fix(hermes): honor Ollama context floor

agents/hermes/config/build-env.ts

@@ -9,6 +9,9 @@ import { isObjectRecord } from "./object-record.ts";
 
 export type HermesWebSearchProvider = "tavily";
 
+/** Minimum context window Hermes accepts for generated configuration. */
+export const MIN_HERMES_CONTEXT_WINDOW = 64_000;
+
 export type HermesBuildSettings = {
   model: string;
   baseUrl: string;
@@ -29,6 +32,7 @@ export type HermesBuildSettings = {
   };
 };
 
+/** Read and validate the environment consumed by the Hermes config generator. */
 export function readHermesBuildSettings(env: NodeJS.ProcessEnv): HermesBuildSettings {
   const model = readRequiredEnv(env, "NEMOCLAW_MODEL");
   const baseUrl = readRequiredEnv(env, "NEMOCLAW_INFERENCE_BASE_URL");
@@ -50,14 +54,24 @@ export function readHermesBuildSettings(env: NodeJS.ProcessEnv): HermesBuildSett
   };
 }
 
-// Parse NEMOCLAW_CONTEXT_WINDOW as a positive integer of tokens. Empty, absent,
-// or malformed values return null so the generated config omits context_length
-// and Hermes keeps auto-detecting from the endpoint's /v1/models. See #6177.
+/**
+ * Parse `NEMOCLAW_CONTEXT_WINDOW` for Hermes config generation.
+ *
+ * Empty, absent, or malformed values return null so Hermes keeps auto-detecting
+ * from the endpoint's `/v1/models`; explicit values below the Hermes floor fail
+ * before writing an unusable config. See #6177.
+ */
 function readContextWindow(env: NodeJS.ProcessEnv): number | null {
   const raw = (env.NEMOCLAW_CONTEXT_WINDOW || "").trim();
   if (!/^[1-9][0-9]*$/.test(raw)) return null;
   const parsed = Number(raw);
-  return Number.isSafeInteger(parsed) && parsed > 0 ? parsed : null;
+  if (!Number.isSafeInteger(parsed) || parsed <= 0) return null;
+  if (parsed < MIN_HERMES_CONTEXT_WINDOW) {
+    throw new Error(
+      `Hermes NEMOCLAW_CONTEXT_WINDOW must be at least ${MIN_HERMES_CONTEXT_WINDOW} tokens, got ${parsed}`,
+    );
+  }
+  return parsed;
 }
 
 function readWebSearchProvider(env: NodeJS.ProcessEnv): HermesWebSearchProvider | null {

docs/inference/configure-model-limits.mdx

@@ -45,15 +45,21 @@ Hermes accepts `NEMOCLAW_CONTEXT_WINDOW` as its model-limit override.
 
 | Variable | Values | Default |
 |---|---|---|
-| `NEMOCLAW_CONTEXT_WINDOW` | Positive integer in tokens | Unset so Hermes auto-detects |
+| `NEMOCLAW_CONTEXT_WINDOW` | Positive integer, at least `64000` tokens | Unset so Hermes auto-detects |
 
 ```bash
 export NEMOCLAW_CONTEXT_WINDOW=65536
 $$nemoclaw onboard
 ```
 
 When onboarding resolves a valid value, NemoClaw writes it as `model.context_length` in `/sandbox/.hermes/config.yaml`.
-If no explicit or probed value is available, the field remains unset so Hermes can auto-detect it from the endpoint.
+For non-Ollama endpoints, the field remains unset when no explicit or probed value is available so Hermes can auto-detect it from the endpoint.
+When NemoClaw starts Local Ollama on macOS or Linux, it requests at least `64000` tokens.
+Fresh onboarding then verifies the loaded model's actual `context_length` through `/api/ps`.
+Resumed onboarding and sandbox rebuilds warm the exact recorded Ollama model and repeat this verification before reusing its route.
+When `NEMOCLAW_CONTEXT_WINDOW` is larger than `64000`, the Ollama runtime must provide at least that larger value.
+Onboarding stops before building the sandbox when Ollama reports a smaller, missing, or malformed value and shows the required `OLLAMA_CONTEXT_LENGTH` value for restarting Ollama.
+Setting `NEMOCLAW_CONTEXT_WINDOW` does not raise the Ollama daemon's runtime context or bypass this check.
 
