@cohortapp/agent-sdk 2.3.0

package/lib/model-router/catalog/anthropic.yaml

@@ -0,0 +1,153 @@
+# lib/model-router/catalog/anthropic.yaml
+# Bundled Anthropic catalog rows (ships with @cohortapp/agent-sdk).
+#
+# Prices are $/MTok, sourced EXACT from
+#   docs/audit/2026-06-10-enterprise/research/model-pricing-matrix.md §1.1
+#   (Anthropic pricing page, fetched 2026-06-10).
+# cache_read = 0.1x input; 5m cache write = 1.25x input; 1h cache write = 2x input.
+# cache_write below records the 1h-write figure (the maestro default cache_ttl is 1h;
+#   the 5m figure is derivable as input * 1.25 when a lane uses 5m TTL).
+# Batch = 50% off input and output.
+#
+# tool_reliability: A across the board — Anthropic is the reference harness
+#   (Claude Code, tool SDKs, Managed Agents are all first-party).
+# cost_provenance.volatile: false — Anthropic prices are stable list prices.
+#
+# DO NOT hand-edit cost fields to track a promo: the pricing-cache layer
+# (state/model-router/pricing-cache.json) updates cost ONLY; capabilities,
+# compat, and tool_reliability are human-curated here.
+
+provider: anthropic
+auth_env: ANTHROPIC_API_KEY
+endpoints:
+  anthropic: https://api.anthropic.com
+  openai: https://api.anthropic.com/v1   # compat shim — test/compare only, no caching
+data_residency: us
+models:
+  # ---- Frontier agentic tier -------------------------------------------------
+  - id: claude-opus-4-8
+    status: available
+    replaced_by: null
+    context_window: 1000000
+    context_tokens: 950000
+    max_tokens: 128000
+    cost:
+      input: 5.0
+      output: 25.0
+      cache_read: 0.5        # 0.1x input
+      cache_write: 10.0      # 1h write = 2x input
+      tiered: []
+    cost_provenance: { source: official, fetched: "2026-06-10", volatile: false }
+    compat:
+      cache_control: anthropic
+      max_tokens_field: max_tokens
+      thinking_format: anthropic
+      requires_reasoning_echo: false
+      rejects_duplicate_tool_names: false
+      omit_temperature: true       # Opus 4.8: sampling params removed, adaptive-thinking-only
+      supports_prompt_cache_key: true
+    tool_reliability: A
+    harness: { session: true, direct: true, batch: true }
+    quality: { swe_verified: 88.6, swe_pro: 69.2, terminal_bench: 74.6 }
+
+  - id: claude-opus-4-7
+    status: deprecated
+    replaced_by: claude-opus-4-8
+    context_window: 1000000
+    context_tokens: 950000
+    max_tokens: 128000
+    cost:
+      input: 5.0
+      output: 25.0
+      cache_read: 0.5
+      cache_write: 10.0
+      tiered: []
+    cost_provenance: { source: official, fetched: "2026-06-10", volatile: false }
+    compat:
+      cache_control: anthropic
+      max_tokens_field: max_tokens
+      thinking_format: anthropic
+      requires_reasoning_echo: false
+      rejects_duplicate_tool_names: false
+      omit_temperature: false
+      supports_prompt_cache_key: true
+    tool_reliability: A
+    harness: { session: true, direct: true, batch: true }
+    quality: { swe_verified: 87.6, swe_pro: 64.3, terminal_bench: 66.1 }
+
+  # ---- Default workhorse tier (1M ctx at standard pricing, no LC premium) -----
+  - id: claude-sonnet-4-6
+    status: available
+    replaced_by: null
+    context_window: 1000000
+    context_tokens: 950000
+    max_tokens: 64000
+    cost:
+      input: 3.0
+      output: 15.0
+      cache_read: 0.3        # 0.1x input
+      cache_write: 6.0       # 1h write = 2x input
+      tiered: []
+    cost_provenance: { source: official, fetched: "2026-06-10", volatile: false }
+    compat:
+      cache_control: anthropic
+      max_tokens_field: max_tokens
+      thinking_format: anthropic
+      requires_reasoning_echo: false
+      rejects_duplicate_tool_names: false
+      omit_temperature: false
+      supports_prompt_cache_key: true
+    tool_reliability: A
+    harness: { session: true, direct: true, batch: true }
+    quality: { swe_verified: null, terminal_bench: null }
+
+  - id: claude-sonnet-4-5
+    status: deprecated
+    replaced_by: claude-sonnet-4-6
+    context_window: 200000
+    context_tokens: 190000
+    max_tokens: 64000
+    cost:
+      input: 3.0
+      output: 15.0
+      cache_read: 0.3
+      cache_write: 6.0
+      tiered: []
+    cost_provenance: { source: official, fetched: "2026-06-10", volatile: false }
+    compat:
+      cache_control: anthropic
+      max_tokens_field: max_tokens
+      thinking_format: anthropic
+      requires_reasoning_echo: false
+      rejects_duplicate_tool_names: false
+      omit_temperature: false
+      supports_prompt_cache_key: true
+    tool_reliability: A
+    harness: { session: true, direct: true, batch: true }
+    quality: {}
+
+  # ---- Utility / fast tier ---------------------------------------------------
+  - id: claude-haiku-4-5
+    status: available
+    replaced_by: null
+    context_window: 200000
+    context_tokens: 190000
+    max_tokens: 64000
+    cost:
+      input: 1.0
+      output: 5.0
+      cache_read: 0.1        # 0.1x input
+      cache_write: 2.0       # 1h write = 2x input
+      tiered: []
+    cost_provenance: { source: official, fetched: "2026-06-10", volatile: false }
+    compat:
+      cache_control: anthropic
+      max_tokens_field: max_tokens
+      thinking_format: anthropic    # no `effort` param; thinking layout still anthropic-native
+      requires_reasoning_echo: false
+      rejects_duplicate_tool_names: false
+      omit_temperature: false
+      supports_prompt_cache_key: true
+    tool_reliability: A
+    harness: { session: true, direct: true, batch: true }
+    quality: { swe_verified: null, terminal_bench: null }

