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

docs/api-reference.md

@@ -1205,13 +1205,18 @@ data}` envelope on the wire — unwrap it the same way). Answers `[]` for a
 boolean` reports whether one is stored, so a client can render "unchanged
   if left blank" without ever holding the real value.
 - `POST /api/v1/model-endpoints` with `{ id, label, baseUrl, apiKey?,
-authStyle?, defaultModelId? }` creates one. `id` must match
+authStyle?, defaultModelId?, defaultContextLength? }` creates one. `id`
+  must match
   `^[a-zA-Z0-9_-]+$`; `authStyle` is `bearer` (default) or `api-key`, never
   both (a real server hung indefinitely when sent both headers on one
-  request); `baseUrl` must be `http(s)`, carry no embedded credentials, and
+  request); `defaultContextLength` is a user-set fallback used when
+  assigning a model that discovery has no context length for (discovery
+  itself never writes it); `baseUrl` must be `http(s)`, carry no embedded
+  credentials, and
   is refused if it points at (or resolves to) a link-local or
   cloud-metadata address. `409 ALREADY_EXISTS` on a duplicate id.
-- `PUT /api/v1/model-endpoints/:id` updates one. An **absent** `apiKey`
+- `PUT /api/v1/model-endpoints/:id` updates one, also accepting
+  `defaultContextLength`. An **absent** `apiKey`
   keeps the stored one rather than clearing it — the client never receives
   the real value to resend deliberately unchanged, so omission is the only
   way to say "leave it alone"; there is no way to clear a key back to unset
@@ -1221,8 +1226,20 @@ authStyle?, defaultModelId? }` creates one. `id` must match
 - `POST /api/v1/model-endpoints/:id/discover-models` fetches the endpoint's
   own `GET /v1/models` and stores the result as `models`, updating
   `lastDiscoveredAt`, plus (best-effort, only for a model llama-swap's own
-  response already reports loaded) `modelContextLengths` and `modelSizesGB`.
-  A `defaultModelId` that no longer appears in the fresh list is dropped
+  response already reports loaded) `modelContextLengths`,
+  `modelSlotCounts` and `modelSizesGB`. A loaded model's context figure
+  comes from llama-swap's own launch command (`--fit-ctx` / `-c` /
+  `--ctx-size`), falling back to `/props` (`n_ctx`, `total_slots`) for
+  whatever the command did not answer; the model's own listing
+  (`context_length`, then `context_window`) only ever fills the gap when
+  neither probe does, and is recorded as a single-slot figure —
+  llama-swap lists the PER-SLOT window, so whenever the listed figure is
+  the one taken the stored slot count is 1. A server with no `status`
+  field on any entry (plain llama.cpp, a non-llama-swap gateway) gets no
+  context/size enrichment at all. `modelContextLengths` falls back to the
+  endpoint's `defaultContextLength` when a session is assigned a model
+  discovery never learned a window for. A `defaultModelId` that no longer
+  appears in the fresh list is dropped
   rather than carried forward invalid. Failures answer `422 OPERATION_FAILED`
   with the underlying connection error, or a named egress refusal if the
   resolved address turned out to be blocked. The same refresh also runs

docs/architecture-invariants.md

@@ -628,7 +628,7 @@ Adding the header changed the preamble, so `CODEMAN_PREAMBLE` was bumped (1.22.0
 
 **Custom Model Endpoint Profiles** (opt-in, `customModelEndpointsEnabled`, SYNCED, default OFF; `docs/custom-model-endpoints.md`, design doc `docs/custom-model-endpoints-plan.md`; full stack — settings-panel CRUD + the Run-menu picker, on top of the backend below): points a session at a user-configured custom OpenAI-compatible endpoint — local (llama.cpp, DGX Spark, Strix Halo) or cloud (Azure AI Foundry, OpenRouter) — instead of its harness's native cloud backend. Endpoints are a read/write-array store (`custom-model-hosts.ts`, `~/.codeman/custom-model-hosts.json`) discovered via `GET <baseUrl>/v1/models`; `CustomModelHost.authStyle` is `'bearer'` (default, `Authorization: Bearer`) or `'api-key'` (Azure's convention) — **never both**, live-tested against a real server: sending both headers on one request reliably hangs it indefinitely, reproduced 3×.
 
