fix: thinking strength now follows the chosen level on Kimi/StepFun/MiniMax coding plans

[Astro-Han]

Problem

Kimi Coding Plan did not expose thinking-strength controls. The root cause was a pre-#2050 orphan: kimi-coding-plan was missing from the PROVIDERS map in scripts/sync-model-metadata.mjs, so its metadata stayed frozen in a hand-written static block from #1164 (k3 pinned to a single max effort, k3-256k absent), and the runtime wire hard-coded max without ever reading the chosen level.

A systematic audit found two more instances of the same class:

• stepfun-step-plan — same orphan shape (upstream segment existed, sync map missing, hand-written static block expanding the stepfun API segment, modelsDevId mislabeled as stepfun).
• minimax-coding-plan — alias debt (declared modelsDevId: minimax and reached metadata through a static alias to the MiniMax API segment; the two models.dev segments could drift apart silently).

Changes (5 commits)

1. kimi-coding-plan (f2fdc7b): sync map gains kimi-coding-plan → kimi-for-coding; the snapshot now declares k3 (toggle + low/high/max) and k3-256k (low/high/max); the hand-written static block is deleted; buildProviderOptions passes the chosen level through both wires (anthropic effort and openai-chat reasoningEffort), defaulting to max when unset; k3-256k joins the K3 adaptive-thinking wire.

2. stepfun-step-plan (d61e178): sync map gains its own models.dev segment; the static block expanding stepfun records is deleted; modelsDevId corrected from the mislabeled stepfun to its own segment, and the import-time model guard now checks the stepfun-step-plan segment.

3. minimax-coding-plan + sync contract (3a98f74): sync map gains its own segment (7 models, MiniMax-M3 toggle preserved); the alias line is deleted; registry modelsDevId and the fallback guard point at its own segment. Adds models-dev-sync-contract.test.ts and declares modelsDevId for every snapshot-backed provider (anthropic, openai, google, deepseek, kimi-coding-plan, zai-coding-plan, MiniMax, MiniMax-cn, gemini-cli, claude-subscription, openai-codex).

4. Three-way sync contract (edef7eb): sync script emits the complete models.dev directory (GENERATED_MODELS_DEV_DIRECTORY); the contract gains a directory direction (registry base-URL host matching a directory provider must declare it); kimi off is explicitly rejected in the wire; fallback guards and intent comments added.

5. External review round (Claude Opus + Codex) (b3f8471): both independent reviews found no P0/P1 but converged on real gaps, all fixed:

   • kimi off rejection was dead code (the entry gate normalized off to undefined before the case guard, silently turning an explicit off into max; a test even locked that in). The case now checks the raw thinkingLevel argument and rejects off with empty options.
   • the sync contract could not see alias-type orphans: a provider declaring a neighbour segment id (the exact pre-fix stepfun/minimax shape) passed all checks, and claude-subscription → openai passed too. Replaced host-based matching with a deterministic completeness check: every provider must have its own segment, a declared alias (xai-oauth/opencode-free/claude-subscription/openai-codex), or a whitelist reason (9 local/user-configured/volcengine entries). Mutation-tested: both bug shapes now fail the suite.
   • main() in the sync script had zero coverage (the place the orphan bug lived): added an end-to-end fixture test over the real main() path (9/9 script tests).
   • the wire-contract sweep only exercised kimi's anthropic branch: added an explicit openai-chat sweep for kimi-coding-plan.
   • GENERATED_MODELS_DEV_DIRECTORY shrank to { api?: string } (name was never consumed); dropped the unused FACTS_BY_PROVIDER alias.
   • stepfun guard now covers all 4 fallback models (the old slice(0, 3) was stale); minimax guard label says MiniMax Coding Plan.

Verification

• core 772/772 pass (incl. the sync contract suite; mutation tests confirm the kimi, alias-mislabel, and neighbour-segment bug shapes all fail)
• runtime 3089 pass, 4 pre-existing macOS /var → /private/var symlink failures (identical on main, unrelated)
• sync script tests 9/9; biome lint / format:check / full-workspace typecheck clean
• Behavior: k3/k3-256k → [low, high, max]; step-3.7-flash → [low, medium, high]; MiniMax-M3 toggle-only, no regression
• StepFun official docs cross-checked: step-router-v1 is a routing model (reasoning: false is correct; the old hand-written true was wrong), and the low/medium/high effort levels match the snapshot

Confirmed external facts

• Kimi Coding Plan API accepts low/high/max for k3 / k3-256k (verified by the user; no live-check script needed). The snapshot's effort levels are safe to pass through; no STATIC narrowing back to ['max'].

