@linxiraos/pi-catalog 1.1.27 → 1.2.1

package/src/compat/rules/classes/anthropic.kdl

@@ -9,7 +9,13 @@
 // are never undercounted; only Opus/Sonnet/Haiku revisions known to predate
 // 4.7 take the standard tier. Separator-collapsed revisions parse as two
 // digits (`claude-opus-45` is Opus 4.5 → 45), so 10–46 is standard too.
+// Claude's prompt cache finds an earlier request's entry only within 20 block
+// positions back from a breakpoint, counting the breakpoint; a run of
+// consecutive tool_use or tool_result blocks is one position. Warm-cache
+// tool-result pruning keeps its rewrites inside this window.
+// https://platform.claude.com/docs/en/build-with-claude/prompt-caching
 class "anthropic" {
+	prompt-cache-lookback 20
 	image-tokenization {
 		regime "anthropic-patch"
 		max-edge 2576
@@ -72,7 +78,7 @@ class "anthropic" {
 			thinking-prefix-binding #true
 		}
 	}
-	// Adaptive Claude (Opus 4.7+, Sonnet/Fable/Mythos 5+) rejects explicit
+	// Adaptive Claude (Opus 4.7+, Sonnet/Fable/Mythos 5+, Haiku 5.5+) rejects explicit
 	// sampling params (`temperature`, `top_p`, `top_k`) with a 400 on every
 	// serving host. The `<10` bound keeps whole-number revisions such as
 	// `claude-opus-45` (Opus 4.5) out of the range.
@@ -96,6 +102,22 @@ class "anthropic" {
 			supports-sampling-params #false
 		}
 	}
+	family "haiku" {
+		revision ">=5.5 <10" {
+			supports-sampling-params #false
+		}
+		// Haiku 5.5 bills every rate 5x for prompts over 100K. Derived from each
+		// host's list price so non-Anthropic hosts get the band too; the Claude
+		// API and Cursor pin absolute rates, and GitHub Copilot opts out because
+		// its long tier is the `-1m` sibling row.
+		// https://platform.claude.com/docs/en/models/haiku-5-5/overview#pricing
+		revision ">=5.5 <6" {
+			long-context-cost {
+				input-threshold 100000
+				multiplier 5.0
+			}
+		}
+	}
 	// On-demand compaction (compact-2026-09-04) works on the Claude API, Vertex,
 	// and Amazon Bedrock's `/anthropic` routes, not Bedrock Converse.
 	// Foundry / Claude Platform on AWS are first-party Anthropic routes whose
@@ -121,6 +143,11 @@ class "anthropic" {
 				supports-server-compaction #true
 			}
 		}
+		family "haiku" {
+			revision ">=5.5 <6" {
+				supports-server-compaction #true
+			}
+		}
 		models "claude-mythos-preview" {
 			supports-server-compaction #true
 		}
@@ -173,6 +200,15 @@ class "anthropic" {
 			thinking-prefix-binding #true
 		}
 	}
+	// Haiku 5.5 binds signed thinking to the exact preceding conversation like
+	// Sonnet 5.5. It accepts a forced `tool_choice` (the response just skips
+	// thinking), so it keeps forced tool selection.
+	// https://platform.claude.com/docs/en/models/haiku-5-5/migration-guide
+	family "haiku" {
+		revision ">=5.5 <6" {
+			thinking-prefix-binding #true
+		}
+	}
 	on "anthropic" "cloudflare-ai-gateway" "google-vertex" {
 		family "fable" {
 			revision ">=5" {
@@ -200,6 +236,13 @@ class "anthropic" {
 				supports-turn-scoped-system #true
 			}
 		}
+		family "haiku" {
+			revision ">=5.5 <6" {
+				supports-mid-conversation-system #true
+				supports-mid-conversation-tool-changes #true
+				supports-turn-scoped-system #true
+			}
+		}
 		family "opus" {
 			revision ">=4.8" {
 				supports-mid-conversation-system #true
@@ -230,6 +273,13 @@ class "anthropic" {
 				supports-per-message-effort #true
 			}
 		}
+		// Haiku 5.5 mirrors Sonnet 5.5; Vertex stays off until verified.
+		family "haiku" {
+			revision ">=5.5 <6" {
+				supports-thinking-binding-controls #true
+				supports-per-message-effort #true
+			}
+		}
 		family "opus" {
 			revision ">=5" {
 				supports-per-message-effort #true
@@ -257,6 +307,14 @@ class "anthropic" {
 		"google-vertex" "opencode-zen" "snowflake" "vercel-ai-gateway" {
 		family "haiku" {
 			thinking-mode "budget"
+			// Haiku 5.5 is the first adaptive Haiku: `budget_tokens` is a 400.
+			// An omitted `thinking` is adaptive, but unlike Sonnet/Opus 5.5 it
+			// accepts `disabled` at effort `high` or below, so Off is explicit.
+			// https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-haiku-5-5#use-effort-to-control-thinking
+			revision ">=5.5 <6" {
+				thinking-mode "anthropic-adaptive"
+				disabled-thinking "disabled"
+			}
 		}
 	}
 	on "amazon-bedrock" "anthropic" "cloudflare-ai-gateway" "github-copilot" "gitlab-duo" \
@@ -365,6 +423,34 @@ class "anthropic" {
 				max-tokens 128000
 			}
 		}
+		// /v1/models reports neither reasoning, vision, pricing nor limits for
+		// Haiku 5.5, so a discovered row would materialize as a text-only,
+		// unpriced, non-reasoning model; pin the model card. Prompts over 100K
+		// input tokens bill the whole request at the long-context rates.
+		// https://platform.claude.com/docs/en/models/haiku-5-5/overview
+		family "haiku" {
+			revision ">=5.5 <6" {
+				thinking-upgrade-neutral #true
+				input-modalities "text" "image"
+				cost-patch {
+					input 0.1
+					output 0.5
+					cache-read 0.01
+					cache-write 0.125
+				}
+				long-context-cost {
+					input-threshold 100000
+					input 0.5
+					output 2.5
+					cache-read 0.05
+					cache-write 0.625
+				}
+				limits-patch {
+					context-window 1000000
+					max-tokens 128000
+				}
+			}
+		}
 		family "mythos" {
 			cost-patch {
 				input 10.0