-⚠️ The actual per-CLI redirect is `capabilities.customModelInjection` on the CLI registry (four kinds: `env` for claude/gemini/deepseek, `configContentEnv` reusing opencode's existing `OPENCODE_CONFIG_CONTENT`, `configDir` for codex/pi/grok/omp — writes an isolated per-session config file, NEVER the user's real `~/.codex`/`~/.pi`/`~/.omp`/grok config — and `unsupported` for antigravity, which has no known mechanism), computed by the pure `custom-model-injection.ts` (mirrors `session-cli-builder.ts`'s no-IO discipline). ⚠️ `PI_CONFIG_DIR` does NOTHING for pi or omp (grepped pi's entire bundled JS source — the string appears nowhere); both hardcode `~/.pi/agent/models.json` / `~/.omp/agent/models.yml` with no dedicated override, so the real redirect for both is the child process's own **`HOME`**, and both need `models` as an ARRAY of `{id}` objects (an object keyed by id silently loads zero models). Grok's real mechanism turned out to be a `config.toml` `[model.<name>]` block redirected via `GROK_HOME` — its original env-var-based recipe was flat-out wrong (produced "Not signed in" against a real binary), not just unverified.
+⚠️ The actual per-CLI redirect is `capabilities.customModelInjection` on the CLI registry (four kinds: `env` for claude/gemini/copilot/deepseek, `configContentEnv` reusing opencode's existing `OPENCODE_CONFIG_CONTENT`, `configDir` for codex/pi/grok/omp — writes an isolated per-session config file, NEVER the user's real `~/.codex`/`~/.pi`/`~/.omp`/grok config — and `unsupported` for antigravity, which has no known mechanism), computed by the pure `custom-model-injection.ts` (mirrors `session-cli-builder.ts`'s no-IO discipline). ⚠️ `PI_CONFIG_DIR` does NOTHING for pi or omp (grepped pi's entire bundled JS source — the string appears nowhere); both hardcode `~/.pi/agent/models.json` / `~/.omp/agent/models.yml` with no dedicated override, so the real redirect for both is the child process's own **`HOME`**, and both need `models` as an ARRAY of `{id}` objects (an object keyed by id silently loads zero models). Grok's real mechanism turned out to be a `config.toml` `[model.<name>]` block redirected via `GROK_HOME` — its original env-var-based recipe was flat-out wrong (produced "Not signed in" against a real binary), not just unverified.
 
 ⚠️ **Two launch paths, chosen by mechanism, not preference**: opencode/codex/gemini/pi/grok/deepseek/omp apply the selection ONE-SHOT, before the session/process ever exists, with no restart at all; claude alone still applies a selection by **restarting the session's CLI process in place** via `Session.restartCli()` — a de-restricted `reattachRemote()` reusing the same `respawn-pane -k` primitive local/remote respawns already share — because every one of these harnesses reads its endpoint config at process start, never per-turn, so there is no live hot-swap; `Session.setCustomModel()` undoes the PREVIOUS selection's env keys (and deletes its old `configDir`) before merging the new ones in, so switching endpoints or clearing back to native cloud never leaves a stale key behind. ⚠️ Deleting a key from `_envOverrides` is NOT enough on its own: `tmux setenv` persists at the tmux-session level and is inherited by `respawn-pane` (measured: `setenv FOO bar` survived two successive `respawn-pane -k`), so the retired keys are queued (`_pendingEnvUnsets`) and ride `RespawnPaneOptions.unsetEnvKeys` into `applyEnvOverrides()`, which `setenv -u`s them BEFORE re-applying the live overrides. ⚠️ `restartCli()` relaunches a CLI in an existing pane, so a CLI whose launch declares a `fallback` chain (claude) gets a conversation pinned as `resumeSessionId` for that one respawn: `--session-id <id>` refuses an id that already has a transcript (`Session ID ... is already in use`), and without the `--resume <id> || --session-id <id>` shape the docker/remote pane commands already use, applying a model killed the pane and lost the session. The dead-pane respawn shares this pin, see [Dead-pane respawn: the resume pin](#dead-pane-respawn-the-resume-pin). ⚠️ pi, omp and grok need the config file AND a `model` launch param (`custom/<id>` for pi/omp, grok's `[model.codeman-custom]` block name): that is the registry's `customModelInjection.launchModel` template, applied onto the respawn options through `legacyConfigField` by `_withCustomModelLaunchModel()`, never by id, and a model id the CLI's `model` token pattern cannot carry is refused with a 400 rather than silently dropped by the argv engine. ⚠️ Remote (SSH) and Docker sessions are REFUSED (400): their `restartCli()` reattaches a durable tmux rather than restarting the agent and the env lands on the local pane, so they used to report `restarted:true` and change nothing. The selection survives a Codeman restart as the disk-only `__customModel` (bookkeeping: env KEYS, config dir, launch model; never the values, which carry the API key and are re-derived from the endpoint store on recovery), the config dir is removed with the session, and every secret-bearing file (`custom-model-hosts.json`, the per-session config dir) is written 0600.
 

