llm-orchestrator 1.0.7 → 1.2.0

package/models/model-thinking-matrix.md

@@ -43,6 +43,7 @@ Benchmark cost index: `100 × observed benchmark cost / $0.50`; GPT-5.6 Sol medi
 | claude-opus-5 | Claude Opus 5 | high | 48 | $3.61 | 722 | 125 | +$1.42 (1.65×); +3 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
 | claude-opus-5 | Claude Opus 5 | xhigh | 50 | $4.88 | 976 | 125 | +$1.27 (1.35×); +2 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
 | claude-opus-5 | Claude Opus 5 | max | 51 | $5.86 | 1172 | 125 | +$0.98 (1.20×); +1 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
+| claude-opus-5-5 | Claude Opus 5.5 | max | 58 | $5.98 | 1196 | 100 | — | Segregate from standalone cross-model comparisons. |
 | claude-fable-5-1 | Claude Fable 5.1 | low | 47 | $2.37 | 474 | 250 | — | Segregate from standalone cross-model comparisons. |
 | claude-fable-5-1 | Claude Fable 5.1 | medium | 49 | $2.98 | 596 | 250 | +$0.61 (1.26×); +2 score | Segregate from standalone cross-model comparisons. |
 | claude-fable-5-1 | Claude Fable 5.1 | high | 51 | $3.91 | 782 | 250 | +$0.93 (1.31×); +2 score | Segregate from standalone cross-model comparisons. |
@@ -88,6 +89,7 @@ Benchmark cost index: `100 × observed benchmark cost / $0.50`; GPT-5.6 Sol medi
 | GPT-6 Astra | $10.000 | $50.000 | AA-observed price; not a live account tariff. |
 | Claude Sonnet 5 | $2.000 | $10.000 | AA-observed price; not a live account tariff. |
 | Claude Opus 5 | $5.000 | $25.000 | AA-observed price; not a live account tariff. |
+| Claude Opus 5.5 | $4.000 | $20.000 | AA-observed price; not a live account tariff. |
 | Claude Fable 5.1 | $10.000 | $50.000 | AA-observed price; not a live account tariff. |
 | Claude Fable 5 | $10.000 | $50.000 | AA-observed price; not a live account tariff. |
 | Claude Haiku 4.5 | $1.000 | $5.000 | AA-observed price; not a live account tariff. |
@@ -114,6 +116,7 @@ Benchmark cost index: `100 × observed benchmark cost / $0.50`; GPT-5.6 Sol medi
 | GPT-6 Astra | `gpt-6-astra` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
 | Claude Sonnet 5 | `claude-sonnet-5` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
 | Claude Opus 5 | `claude-opus-5` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
+| Claude Opus 5.5 | `claude-opus-5-5` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
 | Claude Fable 5.1 | `claude-fable-5-1` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
 | Claude Fable 5 | `claude-fable-5` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
 | Claude Haiku 4.5 | `claude-haiku-4-5` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
@@ -138,6 +141,7 @@ Benchmark cost index: `100 × observed benchmark cost / $0.50`; GPT-5.6 Sol medi
 - [GPT-6 Astra](https://artificialanalysis.ai/models/gpt-6-astra)
 - [Claude Sonnet 5](https://artificialanalysis.ai/models/claude-sonnet-5)
 - [Claude Opus 5](https://artificialanalysis.ai/models/claude-opus-5)
+- [Claude Opus 5.5](https://artificialanalysis.ai/models/claude-opus-5-5) — Only the max-effort configuration is published (Adaptive Reasoning, Max Effort, Default Fallback); low–xhigh are not measured. The score includes the benchmark default fallback.
 - [Claude Fable 5.1](https://artificialanalysis.ai/models/claude-fable-5-1) — All displayed scores include the benchmark default fallback and are not standalone scores.
 - [Claude Fable 5](https://artificialanalysis.ai/models/claude-fable-5) — AA v4.3.2 max result uses Opus 4.8 default fallback. This legacy catalog model is not in the current leaderboard filter; segregate it from standalone cross-model comparisons.
 - [Claude Haiku 4.5](https://artificialanalysis.ai/models/claude-4-5-haiku) — Its API thinking budget is not an effort enum. The disabled score 15 is incomplete and intentionally excluded.