package/src/compat/rules/classes/deepseek.kdl

@@ -98,4 +98,10 @@ class "deepseek" {
 	// providers alike. On hosts that parse it the healer sees no markup.
 	stream-markup-healing-pattern "dsml"
 	thinking-loop-guard "deepseek"
+	// V4 reasoning reaches omp as its plaintext trace on every host, and the
+	// revision pins the weights (flash 4.0 / 4.1, pro 4.0). Earlier lineages
+	// classify too coarsely to carry: `r1` also matches the R1 distills.
+	revision ">=4" {
+		portable-reasoning #true
+	}
 }

package/src/compat/rules/classes/glm.kdl

@@ -65,7 +65,7 @@ class "glm" {
 	// GLM Coding Plan: the selectable 1M-context served ids; pin them so
 	// endpoint discovery or older bundled fallbacks cannot regress to 200k.
 	on "zai" "zhipu-coding-plan" {
-		models "glm-5.2" "glm-5.3" "glm-5.3-flash" {
+		models "glm-5.2" "glm-5.3" "glm-5.3-flash" "glm-5.3-flashx" {
 			limits-patch {
 				context-window 1000000
 				max-tokens 131072
@@ -84,10 +84,9 @@ class "glm" {
 			}
 		}
 	}
-	// glm-5.3-flash ships native image input that upstream metadata still
-	// reports as text-only.
+	// GLM-5.3 Flash and FlashX accept native image input on their coding routes.
 	on "zai" "zhipu-coding-plan" {
-		models "glm-5.3-flash" {
+		models "glm-5.3-flash" "glm-5.3-flashx" {
 			input-modalities "text" "image"
 		}
 	}
@@ -99,17 +98,27 @@ class "glm" {
 			thinking-efforts "minimal" "low" "medium" "high"
 		}
 	}
-	// GLM-5.3+ coding SKUs (base/air/turbo/flash, non-vision) expose a uniform
+	// GLM-5.3+ coding SKUs (base/air/turbo/flash/flashx) expose a uniform
 	// wire-exact low/high/max reasoning-effort ladder on every host, thinking
 	// can no longer be disabled, and the default effort is max.
 	// residue: taxonomy ranks cannot separate these ids from the vision (`…v`)
-	// and flashx/preview shapes that keep host dialects; the explicit priority resolves
+	// and preview shapes that keep host dialects; the explicit priority resolves
 	// the intentional overlap with census host-dialect residues such as
 	// baseten's `zai-org/GLM-5*`.
-	models "*glm-5.3" "*glm-5.3-air" "*glm-5.3-turbo" "*glm-5.3-flash" "*glm-5.3:*" \
+	models "*glm-5.3" "*glm-5.3-air" "*glm-5.3-turbo" "*glm-5.3-flash" "*glm-5.3-flashx" "*glm-5.3:*" \
 		"*glm-5.3-flash-uncensored" "*glm-5.3-free" "*glm-5.3-highspeed" "*glm-5.3-fast" "*glm-5.3-flash-lab" priority=10 {
 		thinking-default-level "max"
 		thinking-efforts "low" "high" "max"
 		thinking-requires-effort #true
 	}
+	// GLM 5+ reasoning reaches omp as its plaintext trace on every host, and
+	// family plus revision pin the weights (glm-5.2, glm-5.3-flash, ...). GLM 4
+	// classifies too coarsely to carry: 4.0 also matches glm-4.5v and glm-4.6v.
+	revision ">=5" {
+		portable-reasoning #true
