feat(enrichment): make LLM output budgets configurable and model-aware

[TeaShaman-cyber]

Problem

AutoMem currently gives classification LLM calls a fixed 50-token output budget. That is enough for a very small deterministic response, but it is too small for some free or reasoning-oriented models: the model can spend the budget before it reaches the final JSON. AutoMem then falls back to a normal type such as Context, so the store succeeds while the quality loss remains quiet.

The same budget problem applies to LLM-assisted enrichment (summary, tags and entities), but classification and enrichment should be configurable separately because their outputs have different sizes.

Observed evidence

In a bounded, non-production AutoMem evaluation through an OpenAI-compatible route:

• BotHub Nemotron produced 4/6 valid JSON responses at 256 tokens and 5/6 at 512 tokens;
• Solar Pro4 produced 6/6 valid JSON responses at 128 tokens;
• Laguna S still produced one parser error in a 10-memory dry-run at both 256 and 512 tokens;
• a local experimental patch with separate budgets passed the focused AutoMem regression subset: 210 passed, 1 skipped;
• the full sanitized research note and aggregate evidence are public: https://github.com/TeaShaman-cyber/theseus-research/blob/main/docs/research/automem-resource-and-helper-evaluation-2026-08-22.md

These are protocol observations, not a claim that one model is universally better. The important point is that output budget is part of the provider/model compatibility contract.

Likely scope

The change could be kept narrow:

• make the classification completion budget configurable;
• give enrichment its own configurable budget, or document a clear shared setting;
• select max_tokens versus max_completion_tokens according to the model/API contract;
• preserve fail-soft storage and make malformed-output/fallback behavior observable;
• add focused tests for short-output, reasoning/free-model, malformed-JSON and fallback cases.

This is separate from quota circuit breaking tracked in #222 and from the already-closed custom-base-URL work in #96.

Acceptance criteria

• A deployment can configure the classification output budget without editing Python source.
• Classification and enrichment budgets do not have to be identical.
• Model families that require max_completion_tokens receive the correct parameter.
• A truncated or malformed response does not make the store fail or silently look like a successful high-confidence classification.
• Existing deployments retain a documented backward-compatible default.
• Focused unit tests cover the request parameters, valid JSON, malformed JSON and fallback counters.

Verification

pytest -q tests/test_classification_stats.py tests/test_enrichment.py
make test

Related

• fix(enrichment): circuit-break exhausted-quota LLM calls #222 — quota circuit breaker for repeated enrichment calls
• Add support for custom LLM base URL (OPENAI_BASE_URL) for classification/enrichment #96 — custom OPENAI_BASE_URL for classification/enrichment
• Public Theseus research note: https://github.com/TeaShaman-cyber/theseus-research/blob/main/docs/research/automem-resource-and-helper-evaluation-2026-08-22.md

No HAPP/VLESS, provider credentials or deployment-specific configuration is required for this feature request.

---

Prepared collaboratively by Semyon Poklad and Svarog, an AI engineering agent and engineering partner. Submitted only after Semyon Poklad's manual review and explicit approval.