package/models/top-models.json

@@ -23,7 +23,7 @@
       "X": {"min": 45, "max": 50},
       "F": {"min": 51, "max": null}
     },
-    "note": "Bands are calibrated on the incumbents' measured range: Luna tops W at 37; Terra spans 38–42 in S; Sol 33–47 and Opus 39–50 straddle into X, whose seats sit at 45–50; Astra 46–53 and Fable 5.1 47–53 cover F from 51 up. They place *candidates*, which sit on no provider ladder. An incumbent's `tier` is its ladder seat and is never re-derived from a score — which is why two incumbent points fall outside their own band: Opus 5 `max` (51, above its `xhigh` policy cap) and Fable 5 `max` (50, its only measured point and above the unmeasured `high` that is the actual F default)."
+    "note": "Bands are calibrated on the incumbents' measured range: Luna tops W at 37; Terra spans 38–42 in S; Sol 33–47 and Opus 5 39–50 straddle into X, whose seats sit at 45–50; Astra 46–53 and Fable 5.1 47–53 cover F from 51 up. They place *candidates*, which sit on no provider ladder. An incumbent's `tier` is its ladder seat and is never re-derived from a score — which is why two incumbent points fall outside their own band: Opus 5 `max` (51, above its `xhigh` policy cap) and Fable 5 `max` (50, its only measured point and above the unmeasured `high` that is the actual F default)."
   },
   "indices_note": "`indices.thinking_cost_index` is 100 × measured $/task ÷ $0.50, the dataset's benchmark cost baseline (`index_baselines.benchmark_cost_usd`). GPT-5.6 Sol `medium` costs exactly $0.50/task, so it is index 100 by construction. `indices.marginal_thinking` holds, for each adjacent pair of fully measured efforts, the displayed-score delta and the cost multiplier of the step — the price of one more notch of thinking. Efforts with no published cost carry a null index and open no marginal pair; nothing is interpolated.",
   "models": [
@@ -491,7 +491,7 @@
         "share_max": null
       },
       "notes": [
-        "Opus 4.8/4.7/4.6 share the $5/$25 price and are fallback-only.",
+        "Fallback for Claude Opus 5.5 on the Claude X seat: dispatched when Opus 5.5 is not exposed, refuses (it runs without Opus 5.5 thinking blocks), or is rolled back. Opus 4.8/4.7/4.6 share the $5/$25 price and are fallback-only below it.",
         "Every measured Opus config is dominated on $/task by a Sol or Astra config; it stays the Claude X seat for single-provider flows and retention-constrained work."
       ],
       "sources": [
@@ -499,6 +499,64 @@
       ],
       "observed_at": "2026-09-22"
     },
+    {
+      "key": "claude-opus-5-5",
+      "display_name": "Claude Opus 5.5",
+      "provider": "anthropic",
+      "ladder": "claude",
+      "admission": "incumbent",
+      "tier": "X",
+      "eligible_tiers": ["X"],
+      "supersedes": "claude-opus-5",
+      "api_ids": {
+        "codex": null,
+        "claude": "claude-opus-5-5",
+        "opencode": null,
+        "kilo": null
+      },
+      "context_window": 1000000,
+      "max_output_tokens": 128000,
+      "thinking": {
+        "control": "effort",
+        "levels": ["low", "medium", "high", "xhigh", "max"],
+        "unmeasured_levels": ["low", "medium", "high", "xhigh"],
+        "always_on": true,
+        "default_for_tier": "high",
+        "note": "Thinking is always on: omit `thinking` or send `adaptive`; `disabled` and `budget_tokens` return 400 at every effort. The API default effort is `medium` (one below Opus 5), so X always sets effort explicitly — `high` for T3, `xhigh` for T4. Forced `tool_choice` any/tool returns 400 — use `auto` + instruction, `strict: true`, or structured outputs. Thinking blocks are model-bound; keep transcripts append-only. Computer use only via `computer_toolset_20260801`."
+      },
+      "price": {
+        "input_usd_per_mtok": 4,
+        "output_usd_per_mtok": 20,
+        "cache_read_usd_per_mtok": 0.2,
+        "cache_write_multiplier": 1.25,
+        "fast_mode_note": "`speed: \"fast\"` reprices Opus 5.5 to $8/$40 (Claude API only). Off by default; it never replaces an F dispatch."
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "max": {"score": 58, "usd_per_task": 5.98}
+      },
+      "measurement_note": "Only `max` is published (Adaptive Reasoning, Max Effort, Default Fallback); the score includes the benchmark default fallback. low–xhigh are unmeasured, so a dispatch at them carries no $/task estimate until Artificial Analysis publishes them.",
+      "indices": {
+        "thinking_cost_index": {
+          "max": 1196
+        },
+        "marginal_thinking": {}
+      },
+      "caps": {
+        "max_effort": "xhigh",
+        "share_max": null
+      },
+      "notes": [
+        "Successor to Claude Opus 5 on the Claude X seat. Every per-token price is lower (Anthropic list: input $4 vs $5, output $20 vs $25, cache read $0.20 vs $0.50, cache write $5 vs $6.25), with the same 1M context, 128K output and tokenizer, so the router ranks it ahead of Opus 5 at the same effort; Opus 5 stays routable as its fallback.",
+        "Only `max` is independently measured (Artificial Analysis 58 @ $5.98, above every other shortlisted config). Anthropic's launch table puts it above Opus 5 on every published benchmark (Terminal-Bench 4.0 66.4 vs 52.3, FrontierCode v1.1 54.4 vs 48.0, CursorBench 4.0 57.8 vs 46.6) and claims ~40% lower cost on typical workloads — vendor figures, kept as notes and never used as $/task estimates.",
+        "Seat stays X: an incumbent tier is never re-derived from a score."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/claude-opus-5-5",
+        "https://www.anthropic.com/claude-opus-5-5"
+      ],
+      "observed_at": "2026-09-22"
+    },
     {
       "key": "claude-fable-5",
       "display_name": "Claude Fable 5",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "llm-orchestrator",
-  "version": "1.0.7",
+  "version": "1.2.0",
   "description": "Write /task once — it plans the work, shards it across parallel subagents, gates every phase and verifies before claiming done. Claude Code, Codex, OpenCode, Kilo.",
   "type": "module",
   "engines": {
@@ -47,6 +47,7 @@
     "LICENSE",
     "NOTICE",
     "skills",
+    "hooks",
     ".claude-plugin"
   ],
   "scripts": {

