fix(thinking): generic reasoning_effort for models with no known thinking family (fixes #252)

[ashif4ai]

Problem

For any model that doesn't match a hardcoded thinking family in src/core/registry.ts (space-bunny-free, big-pickle, hy3, …), the Thinking Effort control appeared in the Copilot model picker — but selecting a level did nothing. The request never carried a reasoning field, so the model always answered at its default effort.

Reproduced on space-bunny-free (Go and Zen): set the per-model Thinking Effort to a non-off level, send a message, and the pick is silently discarded.

Root cause — two independent defects

1. FallbackThinking had no payload path. buildPayload() returned {} for every unknown model, so no reasoning field could ever be emitted — while schema() already used metadata.reasoning to decide whether to show a picker. The UI promised a control the request ignored.

2. buildPayload was never given metadata. All four production call sites called thinkingProviderFor(modelId) with no metadata, so this.metadata was always undefined inside FallbackThinking. Any metadata-driven gate would therefore never have fired, even once #1 was fixed.

This is also why patching a single model name can't work — the model was never the problem, the strategy was.

Fix

models.dev is now the source of truth, with no per-model hardcoding:

• metadata.reasoning decides whether a model gets a reasoning control at all.
• Its own reasoning_options decide which levels; the shared off/low/medium/high/xhigh/max ladder only fills the gap — so a model declaring a level this build has never seen works with no code change.
• The pick is sent as a plain OpenAI-compatible reasoning_effort.
• New opencodego.thinking.generic setting for a workspace-wide default, surfaced in the Set Thinking Effort… command.

FallbackThinking also serves family-less models on other wire formats (claude-* → Anthropic messages, gemini-* → Google, grok-* → Responses). reasoning_effort is meaningless there, so those are deliberately excluded via the registry's endpointKind and keep today's exact no-op behaviour rather than gaining a picker that silently does nothing. The big-pickle repetition penalty (#207) is preserved and composes with the effort field.

Follow-up bug found while testing

Fixing this exposed a second defect in the same feature: prepareChatRequest called resolveThinkingConfig() before metadata was resolved and without passing it, so FallbackThinking.applyOverride discarded the per-model pick outright. The workspace setting kept working because it flows through buildPayload, which does receive metadata — producing a confusing split where the Copilot dropdown did nothing but the same value in settings worked.

metadata is now a required field on ResolveThinkingConfigInput, so the compiler rejects that call shape and it can't silently return. This also restores stripSchemaDefaultEcho (#226) for generic models — without metadata the picker schema is unknown, no default is stripped, and VS Code's "off" baseline would clobber the user's workspace pick.

Testing

512 unit tests pass (495 before), npm run lint 7/7, npm run compile clean. 14 new cases in src/test/thinking.test.ts cover unknown reasoning-capable models, non-reasoning models, models with no metadata at all, non-chat endpoints, penalty coexistence, and a full known-family regression suite.

Verified live against https://models.dev/api.json to confirm which models take each path:

Model	reasoning_options	Path
space-bunny-free	effort: low,medium,high,xhigh,max	metadata levels (the reported bug)
hy4	effort: none,high	metadata levels
big-pickle, nemotron-3-ultra-free, ling-3.0-tiny-free	(empty)	new generic ladder
longcat-2.*, ling-3.1-flash-free	toggle	pre-existing off/on path (unchanged)

Manually confirmed in the Extension Development Host on space-bunny-free and big-pickle — the latter sends { repetition_penalty: 1.2, reasoning_effort: "high" } at high.

Notes for review

• Version bump intentionally omitted. Your history does releases as separate chore(release): 0.7.x commits, so I left package.json at 0.7.8 and only added the ## [0.7.9] CHANGELOG heading — happy to fold the bump in if you'd rather ship it here.
• ARCHITECTURE-MAP.md corrections beyond this fix. The Thinking row had pre-existing drift (glm.ts 53→93, mimo.ts 72→78, provider.ts 78→81) and was missing muse.ts entirely. Since I was updating the row anyway I made it accurate; happy to split that out if you'd prefer a focused diff.
• Endpoint exclusion is a judgment call. Emitting reasoning_effort to Anthropic/Google/Responses would be meaningless or rejected, so those models keep today's behaviour. Covering Claude properly would need an Anthropic-shaped thinking block — worth a separate issue if you want it.

Full write-up: docs/issues/109-20261004-issue252-generic-reasoning-effort.md

🤖 Generated with Claude Code

[ltmoerdani]

Nice one @ashif4ai, this is a clean fix. The two-defect diagnosis (payload returning {} plus the missing metadata on every thinkingProviderFor call site) explains why patching a single model name could never work, and making metadata required on ResolveThinkingConfigInput so the compiler catches future call shapes is a smart move.

I ran the verification locally before reviewing:

• compile and the full lint gate (all 7 checks including unit tests) pass on the branch
• merged against current main (which now includes fix(cache): deterministic Responses ids, stable session anchor, structural Anthropic breakpoints #251's cache-control changes, overlapping file: src/request/anthropic.ts, different hunks), compile and tests still pass, no conflicts
• regression guard looks good: the tests asserting all 8 known families keep their exact payload shapes, and that the generic setting never leaks into them, are exactly the protection this area needs

One small thing I want to keep an eye on (not a blocker): for models that declare reasoning but no reasoning_options in models.dev, we now send xhigh/max from the shared ladder. If any gateway model only accepts low/medium/high, that could 400. Worth watching after release, and easy to trim the ladder later if a report comes in.

Planning to merge with a merge commit to preserve both of your commits. Thanks for the thorough issue doc and changelog entry too.