feat(agent): make MINIMUM_CONTEXT_LENGTH configurable via HERMES_MIN_CONTEXT_LENGTH

[divnjl2]

Problem

agent/agent_init.py hard-rejects any model whose context window is below MINIMUM_CONTEXT_LENGTH = 64_000:

Model X has a context window of 40,000 tokens, which is below the minimum
64,000 required by Hermes Agent. ... or set model.context_length in config.yaml to override.

But the advertised override does not work: an explicit model.context_length below 64K is still rejected (the same constant is also the context_compressor floor). So there is no real escape hatch.

Who this hurts

Operators running local GGUF backends (llama.cpp / llama-swap / ik_llama) where the effective per-request context = -c / --parallel. A 256K model split into 6 slots is a 43K slot; that is a perfectly usable agent context, but it is hard-rejected at init, bricking the whole deployment.

Change

Read the floor from HERMES_MIN_CONTEXT_LENGTH (default 64000 — behaviour unchanged for everyone who doesn't set it). One line, backward-compatible. Lets small-slot operators lower the floor to match their real slot size.

Test

Local llama-swap fleet (35B/9B/4B split across two GPUs, slots 8-43K). Before: every cell dies at init with the 64K refusal. After (HERMES_MIN_CONTEXT_LENGTH=8000): 35B and 9B cells initialize and complete tool-calling turns normally.

[teknium1]

Thanks for identifying the real lower-context startup constraint and for keeping the patch narrow.

This automated hermes-sweeper review is closing the PR under the project's standing configuration policy:

• AGENTS.md:102-107 reserves .env / HERMES_* variables for secrets and requires non-secret behavioral settings to live in config.yaml.
• This patch's substantive mechanism is the new HERMES_MIN_CONTEXT_LENGTH environment variable in agent/model_metadata.py (f25a5f084cf9).
• The current guard is real (agent/agent_init.py:1872-1880), so a future proposal can re-scope this as a validated, documented config.yaml setting with coverage for the startup, compression, and local-runtime paths.

The issue is not the local-model use case; it is the user-facing environment-variable configuration surface.

---

Closed as not-planned per standing maintainer policy (env-var-for-config). This is a design-direction decision, not a code-quality judgment — see the Contribution Rubric in AGENTS.md for what the project is looking for. If you believe this policy was misapplied to your change, comment here and a maintainer will take a look.