docs/custom-model-endpoints.md

@@ -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

src/config/cli-registry/patterns.ts

@@ -21,6 +21,7 @@ export type TokenPattern =
   | 'model'
   | 'model-claude'
   | 'model-plain'
+  | 'model-plain-colon'
   | 'model-pi'
   | 'id'
   | 'id-dotted'
@@ -47,6 +48,10 @@ const PATTERNS: Record<TokenPattern, RegExp> = {
   // copilot --model. Same characters as `model`, but a leading alphanumeric: the value follows a
   // flag, so `--yolo` must not pass as a model id and become its own flag (the allowAll clamp bypass).
   'model-plain': /^[a-zA-Z0-9][a-zA-Z0-9._\-/]*$/,
+  // codex --model, and the value a custom-model `launchModel` is forced onto it. Same
+  // shape as `model-plain` (leading alphanumeric, so a flag cannot ride the flag) plus
+  // `:` — Ollama-style ids like `qwen3:32b` must survive onto the command line.
+  'model-plain-colon': /^[a-zA-Z0-9][a-zA-Z0-9._\-/:]*$/,
   // buildPiCommand — `:` for a thinking suffix (`sonnet:high`), `/` for `provider/id`
   'model-pi': /^[a-zA-Z0-9._\-/:]+$/,
   // opencode --session, codex resume

src/config/cli-registry/stock.ts

@@ -628,7 +628,10 @@ const CODEX: CliEntry = {
     params: {
       bypassApprovals: { type: 'bool' },
       animations: { type: 'bool' },
-      model: { type: 'token', pattern: 'model' },
+      // `model-plain-colon`, not `model`: the custom-model feature forces the selected id
+      // onto this param (`launchModel: '{modelId}'` below), and Ollama-style ids carry a
+      // `:` (`qwen3:32b`). A leading alphanumeric still keeps `--yolo` from riding a flag.
+      model: { type: 'token', pattern: 'model-plain-colon' },
       reasoningEffort: { type: 'enum', values: [...CODEX_REASONING_EFFORTS] },
       resumeId: { type: 'token', pattern: 'id' },
     },
@@ -1922,8 +1925,8 @@ const COPILOT: CliEntry = {
     // custom provider and drops the GitHub sign-in requirement; COPILOT_PROVIDER_TYPE defaults to
     // "openai", which is what a Codeman custom endpoint is (llama.cpp, vLLM, Ollama).
     // COPILOT_MODEL names the model and COPILOT_PROVIDER_MODEL_ID / _WIRE_MODEL default to it.
-    // The OpenAI-style route is `<base>/v1/chat/completions`, so the base gets the `/v1` suffix
-    // the same way deepseek's does. All four already match the COPILOT_ prefix allowlist, so
+    // The OpenAI-style route is `<base>/v1/chat/completions`, so the base gets the `/v1`
+    // suffix via appendV1Suffix below. All four already match the COPILOT_ prefix allowlist, so
     // they are privileged env keys below: a non-granted multi-user owner must not be able to
     // redirect the session's endpoint or credentials through plain envOverrides.
     customModelInjection: {

