" {
9
9
  supports-forced-tool-choice #false
10
+ // Go's x-opencode-session is the gateway's cache-affinity handle; on
11
+ // Responses the body prompt_cache_key carries the same shared key.
12
+ // Opt-in criterion: a cold / different-family / same-family probe
13
+ // (control sent BEFORE the family repeat) must warm on the repeated
14
+ // key and stay cold under the different key. 2026-10-26, 3 trials,
15
+ // cached tokens family vs control: Spark 1.3 Contributor 2/3 vs 0/3,
16
+ // Spark 1.2 Contributor 3/3 vs 1/3, gpt-5.6-luna 3/3 vs 0/3,
17
+ // gpt-6-luna 3/3 vs 0/3, grok-4.6 3/3 vs 0/3. Off: grok-4.7 (control
18
+ // warmed 2/3, key-independent) and the gateway-only standard Spark
19
+ // 1.3/1.2 ids (never lane-verified).
20
+ models "muse-spark-1.3-contributor" "muse-spark-1.2-contributor" "gpt-5.6-luna" "gpt-6-luna" "grok-4.6" {
21
+ shared-prompt-cache #true
22
+ }
23
+ }
24
+ // Chat Completions sends no prompt_cache_key body field here (the gateway
25
+ // leaves the compat flag unset), so the shared key rides only
26
+ // x-opencode-session. Same probe, 2026-10-26: deepseek-v4-pro 3/3 vs
27
+ // 0/3, omen-alpha 4/4 vs 0/4. Off as key-independent or inconsistent:
28
+ // glm-5.1 (control 3/5), glm-5.3-flash (family 2/7, control 1/7),
29
+ // kimi-k2.7-code (control 4/5), mimo-v2.6-flash (control 2/2, family
30
+ // 1/2), deepseek-v4.1-flash (control 2/2); also qwen3.6-plus and
31
+ // gemini-4-argon (no material warm hit) and mimo-v2.5-pro (upstream 400).
32
+ // Anthropic Messages: claude-haiku-5-5 warmed under any key (control
33
+ // 2/2), so no Messages lane opts in.
34
+ on-api "openai-completions" {
35
+ models "deepseek-v4-pro" "omen-alpha" {
36
+ shared-prompt-cache #true
37
+ }
10
38
  }
11
39
 
12
40
  // The Flash lanes accept images despite their IDs lacking a vision suffix;
@@ -37,9 +65,9 @@ provider "opencode-go" {
37
65
  max-tokens 384000
38
66
  }
39
67
  }
40
- // Live Go discovery lists Argon without capabilities; its native Gemini
41
- // endpoint requires a thinking level and accepts LOW, MEDIUM, or HIGH.
42
- models "gemini-4-argon-contributor" {
68
+ // Go rosters can list either Argon ID without capabilities; the native
69
+ // Gemini endpoint requires a thinking level and accepts LOW, MEDIUM, or HIGH.
70
+ models "gemini-4-argon" "gemini-4-argon-contributor" {
43
71
  thinking-upgrade-neutral #true
44
72
  thinking-mode "google-level"
45
73
  thinking-efforts "low" "medium" "high"
@@ -113,6 +141,13 @@ provider "opencode-go" {
113
141
  thinking-mode "effort"
114
142
  }
115
143
  }
144
+ // Step 5 Preview uses low/medium/high on this gateway (#14987).
145
+ class "stepfun" {
146
+ family "step" {
147
+ thinking-mode "effort"
148
+ thinking-efforts "low" "medium" "high"
149
+ }
150
+ }
116
151
  // LongCat exposes a reasoning toggle, not selectable effort tiers.
117
152
  // Off uses the gateway's boolean reasoning switch, never a fake lowest tier.
118
153
  models "longcat-2.5-preview-free" {
@@ -61,6 +61,13 @@ provider "opencode-zen" {
61
61
  thinking-mode "effort"
62
62
  }
63
63
  }
64
+ // Step 5 Preview uses low/medium/high on this gateway (#14987).
65
+ class "stepfun" {
66
+ family "step" {
67
+ thinking-mode "effort"
68
+ thinking-efforts "low" "medium" "high"
69
+ }
70
+ }
64
71
  class "unknown" {
65
72
  thinking-mode "effort"
66
73
  }
@@ -1,7 +1,7 @@
1
1
  // Provider-wire compat for "xiaomi-token-plan-ams"; regenerated from the frozen census using deployment contract selectors.
2
2
 
