Envelope & Tiers

Every tool returns the same envelope

{
  result: <tool-specific>,
  tier_used: "instant" | "workhorse" | "deep" | "embed",
  model: string,
  hardware_profile: string,     // "dev-rtx5080" | "dev-rtx5080-qwen3" | "m5-max"
  tokens_in: number,
  tokens_out: number,
  elapsed_ms: number,
  residency: {
    in_vram: boolean,
    size_bytes: number,
    size_vram_bytes: number,
    evicted: boolean
  } | null
}

The envelope is the audit trail. Every call shows what tier it used, which model served it, how long it took, and whether the model was resident in VRAM.

The four tiers

Tier	Job shape	Token budget	Typical tool
Instant	Short, stable-shape classification/triage	5s timeout	classify, triage_logs, summarize_fast
Workhorse	Code/prose drafting, schema-constrained extraction	20s timeout	draft, extract, chat
Deep	Long-input synthesis, chunk-grounded answers	90s timeout	research, summarize_deep, corpus_answer, briefs, packs
Embed	Vectors only	—	embed, embed_search, corpus index/refresh/search

Tiers are picked by the tool, not by Claude. Pick the job; the tier follows.

Hardware profiles

Profile	Instant	Workhorse	Deep	Embed
dev-rtx5080 (default)	hermes3 8B	hermes3 8B	hermes3 8B	nomic-embed-text
dev-rtx5080-qwen3	qwen3 8B	qwen3 8B	qwen3 14B	nomic-embed-text
m5-max	qwen3 14B	qwen3 14B	qwen3 32B	nomic-embed-text

The default dev profile collapses all three work tiers onto hermes3:8b — the validated Hermes Agent integration path. One model to pull, one residency cost, one set of behavior. dev-rtx5080-qwen3 is the alternate rail for callers who want Qwen 3’s thinking mode via the server’s THINK_BY_SHAPE plumbing. m5-max scales the Qwen 3 ladder for unified memory.

Residency — reading /api/ps

Ollama pages models to disk silently when VRAM is tight. Inference drops 5–10× when this happens, and the only signal is in /api/ps. The envelope carries that signal:

• residency.in_vram: false — not loaded
• residency.evicted: true — was loaded, paged out
• residency.size_vram_bytes < residency.size_bytes — partially resident

When you see any of these during a slow call, trim OLLAMA_MAX_LOADED_MODELS or restart Ollama. This is the main way production slowdowns sneak in.

The NDJSON log

Every call is logged as one line of NDJSON to ~/.ollama-intern/log.ndjson:

{"kind":"call","ts":"2026-04-21T12:00:00Z","tool":"ollama_classify","envelope":{"tier_used":"instant","model":"hermes3:8b","hardware_profile":"dev-rtx5080","tokens_in":87,"tokens_out":12,"elapsed_ms":340,"residency":{"in_vram":true,"evicted":false}}}

No prompts are logged. No inline text is logged. Just the envelope. Nothing leaves the box.

Filter by hardware_profile to keep dev numbers out of publishable benchmark tables — bench scripts already do this.

One-tier overrides

Env vars beat profile picks for one-off swaps:

Env var	Example
INTERN_TIER_INSTANT	hermes3:8b
INTERN_TIER_WORKHORSE	hermes3:8b
INTERN_TIER_DEEP	hermes3:8b
INTERN_EMBED_MODEL	nomic-embed-text