package/lib/model-router/catalog/deepseek.yaml

@@ -0,0 +1,86 @@
+# lib/model-router/catalog/deepseek.yaml
+# Bundled DeepSeek V4 catalog rows (ships with @cohortapp/agent-sdk).
+#
+# Prices are $/MTok, sourced EXACT from
+#   docs/audit/2026-06-10-enterprise/research/model-pricing-matrix.md §2.1
+#   (DeepSeek pricing page, fetched 2026-06-10).
+#
+# cost_provenance.volatile: TRUE — V4-Pro launched with a "limited-time 75%
+#   discount until 2026/05/31"; the page now shows $0.435/$0.87 with no promo
+#   language, but Fireworks lists US-hosted V4-Pro at exactly 4x ($1.74/$3.48),
+#   strongly suggesting $1.74/$3.48 is the undiscounted reference. Because
+#   volatile:true, budget-guard projects spend at the HIGHER steady-state figure
+#   (see cost.tiered[].steady_state and ledger.mjs estimateCost) so a promo
+#   expiry never silently busts a budget. Re-verify before locking past July.
+#
+# endpoints.anthropic is the session-retarget lever: set
+#   ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic and Claude Code runs
+#   on DeepSeek (claude-opus-* -> v4-pro; claude-sonnet-*/haiku-* -> v4-flash).
+#
+# Context caching is AUTOMATIC (no cache_control plumbing) -> compat.cache_control: implicit.
+# cache_write: null (no write premium on automatic caching).
+#
+# tool_reliability: B — strong (MCPAtlas, Terminal-Bench) but session-hosting
+#   requires probe-ledger grade A (§6.6) before any cheap-session rule can fire.
+
+provider: deepseek
+auth_env: DEEPSEEK_API_KEY
+endpoints:
+  anthropic: https://api.deepseek.com/anthropic
+  openai: https://api.deepseek.com
+data_residency: prc
+models:
+  - id: deepseek-v4-flash
+    status: available
+    replaced_by: null
+    context_window: 1000000
+    context_tokens: 900000
+    max_tokens: 384000
+    cost:
+      input: 0.14
+      output: 0.28
+      cache_read: 0.0028     # cheapest cache-hit in the industry (~1/50 of miss)
+      cache_write: null      # automatic caching, no write premium
+      tiered: []
+    cost_provenance: { source: official, fetched: "2026-06-10", volatile: true }
+    compat:
+      cache_control: implicit
+      max_tokens_field: max_tokens
+      thinking_format: deepseek      # explicit extra_body.thinking after first tool call
+      requires_reasoning_echo: true
+      rejects_duplicate_tool_names: true
+      omit_temperature: false
+      supports_prompt_cache_key: false
+    tool_reliability: B
+    harness: { session: true, direct: true, batch: false }
+    quality: {}
+
+  - id: deepseek-v4-pro
+    status: available
+    replaced_by: null
+    context_window: 1000000
+    context_tokens: 900000
+    max_tokens: 384000
+    cost:
+      # Official page price (possibly post-promo); volatile:true means budgeting
+      # projects at the higher steady_state below, not these promo figures.
+      input: 0.435
+      output: 0.87
+      cache_read: 0.003625
+      cache_write: null
+      tiered:
+        # steady_state = the undiscounted reference (Fireworks US-hosted, 4x).
+        # ledger.estimateCost reads this when volatile:true for budget projection.
+        - { kind: steady_state, input: 1.74, output: 3.48, cache_read: 0.145 }
+    cost_provenance: { source: official, fetched: "2026-06-10", volatile: true }
+    compat:
+      cache_control: implicit
+      max_tokens_field: max_tokens
+      thinking_format: deepseek
+      requires_reasoning_echo: true
+      rejects_duplicate_tool_names: true
+      omit_temperature: false
+      supports_prompt_cache_key: false
+    tool_reliability: B
+    harness: { session: true, direct: true, batch: false }
+    quality: { swe_verified: 80.6, terminal_bench: 67.9 }

