feat(models): move manual reasoning efforts into a catalog override

PR #69 added efforts for meta/muse-spark-* directly to the generated
catalog, which the drift check flags and the daily sync job reverts.

Keep src/commandcode-catalog.ts byte-identical to upstream and merge a
separate src/commandcode-catalog-overrides.ts over it at load time. A
test fails as soon as upstream publishes efforts for an overridden model
so the override gets removed instead of shadowing the CLI catalog.

Verified against the live endpoint: pi --thinking xhigh sends
reasoning_effort="xhigh" for meta/muse-spark-1.2-contributor and the
request succeeds; --thinking off sends none.

Closes #69

.agents/skills/refresh-model-catalog/SKILL.md

@@ -32,6 +32,8 @@ npm run sync:commandcode-catalog
 
 Regenerates `src/commandcode-catalog.ts` and bumps the documented CLI version in `README.md`. Review the diff; the catalog also lists reasoning models without selectable efforts.
 
+Never add efforts to the generated file by hand. Manual effort policy for reasoning models that upstream ships without levels lives in `src/commandcode-catalog-overrides.ts` and is merged at load time. When the sync report lists a model from that file under "New effort metadata", remove its override; `tests/test-models.ts` fails until you do.
+
 ### 3. Update display pricing (manual review)
 
 Fetch <https://commandcode.ai/docs/resources/pricing-limits> and compare against `src/pricing.ts`:

CHANGELOG.md

@@ -2,6 +2,7 @@
 
 ## Unreleased
 
+- Expose selectable thinking levels (`minimal`, `low`, `medium`, `high`, `xhigh`) for `meta/muse-spark-1.1`, `meta/muse-spark-1.2`, and `meta/muse-spark-1.2-contributor` through a manual catalog override, so `/thinking` and `Shift+Tab` no longer stay locked on `off` for these reasoning models.
 - Start from the cached model catalog and refresh it in the background instead of blocking host startup on the catalog request; a first start without a cache still waits for the live catalog.
 - Register the `commandcode-custom` API in the `@earendil-works/pi-ai/compat` registry so sibling extensions that stream with the active Command Code model no longer fail with `No API provider registered for api: commandcode-custom` on plain pi.
 - Assert structural catalog invariants in the model tests so the daily catalog sync no longer fails on every upstream change.

README.md

@@ -88,7 +88,7 @@ Open `/model` and select one of the models provided by Command Code. Model avail
 
 ### Reasoning support
 
-Reasoning capability and selectable effort levels follow the official CLI catalog independently. Models can therefore be marked as reasoning-capable even when Command Code chooses their depth automatically; for those models (for example `meta/muse-spark-1.1`, `meta/muse-spark-1.2`, and `meta/muse-spark-1.2-contributor`) the thinking level stays `off` because the CLI itself sends no effort parameter and Command Code has not published selectable levels. Models with explicit effort support register a model-specific `thinkingLevelMap`, so pi and OMP expose only valid levels. Pi's native OpenAI- and Anthropic-compatible providers translate the selected level for Provider API accounts; the existing Command Code generate transport sends the matching `reasoning_effort` for Go accounts.
+Reasoning capability and selectable effort levels follow the official CLI catalog independently. Models can therefore be marked as reasoning-capable even when Command Code chooses their depth automatically. Models with explicit effort support register a model-specific `thinkingLevelMap`, so pi and OMP expose only valid levels. For a few reasoning models the CLI catalog ships no effort levels although the endpoint accepts `reasoning_effort`; `src/commandcode-catalog-overrides.ts` adds a manual level set for those (currently `meta/muse-spark-1.1`, `meta/muse-spark-1.2`, and `meta/muse-spark-1.2-contributor`) on top of the generated catalog, and the tests fail once upstream publishes its own levels so the override gets removed. Pi's native OpenAI- and Anthropic-compatible providers translate the selected level for Provider API accounts; the existing Command Code generate transport sends the matching `reasoning_effort` for Go accounts.
 
 List Command Code models from the terminal:
 

src/commandcode-catalog-overrides.ts

@@ -0,0 +1,23 @@
+import type { CommandCodeReasoningEffort } from "./commandcode-catalog.ts"
+
+/**
+ * Manual reasoning-effort policy for models the official CLI marks as
+ * reasoning-capable without publishing selectable efforts.
+ *
+ * `src/commandcode-catalog.ts` is generated from the CLI package and must stay
+ * byte-identical to upstream so the daily drift check works. Entries here are
+ * merged over the generated catalog at load time and are not touched by
+ * `npm run sync:commandcode-catalog`.
+ *
+ * Add a model only when the effort parameter is known to be accepted by the
+ * Command Code endpoint; remove it once the CLI catalog ships its own efforts.
+ */
+export const MODEL_EFFORT_OVERRIDES: Readonly<
+  Record<string, readonly CommandCodeReasoningEffort[]>
+> = {
+  // Meta Muse Spark: the CLI ships no effort levels, but the endpoint accepts
+  // `reasoning_effort` for these models and other hosts expose the same set.
+  "meta/muse-spark-1.1": ["minimal", "low", "medium", "high", "xhigh"],
+  "meta/muse-spark-1.2": ["minimal", "low", "medium", "high", "xhigh"],
+  "meta/muse-spark-1.2-contributor": ["minimal", "low", "medium", "high", "xhigh"],
+}