src/config/cli-registry/types.ts

@@ -856,19 +856,13 @@ export interface CliCapabilities {
    * reuses that file.
    *
    * `appendV1Suffix` (env kind only): the raw `endpoint.baseUrl` gets `withV1Suffix()`
-   * applied before being written to `baseUrlVar`, instead of being used verbatim.
-   * DeepSeek needs this and claude/gemini must NOT get it — a per-CLI asymmetry confirmed
-   * by reading each SDK's own request-building source, not assumed: DeepSeek Harness's
-   * bundled `@deepseek-ai/dsh-llm-deepseek` concatenates `${connection.baseURL}/chat/
-   * completions` with no `/v1` insertion of its own (its real public API base,
-   * `https://api.deepseek.com`, expects the caller's base URL to already carry any
-   * needed prefix), while llama-swap/llama.cpp only ever serves the OpenAI-conventional
-   * `/v1/chat/completions` — confirmed live: a bare `POST <baseUrl>/chat/completions`
-   * 404s, `POST <baseUrl>/v1/chat/completions` succeeds, and the harness's own error
-   * message template (`DeepSeek API error (HTTP ${status})`) reproduces the exact
-   * `HTTP_404` this feature originally shipped with unexplained. Claude Code's own SDK,
-   * by contrast, was already confirmed working end-to-end against the RAW `baseUrl` with
-   * no suffix — appending one there would be wrong, not just redundant.
+   * applied before being written to `baseUrlVar`, instead of being used verbatim. Only
+   * deepseek and copilot declare it — both speak to `<base>/v1/chat/completions` and a
+   * gateway saves the bare root. claude/gemini must NOT get it: Claude Code's own SDK was
+   * confirmed working end-to-end against the RAW `baseUrl`, so appending a suffix there
+   * would be wrong, not just redundant. The per-session config-file recipes do not need
+   * this flag at all — `renderConfigFile()` applies the suffix inside every template, so
+   * no configDir entry carries a raw base URL.
    */
   customModelInjection:
     | {

src/custom-model-hosts.ts

@@ -102,6 +102,17 @@ export interface CustomModelHost {
   defaultContextLength?: number;
 }
 
+/**
+ * The context window to apply for `modelId` on `host`: the discovered per-model figure when
+ * discovery has one, else the endpoint's user-set `defaultContextLength`, else undefined
+ * (no override applied, never a guess). ONE helper for every read site — the swap route,
+ * the quickstart route and boot recovery each re-write the CLI's config files, and a site
+ * that forgot the fallback silently loses the endpoint default on the next respawn.
+ */
+export function resolveCustomModelContextLength(host: CustomModelHost, modelId: string): number | undefined {
+  return host.modelContextLengths?.[modelId] ?? host.defaultContextLength;
+}
+
 export function customModelHostsPath(configDir: string): string {
   return join(configDir, CUSTOM_MODEL_HOSTS_FILE);
 }

src/custom-model-injection.ts

@@ -175,7 +175,9 @@ function opencodeModelLimit(contextLength: number | undefined): { context: numbe
   const total = Math.trunc(contextLength);
   if (total <= 0) return undefined;
   const output = Math.max(1, Math.min(8192, Math.floor(total / 4)));
-  return { context: total - output, output };
+  // `limit.context` is the FULL window as far as opencode is concerned — it reserves for
+  // the reply itself — so the discovered total is passed through, never a shrunk figure.
+  return { context: total, output };
 }
 
 function renderConfigContent(
@@ -189,10 +191,9 @@ function renderConfigContent(
     case 'opencode-json': {
       // `limit` carries the discovered context window (issue #749): without it opencode assumes
       // a large default window for an unrecognized model id and eventually overflows a much
-      // smaller local server. `context` is what opencode caps the PROMPT at, so the response
-      // headroom from `opencodeModelLimit` is subtracted up front — `output` is opencode's own
-      // cap on the reply, so setting them from the same discovered total keeps a single turn
-      // inside the real window.
+      // smaller local server. `context` is the FULL window and opencode reserves for the reply
+      // inside it itself, so `opencodeModelLimit` passes the real total through and only
+      // `output` gets the capped reservation.
       const limit = opencodeModelLimit(contextLength);
       return JSON.stringify({
         $schema: 'https://opencode.ai/config.json',