package/lib/model-router/catalog/moonshot.yaml

@@ -0,0 +1,81 @@
+# lib/model-router/catalog/moonshot.yaml
+# Bundled Moonshot AI (Kimi) catalog rows (ships with @cohortapp/agent-sdk).
+#
+# Prices are $/MTok, sourced EXACT from
+#   docs/audit/2026-06-10-enterprise/research/model-pricing-matrix.md §3.1
+#   (platform.kimi.ai pricing page, fetched 2026-06-10).
+#
+# K2.6 is the only K2.x on the official first-party price page
+#   (K2/K2.5 retired from first-party; K2.5 still served by Fireworks).
+#   We carry a disabled k2.5 row for catalog completeness / succession.
+#
+# endpoints.anthropic is THE Claude-Code-on-Kimi lever: Moonshot ships an
+#   Anthropic-compatible endpoint, so a Claude Code session can be re-pointed
+#   per-task via ANTHROPIC_BASE_URL + ANTHROPIC_MODEL with no harness change.
+#
+# cost_provenance.volatile: TRUE — Chinese open-weight provider pricing moves;
+#   budget at the same sticker (no known promo on K2.6) but flag for re-verify.
+#
+# Caching: Moonshot exposes cache-hit pricing ($0.16 vs $0.95 miss); cache is
+#   provider-managed -> compat.cache_control: implicit; cache_write: null.
+#
+# tool_reliability: B — among the best open models (tau2-telecom 96, Toolathlon
+#   50); session-hosting still requires probe-ledger grade A (§6.6) + Phase 5 gates.
+
+provider: moonshot
+auth_env: MOONSHOT_API_KEY
+endpoints:
+  anthropic: https://api.moonshot.ai/anthropic
+  openai: https://api.moonshot.ai/v1
+data_residency: prc
+models:
+  - id: kimi-k2.6
+    status: available
+    replaced_by: null
+    context_window: 262144
+    context_tokens: 250000
+    max_tokens: 65536
+    cost:
+      input: 0.95
+      output: 4.0
+      cache_read: 0.16
+      cache_write: null      # provider-managed cache, no write premium
+      tiered: []
+    cost_provenance: { source: official, fetched: "2026-06-10", volatile: true }
+    compat:
+      cache_control: implicit
+      max_tokens_field: max_tokens
+      thinking_format: openai   # OpenAI-format API; thinking/Instant modes
+      requires_reasoning_echo: false
+      rejects_duplicate_tool_names: false
+      omit_temperature: false
+      supports_prompt_cache_key: false
+    tool_reliability: B
+    harness: { session: true, direct: true, batch: false }
+    quality: { swe_verified: 80.2, swe_pro: 58.6, tau2_telecom: 96.0 }
+
+  - id: kimi-k2.5
+    status: deprecated
+    replaced_by: kimi-k2.6
+    context_window: 262144
+    context_tokens: 250000
+    max_tokens: 65536
+    cost:
+      # Retired from first-party; Fireworks still serves at $0.60/$0.10 cached/$3.00.
+      input: 0.60
+      output: 3.0
+      cache_read: 0.10
+      cache_write: null
+      tiered: []
+    cost_provenance: { source: secondary, fetched: "2026-06-10", volatile: true }
+    compat:
+      cache_control: implicit
+      max_tokens_field: max_tokens
+      thinking_format: openai
+      requires_reasoning_echo: false
+      rejects_duplicate_tool_names: false
+      omit_temperature: false
+      supports_prompt_cache_key: false
+    tool_reliability: B
+    harness: { session: true, direct: true, batch: false }
+    quality: {}