src/commandcode-catalog.ts

@@ -133,9 +133,6 @@ export const MODEL_EFFORTS: Readonly<Record<string, readonly CommandCodeReasonin
   "gpt-5.6-sol": ["low", "medium", "high", "xhigh", "max"],
   "gpt-5.6-terra": ["low", "medium", "high", "xhigh", "max"],
   "moonshotai/Kimi-K3": ["low", "high", "max"],
-  "meta/muse-spark-1.1": ["minimal", "low", "medium", "high", "xhigh"],
-  "meta/muse-spark-1.2": ["minimal", "low", "medium", "high", "xhigh"],
-  "meta/muse-spark-1.2-contributor": ["minimal", "low", "medium", "high", "xhigh"],
   "Qwen/Qwen3.8-27B": ["low", "medium", "xhigh"],
   "Qwen/Qwen3.8-Flash": ["low", "medium", "xhigh"],
   "Qwen/Qwen3.8-Max": ["low", "medium", "xhigh"],

src/models.ts

@@ -1,8 +1,9 @@
 import { mkdir, readFile, rename, rm, writeFile } from "node:fs/promises"
 import { dirname } from "node:path"
 
+import { MODEL_EFFORT_OVERRIDES } from "./commandcode-catalog-overrides.ts"
 import {
-  MODEL_EFFORTS,
+  MODEL_EFFORTS as CATALOG_MODEL_EFFORTS,
   MODEL_INPUT_MODALITIES,
   MODEL_MAX_OUTPUT_TOKENS,
   MODEL_REASONING,
@@ -10,7 +11,13 @@ import {
   type CommandCodeReasoningEffort,
 } from "./commandcode-catalog.ts"
 
-export { MODEL_EFFORTS, MODEL_INPUT_MODALITIES, MODEL_MAX_OUTPUT_TOKENS, MODEL_REASONING }
+/** Upstream CLI efforts with the manual overrides merged over them. */
+export const MODEL_EFFORTS: Readonly<Record<string, readonly CommandCodeReasoningEffort[]>> = {
+  ...CATALOG_MODEL_EFFORTS,
+  ...MODEL_EFFORT_OVERRIDES,
+}
+
+export { MODEL_INPUT_MODALITIES, MODEL_MAX_OUTPUT_TOKENS, MODEL_REASONING }
 export type { CommandCodeInputType }
 
 export const DEFAULT_PROVIDER_API_BASE = "https://api.commandcode.ai/provider/v1"

tests/test-models.ts

@@ -4,7 +4,11 @@ import { tmpdir } from "node:os"
 import { join } from "node:path"
 import { describe, it } from "node:test"
 
-import { COMMAND_CODE_CLI_VERSION } from "../src/commandcode-catalog.ts"
+import { MODEL_EFFORT_OVERRIDES } from "../src/commandcode-catalog-overrides.ts"
+import {
+  COMMAND_CODE_CLI_VERSION,
+  MODEL_EFFORTS as CATALOG_MODEL_EFFORTS,
+} from "../src/commandcode-catalog.ts"
 import {
   apiForModelId,
   baseUrlForModel,
@@ -192,6 +196,26 @@ describe("commandCodeModelsFromApiResponse()", () => {
     }
   })
 
+  it("merges manual effort overrides over the generated catalog", () => {
+    const validEfforts = new Set(["minimal", "low", "medium", "high", "xhigh", "max"])
+    assert.ok(Object.keys(MODEL_EFFORT_OVERRIDES).length > 0)
+    for (const [modelId, efforts] of Object.entries(MODEL_EFFORT_OVERRIDES)) {
+      assert.equal(MODEL_REASONING[modelId], true, `${modelId} override needs a reasoning flag`)
+      assert.equal(
+        CATALOG_MODEL_EFFORTS[modelId],
+        undefined,
+        `${modelId} now has upstream efforts; drop the manual override`,
+      )
+      assert.ok(efforts.length > 0)
+      assert.ok(efforts.every((effort) => validEfforts.has(effort)))
+      assert.deepEqual(MODEL_EFFORTS[modelId], efforts)
+      assert.deepEqual(thinkingMetadataForModel(modelId)?.thinking?.efforts, efforts)
+    }
+    for (const [modelId, efforts] of Object.entries(CATALOG_MODEL_EFFORTS)) {
+      assert.deepEqual(MODEL_EFFORTS[modelId], efforts, `${modelId} upstream efforts changed`)
+    }
+  })
+
   it("builds separate canonical pi and OMP metadata", () => {
     for (const [modelId, efforts] of Object.entries(MODEL_EFFORTS)) {
       const metadata = thinkingMetadataForModel(modelId)