fix(custom-model): core review fixes for context-length discovery and injection

- share one context-length resolution (discovered figure, endpoint default fallback) across the routes and boot recovery, with a recovery test
- record listed per-slot context figures with a slot count of 1 and force the slot-count fallback once per rediscovery
- open up the codex model launch token for qwen3-style ids (model-plain-colon), pinned by a spawn-golden case
- pass the full discovered window to opencode as limit.context
- translate the picker search strings and document the discovery precedence
This commit is contained in:
Devvyn committed 2026-11-01 21:26:54 +08:00
1 parent dbd1b6e587
commit 1b8ac40baf
16 files changed
+256 -47

No files matched your search

+19 -9
View File
@@ -67,15 +67,23 @@ configured, `PUT`/`DELETE /api/model-endpoints/:id` update or remove one.
Endpoint management is admin-only in multi-user mode, same as remote/docker
hosts — these are machine-level infra, not per-user settings.
**Context length is discovered too, opportunistically and safely.** The plain
`GET /v1/models` response has no context-window field. Discovery only ever
looks for one for a model llama-swap's own response already reports
**Context length is discovered too, opportunistically and safely.** Older
servers state no context-window field in the plain `GET /v1/models`
response at all; current llama-swap states one per model (both
`context_length` and `context_window`) and gateways like Bifrost pass
`context_window` through — that listed figure is the *lowest*-precedence
source, used only when neither probe below answers, and because llama-swap
documents it as the PER-SLOT window, a model whose figure comes from the
listing is recorded with a slot count of 1 so the slot check never divides a
per-slot figure a second time. The probes are only ever looked for on a model
llama-swap's own response already reports
`status.value === "loaded"` for — never for an unloaded one, because
llama-swap treats `?model=` as a routing hint and asking about a model that
isn't loaded risks triggering an actual (slow, GPU-swapping) load as a side
effect of what should be read-only discovery. A server with no `status` field
on any entry at all (not llama-swap) gets no context-length enrichment,
rather than guessing. A model's previously-learned context length survives a
on any entry at all (not llama-swap) gets no launch-command/`/props`
enrichment, only whatever its listings state — rather than guessing. A
model's previously-learned context length survives a
later cycle where it wasn't the loaded one; it's dropped only once the model
disappears from the endpoint's list entirely. Stored per model in
`modelContextLengths` and applied automatically (see "Applying a model to a
@@ -293,14 +301,16 @@ agent, so the selection would report success and change nothing.
**Claude gets two more env vars when known/applicable, both declared on its
registry entry (`contextLengthVar`/`configDirVar`), not hardcoded here:**
- `CLAUDE_CODE_MAX_CONTEXT_TOKENS` is set to `modelId`'s discovered context
length (see the discovery section above) whenever one is known. Without
- `CLAUDE_CODE_MAX_CONTEXT_TOKENS` is set to `modelId`'s context length —
the discovered figure (see the discovery section above), falling back to
the endpoint's user-set `defaultContextLength` — whenever one is known.
Without
it, Claude Code assumes a large (200k) window for any unrecognized custom
model id and never compacts, which reliably overflows a much smaller real
local context — confirmed live: a stock ~33.7K-token system prompt against
a 16384-token llama-swap model failed with `exceeds the available context
size`. No entry for the model in `modelContextLengths` means the var is
simply omitted, never a guess. ⚠️ **This var only affects when Claude
size`. Neither a `modelContextLengths` entry nor a `defaultContextLength`
for the model means the var is simply omitted, never a guess. ⚠️ **This var only affects when Claude
Code compacts conversation _history_ — it cannot fix a model whose real
context is smaller than Claude Code's own fixed per-turn overhead**
(system prompt + tool schemas, empirically ~36.4K tokens, confirmed live