3
3
  provider "xiaomi-token-plan-ams" {
4
- default-model "mimo-v2.5"
4
+ default-model "mimo-v2.6-flash"
5
5
  env "XIAOMI_TOKEN_PLAN_AMS_API_KEY"
6
6
 
7
7
  class "mimo" {
@@ -1,7 +1,7 @@
1
1
  // Provider-wire compat for "xiaomi-token-plan-cn"; regenerated from the frozen census using deployment contract selectors.
2
2
 
3
3
  provider "xiaomi-token-plan-cn" {
4
- default-model "mimo-v2.5"
4
+ default-model "mimo-v2.6-flash"
5
5
  env "XIAOMI_TOKEN_PLAN_CN_API_KEY"
6
6
 
7
7
  // Xiaomi's credential-scoped roster returns bare ids, so these documented
@@ -1,7 +1,7 @@
1
1
  // Provider-wire compat for "xiaomi-token-plan-sgp"; regenerated from the frozen census using deployment contract selectors.
2
2
 
3
3
  provider "xiaomi-token-plan-sgp" {
4
- default-model "mimo-v2.5"
4
+ default-model "mimo-v2.6-flash"
5
5
  env "XIAOMI_TOKEN_PLAN_SGP_API_KEY"
6
6
 
7
7
  class "mimo" {
@@ -1,7 +1,7 @@
1
1
  // Provider-wire compat for "xiaomi"; regenerated from the frozen census using deployment contract selectors.
2
2
 
3
3
  provider "xiaomi" {
4
- default-model "mimo-v2.5"
4
+ default-model "mimo-v2.6-flash"
5
5
  env "XIAOMI_API_KEY"
6
6
  discovery label="Xiaomi"
7
7
 
@@ -26,6 +26,11 @@ behavior {
26
26
  // from seeding cross-provider bare-id enrichment references (#10932).
27
27
  reference-isolated-providers "cline-pass"
28
28
 
29
+ // These hosts reuse an unchanged message prefix across turns; history
30
+ // rewrites otherwise discard their warm prompt cache.
31
+ append-only-context-providers "deepseek" "ollama" "ollama-cloud" "lm-studio" "llama.cpp" \
32
+ "google-antigravity" "google" "google-vertex"
33
+
29
34
  // Discovered Responses/Codex models receive these additional operations.
30
35
  model-operations provider="openai" {
31
36
  exact "o3"
@@ -48,9 +53,11 @@ behavior {
48
53
  cursor-model-parameter model="composer-2.5" id="fast" value="false"
49
54
 
50
55
  // Cursor's dashboard splits native Grok/Composer traffic from all other
51
- // models. Substring fallbacks cover newly discovered family revisions.
56
+ // models; Auto (`default`) bills to the Cursor-models pool on both legacy
57
+ // plans and the X Subscription. Substring fallbacks cover newly discovered
58
+ // family revisions.
52
59
  quota-tiers provider="cursor" default="api" {
53
- tier "auto" "composer-2.5"
60
+ tier "auto" "composer-2.5" "default"
54
61
  fallback "auto" substring="grok"
55
62
  fallback "auto" substring="composer"
56
63
  }
@@ -104,14 +111,17 @@ behavior {
104
111
  "gpt-5.4-mini-fast" "gpt-5.3-codex" "gpt-5.3-codex-fast" "gpt-5.2" "garnet-07-15" \
105
112
  "gemini-3.1-pro-preview" "gemini-3.8-flash" "gemini-3.7-flash" "gemini-3.6-flash" \
106
113
  "gemini-3.5-flash" "gemini-3-flash-preview" "grok-4.7" "grok-4.6" "grok-4.5" "atlas-07-21" \
107
- "aster-07-15"
114
+ "aster-07-15" "gpt-image-2.5-sunburst" "gpt-image-2.5-flare" "gemini-3-pro-image-preview"
108
115
  }
109
116
  // Subscription eligibility for model-aware Codex credential routing.
117
+ // GPT-6.1 Terra launched for Plus, Pro, Enterprise, and Edu, not Free/Go
118
+ // (https://openai.com/index/introducing-gpt-6-1-terra/, 2026-10-26).
110
119
  plan-requirement provider="openai-codex" {
111
120
  tier "pro" substring="-spark"
112
121
  tier "paid" exact="gpt-5.6" exact="gpt-5.6-sol" exact="gpt-5.6-sol-pro" \
113
122
  exact="gpt-5.6-luna" exact="gpt-5.6-luna-pro" glob="*/gpt-5.6" glob="*/gpt-5.6-sol" \
114
- glob="*/gpt-5.6-sol-pro" glob="*/gpt-5.6-luna" glob="*/gpt-5.6-luna-pro"
123
+ glob="*/gpt-5.6-sol-pro" glob="*/gpt-5.6-luna" glob="*/gpt-5.6-luna-pro" \
124
+ glob="gpt-6.1-terra*" glob="*/gpt-6.1-terra*"
115
125
  }
116
126
 
