Codeman/docs/custom-model-endpoints.md

# Custom Model Endpoint Profiles

Point any Codeman-supported harness — Claude, opencode, Codex, Gemini, Pi,
Grok, DeepSeek, or OMP — at a custom OpenAI-compatible endpoint instead of
its native cloud backend, for a given session. "Custom endpoint" covers both
**local** hardware (llama.cpp, Ollama, vLLM, a home GPU rig, or purpose-built
boxes like NVIDIA DGX Spark or AMD Strix Halo mini-PCs) and **cloud**
services (Azure AI Foundry's OpenAI-compatible endpoint, OpenRouter, a
company gateway) — anything answering `GET /v1/models` and
`POST /v1/chat/completions` in the standard shape. Design doc, per-CLI
recipe confidence table, and security reasoning:
[`custom-model-endpoints-plan.md`](custom-model-endpoints-plan.md).

> **Status**: fully wired end to end — registry capability, the injection
> engine, the endpoint store + discovery route, the session restart route,
> a settings-panel CRUD surface, and the Run-menu picker described below.
> Antigravity has no known custom-endpoint mechanism and is not supported.
> The HTTP API (examples below) still works directly and is what the picker
> itself calls under the hood.

## Turning it on

App Settings → Models → **Custom model endpoints** (synced setting
`customModelEndpointsEnabled`, default **OFF**). Turning it on does two
things: it reveals the endpoint list/add/edit/discover panel in that same
settings section, and it makes the Run menu offer a generated entry per
(harness, endpoint) pair — see "The Run-menu picker" below. The API
equivalent:

```bash
curl -sk -X PUT https://localhost:3000/api/settings \
  -H 'Content-Type: application/json' \
  -d '{"customModelEndpointsEnabled": true}'
```

## Adding an endpoint

Via App Settings → Models → Custom model endpoints → **+ Add endpoint**, or
directly:

```bash
curl -sk -X POST https://localhost:3000/api/model-endpoints \
  -H 'Content-Type: application/json' \
  -d '{"id": "llama-box", "label": "Home llama.cpp", "baseUrl": "http://192.168.1.50:8080"}'
```

`apiKey` is optional (most local servers don't check it). `authStyle`
(`bearer` | `api-key`, default `bearer`) controls which auth header
convention discovery uses: `bearer` is `Authorization: Bearer <key>`
(llama.cpp, OpenAI-compatible servers, most gateways), `api-key` is the
`api-key: <key>` header Azure AI Foundry wants. There is deliberately no
"send both" option: measured against a real llama-swap server, a request
carrying both headers hung indefinitely. `baseUrl` must be `http(s)`, carry
no embedded credentials, and may not point at a link-local or cloud-metadata
address; discovery re-checks the address the name actually resolves to.

Discover its available models:

```bash
curl -sk -X POST https://localhost:3000/api/model-endpoints/llama-box/discover-models
```

This calls the endpoint's own `GET /v1/models` and stores the returned list
on the endpoint record; `GET /api/model-endpoints` lists everything
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, but llama.cpp's
llama-swap-proxied `GET /props?model=<id>` does (`n_ctx`). Discovery only ever
calls it for 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
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
session" below) so a CLI that would otherwise assume a large default context
window for an unrecognized model id stops silently overflowing a much
smaller real one.

**File size is discovered too, when the server states one.** llama-swap
writes a GB figure into an auto-discovered model's own `description`
(`"Auto-discovered 16.35 GB - parameters auto-fitted by llama.cpp"`), parsed
into `modelSizesGB` — unlike context length, this needs no `/props` probe
(the figure is right there in the `/v1/models` response) and so is populated
for every model regardless of loaded state. A hand-configured profile's own
description has no such figure and correctly gets no entry, never a guess.
Used only to label the Run-menu picker's "loading model" banner with a
rough, UNMEASURED expected-time estimate (`_estimateModelLoad()` in
session-ui.js, based on typical local NVMe/SSD throughput — not benchmarked
against any real endpoint's actual hardware/storage) and to scale that same
banner's own give-up timeout for a very large model; never anything a
server-side check relies on.

