@linxiraos/pi-catalog 1.1.13 → 1.1.15

package/src/compat/rules/README.md

@@ -207,6 +207,43 @@ The three value shapes are:
 A rule cannot assign the same resolved axis twice in one block.
 One object axis carries a computed form: `long-context-cost` accepts either the absolute rates (`input-threshold` + `input`/`output`/`cache-read`/`cache-write`) or `input-threshold` + `multiplier` (with optional `input-threshold-inclusive`), which derives the tier from the row's live base price at build time so the rule tracks upstream list-price updates (xAI's SuperGrok 200K tier). Rows without a token price carry no tier.
 
+### Time-based pricing
+
+The catalog object axis `time-based-cost` materializes into `ModelCost.timeBased`. It uses **named child objects**, not KDL arrays or repeated anonymous windows. For example, inside a matching provider/model scope:
+
+```kdl
+time-based-cost {
+    off-peak-multiplier 0.5
+    peak-windows {
+        morning {
+            weekdays "1,2,3,4,5"
+            start-minute 60
+            end-minute 240
+        }
+        afternoon {
+            weekdays "1,2,3,4,5"
+            start-minute 360
+            end-minute 600
+        }
+    }
+    effective-rates {
+        flash-pricing {
+            effective-from "2026-09-14T04:00:00Z"
+            input 0.30
+            output 1.20
+            cache-read 0.006
+            cache-write 0
+        }
+    }
+}
+```
+
+`morning`, `afternoon`, and `flash-pricing` are arbitrary unique object names, discarded when the payload is normalized into arrays. `weekdays` is a comma-separated string of distinct UTC weekday numbers (`0` = Sunday through `6` = Saturday), without spaces. Each window has integer minutes with `0 <= start-minute < end-minute <= 1440`; its start is inclusive and end exclusive. Split overnight windows across days. Outside the union of peak windows, the nonnegative `off-peak-multiplier` applies to token costs.
+
+`effective-rates` is optional. Each entry requires a distinct valid ISO UTC `effective-from` (`YYYY-MM-DDTHH:mm:ssZ` or with three fractional-second digits) and all four nonnegative per-million-token rates. The latest entry at or before the request timestamp replaces the entire base card; before the first entry, the base card applies. An entry may contain a `long-context` object with `input-threshold`, optional `input-threshold-inclusive`, and all four absolute rates. Effective rates do not inherit the base card's long-context tier. Selection order is effective card, context tier, then tariff multiplier.
+
+The recurring schedule and dated DeepSeek transition above come from [DeepSeek's official pricing](https://api-docs.deepseek.com/quick_start/pricing); see `providers/deepseek.kdl` for the complete rules and [the catalog API](../../../README.md#cost-calculation) for timestamp semantics. This is catalog policy metadata, not a supported schedule syntax for the coding agent's `models.yml`.
+
 ### Precedence and ambiguity
 
 Rules resolve independently per axis. A matching rule is ranked by:

package/src/compat/rules/auth/_order.kdl

@@ -5,7 +5,7 @@ login-order "openai-codex" "anthropic" "zai" "zai-coding-plan" "kimi-code" "open
 	"cursor" "devin" "google-antigravity" "google-gemini-cli" "openai-codex-device" "xai" "xai-oauth" \
 	"gitlab-duo" "gitlab-duo-agent" "alibaba-coding-plan" "alibaba-token-plan" "aiand" "abliteration" "zhipu-coding-plan" \
 	"umans" "qwen-portal" "sakana" "minimax-code" "minimax-code-cn" "xiaomi" "xiaomi-token-plan-sgp" \
-	"xiaomi-token-plan-ams" "xiaomi-token-plan-cn" "firepass" "cline-pass" "deepseek" "muse-code" "meta" "moonshot" \
+	"xiaomi-token-plan-ams" "xiaomi-token-plan-cn" "firepass" "cline-pass" "commandcode" "charm-hyper" "deepseek" "muse-code" "meta" "moonshot" \
 	"cerebras" "baseten" "fireworks" "together" "nvidia" "novita" "deepinfra" "huggingface" "perplexity" \
 	"qianfan" "venice" "siliconflow" "siliconflow-cn" "synthetic" "nanogpt" "wafer-serverless" "coreweave" \
 	"vercel-ai-gateway" "cloudflare-ai-gateway" "litellm" "kilo" "zenmux" "opencode-zen" "opencode-go" \