 </AgentOnly>
 

docs/inference/set-up-ollama.mdx

@@ -85,6 +85,15 @@ If the validation probe times out, NemoClaw retries with a larger timeout before
 Each Ollama-backed OpenClaw passthrough checks whether the selected model is still loaded and sends a bounded warm-up request when necessary.
 </AgentOnly>
 
+<AgentOnly variant="hermes">
+Hermes requires at least `64000` tokens, so NemoClaw requests that host-side context length when it starts Ollama on macOS or Linux.
+Fresh onboarding verifies after model warm-up that `/api/ps` reports at least `64000` for the loaded model.
+Resumed onboarding and sandbox rebuilds warm the exact recorded Ollama model and repeat this verification before reusing its route.
+When `NEMOCLAW_CONTEXT_WINDOW` is unset, NemoClaw writes the verified runtime value as `model.context_length` in `/sandbox/.hermes/config.yaml`.
+An explicit `NEMOCLAW_CONTEXT_WINDOW` must be at least `64000`; NemoClaw writes that value only when the loaded model reports at least the same context length.
+If an existing or unmanaged daemon reports less, omits the value, or returns a malformed value, onboarding stops before building the sandbox and shows the required `OLLAMA_CONTEXT_LENGTH` value for restarting Ollama.
+</AgentOnly>
+
 ## Use Windows-Host Ollama from WSL
 
 When NemoClaw runs in WSL, the provider menu can offer these Windows-host actions:
@@ -201,6 +210,12 @@ It does not make the larger models usable on N1X, reject an explicitly selected
 When Ollama reports a context length below `16384` and `NEMOCLAW_CONTEXT_WINDOW` is unset, NemoClaw writes a `contextWindow` of `16384` so the agent prompt and tool definitions fit better than the stock daemon default.
 </AgentOnly>
 
+<AgentOnly variant="hermes">
+When `NEMOCLAW_CONTEXT_WINDOW` is unset, NemoClaw writes a valid loaded-model context length of at least `64000` as `model.context_length`.
+When you set a larger value explicitly, the loaded model must report at least that value before NemoClaw writes the explicit value.
+When the runtime value is lower or cannot be verified, onboarding stops and tells you which `OLLAMA_CONTEXT_LENGTH` value to use when restarting the host daemon before retrying.
+</AgentOnly>
+
 If the initial validation times out during a cold load, NemoClaw retries once with a 300-second probe budget.
 This retry also applies to tight-VRAM hosts where model warm-up can spill from GPU to CPU.
 

docs/reference/commands.mdx

@@ -3320,6 +3320,7 @@ Hermes-specific onboarding configuration:
 | `NEMOCLAW_NOUS_AUTH_METHOD` | same as `NEMOCLAW_HERMES_AUTH_METHOD` | Nous-specific alias for Hermes Provider authentication selection. |
 | `NEMOCLAW_HERMES_TOOL_GATEWAYS` | comma-separated list | Selects managed Hermes tool gateways in non-interactive onboarding. Valid values are `nous-web`, `nous-image`, `nous-audio`, `nous-browser`, and `nous-code`; the `nous-` prefix is optional. Unknown values fail before sandbox creation. |
 | `NEMOCLAW_HERMES_TOOL_GATEWAY_PRESETS` | comma-separated list | Back-compatible alias for `NEMOCLAW_HERMES_TOOL_GATEWAYS`. |
+| `NEMOCLAW_CONTEXT_WINDOW` | positive integer, at least `64000` tokens | Overrides `model.context_length` in the built Hermes config. Fresh and resumed Local Ollama onboarding, including sandbox rebuilds, must verify a loaded runtime context at least as large as this value. |
 | `NEMOCLAW_EXTRA_PLACEHOLDER_KEYS` | whitespace- or comma-separated list of upper-snake env keys | Adds operator-supplied OpenShell provider rows so per-profile credentials such as `TELEGRAM_BOT_TOKEN_AGENT_A` flow through the same out-of-process placeholder injection that the canonical channel tokens use, instead of being baked into each Hermes profile `.env` as raw text. Refer to [Extra placeholder keys](#extra-placeholder-keys) for the entry shape and validation rules. |
 