src/web/public/i18n.js

@@ -630,6 +630,9 @@
     'Last used': '上次使用',
     'That endpoint no longer exists': '该端点已不存在',
     'No models discovered for this endpoint yet': '此端点尚未发现任何模型',
+    // Model picker search box (index.html placeholder/aria-label and the no-matches line)
+    'Search models': '搜索模型',
+    'No models match': '没有匹配的模型',
     'Subagent Options': '子智能体选项',
     'Enable Tracking': '启用跟踪',
     'Active Tab Only': '仅活动标签页',

src/web/routes/custom-model-routes.ts

@@ -371,6 +371,9 @@ async function discoverModels(
   // itself calls loaded. Listed figures above are unaffected: reading them triggers nothing.
   const hasStatusField = entries.some((m) => m && typeof m === 'object' && 'status' in m);
   const slotCounts: Record<string, number> = {};
+  // Ids whose figure THIS round's loaded-model probe answered — the listed figure remains
+  // the source only for the ids NOT in this set (see the per-slot correction below).
+  const ctxFromProbe = new Set<string>();
   if (hasStatusField) {
     const loadedIds = entries
       .filter((m) => m.status && typeof m.status === 'object' && (m.status as { value?: unknown }).value === 'loaded')
@@ -407,11 +410,27 @@ async function discoverModels(
         // Assigning only when a probe answered keeps the listed figure (if any) when
         // neither the launch command nor /props stated one, and replaces it with the
         // stronger source whenever they did: cmd > /props > listed, never the reverse.
-        if (ctx !== undefined) contextLengths[id] = ctx;
+        if (ctx !== undefined) {
+          contextLengths[id] = ctx;
+          ctxFromProbe.add(id);
+        }
         if (slots !== undefined) slotCounts[id] = slots;
       }
     }
   }
+  // llama-swap documents the LISTED `context_length` as the PER-SLOT window, so whenever
+  // the listed figure is the one taken this round — the model was not loaded, or the
+  // probe answered nothing — its slot count is 1 by definition: the floor check must NOT
+  // divide a per-slot figure a second time. Without recording 1 here, the count a loaded
+  // round stored survives the merge (applyDiscoveredModels keeps what a later round does
+  // not overwrite): 262144 @ 4 slots stored, model unloaded, listed 65536 → 65536 with 4
+  // slots → a false 16384-per-slot warning. Even a count this round's launch command
+  // parsed must yield — pairing a per-slot listed figure with the cmd's --parallel is the
+  // same double division. Same rule the /props branch above applies. Ids whose figure the
+  // probe DID answer keep their probe-derived count untouched.
+  for (const id of Object.keys(contextLengths)) {
+    if (!ctxFromProbe.has(id)) slotCounts[id] = 1;
+  }
   return { models, contextLengths, slotCounts, sizesGB };
 }
 

src/web/routes/session-routes.ts

@@ -70,7 +70,7 @@ import {
   SessionWaitOutputQuerySchema,
   CustomModelSelectionSchema,
 } from '../schemas.js';
-import { readCustomModelHosts } from '../../custom-model-hosts.js';
+import { readCustomModelHosts, resolveCustomModelContextLength } from '../../custom-model-hosts.js';
 import { applyCustomModelInjection, removeConfigDir } from '../../custom-model-injection-apply.js';
 import {
   getLlamaSwapStatus,
@@ -1544,7 +1544,7 @@ export function registerSessionRoutes(
     // Discovery's per-model figure wins; the endpoint's user-set default covers models
     // whose size discovery could not determine (a gateway that states none, say). With
     // neither, the existing "no override, CLI assumes its own default" behaviour stands.
-    const contextLength = endpoint.modelContextLengths?.[body.modelId] ?? endpoint.defaultContextLength;
+    const contextLength = resolveCustomModelContextLength(endpoint, body.modelId);
     const slotCount = endpoint.modelSlotCounts?.[body.modelId];
 