package/policies/routing.md

@@ -76,7 +76,7 @@ the incumbents below are role mappings, not guarantees of availability.
 | --- | --- | --- | --- |
 | **W** worker | local, mechanical, repetitive, well-defined: code search, classification, extraction, small edits, boilerplate, simple tests, consistency checks, scoped transforms | Haiku 4.5 (`claude-haiku-4-5`) | `gpt-5.6-luna` |
 | **S** standard | default software-engineering model: normal implementation, frontend/backend, moderate debugging, tests, reasonable multi-file refactors, codebase analysis, tool use | Sonnet 5 (`claude-sonnet-5`) | `gpt-5.6-terra` |
-| **X** senior | hard debugging, architecture, concurrency, migrations, security, auth, payments, billing, backwards compatibility, critical code review, many invariants | Opus 5 (`claude-opus-5`) | `gpt-5.6-sol` |
+| **X** senior | hard debugging, architecture, concurrency, migrations, security, auth, payments, billing, backwards compatibility, critical code review, many invariants | Opus 5.5 (`claude-opus-5-5`), falling back to Opus 5 (`claude-opus-5`) | `gpt-5.6-sol` |
 | **F** frontier | exceptional escalation: very ambiguous, long-horizon, cross-system, major architecture, very large codebase, planning under heavy constraints, or when X fails to produce a solid solution | Fable 5 (`claude-fable-5`) — default F. Fable 5.1 (`claude-fable-5-1`) is a hard-capped exception: **≤2% of all dispatches**, explicit request or documented F-T4 failure on Fable 5 only | GPT-6 Astra (`gpt-6-astra`) — a real single-agent frontier tier, no decomposition workaround needed |
 