**The loading banner shows a live countdown against that same timeout, and
treats a real timeout as a failure, not a shrug.** It checks llama-swap's
own `/running` every second (`GET /api/model-endpoints/:id/running-status`)
and counts down against the size-scaled (or flat 5-minute) timeout live; if
the countdown reaches zero with the target model still not ready, the
banner turns into a sticky error naming the llama-swap server's own logs as
where to look, and the session the load was for is closed automatically —
a console left open and pointed at a model that never finished loading is
worse than no console at all.

`defaultModelId` names which discovered model the picker pre-marks for that
endpoint — the settings panel's Edit form exposes it as a select populated
from the endpoint's own discovered `models`, and the route refuses a value
that isn't one of them. It is applied automatically only when the endpoint
has exactly one discovered model (nothing to choose); with two or more it
is a pre-selection in the model-picker dialog below, never a silent default.
Re-discovering drops a default that no longer appears in the fresh list
rather than carrying an invalid one forward.

**Model lists refresh themselves.** A background sweep (`server.ts`,
`CUSTOM_MODEL_REDISCOVER_INTERVAL_MS`, every 5 minutes) re-discovers every
saved endpoint the same way the manual `POST .../discover-models` route
does, best-effort per endpoint — one being unreachable on a given cycle
never blocks the others. Off under `npm test`, same reasoning as the Codex
plan-usage poll it sits beside: no real network to hit, no server instance
to keep the timer alive for.

## The Run-menu picker

With the setting on and at least one endpoint carrying a discovered model,
the toolbar's Run dropdown grows a **Custom Endpoints** section: one entry
per (harness that can redirect to a custom endpoint, saved endpoint) pair,
e.g. "Claude Code (llama.cpp)". The harness list is read off the CLI
registry's own `capabilities.customModelInjection` at page render
(`window.__codemanCustomModelClis`, `server.ts`) — never a hardcoded id list
in the frontend — so a CLI whose injection recipe lands later shows up with
no frontend change, and Antigravity (`unsupported`) never does.

Picking an entry re-fetches the endpoint (`selectCustomModelEntry()`,
`session-ui.js`) rather than trusting anything cached from the dropdown's
own render — the model list can have changed via the 5-minute sweep above
or a settings-panel edit since the menu opened. With exactly one discovered
model it runs straight away; with two or more, a small modal
(`#customModelPickModal`) lists them and asks which one to use for this
launch, with the endpoint's `defaultModelId` marked but not auto-chosen —
the point of asking is letting one launch deliberately differ from the
saved default, not just confirming it.

**How the launch itself applies the endpoint depends on the harness.** For
opencode, Codex, Gemini, Pi, Grok, DeepSeek and OMP (`runCustomModelEntry` →
`_runCustomModelEntryOneShot`), the endpoint/model is folded into the SAME
`POST /api/quick-start` call that creates the session (`customModel` field),
so the session launches directly on the endpoint — no restart, no visible
relaunch. Claude (`_runCustomModelEntryViaRestart`) still uses the original
two-step design: the launch runs a single native session exactly the way its
own Run-menu entry would, then **waits for the new session to go idle**
(`GET .../wait?until=idle`, bounded at 20s — a normal 200 either way, never
an error, per the wait endpoint's own contract) before applying the endpoint
via the restart route below. That wait exists because a freshly launched CLI
reports itself as `busy` for its own startup (a boot spinner, a
workspace-trust check) well before the apply call would otherwise reach it,
and the apply route correctly refuses to restart a session mid-turn — a
fresh boot looks exactly like one from the outside. A session still busy
after the wait reaches the apply call anyway and gets that route's own
honest `SESSION_BUSY` error, now visible as a sticky toast with a close
button rather than a generic message that vanished in three seconds. Claude
stays on this path because its own restart (`--resume`-based, keeping the
conversation) is far less jarring than the other seven's, and `runClaude()`'s
multi-tab launch and docker-config-drift confirm/retry loop make folding it
into the one-shot path separate work. It is a
one-off "try this endpoint" action, not a sticky mode: the plain Run button
still means "this harness, native cloud" afterward. Entries are hidden
entirely for a remote or Docker active case, since the apply route refuses
both (see the next section).