package/lib/model-router/catalog/qwen.yaml

@@ -0,0 +1,114 @@
+# lib/model-router/catalog/qwen.yaml
+# Bundled Alibaba Qwen catalog rows (ships with @cohortapp/agent-sdk).
+#
+# Prices are $/MTok, sourced EXACT from
+#   docs/audit/2026-06-10-enterprise/research/model-pricing-matrix.md §5.1
+#   (Alibaba Model Studio international/Singapore pricing, fetched 2026-06-10).
+#
+# ROUTER HAZARD (§5.3): Qwen prices are TIERED BY INPUT PROMPT SIZE. A 140K-token
+#   prompt to a tiered model costs multiples of the 30K-token rate. The
+#   cost.tiered[] array encodes these ranges; ledger.estimateCost picks the tier
+#   whose [min_input_tokens, max_input_tokens) bracket contains the actual input
+#   token count, falling back to cost.input/output when tiered is empty.
+#   tiered[].kind == "input_range" distinguishes these from the volatile
+#   "steady_state" tier used by DeepSeek.
+#
+# cost_provenance.volatile: TRUE — Chinese-provider pricing + promo cadence.
+# Caching: implicit (DashScope), modest discount; we leave cache_read null on
+#   the utility rows where the page does not separately publish a cache-hit rate.
+#
+# tool_reliability: B for the flagship hybrid (3.5-plus), C for the cheapest
+#   utility (3.5-flash) — DashScope function calling works but is a tier below
+#   the agentic leaders; reserve flash for text-only / direct lanes.
+
+provider: qwen
+auth_env: DASHSCOPE_API_KEY
+endpoints:
+  openai: https://dashscope-intl.aliyuncs.com/compatible-mode/v1
+data_residency: sg
+models:
+  - id: qwen3.5-flash
+    status: available
+    replaced_by: null
+    context_window: 1000000
+    context_tokens: 900000
+    max_tokens: 65536
+    cost:
+      input: 0.10
+      output: 0.40
+      cache_read: null
+      cache_write: null
+      tiered: []      # flat rate (no prompt-size tiering on the flash utility tier)
+    cost_provenance: { source: official, fetched: "2026-06-10", volatile: true }
+    compat:
+      cache_control: implicit
+      max_tokens_field: max_tokens
+      thinking_format: openai
+      requires_reasoning_echo: false
+      rejects_duplicate_tool_names: false
+      omit_temperature: false
+      supports_prompt_cache_key: false
+    tool_reliability: C
+    harness: { session: false, direct: true, batch: false }
+    quality: {}
+
+  - id: qwen3.5-plus
+    status: available
+    replaced_by: null
+    context_window: 1000000
+    context_tokens: 900000
+    max_tokens: 65536
+    cost:
+      # Tiered by prompt size: <=256K input is the cheap bracket, >256K steps up.
+      # input/output below mirror the first (cheapest) bracket so cache-blind /
+      # legacy callers still get a sane base; estimateCost prefers tiered[].
+      input: 0.40
+      output: 2.40
+      cache_read: null
+      cache_write: null
+      tiered:
+        - { kind: input_range, min_input_tokens: 0,      max_input_tokens: 256000, input: 0.40, output: 2.40 }
+        - { kind: input_range, min_input_tokens: 256000, max_input_tokens: null,   input: 0.50, output: 3.00 }
+    cost_provenance: { source: official, fetched: "2026-06-10", volatile: true }
+    compat:
+      cache_control: implicit
+      max_tokens_field: max_tokens
+      thinking_format: openai
+      requires_reasoning_echo: false
+      rejects_duplicate_tool_names: false
+      omit_temperature: false
+      supports_prompt_cache_key: false
+    tool_reliability: B
+    harness: { session: false, direct: true, batch: false }
+    quality: { swe_verified: 76.2, swe_pro: 50.9, terminal_bench: 52.5 }
+
+  - id: qwen3-coder-plus
+    status: available
+    replaced_by: null
+    context_window: 1000000
+    context_tokens: 900000
+    max_tokens: 65536
+    cost:
+      # 4-tier prompt-size pricing ($1.00 -> $6.00 input; $5 -> $60 output).
+      # Boundaries follow the documented 32K/128K/256K class steps.
+      input: 1.00
+      output: 5.00
+      cache_read: null         # cache: 20% implicit / 10% explicit (page note)
+      cache_write: null
+      tiered:
+        - { kind: input_range, min_input_tokens: 0,      max_input_tokens: 32000,  input: 1.00, output: 5.00 }
+        - { kind: input_range, min_input_tokens: 32000,  max_input_tokens: 128000, input: 1.80, output: 9.00 }
+        - { kind: input_range, min_input_tokens: 128000, max_input_tokens: 256000, input: 3.00, output: 15.00 }
+        - { kind: input_range, min_input_tokens: 256000, max_input_tokens: null,   input: 6.00, output: 60.00 }
+    cost_provenance: { source: official, fetched: "2026-06-10", volatile: true }
+    compat:
+      cache_control: implicit
+      max_tokens_field: max_tokens
+      thinking_format: openai
+      requires_reasoning_echo: false
+      rejects_duplicate_tool_names: false
+      omit_temperature: false
+      supports_prompt_cache_key: false
+    tool_reliability: B
+    harness: { session: false, direct: true, batch: false }
+    quality: {}