+		// residue: fine-tuned weights that classify as the base GLM revision.
+		models "*-uncensored" "*-cybersecurity" {
+			portable-reasoning #false
+		}
+	}
 }

package/src/compat/rules/classes/kimi.kdl

@@ -123,4 +123,12 @@ class "kimi" {
 			}
 		}
 	}
+	// K3 reasoning reaches omp as its plaintext trace on every host, and K3 was
+	// trained to read all of its earlier reasoning back (preserved thinking).
+	// A recall probe showed the reasoning slot of moonshot, factory-droid,
+	// openrouter and devin reaches the model, so K3 reasoning moves between
+	// those hosts natively instead of as visible `<think>` text.
+	family "k3" {
+		portable-reasoning #true
+	}
 }

package/src/compat/rules/classes/minimax.kdl

@@ -51,7 +51,7 @@ class "minimax" {
 			thinking-efforts "minimal" "low" "medium" "high" "xhigh"
 		}
 	}
-	on "minimax" "minimax-cn" "vercel-ai-gateway" {
+	on "minimax" "minimax-cn" "minimax-code" "minimax-code-cn" "vercel-ai-gateway" {
 		thinking-mode "anthropic-adaptive"
 		family "m2" {
 			// Replaces the Anthropic-route MiniMax effort remap table.
@@ -71,6 +71,26 @@ class "minimax" {
 			thinking-efforts "low" "medium" "high"
 		}
 	}
+	// MiniMax-M3.1-Flash-Preview always thinks (`thinking.type: "disabled"` and
+	// effort `none` 400 with error 2013) and tunes depth through
+	// `output_config.effort` from `low` to `max` (the default when omitted).
+	// The identity map replaces the M3 family's adaptive-tag-only map so each
+	// level reaches the wire. Like the M2 family, thinking cannot be turned off:
+	// `thinking-requires-effort` hides the off level and clamps forced-off
+	// requests to the lowest effort instead of sending `disabled`.
+	on "minimax" "minimax-cn" "minimax-code" "minimax-code-cn" {
+		models "MiniMax-M3.1-Flash-Preview" {
+			thinking-requires-effort #true
+			thinking-efforts "low" "medium" "high" "xhigh" "max"
+			thinking-effort-map {
+				low "low"
+				medium "medium"
+				high "high"
+			}
+		}
+	}
+	// Token Plan rows ride the Anthropic route; these OpenAI-only wire axes
+	// serve the legacy lightning rows that stay on the OpenAI-compatible one.
 	on "minimax-code" "minimax-code-cn" {
 		supports-reasoning-effort #false
 		supports-store #false
@@ -78,4 +98,14 @@ class "minimax" {
 	on "minimax-code" "minimax-code-cn" "wafer-serverless" {
 		reasoning-content-field "reasoning_content"
 	}
+	// M3 reasoning reaches omp as its plaintext trace on every host, and
+	// MiniMax asks agents to pass all earlier reasoning back.
+	family "m3" {
+		portable-reasoning #true
+		// residue: M3.1 Flash Preview classifies as family m3 (no revision is
+		// extracted), but it is a different model.
+		models "*m3.1*" {
+			portable-reasoning #false
+		}
+	}
 }

