Greptile Summary

This PR centralizes model metadata handling into a new src/agents/model-metadata.ts module and routes model catalog, context-window resolution, and status/session display through shared provider-aware helpers. It also ships two real bug fixes: (1) resolveContextTokensForModel now passes params.provider instead of the previously inferred ref.provider to config lookups, preventing wrong-provider context window resolution when a model ID contains a slash; (2) stale persisted session contextTokens no longer serve as the contextTokensOverride in status/session displays.

Key changes:

• New model-metadata.ts module with ModelMetadata type and helpers — eliminates four separately maintained inline model-config parsing loops across model-catalog.ts, model-selection.ts, context-window-guard.ts, and context.ts.
• applyConfiguredContextWindows now writes provider-scoped cache keys in addition to the legacy bare key; configured values win over discovery for scoped lookups.
• session-reset-service.ts resets contextTokens: undefined so context window is recalculated from current model metadata after a reset.
• followup-runner.ts correctly resolves providerUsed from agentMeta.provider rather than relying only on fallbackProvider.
• normalizeModelCost silently drops a cost object if any of the four required fields is missing — a subsystem log warning would help diagnose partial cost configurations.
• mergeModelMetadata calls normalizePositiveInteger twice per field — minor inefficiency.

Confidence Score: 5/5

Safe to merge; the refactoring is well-tested and all remaining findings are P2 style suggestions.

Both bug fixes are correct and well-motivated. The new model-metadata module is clearly typed and thoroughly tested. No P0 or P1 issues found; the two P2 comments address a silent drop of partial cost configs and a minor double-call inefficiency in mergeModelMetadata, neither of which affects correctness.

src/agents/model-metadata.ts — normalizeModelCost silent drop and mergeModelMetadata double-call; otherwise no files require special attention.

This is a comment left during a code review.
Path: src/agents/model-metadata.ts
Line: 50-68

Comment:
**Silent drop for partial cost configs**

`normalizeModelCost` silently returns `undefined` if any of the four required fields (`input`, `output`, `cacheRead`, `cacheWrite`) is absent, with no log warning. Since this is new functionality that users can configure in `models.providers.*.models[].cost`, a user who defines only `{ input: 1, output: 2 }` (without cache pricing) will have their cost entry silently ignored.

Consider emitting a subsystem log warning when a cost object is provided but rejected due to missing fields, so misconfiguration is easier to diagnose.

How can I resolve this? If you propose a fix, please make it concise.

---

This is a comment left during a code review.
Path: src/agents/model-metadata.ts
Line: 1069-1075

Comment:
**`normalizePositiveInteger` called twice per field in `mergeModelMetadata`**

`normalizePositiveInteger(overlay.contextWindow)` and `normalizePositiveInteger(overlay.maxTokens)` are each invoked twice — once in the condition and once in the spread value — making two redundant calls per field on every merge. Consider caching the result:

```
const contextWindow = normalizePositiveInteger(overlay.contextWindow);
const maxTokens = normalizePositiveInteger(overlay.maxTokens);
return {
  ...base,
  ...(contextWindow === undefined ? {} : { contextWindow }),
  ...(maxTokens === undefined ? {} : { maxTokens }),
  ...
};
```

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "refactor(models): unify configured model..." | Re-trigger Greptile}

Thanks for the model metadata pass. Closing this because it is a stale dirty XL refactor across docs, gateway, commands, agents, and OpenAI extension code; please resubmit a narrow owner-backed slice.