117
127
  // Z.AI and Zhipu emit timezone-naive account-reset wall clocks in Beijing
@@ -144,17 +154,19 @@ behavior {
144
154
  exclude-discovery-modes "audio_speech" "audio_transcription" "batch" "embedding" "guardrail" \
145
155
  "image_edit" "image_generation" "moderation" "ocr" "rerank" "search" "vector_store" \
146
156
  "video_generation" provider="litellm"
147
- // models.dev rows ZETA cannot route successfully: AWS does not document
157
+ // models.dev rows OMP cannot route successfully: AWS does not document
148
158
  // the jp. Opus 5 profile, and the openai.gpt-5.x rows are Mantle-only
149
159
  // ids that Bedrock rejects or misroutes.
150
160
  exclude-models provider="amazon-bedrock" exact="jp.anthropic.claude-opus-5" exact="openai.gpt-5.4" \
151
161
  exact="openai.gpt-5.5" exact="openai.gpt-5.6-luna" exact="openai.gpt-5.6-sol" exact="openai.gpt-5.6-terra"
152
162
  // [1m] is a Claude Code selector convention, not an inference id.
153
163
  exclude-models provider="zai" glob="*[1m]"
154
- // Anthropic /v1/models rows ZETA does not route: pre-tool-era snapshots and
155
- // the 3.5/3.7 lanes Claude Code retired from the public roster.
156
- exclude-models provider="anthropic" prefix="claude-3-5-haiku" prefix="claude-3-7-sonnet" \
157
- exact="claude-3-opus-20240229" exact="claude-3-sonnet-20240229"
164
+ // Anthropic's retired snapshots and their aliases reject inference requests
165
+ // (https://platform.claude.com/docs/en/about-claude/model-deprecations).
166
+ exclude-models provider="anthropic" prefix="claude-3-5-haiku" prefix="claude-3-5-sonnet" \
167
+ prefix="claude-3-7-sonnet" prefix="claude-opus-4-1" exact="claude-3-haiku-20240307" \
168
+ exact="claude-3-opus-20240229" exact="claude-3-sonnet-20240229" exact="claude-opus-4-0" \
169
+ exact="claude-opus-4-20250514" exact="claude-sonnet-4-0" exact="claude-sonnet-4-20250514"
158
170
  // Control-plane resource ids are not accepted by the inference API.
159
171
  exclude-models provider="fireworks" prefix="accounts/fireworks/"
160
172
  exclude-models provider="firepass" prefix="accounts/fireworks/"
@@ -273,7 +285,7 @@ behavior {
273
285
  // responses pins are gateway-verified (deepseek-v4-flash 2026-08-08;
274
286
  // muse-spark per opencode.ai/docs/go/#endpoints, #8957, #10610).
275
287
  api-routes provider="opencode-go" {
276
- route "google-generative-ai" exact="gemini-4-argon-contributor"
288
+ route "google-generative-ai" exact="gemini-4-argon" exact="gemini-4-argon-contributor"
277
289
  route "anthropic-messages" exact="union-alpha"
278
290
  route "openai-responses" prefix="muse-spark-" exact="deepseek-v4-flash" exact="muse-spark-1.2" \
279
291
  exact="muse-spark-1.2-contributor" exact="muse-spark-1.3" exact="muse-spark-1.3-contributor"
@@ -296,10 +308,11 @@ behavior {
296
308
  exact="gpt-5.6-sol" exact="gpt-5.6-sol-fast" exact="gpt-5.6-terra" exact="gpt-5.6-luna" \
297
309
  exact="gpt-5.5" exact="gpt-5.5-fast" exact="gpt-5.5-pro" exact="gpt-5.4" exact="gpt-5.4-fast" \
298
310
  exact="gpt-5.4-mini" exact="gpt-5.4-mini-fast" exact="gpt-5.3-codex" exact="gpt-5.3-codex-fast" \
299
- exact="gpt-5.2" exact="grok-4.7" exact="grok-4.6" exact="grok-4.5"
311
+ exact="gpt-5.2" exact="grok-4.7" exact="grok-4.6" exact="grok-4.5" exact="gpt-image-2.5-sunburst" \
312
+ exact="gpt-image-2.5-flare"
300
313
  route "google-generate" exact="garnet-07-15" exact="gemini-3.1-pro-preview" exact="gemini-3.8-flash" \
301
314
  exact="gemini-3.7-flash" exact="gemini-3.6-flash" exact="gemini-3.5-flash" \
302
- exact="gemini-3-flash-preview"
315
+ exact="gemini-3-flash-preview" exact="gemini-3-pro-image-preview"
303
316
  route "openai-completions" exact="inkling" exact="mistral-large-4" exact="mistral-medium-3.5" exact="glm-5.3-flash" \
304
317
  exact="glm-5.3" exact="glm-5.2" exact="glm-5.2-fast" exact="kimi-k3" exact="qwen3.8-max" \
305
318
  exact="nemotron-3-ultra" exact="deepseek-v4.1-flash" exact="deepseek-v4-flash-0731" \
@@ -325,15 +338,18 @@ behavior {
325
338
  }
326
339
  // Kimi Code's /coding/v1/models envelope carries no output-limit field;
327
340
  // these are the documented per-family caps (models.dev `kimi-for-coding`
328
- // and `moonshotai` kimi-k3 entries, #6711). Legacy K2 rows keep their
341
+ // and `moonshotai` kimi-k3 entries, #6711; K3.1 Pro's documented
342
+ // `max_completion_tokens` default). Legacy K2 rows keep their
329
343
  // discovery-reported values.
330
344
  model-limits provider="kimi-code" {
331
345
  limits "k3" max-tokens=131072
332
346
  limits "k3-256k" max-tokens=131072
347
+ limits "k3.1-pro" max-tokens=131072
348
+ limits "k3.1-pro-highspeed" max-tokens=131072
333
349
  limits "kimi-for-coding" max-tokens=32768
334
350
  limits "kimi-for-coding-highspeed" max-tokens=32768
335
351
  }
336
- // Bedrock lists conversational SKUs ZETA cannot drive over Converse
352
+ // Bedrock lists conversational SKUs OMP cannot drive over Converse
337
353
  // tool-calling: Jamba, Titan Text Express, and Mistral 7B v0.
338
354
  exclude-models provider="amazon-bedrock" prefix="ai21.jamba" prefix="amazon.titan-text-express" \
339
355
  prefix="mistral.mistral-7b-instruct-v0"
@@ -19,6 +19,13 @@ class "kimi" {
19
19
  family "k3" glob="k3"
20
20
  family "k3" glob="k3-256k"
21
21
 
