feat(models): correct under-reported Copilot/Codex per-model context and output limits

[arminanton]

Problem

Hermes reads per-model context_window and max_output from the community models.dev catalog. For several providers that catalog under-reports the real paid-tier limits — most visibly github-copilot, where claude-opus-4.x is listed as 200k / 64k while the actual /v1/messages wire path serves 1,000,000 / 128,000.

Because hermes /models and the context manager rely on these numbers, the under-report makes the UI display wrong limits and can trigger premature context compression even when the endpoint would happily accept far more.

Fix

A probe-verified override layer in agent/models_dev.py:

• _PROBE_VERIFIED_OVERRIDES — a (provider, model) → {context_window, max_output} table carrying the real wire-path limits for the provider+model pairs models.dev gets wrong (github-copilot, openai-codex, google, anthropic).
• _resolve_probe_override / _canonicalize_model_id — normalization so a Hermes provider id (copilot) and either version shape resolve to the right entry:
  • dot vs dash version equivalence (claude-opus-4.8 ⇄ claude-opus-4-8),
  • vendor-prefix stripping (anthropic/claude-opus-4.7 → claude-opus-4.7),
  • date-stamp stripping, and a graceful family-prefix fallback.
• get_model_info now applies a matching override on top of the models.dev base — replacing only the numeric limits while preserving cost / modalities / capability fields — and synthesizes a minimal honest entry when models.dev has no row but an override exists (better than returning None).

Scope

Numbers/limits only. This does not change routing or endpoint selection. It's the data-correctness layer: when the catalog and the real wire path disagree, prefer the verified number.

Tests

tests/agent/test_models_dev_probe_overrides.py — 15 tests covering exact match, provider-alias normalization, dot/dash equivalence, vendor-prefix stripping, family-shrink fallback, the override-merge (limits replaced, other fields preserved), and the synthesis-when-missing path. ruff clean.

Prior art

Superset of #29146 / #29147, which cover only the narrow max_context_window_tokens preference. This corrects the full per-model context_window + max_output catalog across providers.

Relationship to #49184

Complements #49184 (route Claude-on-Copilot to /v1/messages). This PR corrects the catalog numbers so hermes /models and the context manager report the true limits; #49184 establishes the /v1/messages path on which those larger windows are actually usable. The two are independent (this one is the data layer, that one is routing) but land best together: the corrected numbers are most meaningful once the routing in #49184 is in place.

[alt-glitch]

Related: #42632 / #29146 / #29147 — the existing Copilot-only context-underreport cluster (which corrects max_prompt_tokens for Copilot Claude 1M models in the hermes_cli path).

This PR is a broader, multi-provider mechanism: a probe-verified override table in agent/models_dev.py covering github-copilot, openai-codex, google, and anthropic. Same goal (correct under-reported context/output limits → avoid premature compression) but a different code path and wider scope, so related rather than duplicate.

[arminanton]

Thanks @alt-glitch — agreed, this is the broader multi-provider probe-verified override table in agent/models_dev.py (github-copilot, openai-codex, google, anthropic), related to but wider than the Copilot-only #42632/#29146/#29147 cluster. Same goal, different code path + scope.

[teknium1]

Thanks for tackling the models.dev under-reporting problem. The current patch needs rework to reach the operational context path it describes.

Problems

• agent/model_metadata.py:2369-2384 drives compression context via lookup_models_dev_context(), but the PR changes only get_model_info(). The raw lookup remains at agent/models_dev.py:321-377, so the proposed table does not prevent premature compression.
• get_model_capabilities() separately reads raw limit.context/limit.output at agent/models_dev.py:450-508, leaving another metadata path uncorrected.
• The proposed openai-codex/gpt-5.5 = 1.05M conflicts with the existing Codex OAuth resolver: agent/model_metadata.py:1826-1942 treats Codex /models as authoritative and records 272K; that resolver is selected at agent/model_metadata.py:2309-2317.

Suggested changes

• Centralize verified overrides in the provider-aware context resolver (and share them with capability/display metadata), then add integration coverage through get_model_context_length().
• Reconcile Codex values against the endpoint-specific source already used on main before adding a static higher cap.

Automated hermes-sweeper review.