     // Some CLIs (today: only claude) carry enough of their own fixed system-prompt/tool-
@@ -4452,7 +4452,7 @@ export function registerSessionRoutes(
       if (!cmEndpoint) return createErrorResponse(ApiErrorCode.NOT_FOUND, 'Model endpoint not found');
       // Per-model discovery figure first, user-set endpoint default second, nothing third
       // (see the dedicated route's own comment above).
-      const cmContextLength = cmEndpoint.modelContextLengths?.[customModel.modelId] ?? cmEndpoint.defaultContextLength;
+      const cmContextLength = resolveCustomModelContextLength(cmEndpoint, customModel.modelId);
       const cmSlotCount = cmEndpoint.modelSlotCounts?.[customModel.modelId];
 
       // See the dedicated route's own comment for the full reasoning: some CLIs' own fixed

src/web/server.ts

@@ -80,7 +80,7 @@ import { attachmentRegistry, buildFileThumbnailRoute, registerExternalAttachment
 import { folderGrants } from '../folder-listing.js';
 import { getCli, enabledClis, listClis } from '../config/cli-registry/registry.js';
 import { isCliEntryInstalled, probeStockCliAvailability } from '../utils/cli-installed-probes.js';
-import { readCustomModelHosts } from '../custom-model-hosts.js';
+import { readCustomModelHosts, resolveCustomModelContextLength } from '../custom-model-hosts.js';
 import { applyCustomModelInjection, customModelConfigDir, removeConfigDir } from '../custom-model-injection-apply.js';
 import type { CustomModelBookkeeping } from '../types/session.js';
 import { registerGeneratedArtifactAttachment } from '../generated-artifact-attachments.js';
@@ -1470,7 +1470,7 @@ export class WebServer extends EventEmitter {
         endpoint,
         saved.modelId,
         session.id,
-        endpoint.modelContextLengths?.[saved.modelId]
+        resolveCustomModelContextLength(endpoint, saved.modelId)
       )?.envOverrides;
     } catch (err) {
       console.warn('[WebServer] Failed to rebuild custom-model env on recovery:', err);

test/cli-registry-spawn-golden.test.ts

@@ -171,6 +171,15 @@ describe('codex', () => {
     expect(cx({ model: 'gpt-5', resumeSessionId: 'roll_42' })).toBe('codex --model gpt-5 resume roll_42');
   });
 
+  it('carries an Ollama-style `qwen3:32b` model id, and still drops a leading-flag shape', () => {
+    // `model-plain-colon`: the custom-model feature forces the selected id onto this param,
+    // and ids like `qwen3:32b` carry a colon. The leading-alphanumeric rule is what keeps a
+    // flag from riding the flag, so `:x` and `-x` must still drop out.
+    expect(cx({ model: 'qwen3:32b' })).toBe('codex --model qwen3:32b');
+    expect(cx({ model: ':leading' })).toBe('codex');
+    expect(cx({ model: '-yolo' })).toBe('codex');
+  });
+
   it('sends reasoning effort as one model_reasoning_effort config value, for every level', () => {
     for (const level of CODEX_REASONING_EFFORTS) {
       expect(cx({ reasoningEffort: level })).toBe(`codex --config model_reasoning_effort=${level}`);