 Escalation order within a provider: **W → S → X → F**.
@@ -87,15 +87,20 @@ Verified prices (Sep 2026, provider pricing pages), USD in/out per MTok:
 | --- | --- | --- | --- | --- |
 | W | Haiku 4.5 (200K ctx) | $1 / $5 | `gpt-5.6-luna` (1.05M ctx) | $0.20 / $1.20 |
 | S | Sonnet 5 (1M ctx) | $2 / $10 | `gpt-5.6-terra` (1.05M ctx) | $2 / $12 |
-| X | Opus 5 (1M ctx) | $5 / $25 | `gpt-5.6-sol` (1.05M ctx) | $4 / $20 |
+| X | Opus 5.5 (1M ctx); fallback Opus 5 $5 / $25 | $4 / $20 | `gpt-5.6-sol` (1.05M ctx) | $4 / $20 |
 | F | Fable 5 (1M ctx) | $10 / $50 | `gpt-6-astra` (1.05M ctx) | $10 / $50 list — **but the cheapest F per completed task of any model here** |
 | F+ (≤2%) | Fable 5.1 (1M ctx) | $10 / $50 list — **effective cost significantly higher** (always-on thinking, longer turns, more output tokens per task) | — (Astra covers F) | — |
 
 Claude notes: Fable 5.1 shares Fable 5's list price but costs significantly more per completed task
 — judge it on cost per task, not per token. **When F is needed, use Fable 5; Fable 5.1 is capped at
 ≤2% of dispatches.** Fable 5.1 cache reads bill at $0.25/MTok; Fable 5 uses the standard
-10%-of-input cache-read rate. Opus 5 fast mode (`speed: "fast"`) reprices Opus to $10 / $50 —
-Fable-tier cost for Opus-tier output, so it never replaces an F dispatch and is off by default.
+10%-of-input cache-read rate. Opus 5.5 is at or below Opus 5 on every per-token price (input $4 vs $5,
+output $20 vs $25, cache read $0.20 vs $0.50, cache write $5 vs $6.25); only its `max` effort is
+independently measured so far, so the router ranks it ahead of Opus 5 by `supersedes` price
+succession — never by an estimated $/task — and keeps Opus 5 as the fallback when 5.5 is not exposed
+or refuses. Opus 5.5's API default effort is `medium`, one below Opus 5, so X always sets effort
+explicitly. Fast mode (`speed: "fast"`) reprices Opus 5.5 to $8 / $40 and Opus 5 to $10 / $50 —
+never a replacement for an F dispatch, off by default.
 Legacy fallbacks: Opus 4.8 / 4.7 / 4.6 $5 / $25, Sonnet 4.6 $3 / $15 (more expensive than Sonnet 5 —
 never pick it for cost).
 

package/protocol.md

@@ -33,7 +33,12 @@ can target it.
 