package/src/compat/rules/auth/charm-hyper.kdl

@@ -0,0 +1,19 @@
+// Validation targets `/v1/credits`, not `/v1/models`: the models endpoint is
+// public and answers 200 with no credentials or a bogus key (verified
+// 2026-09-11), so a models-endpoint probe against it could never reject a bad
+// key. `/v1/credits` is an authenticated, non-billable GET that returns the
+// account balance on success and 401 `{"error":"authentication failed"}`
+// otherwise — the same trick cline-pass uses to validate against its own
+// `/users/me`, and cheaper than a chat-completions probe that would bill the
+// key on every login.
+auth "charm-hyper" {
+	name "Charm Hyper"
+	env "CHARM_HYPER_API_KEY" "HYPER_API_KEY"
+	login "api-key" {
+		auth-url "https://hyper.charm.land/"
+		instructions "Create or copy an API key from the Charm Hyper dashboard"
+		prompt "Paste your Charm Hyper API key" placeholder="sk-hyper-..."
+		normalize "strip-bearer"
+		validate "models-endpoint" label="Charm Hyper" url="https://hyper.charm.land/v1/credits"
+	}
+}

package/src/compat/rules/auth/commandcode.kdl

@@ -0,0 +1,13 @@
+// No `validate` probe on purpose: GET /provider/v1/models is public (200
+// with no or bogus credentials, verified 2026-09-09), so a models-endpoint
+// probe could never reject a bad key. A chat-completions probe would bill
+// the key and 403 for Go-plan keys that are still valid CLI credentials.
+auth "commandcode" {
+	name "Command Code"
+	env "COMMAND_CODE_API_KEY" "COMMANDCODE_API_KEY"
+	login "api-key" {
+		auth-url "https://commandcode.ai/studio"
+		instructions "Create or copy a Provider API key from Command Code Studio"
+		prompt "Paste your Command Code API key" placeholder="user_..."
+	}
+}

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

@@ -7,6 +7,30 @@ class "anthropic" {
 			thinking-prefix-binding #true
 		}
 	}
+	// Server-side compaction (compact-2026-01-12) is a lineage capability: the
+	// documented model list is the adaptive-thinking generation onward (Opus
+	// 4.6+, Sonnet 4.6+, Fable/Mythos 5+). Whether a host delivers the beta is
+	// decided at request time from the effective endpoint, not here.
+	family "fable" {
+		revision ">=5" {
+			supports-server-compaction #true
+		}
+	}
+	family "mythos" {
+		revision ">=5" {
+			supports-server-compaction #true
+		}
+	}
+	family "opus" {
+		revision ">=4.6" {
+			supports-server-compaction #true
+		}
+	}
+	family "sonnet" {
+		revision ">=4.6" {
+			supports-server-compaction #true
+		}
+	}
 	on "anthropic" "cloudflare-ai-gateway" "google-vertex" {
 		family "fable" {
 			revision ">=5" {
@@ -145,6 +169,23 @@ class "anthropic" {
 				cache-read 1.0
 				cache-write 12.5
 			}
+			// Fable 5.1 cut cache reads to $0.25/MTok, 0.025x input, a 75% rate
+			// unique to the 5.1 generation; Fable 5 stays at the 0.1x family
+			// default. https://www.anthropic.com/claude/fable
+			//
+			// `cost-patch` is an object axis, so the winning rule supplies the
+			// whole payload rather than merging field-wise over the family
+			// patch above. Restate input/output/cache-write, or a
+			// discovery-sourced 5.1 row (Anthropic's /v1/models omits pricing)
+			// would keep only the cache-read correction.
+			revision "=5.1" {
+				cost-patch {
+					input 10.0
+					output 50.0
+					cache-read 0.25
+					cache-write 12.5
+				}
+			}
 			limits-patch {
 				context-window 1000000
 				max-tokens 128000