 </AgentOnly>
@@ -3432,7 +3433,7 @@ Set them before running `$$nemoclaw onboard`.
 | Variable | Format | Effect |
 |----------|--------|--------|
 | `NEMOCLAW_YES` | `1` to enable | Auto-accepts confirmation prompts (`--yes` equivalent) including in helpers like the Ollama proxy auth setup, but does not accept managed-vLLM storage warnings. |
-| `NEMOCLAW_OLLAMA_NO_AUTOSTART` | `1` to enable | Skips the wizard's eager Ollama auto-start during inference-provider selection (equivalent to passing `--no-ollama-autostart`). When set and Ollama is not running on `localhost:11434`, the `$$nemoclaw onboard` Local Ollama path prints a warning and selects the default fallback model instead of spawning `ollama serve`. The flag covers only the provider-selection step; later setup steps (auth proxy, validation, model warm) still expect a reachable Ollama. On Linux hosts with a systemd Ollama unit, the loopback-override path may still restart the daemon before this gate runs. |
+| `NEMOCLAW_OLLAMA_NO_AUTOSTART` | `1` to enable | Skips the wizard's eager Ollama auto-start during inference-provider selection (equivalent to passing `--no-ollama-autostart`). When set and Ollama is not running on `localhost:11434`, an agent that uses the legacy `16384`-token context floor, currently OpenClaw, prints a warning and selects the default fallback model instead of spawning `ollama serve`. An agent that requires a larger verified runtime context, currently Hermes at `64000` tokens, returns to interactive provider selection or exits when the Ollama provider is pinned or onboarding is non-interactive. The flag covers only the provider-selection step; later setup steps (auth proxy, validation, model warm) still expect a reachable Ollama. On Linux hosts with a systemd Ollama unit, the loopback-override path may still restart the daemon before this gate runs. |
 | `NEMOCLAW_NON_INTERACTIVE_SUDO_MODE` | `prompt` or empty/unset | When set to `prompt`, allows non-interactive onboarding to use prompt-capable `sudo` for host setup steps that require elevation, which can ask for a password. Empty/unset is the default and uses `sudo -n`, which fails instead of asking for a password. Any other value is rejected. |
 | `NEMOCLAW_NO_EXPRESS` | `1` to enable | Installer-only. Skips the DGX Spark, DGX Station, and Windows WSL express install prompt and continues with the normal interactive onboarding flow. |
 | `NEMOCLAW_EXPERIMENTAL` | `1` to enable | Surfaces experimental providers and flows in onboarding. |

docs/reference/troubleshooting.mdx

@@ -2444,6 +2444,9 @@ If the process exists but the endpoint is unreachable, use the restart action wh
 ### Ollama inference fails or hangs in WSL
 
 Ollama configures context length based on your hardware.
+
+<AgentOnly variant="openclaw">
+
 On some GPUs (for example RTX 3500), the default context length is not sufficient for OpenClaw.
 During onboarding, NemoClaw raises loaded-model context lengths below `16384` to `16384` when `NEMOCLAW_CONTEXT_WINDOW` is unset.
 Set the variable manually when you need a different value or when you run Ollama outside the managed onboarding path.
@@ -2454,6 +2457,26 @@ pkill -f 'ollama serve'
 OLLAMA_CONTEXT_LENGTH=16384 ollama serve
 ```
 
+</AgentOnly>
+
+<AgentOnly variant="hermes">
+
+Hermes requires at least `64000` tokens.
+During onboarding, NemoClaw verifies the loaded model's actual `context_length` through Ollama's `/api/ps` endpoint.
+Resumed onboarding and sandbox rebuilds warm the exact recorded Ollama model and repeat this check before reusing its route.
+Resume stops when that recorded model is missing, Ollama is unreachable, model warm-up fails, or the runtime context cannot be verified.
+If you set `NEMOCLAW_CONTEXT_WINDOW` above `64000`, the loaded model must provide at least that larger value.
+If the runtime value is too small, missing, or malformed, onboarding stops before sandbox creation and asks you to restart the host daemon with the required context length.
+`NEMOCLAW_CONTEXT_WINDOW` controls Hermes prompt budgeting; it does not change the Ollama daemon or bypass this runtime check.
+Force a larger context length:
+
+```bash
+pkill -f 'ollama serve'
+OLLAMA_CONTEXT_LENGTH=64000 ollama serve
+```
+
+</AgentOnly>
+
 Verify that Ollama inference works:
 
