Greptile Summary

This PR honors per-model context-token overrides (contextTokens and context1m params) during model selection in the auto-reply pipeline. It extracts the logic into a shared resolveContextTokens helper and ensures consistent usage across both get-reply-directives.ts and directive-handling.persist.ts.

Key changes:

• model-selection.ts: resolveContextTokens now accepts provider and resolves a per-model entry from agentCfg.models (direct key lookup first, then normalized fallback). Priority order is: global agentCfg.contextTokens → per-model params.contextTokens → per-model context1m → catalog lookup → DEFAULT_CONTEXT_TOKENS.
• get-reply-directives.ts: Passes the previously-missing provider to resolveContextTokens; the resolved provider from modelState is correctly used before the call.
• directive-handling.persist.ts: Replaces the inline ad-hoc expression with the shared helper, removing the now-redundant lookupContextTokens / DEFAULT_CONTEXT_TOKENS imports.
• model-selection.test.ts: Four new unit tests cover the main priority scenarios.

Observation — untested edge case: There is no test for when per-model params has both contextTokens set to an invalid value (e.g., 0 or negative) and context1m: true. Under the current logic this correctly falls through to use CONTEXT_1M_TOKENS, but an explicit test would help prevent regressions if the ternary is ever refactored.

Confidence Score: 5/5

• This PR is safe to merge — the changes are additive, well-tested, and the priority fallback chain is consistent with existing behaviour.
• The logic is straightforward: a new provider parameter is threaded through a single helper and the two call-sites are updated consistently. Input validation (finite, positive, truncated) is sound. Four targeted unit tests cover all priority tiers. No existing tests are broken and the behavioural change (per-model override) is additive, not breaking.
• No files require special attention.

_{Last reviewed commit: e748da9}