## Launching directly on an endpoint (no restart)

```bash
curl -sk -X POST https://localhost:3000/api/quick-start \
  -H 'Content-Type: application/json' \
  -d '{"caseName": "myapp", "mode": "codex", "customModel": {"endpointId": "llama-box", "modelId": "qwen3"}}'
```

`POST /api/quick-start`'s `customModel` field (`{endpointId, modelId,
confirmed?}`) computes the same injection the restart route below does, but
BEFORE the session exists — the session is minted its own id up front
(`crypto.randomUUID()`), the injection (env vars, and for a `configDir`-kind
CLI, the written config file) targets that real id, and the session launches
already pointed at the endpoint. No restart, because there was never a
native-backend launch to restart away from. Runs the same llama-swap
conflict check as the restart route (below) — a `409`-shaped
`{requiresConfirmation, currentlyLoadedModel, affectedSessions}` response
with no session created, resolved by retrying with `confirmed: true` — and
is refused the same way for a remote or Docker case. This is what the
Run-menu picker uses for opencode, Codex, Gemini, Pi, Grok, DeepSeek and OMP;
Claude still uses the restart route below (see "The Run-menu picker" above
for why).

## Applying a model to an ALREADY-RUNNING session

```bash
curl -sk -X POST https://localhost:3000/api/sessions/<sessionId>/custom-model \
  -H 'Content-Type: application/json' \
  -d '{"endpointId": "llama-box", "modelId": "qwen3"}'
```

This computes the CLI-specific env vars / config for that session's mode
(see the recipe table in `custom-model-endpoints-plan.md`) and **restarts the session's
CLI process in place** — same pane, same tmux session, fresh env. That
restart is necessary, not incidental: every supported harness reads its
endpoint config at process start, not per-turn, so there is no live
hot-swap. A Claude session is relaunched with `--resume <conversation> ||
--session-id <id>`, so it continues the conversation it was on; pi, omp and
grok are relaunched with the `--model` value that selects the injected
provider (`custom/<modelId>` for pi and omp, `codeman-custom` for grok),
since for those three the config file alone does not switch the model.
**Remote (SSH) and Docker sessions are refused** (400) for now: their restart
reattaches the durable remote/in-container tmux rather than relaunching the
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
  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.
- `CLAUDE_CONFIG_DIR` is pointed at the same isolated per-session directory
  the `configDir`-kind CLIs use (empty, no files written into it), so the
  injected `ANTHROPIC_API_KEY` never shares a directory with a stored
  claude.ai OAuth login. Claude Code still prints "Both claude.ai and
  ANTHROPIC_API_KEY set" when the two coexist in the same config directory —
  cosmetic (confirmed live: the API key wins for actual requests either way,
  visible in the terminal's own `API Usage Billing` line) but worth
  eliminating rather than living with. The directory's `projects`
  subdirectory is symlinked (a junction on Windows) back to the real
  `~/.claude/projects` so the response viewer, subagent windows and Read My
  Mind keep working for that session — the same trade-off and fix documented
  for a manually-set `CLAUDE_CONFIG_DIR` in
  [`docs/wiki/Agent-CLIs.md`](wiki/Agent-CLIs.md), just applied
  automatically here. Best-effort: a platform that refuses the symlink keeps
  the pre-existing blind-response-viewer side effect rather than failing the
  whole custom-model apply over it.

**That isolated directory needed one more fix to actually be usable
non-interactively.** An otherwise-empty `CLAUDE_CONFIG_DIR` has none of a
real profile's prior "Detected a custom API key — use it?" approvals, so
without more, Claude Code stops and asks that on *every single launch* —
confirmed live, and with nobody at a TTY to answer, its own default answer
("No") silently refuses the very key this feature just injected, which
looks like the endpoint being ignored entirely. `customModelInjection`'s
`apiKeyTrustFile` (`{ relPath: '.claude.json', shape:
'claude-api-key-responses' }` on claude's entry) pre-seeds that exact
approval: the apply step merges `customApiKeyResponses.approved: [apiKey]`
into `<configDir>/.claude.json`, the same field a real answered prompt
itself writes to (confirmed against a real file after answering by hand
once) — this answers the prompt in advance rather than bypassing it. The
merge preserves whatever else the CLI already wrote into that file on an
earlier launch in the same isolated directory (`userID`, `numStartups`,
earlier approved keys), and a missing or corrupt file is treated as empty
rather than failing the apply.