package/src/compat/rules/classes/openai.kdl

@@ -49,10 +49,12 @@ class "openai" {
 		delegation-bias "restrained"
 		supports-steering #true
 	}
-	// GPT-6 Astra changes reasoning effort mid-conversation through a
+	// The GPT-6 family changes reasoning effort mid-conversation through a
 	// `configuration_update` input item so the request-level effort (and the
-	// cached prompt prefix) stays put; every other SKU 400s on the item type.
-	models "gpt-6-astra" {
+	// cached prompt prefix) stays put. OpenAI's reasoning guide ("Change
+	// reasoning mid-conversation") documents it for the family, and Codex's
+	// `/models` reports `supports_reasoning_effort_updates` on every GPT-6 SKU.
+	revision ">=6 <7" {
 		supports-configuration-update #true
 	}
 	// Codex SKUs: the advertised 400K figure includes the output budget; the

package/src/compat/rules/classes/qwen.kdl

@@ -1,7 +1,8 @@
 // Model-lineage compat for "qwen"; regenerated from the frozen census using class taxonomy selectors.
 
 class "qwen" {
-	on "aiand" "aimlapi" "alibaba-coding-plan" "amazon-bedrock" "coreweave" "groq" "huggingface" \
+	// Groq owns its Qwen effort ladder in providers/groq.kdl.
+	on "aiand" "aimlapi" "alibaba-coding-plan" "amazon-bedrock" "coreweave" "huggingface" \
 		"kilo" "nanogpt" "novita" "nvidia" "ollama-cloud" "openrouter" "synthetic" \
 		"together" "venice" "wafer-serverless" "zenmux" {
 		thinking-efforts "minimal" "low" "medium" "high"

package/src/compat/rules/providers/abliteration.kdl

@@ -37,4 +37,11 @@ provider "abliteration" {
 
 	include-encrypted-reasoning #false
 	stream-idle-timeout-ms 0
+	// Abliterated weights classify as the GLM revision they derive from, but
+	// they are modified weights: their reasoning is not carried to or from GLM.
+	class "glm" {
+		revision ">=5" {
+			portable-reasoning #false
+		}
+	}
 }

package/src/compat/rules/providers/anthropic.kdl

@@ -34,6 +34,12 @@ provider "anthropic" {
 			cost input=2 output=10 cache-read=0.2 cache-write=2.5
 			limits context=1000000 max-tokens=128000
 		}
+		model "claude-haiku-5-5" name="Claude Haiku 5.5" {
+			reasoning #true
+			input "text" "image"
+			cost input=0.1 output=0.5 cache-read=0.01 cache-write=0.125
+			limits context=1000000 max-tokens=128000
+		}
 		model "claude-fable-5" name="Claude Fable 5" {
 			reasoning #true
 			input "text" "image"
@@ -68,6 +74,13 @@ provider "anthropic" {
 	// combines this fact with a per-request effective-URL check, so it never
 	// authorizes a block on its own.
 	first-party-provider #true
+	// Messages requests cap at 32 MB. Reserve 8 MB for text, tools, and JSON
+	// framing when retaining inline base64 images in a long conversation.
+	// Applied only on the official endpoint; a custom baseUrl has its own limit.
+	// https://platform.claude.com/docs/en/api/overview#request-size-limits
+	on-api "anthropic-messages" {
+		inline-image-byte-budget 24000000
+	}
 	class "anthropic" {
 		revision ">=3.7 <4.6" {
 			thinking-efforts "minimal" "low" "medium" "high" "xhigh"

