Give Muse Spark its own reasoning effort ladder

[TheGreatAxios]

Closes CL-7867.

muse-spark-1.3-contributor fell through to UNKNOWN_MODEL_EFFORTS, so minimal was unreachable from the model picker, --reasoning-effort, and agent profiles.

Why it matters

Reasoning tokens are ~95% of every completion on this model — a 58-token answer ships behind 500–2000 reasoning tokens — so the effort rung is the largest single lever on token spend and latency.

Direct /responses probes, store: false, deterministic regex graders, 2–3 repeats per cell:

effort	bugfix task	concurrency task	avg reasoning (concurrency)
minimal	3/3	3/3	328
low	3/3	3/3	787
medium	3/3	3/3	2036
high	3/3	3/3	1933

6.2x the reasoning spend at medium versus minimal, identical pass rates. high is not monotonic with medium — it reasons less while passing the same, which is what a saturated rung looks like.

The gateway accepts minimal and rejects none with HTTP 400 on reasoning.effort, so the rung set is minimal/low/medium/high. Family default is low, not minimal — both saturate the graded tasks, and low still costs a quarter of medium.

Max-tokens floor

Reasoning consumes max_output_tokens before any answer text:

effort	512	1024	2048+
minimal	2/2	2/2	2/2
low	1/2	2/2	2/2
medium	0/2	2/2	2/2
high	0/2	2/2	2/2

At 512 with medium, the model spent 397 tokens reasoning and returned 3 tokens of answer. SOURCE_MAX_TOKENS is 16384 with no override path, so nothing is broken today — the added test is what stops a future per-source override dropping below the floor.

Testing

bun run typecheck, bun run lint, full suite: 8893 pass. The 3 failures are pre-existing on main (pack-inflation-guard-delta, and an order-dependent default-sink vendor test that passes in isolation).

Confirmation gate (--repeats 5 against baseline-0286.json) is the remaining unchecked Outcome box.

[linear-code]

CL-7867

[TheGreatAxios]

Critic review · Comment

Gives Muse Spark a minimal/low/medium/high ladder with a low family default, matched by /^muse-spark/i prefix so all five catalog ids land on it. Verified against PR head 5ff7f6163.

Evidence checked

• src/provider/reasoning-effort.ts — supportedEfforts, validateEffort, defaultEffortForModel all funnel through the one new isMuseSparkModel branch, placed after the Codex/OpenAI/grok/GLM branches (no overlap with any of them) and before the unknown-model fallback. Picker (cycleReasoningEffort in src/tui/runner/wiring.ts:355), agent profiles (validateEffort in src/subagent/agent-fleet.ts:1011), and request wiring (resolveSessionEffort in src/config/inference-sources.ts:86) all read this single choke point, so minimal is reachable from every surface named in the PR body. none stays rejected, matching the reported gateway 400.
• Five-id claim checks out: packages/opencode-go/src/models.ts ships the two -contributor ids, packages/zen/src/models.ts:69-71 ships muse-spark-1.3, muse-spark-1.2, muse-spark-1.3-contributor-free. Prefix covers all; the test.each pins all five plus a not-muse-spark-1.3 negative.
• Ran on PR head: bun test src/provider/reasoning-effort.test.ts src/config.test.ts — 164 pass, 0 fail. bun run typecheck — clean. (main typechecks clean too, so no drift.)
• SOURCE_MAX_TOKENS is 16384 (src/config/index.ts:106); the new floor test is a real assertion on the exported constant.

Should-fix (comment-only, no behavior change)

• src/provider/reasoning-effort.ts:82 — the comment claims the prefix "matches isMuseSparkLeafProvider in src/subagent/provider-family.ts, which already keyed off the same prefix." That symbol does not exist anywhere in the repo (git grep on PR head hits only this comment; provider-family.ts defines grok/kimi helpers only). Same claim is in the patch-2 commit message. A future reader grepping for it finds nothing. Reword to cite the file's actual prefix convention (e.g. the /^grok/i + trim() style already used there) or drop the sentence.

Nits (non-blocking)

• src/config.test.ts — the truncation-floor test asserts on a config constant but lives inside describe("buildOpenAISource"); it belongs next to the constant's own tests, if any.
• isMuseSparkModel trims and case-folds while sibling matchers (GLM_53_MODELS.includes, startsWith("grok")) do neither. Harmless — no caller passes padded ids, and it matches provider-family.ts style. Genuinely fine as-is.

Gaps (not covered)

• Live gateway behavior (minimal→200, none→400) is author-measured, not repo-testable; taken on trust.
• Full suite and lint not re-run here; author reports 8893 pass with 3 pre-existing main failures. Touched files are fully green above.

Verdict

No functional blockers — merge-ready once the isMuseSparkLeafProvider comment reference is corrected. Suggested keeper for testsmith (optional): a cycleReasoningEffort case for an unset muse-spark model pinning low→medium.