remove-trained-context-clamp

[androidand]

Proposal

Problem

A hard clamp on --ctx-size was added in both llama-skein and opencode that
limits context window to the model's trained context (from GGUF n_ctx_train
or max_position_embeddings). The clamp is applied in three places:

• llama-skein/internal/fit/fit.go — MaxFitCtx = min(vramMaxCtx, trainedCtx)
• llama-skein/internal/server/apiconfig.go — PATCH clamps user ctx to MaxFitCtx
• opencode-skein reads max_fit_ctx from llama-skein's fit API and uses it
  for VRAM-based clamping in the TUI and overflow recovery

The justification was that above the trained context "RoPE extrapolation degrades
quality." While extrapolation behavior is worth knowing about as guidance, a
hard clamp is overkill. Models are not "totally unable" to use larger contexts —
they extrapolate with varying quality. The user should decide.

Solution

Remove the trained-context clamp from all three locations. VRAM limits remain
enforced (they are physical constraints). The user's context choice is respected.

Changes

llama-skein (Go)

• internal/fit/fit.go — MaxFitCtx is now VRAM-only, no min(vram, trainedCtx).
• internal/server/apiconfig.go — PATCH writes --ctx-size verbatim, no clamp.
• contracts/llama-skein.openapi.json + pkg/apicontract/llama_skein.gen.go —
  Updated max_fit_ctx description.

opencode-skein (TypeScript)

• packages/tui/src/local/model-fit.ts — Updated comment.
• packages/tui/test/model-fit.test.ts — Updated test comments.
• packages/opencode/src/server/routes/instance/httpapi/handlers/local.ts —
  Updated comment.

opencode-skein had no independent trained-context clamp — it flows through
llama-skein's fit data.

Impact

• Users can set --ctx-size above the model's trained context.
• The inference engine (llama.cpp) still enforces VRAM limits at load time.
• Quality degradation above trained context is the user's call.

Disposition (2026-09-18)

Shipped. This repo's share (comments/model-fit) shipped; remaining items are llama-skein commits.

Tasks

Tasks

• Remove min(vramMaxCtx, trainedCtx) clamp from llama-skein/internal/fit/fit.go MaxFitCtx
• Remove ctx-size clamping from llama-skein/internal/server/apiconfig.go PATCH handler
• Update llama-skein/contracts/llama-skein.openapi.json max_fit_ctx description
• Update llama-skein/pkg/apicontract/llama_skein.gen.go MaxFitCtx comment
• Update opencode-skein/packages/tui/src/local/model-fit.ts comment
• Update opencode-skein/packages/tui/test/model-fit.test.ts test comments
• Update opencode-skein/packages/opencode/src/server/routes/instance/httpapi/handlers/local.ts comment
• Review changes
• Commit and push to master
• Deploy llama-skein to all local providers
• Deploy opencode to path

Unchecked items above: see Disposition in proposal.md (2026-09-18).

Plan changes

7 done

[androidand]

Archived 2026-09-18 — see the change's Disposition section.