test/custom-model-endpoint-rediscovery.test.ts

@@ -481,6 +481,52 @@ describe('refreshAllCustomModelHosts: parallel-slot enrichment (llama.cpp splits
     const [updated] = await readCustomModelHosts(dir);
     expect(updated.modelSlotCounts).toEqual({ a: 4 });
   });
+
+  it('a round where only the listed figure answers resets the slot count a loaded round learned', async () => {
+    // The false-warning scenario, measured on a real llama-swap: round 1 has the model
+    // loaded under `--ctx-size 262144 --parallel 4` — a TOTAL window shared across slots,
+    // correctly stored with its count. Round 2 finds it unloaded and takes the LISTED
+    // `context_length`, which llama-swap documents as the PER-SLOT window. If the stored 4
+    // survived, the floor check would divide the per-slot figure a second time
+    // (65536/4 = 16384) and warn about a model that serves 65536 to every request.
+    const dir = getDataDir();
+    await writeCustomModelHosts(dir, [host({ id: 'ep', baseUrl: 'http://localhost:8080' })]);
+    fetchMock.mockImplementation(async (url: URL) => {
+      if (url.pathname === '/v1/models') {
+        return new Response(JSON.stringify({ data: [{ id: 'm', status: { value: 'loaded' } }] }), { status: 200 });
+      }
+      if (url.pathname === '/running') {
+        return new Response(
+          JSON.stringify({
+            running: [{ model: 'm', state: 'ready', cmd: 'llama-server -m /models/m.gguf --ctx-size 262144 --parallel 4' }],
+          }),
+          { status: 200 }
+        );
+      }
+      throw new Error(`unexpected request: ${url.href}`); // the command answers both questions
+    });
+
+    await refreshAllCustomModelHosts();
+
+    let [round1] = await readCustomModelHosts(dir);
+    expect(round1.modelContextLengths).toEqual({ m: 262144 });
+    expect(round1.modelSlotCounts).toEqual({ m: 4 });
+
+    // Round 2: unloaded, no /running or /props probe in order, listing its per-slot figure.
+    fetchMock.mockResolvedValue(
+      new Response(JSON.stringify({ data: [{ id: 'm', status: { value: 'unloaded' }, context_length: 65536 }] }), {
+        status: 200,
+      })
+    );
+
+    await refreshAllCustomModelHosts();
+
+    const [round2] = await readCustomModelHosts(dir);
+    expect(round2.modelContextLengths).toEqual({ m: 65536 });
+    // 1 slot, not the stale 4: 65536 per slot with 1 recorded slot is the honest pair, so
+    // the floor check reads 65536/request — exactly what the server gives a single request.
+    expect(round2.modelSlotCounts).toEqual({ m: 1 });
+  });
 });
 
 describe('refreshAllCustomModelHosts: model-size enrichment (parsed from /v1/models description)', () => {

test/custom-model-injection.test.ts

@@ -112,20 +112,20 @@ describe('buildCustomModelInjection', () => {
     expect(parsed.provider.custom.models.qwen3).toEqual({});
   });
 
-  it('opencode: threads a discovered context length into the model limit with response headroom (issue #749)', () => {
+  it('opencode: threads a discovered context length into the model limit as the full window (issue #749)', () => {
     const result = buildCustomModelInjection(entryOrThrow('opencode'), endpoint, 'qwen3', 16384);
     if (result.kind !== 'env') throw new Error('unreachable');
     const parsed = JSON.parse(result.envOverrides.OPENCODE_CONFIG_CONTENT);
-    // One quarter held back for the reply: context 16384-4096, output 4096 — the sum is the
-    // real window, so a full-size prompt plus a max-length reply never exceeds the server.
-    expect(parsed.provider.custom.models.qwen3).toEqual({ limit: { context: 12288, output: 4096 } });
+    // `limit.context` IS the full window to opencode — it reserves for the reply itself —
+    // so the discovered total passes through; `output` only caps the reply at a quarter.
+    expect(parsed.provider.custom.models.qwen3).toEqual({ limit: { context: 16384, output: 4096 } });
   });
 
   it('opencode: caps the response headroom reservation at 8192 for big windows', () => {
     const result = buildCustomModelInjection(entryOrThrow('opencode'), endpoint, 'qwen3', 262144);
     if (result.kind !== 'env') throw new Error('unreachable');
     const parsed = JSON.parse(result.envOverrides.OPENCODE_CONFIG_CONTENT);
-    expect(parsed.provider.custom.models.qwen3).toEqual({ limit: { context: 253952, output: 8192 } });
+    expect(parsed.provider.custom.models.qwen3).toEqual({ limit: { context: 262144, output: 8192 } });
   });
 
   it('opencode: never invents a limit for an unknown or non-finite context length (issue #749 regression)', () => {