package/src/compat/rules/providers/apple.kdl

@@ -5,4 +5,5 @@
 provider "apple" {
 	default-model "on-device"
 	allow-unauthenticated #true
+	automatic-default #false
 }

package/src/compat/rules/providers/cloudflare-workers-ai.kdl

@@ -0,0 +1,34 @@
+// Cloudflare Workers AI decision models. Clef and Clef Flash answer typed
+// questions over JSON state with the TypeSafe System One body
+// (`{model, state, questions}`) at
+// `POST /client/v4/accounts/{account_id}/ai/run/@cf/cloudflare/<model>`.
+// `<account>` in the base URL is resolved per request from the stored login
+// credential (else `CLOUDFLARE_ACCOUNT_ID`).
+
+provider "cloudflare-workers-ai" {
+	default-model "@cf/cloudflare/clef"
+	env "CLOUDFLARE_API_TOKEN" "CLOUDFLARE_AUTH_TOKEN"
+
+	// developers.cloudflare.com/workers-ai/models/clef and /clef-flash: both
+	// take text and image input and bill input tokens only.
+	seed api="workers-ai-systemone" base-url="https://api.cloudflare.com/client/v4/accounts/<account>/ai/run" bundle="always" {
+		model "@cf/cloudflare/clef" name="Clef" {
+			reasoning #false
+			input "text" "image"
+			cost input=0.24 output=0 cache-read=0 cache-write=0
+			limits context=65536
+			supports-tools #false
+		}
+		model "@cf/cloudflare/clef-flash" name="Clef Flash" {
+			reasoning #false
+			input "text" "image"
+			cost input=0.038 output=0 cache-read=0 cache-write=0
+			limits context=24576
+			supports-tools #false
+		}
+	}
+
+	models "*" {
+		kind "judge"
+	}
+}

package/src/compat/rules/providers/coralbricks.kdl

@@ -0,0 +1,62 @@
+// CoralBricks inference gateway deployment contract
+// (https://www.coralbricks.ai/docs.md; catalog snapshot 2026-10-08).
+//
+// Coral serves open-weight models behind an OpenAI-compatible wire at
+// https://inference.coralbricks.ai/v1 with a single bearer API key. The
+// key-protected `GET /v1/models` publishes each row's context window,
+// per-million pricing, and capability flags, and Coral documents those
+// flags as authoritative (unsupported content answers
+// `400 unsupported_content_type` rather than being dropped), so live
+// discovery is the deployment truth and same-id rows on other hosts must
+// not backfill capabilities or limits into this provider's rows.
+provider "coralbricks" {
+	default-model "glm-5.3-fast"
+	// Official docs authenticate via `CORAL_API_KEY`
+	// (https://www.coralbricks.ai/docs.md#authentication); `CORALBRICKS_API_KEY`
+	// is the alias the Coral pi provider extension shipped with, kept as a
+	// compatibility fallback in precedence order.
+	env "CORAL_API_KEY" "CORALBRICKS_API_KEY"
+	dynamic-models-authoritative #true
+	skip-cross-provider-reference-fills #true
+	discovery label="CoralBricks"
+
+	// Reviewed snapshot of the public catalog
+	// (https://www.coralbricks.ai/api/public/models). Coral retired GLM 5.3
+	// Flash on 2026-10-07 and renamed the remaining slugs (`glm-5.3-fast`,
+	// `deepseek-v4.1-flash-fast`). The public rows list input/output rates
+	// only; cache-write rates come from Coral's published pricing and cached
+	// reads are $0 on every model. `/v1/models` needs a key, so generation
+	// without credentials bundles these rows; successful runtime discovery
+	// replaces them.
+	seed api="openai-completions" base-url="https://inference.coralbricks.ai/v1" bundle="always" {
+		model "glm-5.3-fast" name="GLM 5.3" {
+			reasoning #true
+			input "text"
+			cost input=1.12 output=4.4 cache-read=0 cache-write=1.68
+			limits context=1048576 max-tokens=131072
+		}
+		model "deepseek-v4.1-flash-fast" name="DeepSeek V4.1 Flash" {
+			reasoning #true
+			input "text" "image"
+			cost input=0.3 output=1.2 cache-read=0 cache-write=0.09
+			limits context=1048576 max-tokens=131072
+		}
+	}
+
+	// DeepSeek V4.1 Flash is the one CoralBricks SKU that advertises a genuine
+	// off value (`none` in its reasoning-effort vocabulary, per the reasoning
+	// table at https://www.coralbricks.ai/docs.md#reasoning), so a
+	// thinking-off request disables reasoning instead of clamping to the
+	// lowest effort like the GLM SKUs.
+	models "deepseek-v4.1-flash-fast" {
+		reasoning-disable-mode "none-effort"
+	}
+	// Gateway request dialect, matching Coral's plain Chat Completions
+	// surface ("unchanged from what your existing SDK expects"): output caps
+	// ride `max_tokens`, system prompts ride the `system` role, and no
+	// `store` field is sent. Coral accepts unknown fields, so acceptance is
+	// not evidence the semantics are honored (same call as charm-hyper).
+	supports-developer-role #false
+	supports-store #false
+	max-tokens-field "max_tokens"
+}

