[quickbeard]

What

Ports the hub's per-model limit scheme into packages/codev-gateway, replacing the flat GATEWAY_CONTEXT_WINDOW = 196608 that stamped every gateway model with the same window.

The flat constant claimed to be "identical to the hub's constants so a hub-written and an agent-written codev.json agree byte-for-byte" — no longer true since the hub moved to per-model limits (quickbeard/codev-hub#259 keeps the two in lockstep). A config written by the TUI /configure dialog, codev configure, or the desktop gateway sign-in gave every model the same 196,608-token window and the same compaction point.

How

• model-limits.ts (new): ModelLimits { context, trigger, output? } resolved remote → table → default like the hub — gateway-reported windows first, then the static window table (MiniMax/MiniMax-M3 at 262,144 — the deployment's resource cap, not the model's native window — and zai-org/GLM-4.7-cc at 200K), then the safe 200K default for unknown ids. The trigger is uniformly 90% of the window, whatever its source. declaredInput() implements the hub's limit.input = trigger + reserved trick (clamped to context), which lands each model's own compaction trigger against the single global compaction.reserved: 40000 — opencode's schema has no per-model compaction knob (session/overflow.ts: usable = limit.input − reserved).
• config-block.ts: buildGatewayConfig / buildCustomProviderConfig now write per-model limit.context/input/output via limitsFor (M3 gets context: 262144; its 90% target + reserve exceeds the window, so the clamp pins input: 262144 — effective trigger 222,144); compaction.reserved becomes 40000, matching the hub's writer.
• backend.ts: fetchModelWindows() reads LiteLLM's /model_group/info at the gateway root — never throws, returns {} when the gateway reports nothing (today's deployment: every max_input_tokens is null, so the static table carries all models; the moment an admin fills the field in, that model becomes gateway-driven with no release).
• auth-file.ts: model_limits cache in the shared auth file — the same key the hub's saveModelLimits writes, so hub-cached windows are honored here and vice versa.
• Wiring: the TUI configure dialog and the desktop gateway-sso controller cache windows right after their existing fetchModels step, non-fatally.

The custom-provider block deliberately still writes no compaction key (connecting a provider must not change global compaction for unrelated ones); declaredInput's clamp keeps its declared budget inside the true window.

All touched files are fork-only, so no DIVERGENCES.md rows.

Tests

• packages/codev-gateway: 52 pass (new model-limits.test.ts; fetchModelWindows endpoint/skip/never-throw cases; config-block expectations updated + hermetic CODEV_TEST_HOME so a hub-cached model_limits on a dev machine can't leak in).
• packages/desktop gateway-sso controller: 14 pass (asserts windows are cached during the models step).
• packages/tui custom-provider view test: 6 pass.
• bun typecheck clean in codev-gateway, tui, desktop, app, opencode.

🤖 Generated with Claude Code