Proposal: Show the enforced context limit, and overflow-patch to a safe size #22

Why

opencode's context display and its 413-overflow auto-patch both use the wrong
"context size", which hid a fleet-wide misconfiguration for weeks (audit
2026-07-11).

1. The status bar shows capacity, not the enforced wall. The sidebar
   denominator (packages/tui/src/feature-plugins/sidebar/context.tsx) reads
   model.limit.context, which discovery sets to /api/fit max_safe_ctx
   (or, when fit is unavailable, /v1/models context_length, or ultimately the
   models.dev catalog's native value). None of those is the enforced
   --ctx-size. A fleet model pinned at --ctx-size 3072 displayed as
   "… / 467k" (the catalog native), so requests 413'd at 3072 while the UI
   implied a 467k window. opencode already fetches the enforced value as a
   separate contextMax (configured_ctx) but never shows it.

2. Setting ctx via opencode reverts. setModelCtxSize
   (server/routes/instance/httpapi/handlers/local.ts) PATCHes the backend
   (correct) and optimistically writes limit.context in memory as the raw
   ctx_size. But provider.list() never re-discovers, so on the next
   discovery limit.context is repopulated from max_safe_ctx — a different
   quantity — and the set value appears to "not stick".

3. The 413-overflow handler patches to the native max.
   adjustLocalContextOnOverflow (packages/opencode/src/provider/provider.ts)
   reacts to a 413 by PATCHing the backend ctx_size to the reported max_ctx
   (the model's native context). On a VRAM-constrained host that value cannot
   load — observed live: it bumped one host's 35B to 262144 and another's 35B to
   110592, both unloadable. It "fixes" overflow by creating an OOM.

What

• Display the enforced limit. The context denominator MUST be the enforced
  per-request context (contextMax / configured_ctx), not max_safe_ctx and
  never the catalog native. Where both are useful, show used / safe-budget with
  the enforced ceiling in parentheses; never present a capacity number the
  backend will reject against.
• Make a set stick. After setModelCtxSize succeeds, re-run discovery for
  that provider (or re-fetch /api/fit) so limit.context reflects the new
  enforced value with consistent semantics, instead of an in-memory value that
  reverts on the next discovery.
• Overflow-patch to a safe size. adjustLocalContextOnOverflow MUST patch
  ctx_size to a VRAM-safe value derived from /api/fit (the achievable
  ceiling for current VRAM), not the raw native max_ctx. If no safe value is
  available it MUST surface the overflow rather than set an unloadable ctx.

Non-goals

• No routing/placement changes (separate change
  ctx-aware-subagent-placement).
• No change to llama-skein's /api/fit math (separate change
  bound-max-safe-ctx in the llama-skein repo).

Impact

• packages/tui/src/feature-plugins/sidebar/context.tsx and the app metrics
  helper: denominator source.
• packages/opencode/src/provider/provider.ts: adjustLocalContextOnOverflow
  target value; discovery-refresh after set.
• packages/opencode/src/server/routes/instance/httpapi/handlers/local.ts:
  setModelCtxSize triggers a provider re-discovery.

Disposition (2026-09-18)

Shipped. Shipped: sidebar shows limit.contextMax ?? limit.context; provider.test.ts covers adjustLocalContextOnOverflow.