**llama-swap gets two more fixes on top of the context-length/config-dir
ones above, both from watching a real switch live.** llama.cpp only ever
runs one model at a time; llama-swap swaps the backing process on demand,
which can take anywhere from a few seconds to well over a minute:

- **The conflict check.** Both apply routes (the restart one here and the
  one-shot `POST /api/quick-start` above) call llama-swap's own
  `GET /running` first — feature-detected, so a plain llama.cpp/OpenAI-
  compatible server (no such endpoint) is simply never checked. If a
  *different* model is currently loaded and ready, and another **live
  session's own selection** is using it, the apply returns
  `{requiresConfirmation: true, currentlyLoadedModel, affectedSessions}`
  instead of silently switching — nothing is applied or created yet.
  Retrying with `confirmed: true` skips the check. Switching with nothing
  else affected proceeds immediately; this is a warning about disrupting
  another session, never a gate on the switch itself.
- **Actually starting the load.** llama-swap has no "switch model" admin
  call — the only thing that starts a swap is a real inference request
  naming the model, and confirmed live: applying a selection alone never
  reached llama-swap at all (nothing in its own server logs), since nothing
  had actually asked it to load anything yet. Both apply routes now also
  send the smallest real request that will — `POST <baseUrl>/v1/chat/
  completions` with `max_tokens: 1` and one throwaway message — whenever the
  target model isn't already the one loaded and ready, fire-and-forget (its
  response is never read; `GET /api/model-endpoints/:id/running-status`,
  polled client-side, is what actually confirms readiness). The response
  also carries `modelSwapInProgress: true` in that case, which is what
  drives the Run-menu picker's own "loading model" status banner.

Clear back to the harness's native cloud default with:

```bash
curl -sk -X POST https://localhost:3000/api/sessions/<sessionId>/custom-model \
  -H 'Content-Type: application/json' -d '{"clear": true}'
```

Clearing also removes the env vars the selection injected from the tmux
session (they persist there and would otherwise be inherited by the
relaunched CLI) and deletes the per-session config directory
(`~/.codeman/custom-model-configs/<sessionId>`, written 0600 because pi and
omp embed the API key in it). That directory is also removed when the
session is deleted. The selection survives a Codeman restart: the endpoint
id, model and injected key NAMES are persisted, the values are re-derived
from the endpoint store on recovery, and the pane keeps running against the
endpoint in between because tmux retains its environment.

**New sessions always default back to the harness's native backend.** A
custom-endpoint selection is a per-session choice, never a sticky global
default — starting a fresh session doesn't inherit whatever the last one was
pointed at.

## Confidence per harness

Every harness except Antigravity has now been run end-to-end against a real
llama-swap server via `scripts/test-local-llm-harnesses.ts` (a dynamic
script that reads the live CLI registry, so a registry change is picked up
automatically). Results:

- **Claude, opencode, Pi, Grok, OMP** — verified: a real "hello world" reply
  came back through the endpoint.
- **Codex** — the config is structurally correct, but Codex only speaks the
  Responses API since Feb 2026, which llama.cpp/llama-swap don't implement.
  This is a real protocol incompatibility, not a bug here; Codex support
  needs a Responses-API-compatible endpoint.
- **Gemini** — fails with `Invalid auth method selected`, traced to an
  undocumented `GATEWAY` auth path gemini-cli selects once
  `GOOGLE_GEMINI_BASE_URL` is set. Unresolved after real investigation
  (several auth workarounds were tried and ruled out); do not rely on
  Gemini support yet.
- **DeepSeek** — the request reaches the server (env vars are read) but
  gets a consistent `HTTP_404`. Root cause not identified; best-effort only.
- **Antigravity** — no known custom-endpoint mechanism at all; unsupported.

See the confidence table in `custom-model-endpoints-plan.md` for the full detail behind
each result. `scripts/test-local-llm-harnesses.ts` is the standalone script
used to check a harness against a real endpoint outside the web UI
entirely; see its own `--help` for usage.

## Security note

Every env var this feature can set that redirects a session's traffic
(`ANTHROPIC_BASE_URL`, `GOOGLE_GEMINI_BASE_URL`, `CODEX_HOME`, etc.) is
listed in that CLI's `privilegedEnvKeys` in the CLI registry, so a
non-granted multi-user owner cannot set one directly via the generic
`envOverrides` API field — only through this feature's own route, which
computes the value from an admin-configured, SSRF-guarded endpoint rather
than trusting arbitrary client input. See the "Multi-user security
hardening" section of `custom-model-endpoints-plan.md` for the full reasoning; several
of these were reachable via the generic `envOverrides` field even before
this feature existed, and building this surfaced and closed that gap.