 ```bash
@@ -2470,11 +2493,24 @@ sudo systemctl status ollama
 
 If it is active, stop it first, then start with the custom context length:
 
+<AgentOnly variant="openclaw">
+
 ```bash
 sudo systemctl stop ollama
 OLLAMA_CONTEXT_LENGTH=16384 ollama serve
 ```
 
+</AgentOnly>
+
+<AgentOnly variant="hermes">
+
+```bash
+sudo systemctl stop ollama
+OLLAMA_CONTEXT_LENGTH=64000 ollama serve
+```
+
+</AgentOnly>
+
 For additional troubleshooting, refer to the [Windows Setup](../get-started/prerequisites/windows-preparation) page.
 <AgentOnly variant="openclaw">
 For first-time OpenClaw setup, refer to the [Quickstart](../get-started/quickstart).

src/lib/inference/local.ts

@@ -29,10 +29,16 @@ import {
   OLLAMA_MODEL_REGISTRY,
   SMALLEST_OLLAMA_MODEL_TAG,
 } from "./ollama-model-registry";
-import type { OllamaRuntimeModelStatus } from "./ollama-runtime-context";
+import type {
+  ApplyOllamaRuntimeContextWindowOptions,
+  ApplyOllamaRuntimeContextWindowResult,
+  OllamaRuntimeModelStatus,
+} from "./ollama-runtime-context";
 import {
   applyOllamaRuntimeContextWindow as applyOllamaRuntimeContextWindowWithHost,
+  getOllamaContextWindowFloorForAgent,
   MAX_AUTODETECTED_OLLAMA_CONTEXT_WINDOW,
+  MIN_HERMES_OLLAMA_CONTEXT_WINDOW,
   parsePositiveInteger,
   probeOllamaRuntimeModelStatus as probeOllamaRuntimeModelStatusWithHost,
   resetOllamaRuntimeContextWindowAutoState,
@@ -876,7 +882,12 @@ export function parseOllamaTags(output: string | null | undefined): string[] {
   }
 }
 
-export { MAX_AUTODETECTED_OLLAMA_CONTEXT_WINDOW, parsePositiveInteger };
+export {
+  getOllamaContextWindowFloorForAgent,
+  MAX_AUTODETECTED_OLLAMA_CONTEXT_WINDOW,
+  MIN_HERMES_OLLAMA_CONTEXT_WINDOW,
+  parsePositiveInteger,
+};
 
 export function probeOllamaRuntimeModelStatus(
   model: string,
@@ -900,8 +911,12 @@ export function resolveOllamaRuntimeContextWindow(
 
 export { resetOllamaRuntimeContextWindowAutoState };
 
-export function applyOllamaRuntimeContextWindow(selectedModel: string): void {
-  applyOllamaRuntimeContextWindowWithHost(selectedModel, getResolvedOllamaHost);
+/** Apply Ollama runtime context-window adoption using the resolved local host. */
+export function applyOllamaRuntimeContextWindow(
+  selectedModel: string,
+  options: Pick<ApplyOllamaRuntimeContextWindowOptions, "contextWindowFloor"> = {},
+): ApplyOllamaRuntimeContextWindowResult {
+  return applyOllamaRuntimeContextWindowWithHost(selectedModel, getResolvedOllamaHost, options);
 }
 
 export function applyVllmRuntimeContextWindow(