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

@@ -73,6 +73,17 @@ class "deepseek" {
 	models token="vision" {
 		strip-image-input #false
 	}
+	// DeepSeek V4.1 Flash is natively multimodal (image-text-to-text), but its
+	// id carries no `vision` token, so the class-wide strip above dropped
+	// attachments the model reads on every deployment without its own carve-out.
+	// The glob also spans `-vision`/`-ocr` ids, which already set the same value
+	// below, so the explicit priority resolves that equal-rank tie instead of
+	// throwing AmbiguousOverlapError. Only the strip is disabled; the declared
+	// modality stays host-derived, so endpoints whose OpenAI-compatible surface
+	// rejects `image_url` keep their text-only contract.
+	models "*v4.1-flash*" priority=10 {
+		strip-image-input #false
+	}
 	on "ollama-cloud" "nvidia" "deepseek" "fireworks" "nanogpt" "opencode-go" "openrouter" {
 		stream-markup-healing-pattern "dsml"
 	}

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

@@ -28,8 +28,10 @@ class "minimax" {
 	}
 	// MiniMax-M3: 512K is the standard pricing-tier boundary leaked into
 	// upstream metadata, not the model ceiling (documented: 1M in / 128K out).
-	on "minimax" "minimax-cn" "minimax-code" "minimax-code-cn" {
-		models "MiniMax-M3" {
+	// Exact selectors are case-sensitive, so both served spellings are listed:
+	// MiniMax hosts publish `MiniMax-M3`, Charm Hyper publishes `minimax-m3`.
+	on "charm-hyper" "minimax" "minimax-cn" "minimax-code" "minimax-code-cn" {
+		models "MiniMax-M3" "minimax-m3" {
 			limits-patch {
 				context-window 1000000
 				max-tokens 128000

package/src/compat/rules/providers/amazon-bedrock.kdl

@@ -15,6 +15,8 @@ provider "amazon-bedrock" {
 	}
 	class "openai" {
 		thinking-mode "effort"
+		// OpenAI models reject Converse image blocks nested inside toolResult content.
+		requires-tool-result-image-hoisting #true
 	}
 	class "deepseek" {
 		requires-reasoning-content-for-all-assistant-turns #true
@@ -28,6 +30,11 @@ provider "amazon-bedrock" {
 	class "unknown" {
 		thinking-efforts "minimal" "low" "medium" "high"
 		thinking-mode "budget"
+		// Opaque application-inference-profile ARNs carry no vendor signal, so an
+		// OpenAI-backed profile classifies here and would otherwise nest images in
+		// toolResult content and hit the same rejection. Hoisting is accepted by
+		// every Converse model measured, so it is the safe default for unknowns.
+		requires-tool-result-image-hoisting #true
 	}
 	// residue: taxonomy ranks and exact globs do not isolate these models.
 	models "moonshot.kimi-k2-thinking" {

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

@@ -1,6 +1,14 @@
 // Provider-wire compat for "anthropic"; regenerated from the frozen census using deployment contract selectors.
 
 provider "anthropic" {
+	// First-party provider: models served through this provider reach the
+	// official API unless rerouted. A dedicated axis rather than
+	// `official-endpoint`, which stays URL-derived: a custom `baseUrl` on this
+	// provider must keep `officialEndpoint: false` (SDK `X-Api-Key`
+	// suppression and friends read that flag). The compaction transport
+	// combines this fact with a per-request effective-URL check, so it never
+	// authorizes a block on its own.
+	first-party-provider #true
 	class "anthropic" {
 		revision ">=3.7 <4.6" {
 			thinking-efforts "minimal" "low" "medium" "high" "xhigh"