 ## Pre-Evaluation (MANDATORY for every request, auto, <30s)
 
-No dispatch, edit or shell command may precede this object.
+No dispatch, edit or shell command may precede this object. Opening the run
+(`llm-orchestrator run start --type <TASK_TYPE>`) follows it immediately.
+
+**Trivial tasks.** A one-line, obviously scoped change (a typo, a version bump) may skip the full
+flow, but only by declaring it: `llm-orchestrator run start --trivial "<reason>"`. The
+declaration and its reason are recorded; an undeclared skip is recorded as a skipped flow.
 
 ```json
 {

package/registries/routing-matrix.json

@@ -18,7 +18,8 @@
     "X": {
       "name": "senior",
       "responsibility": "Hard debugging, architecture, concurrency, migrations, security, auth, payments, billing, backwards compatibility, critical code review, many invariants.",
-      "incumbents": {"claude": "claude-opus-5", "codex": "gpt-5.6-sol"}
+      "incumbents": {"claude": "claude-opus-5-5", "codex": "gpt-5.6-sol"},
+      "fallbacks": {"claude": ["claude-opus-5"]}
     },
     "F": {
       "name": "frontier",
@@ -97,15 +98,15 @@
       "codex": {"model": "gpt-5.6-terra", "effort": "high"}
     },
     "X T2": {
-      "claude": {"model": "claude-opus-5", "effort": "medium"},
+      "claude": {"model": "claude-opus-5-5", "effort": "medium"},
       "codex": {"model": "gpt-5.6-sol", "effort": "medium"}
     },
     "X T3": {
-      "claude": {"model": "claude-opus-5", "effort": "high"},
+      "claude": {"model": "claude-opus-5-5", "effort": "high"},
       "codex": {"model": "gpt-5.6-sol", "effort": "high"}
     },
     "X T4": {
-      "claude": {"model": "claude-opus-5", "effort": "xhigh"},
+      "claude": {"model": "claude-opus-5-5", "effort": "xhigh"},
       "codex": {"model": "gpt-5.6-sol", "effort": "high", "independent_second_reviewer": {"model": "gpt-5.6-sol", "effort": "high"}}
     },
     "F T3": {
@@ -279,8 +280,8 @@
       {"model": "claude-sonnet-5", "effort": "low"},
       {"model": "claude-sonnet-5", "effort": "medium"},
       {"model": "claude-sonnet-5", "effort": "high"},
-      {"model": "claude-opus-5", "effort": "high"},
-      {"model": "claude-opus-5", "effort": "xhigh"},
+      {"model": "claude-opus-5-5", "effort": "high"},
+      {"model": "claude-opus-5-5", "effort": "xhigh"},
       {"model": "claude-fable-5", "effort": "high"},
       {"model": "claude-fable-5", "effort": "xhigh"},
       {"model": "claude-fable-5", "effort": "max"},

package/schemas/flow-ledger.schema.json

@@ -0,0 +1,60 @@
+{
+  "_attribution": "llm-orchestrator · created by Bogdan-Gabriel Torcescu · https://www.linkedin.com/in/bogdantorcescu/ · keep this credit when copying or deriving",
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://raw.githubusercontent.com/tbogdan/llm-orchestrator/main/schemas/flow-ledger.schema.json",
+  "title": "Flow adherence ledger (.orchestrator-run/)",
+  "description": "Per-session state and closed-run history written by `llm-orchestrator gate`. Ids, task types, counts and timestamps only — never prompts, tool inputs or file contents.",
+  "$defs": {
+    "run": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["task_id", "task_type", "trivial", "reason", "opened_at", "planned_shards", "subagents_started", "started_outside_flow"],
+      "properties": {
+        "task_id": { "type": "string", "minLength": 1 },
+        "task_type": { "type": ["string", "null"], "enum": ["INCIDENT", "FEATURE", "BUG_FIX", "REFACTOR", "INVESTIGATION", "DEPLOY", "CONFIG", "REVIEW", "RESEARCH", null] },
+        "trivial": { "type": "boolean" },
+        "reason": { "type": ["string", "null"] },
+        "opened_at": { "type": "number" },
+        "planned_shards": { "type": ["integer", "null"], "minimum": 1 },
+        "subagents_started": { "type": "integer", "minimum": 0 },
+        "started_outside_flow": { "type": "boolean" }
+      }
+    },
+    "session": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["version", "session", "run", "prompt_started_at", "nudged", "worked_without_run", "subagents_without_run", "entrypoint_loaded", "seen"],
+      "properties": {
+        "version": { "const": 1 },
+        "session": { "type": "string", "minLength": 1 },
+        "run": { "anyOf": [{ "type": "null" }, { "$ref": "#/$defs/run" }] },
+        "prompt_started_at": { "type": ["number", "null"] },
+        "nudged": { "type": "boolean" },
+        "worked_without_run": { "type": "boolean" },
+        "subagents_without_run": { "type": "integer", "minimum": 0 },
+        "entrypoint_loaded": { "type": "boolean" },