22
+ // K3.1 Pro is a post-trained K3 with K3's full request contract (Moonshot
23
+ // documents identical parameter constraints), so it keeps family `k3` and
24
+ // inherits every K3 lineage rule. The revision gives it its own weights
25
+ // identity, so K3 reasoning never replays natively into K3.1 Pro. Kimi
26
+ // Code's bare `k3.1-pro*` ids reach no class matcher without these.
27
+ override id="kimi-k3-1-pro" glob="*k3.1-pro" class="kimi" family="k3" revision="3.1" rationale="K3.1 Pro is a post-trained K3 sharing K3's parameter contract" provenance="https://platform.kimi.ai/docs/api/models-overview"
28
+ override id="kimi-k3-1-pro-highspeed" glob="*k3.1-pro-highspeed" class="kimi" family="k3" revision="3.1" rationale="K3.1 Pro Highspeed serves K3.1 Pro faster with identical parameter constraints" provenance="https://platform.kimi.ai/docs/api/models-overview"
22
29
  override id="kilo-kimi-k2-thinking-product" provider="kilo" model="kimi-k2-thinking" logical="moonshotai/kimi-k2-thinking" class="kimi" rationale="Oracle exposes this independently priced product rather than an effort route" provenance="fixtures/llm-oracle/catalog/models.normalized.json"
23
30
  override id="openrouter-kimi-k2-thinking-product" provider="openrouter" model="kimi-k2-thinking" logical="moonshotai/kimi-k2-thinking" class="kimi" rationale="Oracle exposes this independently priced product rather than an effort route" provenance="fixtures/llm-oracle/catalog/models.normalized.json"
24
31
  override id="vercel-kimi-k2-thinking-product" provider="vercel-ai-gateway" model="kimi-k2-thinking" logical="moonshotai/kimi-k2-thinking" class="kimi" rationale="Oracle exposes this independently priced product rather than an effort route" provenance="fixtures/llm-oracle/catalog/models.normalized.json"