package/src/compat/rules/providers/charm-hyper.kdl

@@ -0,0 +1,82 @@
+// Charm Hyper gateway deployment contract.
+//
+// Capability data is NOT restated here: `GET /v1/models` publishes each row's
+// context window, output cap, vision flag, effort vocabulary and per-million
+// tariff, and `charmHyperModelManagerOptions` maps all of it verbatim —
+// including `reasoningDisableMode`, which is derived per model from whether
+// the advertised vocabulary contains `none`. Only facts the gateway does not
+// publish, or publishes wrongly, belong in this file.
+provider "charm-hyper" {
+	// OpenAI-dialect gateway in front of open-weight and third-party models.
+	// `developer` and `store` are accepted (HTTP 200) but the gateway accepts
+	// every unknown field it is handed, so acceptance is not evidence the
+	// semantics are honored; both stay off and system prompts ride the
+	// `system` role, matching every sibling gateway. The wire takes
+	// `max_tokens` (verified 2026-09-11).
+	supports-developer-role #false
+	supports-store #false
+	max-tokens-field "max_tokens"
+	thinking-mode "effort"
+	// Reasoning arrives as `reasoning_content` on both the streamed delta and
+	// the final message (verified 2026-09-11).
+	reasoning-content-field "reasoning_content"
+	// ---------------------------------------------------------------------------
+	// Blockless rows are left non-reasoning on purpose.
+	//
+	// Four ids serve a thinking model while publishing no `reasoning` block:
+	// glm-5, glm-5.1, kimi-k2-thinking and minimax-m2.7 each returned
+	// populated `reasoning_content` and non-zero `reasoning_tokens` on a
+	// multi-step prompt (2026-09-11). The other blockless rows
+	// (gemma-4-26b-a4b-it, kimi-k2.5, kimi-k2.7-code, llama-3.3-70b-instruct,
+	// llama-4-maverick-17b-128e-instruct-fp8,
+	// qwen3-coder-480b-a35b-instruct-int4-mixed-ar, qwen3-next-80b-a3b-instruct)
+	// returned none.
+	//
+	// No rule restores them, because none can be written honestly. Those four
+	// expose no effort dial, and `ThinkingConfig.efforts` is documented as
+	// never empty — a reasoning model without a controllable surface carries
+	// `thinking: undefined`, which no KDL axis can express (the only upgrade
+	// path is an exact `thinking-efforts` ladder). Synthesizing a ladder would
+	// add silent no-op rungs: the gateway answers 200 to efforts it never
+	// advertised and simply ignores them, and a repeat measurement of effort
+	// sensitivity on these ids came back non-monotonic, i.e. sampling noise
+	// rather than a hidden dial.
+	//
+	// The cost is narrow and was checked: streamed `reasoning_content` is
+	// parsed and displayed with no `model.reasoning` gate, the effort
+	// parameter is moot without a dial, and the developer-role and
+	// zai-replay paths are already disabled for this provider.
+	// ---------------------------------------------------------------------------
+	// Residue: misreported output caps. The gateway derives
+	// `max_output_tokens` as a flat fraction of the context window rather
+	// than reporting a real ceiling — 10% on 9 rows, the dominant
+	// convention — and the values are advisory, not enforced: glm-5.1
+	// publishes 3276 yet produced 14066 tokens, and minimax-m2.7 publishes
+	// 6553 yet produced 8418, both stopping on `finish_reason: "stop"`
+	// rather than `length`. An over-stated cap is never rejected either
+	// (`max_tokens: 200000` returns 200), so the risk is one-sided:
+	// over-stating costs only the context share 10% already reserves
+	// everywhere else, while under-stating truncates real output.
+	//
+	// Both corrections take a same-window peer's published value rather than
+	// inventing a ceiling, and both clear the measured floor:
+	//   glm-5.1      (202750) is 1.6%; sibling glm-5 (202752) publishes 20275.
+	//   minimax-m2.7 (262100) is 2.5%, the lone outlier in its window class;
+	//                kimi-k2.5 (262144), kimi-k2.6 (262000) and
+	//                qwen3-next-80b (262144) all publish 26214.
+	//
+	// deepseek-v4.1-flash is the only other 2.5% row and is deliberately left
+	// alone: its 1M window has no peer group that agrees on a fraction, and a
+	// long generation stopped naturally at 8504, well inside the published
+	// 26214. Measured 2026-09-11.
+	models "glm-5.1" {
+		limits-patch {
+			max-tokens 20275
+		}
+	}
+	models "minimax-m2.7" {
+		limits-patch {
+			max-tokens 26214
+		}
+	}
+}