package/src/compat/rules/providers/cursor.kdl

@@ -36,6 +36,10 @@ provider "cursor" {
 				input-modalities "text" "image"
 				context-window-floor 1000000
 			}
+			// Cursor's agent protocol relays thinking opaquely, so it is not
+			// verified as K3's full trace: Cursor K3 turns are not carried to
+			// other K3 hosts (Cursor itself refuses foreign history).
+			portable-reasoning #false
 		}
 		family "k2.7-code" {
 			// Cursor advertised Kimi K2.7 Code Context window 262k.
@@ -167,6 +171,24 @@ provider "cursor" {
 			cache-write 12.5
 		}
 	}
+	// Claude Haiku 5.5: $0.10/$0.50, cache $0.01/$0.125. Requests above
+	// 100k input tokens bill every leg at 5x ($0.50/$2.50, cache
+	// $0.05/$0.625), on the 1M lane too.
+	models "claude-haiku-5-5" "claude-haiku-5-5-1m" {
+		cost-patch {
+			input 0.1
+			output 0.5
+			cache-read 0.01
+			cache-write 0.125
+		}
+		long-context-cost {
+			input-threshold 100000
+			input 0.5
+			output 2.5
+			cache-read 0.05
+			cache-write 0.625
+		}
+	}
 	// Claude 4.7 Opus: $5/$25, cache $0.50/$6.25. Up to 1M same rates.
 	models "claude-opus-4-7-high" "claude-opus-4-7-low" "claude-opus-4-7-max" "claude-opus-4-7-medium" "claude-opus-4-7-xhigh" {
 		cost-patch {

package/src/compat/rules/providers/devin.kdl

@@ -14,8 +14,8 @@ provider "devin" {
 	// bundle would misstate every other account's entitlements and leave zombie
 	// rows behind (see CREDENTIAL_SCOPED_PROVIDERS in generate-models.ts). Both
 	// SWE-1.6 lanes are verified live against `GetCliModelConfigs`; the
-	// seed's `default-model` (`swe-1-6`) must resolve synchronously at
-	// boot, before credential-scoped runtime discovery replaces the seed. Field
+	// seed's `default-model` (`swe-1-6`) must resolve synchronously at boot;
+	// discovery then replaces the seed and the account default it marks wins. Field
 	// shape mirrors `devinModelSpec` so seeded and discovered rows are
 	// indistinguishable downstream.
 	seed api="devin-agent" base-url="https://server.codeium.com" bundle="always" {