+        "seen": { "type": "array", "items": { "type": "string" } }
+      }
+    },
+    "historyLine": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["session", "task_id", "task_type", "trivial", "reason", "opened_at", "closed_at", "duration_s", "planned_shards", "subagents_started", "started_outside_flow", "skipped_flow", "closed_by"],
+      "properties": {
+        "session": { "type": "string", "minLength": 1 },
+        "task_id": { "type": ["string", "null"] },
+        "task_type": { "type": ["string", "null"], "enum": ["INCIDENT", "FEATURE", "BUG_FIX", "REFACTOR", "INVESTIGATION", "DEPLOY", "CONFIG", "REVIEW", "RESEARCH", null] },
+        "trivial": { "type": "boolean" },
+        "reason": { "type": ["string", "null"] },
+        "opened_at": { "type": ["string", "null"] },
+        "closed_at": { "type": "string" },
+        "duration_s": { "type": ["integer", "null"], "minimum": 0 },
+        "planned_shards": { "type": ["integer", "null"], "minimum": 1 },
+        "subagents_started": { "type": "integer", "minimum": 0 },
+        "started_outside_flow": { "type": "boolean" },
+        "skipped_flow": { "type": "boolean" },
+        "closed_by": { "enum": ["run_close", "next_prompt", "succession"] }
+      }
+    }
+  }
+}

package/schemas/installation-manifest.schema.json

@@ -89,7 +89,8 @@
               "bridge",
               "native-command",
               "codex-prompt",
-              "agent-file"
+              "agent-file",
+              "flow-plugin"
             ]
           }
         }
@@ -146,6 +147,25 @@
           }
         }
       }
+    },
+    "flow_hooks": {
+      "type": "boolean",
+      "description": "Whether flow-adherence hooks are installed; a later install without --flow-hooks / --no-flow-hooks keeps this choice."
+    },
+    "json_entries": {
+      "type": "array",
+      "description": "Flow-hook groups merged into a user-owned hooks JSON file (.claude/settings.json, .codex/hooks.json). Only the marked groups are owned; the rest of the file is the user's.",
+      "items": {
+        "type": "object",
+        "additionalProperties": false,
+        "required": ["scope", "path", "hash", "created_file"],
+        "properties": {
+          "scope": { "const": "project" },
+          "path": { "type": "string", "minLength": 1 },
+          "hash": { "type": "string", "pattern": "^[0-9a-f]{64}$" },
+          "created_file": { "type": "boolean" }
+        }
+      }
     }
   }
 }

package/schemas/routing-matrix.schema.json

@@ -51,7 +51,16 @@
                 "codex": { "type": "string" }
               }
             },
-            "capped_exception": { "type": "object" }
+            "capped_exception": { "type": "object" },
+            "fallbacks": {
+              "type": "object",
+              "description": "Per-ladder models dispatched when the incumbent is not exposed, refuses, or is rolled back.",
+              "properties": {
+                "claude": { "type": "array", "items": { "type": "string" } },
+                "codex": { "type": "array", "items": { "type": "string" } }
+              },
+              "additionalProperties": false
+            }
           },
           "additionalProperties": false
         }

package/schemas/top-models.schema.json

@@ -105,7 +105,12 @@
               "levels": { "type": "array", "minItems": 1, "items": { "type": "string" } },
               "always_on": { "type": ["boolean", "null"] },
               "default_for_tier": { "type": ["string", "null"] },
-              "note": { "type": "string" }
+              "note": { "type": "string" },
+              "unmeasured_levels": {
+                "type": "array",
+                "items": { "type": "string" },
+                "description": "Efforts the API supports that the benchmark has not measured. Declared explicitly so a dispatch at one of them carries est_usd_per_task null instead of borrowing a measured point."
+              }
             },
             "additionalProperties": false
           },
@@ -149,6 +154,10 @@
             }
           },
           "measurement_note": { "type": "string" },
+          "supersedes": {
+            "type": ["string", "null"],
+            "description": "Key of the same-ladder, same-tier model this one succeeds. The router ranks the successor directly ahead of it only while every per-token price (input, output, cache read) is at or below the predecessor's; the predecessor stays routable as the fallback."
+          },
           "indices": {
             "type": "object",
             "required": ["thinking_cost_index", "marginal_thinking"],