Deferred (tracked, not blocking)

• kimi protocol-decision consolidation onto resolveModelRuntime + provider-options camelCase cleanup (runtime refactor, suggested as a separate PR)
• kimi-for-coding sends effort: 'max' despite having no declared reasoning_options (pre-existing behavior, not a regression; models.dev declares none, the UI offers no menu, and default-max is the sensible wire)
• openrouter maxOutputTokens drift for 4 models (snapshot refresh, noted in the generated diff)
• zai-coding-plan glm-4.6 and tencent-token-plan 8/9 models still lack metadata (upstream segments do not cover them; would need local pins + live checks)

[Astro-Han]

Review round (edef7eb) — internal subagent review

An internal multi-angle review found three gaps in the initial fix; all addressed in this commit:

1. The sync contract test was hollow (H1/P2): it only covered the already-declared universe, so the exact kimi bug shape (registry entry exists, no modelsDevId, no sync-map entry) was invisible to both directions, and the forward check was tautological (every modelsDevId was assigned from the same generated facts object it compared against). → The sync script now emits the complete models.dev directory (GENERATED_MODELS_DEV_DIRECTORY, 178 providers); forward became a directory-existence check; a new directory direction requires any registry provider whose base-URL host matches a directory provider to declare modelsDevId. Discrimination verified: removing kimi's modelsDevId turns both reverse and directory checks red.
2. kimi off silently upgraded to max (M3/P4): the case guard was unreachable and the fallback would turn a future none declaration into max. → The case explicitly rejects off (empty options); tests lock the entry-gate semantics.
3. openai-chat high never asserted (H2): → added, covering low/max/high/off on the openai-chat path.

Also: kimi fallback import-time guard (matching the stepfun/minimax convention), claude-subscription/openai-codex declare their source segments, highspeed-branch intent comment, camelCase alias dependency comment, step-router-v1 no-variants intent locked.

Verified: core 772/772 (the new directory check caught a real lm-studio false positive on its first run, fixed by excluding localhost), runtime 3089 pass / 4 pre-existing macOS symlink failures, scripts 8/8, lint/format/typecheck clean.

[Astro-Han]

External review round (Claude Opus + Codex) — b3f8471

Two independent external reviews (Claude Opus, 55 turns; Codex) both found no P0/P1, but converged on three real gaps in the previous round; all fixed:

1. kimi off rejection was dead code (Codex P2): the entry gate normalized off to undefined before the case guard, so an explicit off silently became max — a test even locked that behavior in. → The case now checks the raw thinkingLevel argument; off returns empty options. Tests updated: off → {} on both protocols, undefined → max preserved.
2. The sync contract could not see alias-type orphans (both reviews P2; Codex proved by mutation that claude-subscription → openai passed all three checks). A provider declaring a neighbour segment id (the exact pre-fix stepfun/minimax shape) was indistinguishable from a legitimate alias. → Replaced host-based matching with a deterministic completeness check: every provider must have its own snapshot segment, a declared alias (xai-oauth / opencode-free / claude-subscription / openai-codex), or a whitelist reason (9 local / user-configured / volcengine entries). Mutation-verified: both the alias-mislabel shape and the neighbour-segment shape now fail the suite.
3. main() in the sync script had zero coverage (Claude A3 — the place the orphan bug lived): → added an end-to-end fixture test over the real main() path (mapped provider generates a segment, unmapped provider appears only in the directory); exported PROVIDERS / main(argv) to make it testable.

Additional fixes: wire-contract sweep gained an explicit kimi openai-chat variant (Claude B2); GENERATED_MODELS_DEV_DIRECTORY shrank to { api?: string } (name was never consumed, Codex P3); dropped the unused FACTS_BY_PROVIDER alias; stepfun guard now covers all 4 fallback models (the slice(0, 3) was stale, Claude C1); minimax guard label says MiniMax Coding Plan (C2).

Deferred with reasons: k3 low/high live check → now confirmed closed (the Kimi API accepts low/high/max; see the follow-up comment), protocol-decision consolidation onto resolveModelRuntime + camelCase namespace cleanup (runtime refactor, separate PR), kimi-for-coding forced max (pre-existing, UI-unreachable).

Verified: core 772/772, runtime 3077 pass / 4 pre-existing macOS symlink failures, scripts 9/9, lint/format/typecheck clean. StepFun official docs cross-checked: step-router-v1 is a routing model (reasoning: false correct).

[Astro-Han]

B1 confirmed closed — external fact verified

The Kimi Coding Plan API accepts low / high / max thinking-effort levels for k3 / k3-256k (verified by the user; no live-check script needed). This means:

• k3's ['low', 'high', 'max'] levels from the models.dev snapshot are safe to pass through — no STATIC narrowing back to ['max'];
• all "Live-check pending" notes are resolved; there is no runtime-400 risk on non-max levels;
• the extra live-check script is unnecessary.

Remaining deferred items (recorded in the review rounds, not blocking merge):

• kimi protocol-decision consolidation onto resolveModelRuntime + provider-options camelCase cleanup (runtime refactor, suggested as a separate PR)
• kimi-for-coding sends effort: 'max' with no declared reasoning_options (pre-existing behavior, not a regression; models.dev declares none, UI offers no menu, default-max is the sensible wire)
• openrouter maxOutputTokens drift for 4 models: noted in the generated snapshot diff, updates naturally on the next sync