test/custom-model-recovery.test.ts

@@ -0,0 +1,91 @@
+/**
+ * @fileoverview Boot recovery rebuilds a saved custom-model selection through the SAME
+ * context-length resolution the routes use: the model's discovered figure, falling back to
+ * the endpoint's `defaultContextLength` (`resolveCustomModelContextLength`). Before the
+ * shared helper, `_rebuildCustomModelEnv` read `modelContextLengths` alone — so a session
+ * on a model that only the endpoint default covers came back from a Codeman restart with
+ * `CLAUDE_CODE_MAX_CONTEXT_TOKENS` (or a config file's context figure) silently gone,
+ * re-arming exactly the window overflow the var exists to prevent.
+ *
+ * `_rebuildCustomModelEnv` reads no `this` state (it re-derives everything from the endpoint
+ * store and the Session it is handed), so it is driven straight off the prototype with an
+ * empty receiver — `restoreMuxSessions()` itself cannot be reached under vitest, the same
+ * constraint recorded in test/session-model-recovery.test.ts.
+ */
+import { afterEach, describe, expect, it } from 'vitest';
+import { getDataDir } from '../src/config/instance.js';
+import { writeCustomModelHosts, type CustomModelHost } from '../src/custom-model-hosts.js';
+import { customModelConfigDir, removeConfigDir } from '../src/custom-model-injection-apply.js';
+import { Session } from '../src/session.js';
+import { WebServer } from '../src/web/server.js';
+import type { CustomModelBookkeeping } from '../src/types/session.js';
+
+type RebuildFn = (session: Session, saved: CustomModelBookkeeping) => Promise<Record<string, string> | undefined>;
+
+/** The real method, with an empty receiver — see the file header for why that is sound. */
+function rebuildCustomModelEnv(session: Session, saved: CustomModelBookkeeping) {
+  const fn = (WebServer.prototype as unknown as { _rebuildCustomModelEnv: RebuildFn })._rebuildCustomModelEnv;
+  expect(typeof fn, '_rebuildCustomModelEnv vanished from WebServer.prototype').toBe('function');
+  return fn.call({}, session, saved);
+}
+
+function endpoint(overrides: Partial<CustomModelHost> = {}): CustomModelHost {
+  return {
+    id: 'ep',
+    label: 'Ep',
+    baseUrl: 'http://localhost:8080',
+    apiKey: 'recovery-probe-key',
+    models: ['qwen3'],
+    ...overrides,
+  };
+}
+
+describe('_rebuildCustomModelEnv: context-length fallback on boot recovery', () => {
+  const configs: string[] = [];
+
+  afterEach(() => {
+    for (const id of configs.splice(0)) removeConfigDir(customModelConfigDir(id));
+  });
+
+  it('falls back to the endpoint defaultContextLength when discovery recorded no per-model figure', async () => {
+    await writeCustomModelHosts(getDataDir(), [endpoint({ defaultContextLength: 32768 })]);
+    const session = new Session({ workingDir: '/tmp', mode: 'claude' });
+    configs.push(session.id);
+
+    const env = await rebuildCustomModelEnv(session, { endpointId: 'ep', modelId: 'qwen3', envKeys: [] });
+
+    expect(env?.CLAUDE_CODE_MAX_CONTEXT_TOKENS).toBe('32768');
+  });
+
+  it('a discovered per-model figure still wins over the endpoint default', async () => {
+    await writeCustomModelHosts(getDataDir(), [
+      endpoint({ modelContextLengths: { qwen3: 131072 }, defaultContextLength: 32768 }),
+    ]);
+    const session = new Session({ workingDir: '/tmp', mode: 'claude' });
+    configs.push(session.id);
+
+    const env = await rebuildCustomModelEnv(session, { endpointId: 'ep', modelId: 'qwen3', envKeys: [] });
+
+    expect(env?.CLAUDE_CODE_MAX_CONTEXT_TOKENS).toBe('131072');
+  });
+
+  it('states nothing rather than guessing when neither source knows a window', async () => {
+    await writeCustomModelHosts(getDataDir(), [endpoint()]);
+    const session = new Session({ workingDir: '/tmp', mode: 'claude' });
+    configs.push(session.id);
+
+    const env = await rebuildCustomModelEnv(session, { endpointId: 'ep', modelId: 'qwen3', envKeys: [] });
+
+    expect(env?.CLAUDE_CODE_MAX_CONTEXT_TOKENS).toBeUndefined();
+  });
+
+  it('drops the selection to undefined (kept for clearing) when the endpoint no longer exists', async () => {
+    await writeCustomModelHosts(getDataDir(), []);
+    const session = new Session({ workingDir: '/tmp', mode: 'claude' });
+    configs.push(session.id);
+
+    await expect(rebuildCustomModelEnv(session, { endpointId: 'gone', modelId: 'qwen3', envKeys: [] })).resolves.toBe(
+      undefined
+    );
+  });
+});