Act mode: three-tier system prompts (frontier / mid / small) instead of compact-vs-full binary

[Ayush7614]

Problem

Act mode currently routes system prompts with a binary choice: full SYSTEM_PROMPT_ACT vs SYSTEM_PROMPT_ACT_COMPACT (via provider useCompactPrompt + compact/mid/full tool tiers).

TODOs.md §1 documents a deeper issue: small local models need more explicit guidance, not a shorter prompt. The compact prompt strips worked examples and nuance that small models rely on — while still paying the token cost of a large tool schema on mid/full tiers.

Evidence cited in TODOs: qwen3.6-27b trace re-downloading the same files because scratchpad/download state fell out of attention — compact prompt would have made this worse by removing SCRATCHPAD guidance.

Current mismatch:

User cohort	What they need	What they get
Frontier cloud (Sonnet, GPT-4o)	Trimmed rules, trust planning	Full ~7K prompt
Mid local (32B–70B)	Rules + 1–2 examples per rule	Full prompt
Small local (7B–27B)	More examples + failure-mode reminders	Often compact prompt (too thin) OR full prompt (token-heavy)

Tool tier (compact / mid / full) and prompt tier are orthogonal but today only loosely coupled via useCompactPrompt.

Proposed solution

Introduce three prompt tiers routed in _getActPrompt():

Tier	Target models	Prompt	Approx size
frontier	Sonnet, Opus, GPT-4o, Gemini Pro	Trim examples, keep rules	~3K tokens
mid	Llama 70B, Qwen 35B, GPT-4o-mini	Full rules + selective examples	~5K tokens
small	7B–30B local (Ollama, llama.cpp, LM Studio)	Full rules + more examples + imperative scaffolding	~6–7K tokens

Implementation sketch

1. Provider config — replace or extend useCompactPrompt with promptTier: 'frontier' | 'mid' | 'small' (default heuristics: cloud flagship → frontier, OpenRouter mid → mid, local servers → small).
2. New prompt bodies in tools.js (Chrome + Firefox):
   • SYSTEM_PROMPT_ACT_FRONTIER
   • Keep SYSTEM_PROMPT_ACT as mid default
   • SYSTEM_PROMPT_ACT_SMALL (expanded, not shrunk — per TODOs analysis)
3. _getActPrompt() — route by promptTier, not binary compact flag.
4. Settings (optional v1) — advanced dropdown “Prompt detail: Auto / Frontier / Mid / Small” overriding heuristic.
5. Token budget guard — if promptTier === 'small' and contextWindow < 12k, warn in verbose mode or auto-compact tools to compact tier.

Verification

• Re-run qwen3.6-27b / small-model scenario from TODOs — no re-download loop when scratchpad pins downloads
• Frontier model still gets shorter prompt (token savings)
• npm test — tier routing unit tests
• Chrome + Firefox parity
• LLM scenario spot-check on frozen harness (no regression on frontier cases)

Why this matters now

WebBrain is on the Chrome Web Store and Firefox AMO — many new users try local LLMs first. Small-model experience directly drives ratings and “doesn’t work” reports.

Scope

Medium–large (~400–600 lines across prompts + routing + tests + settings). Can be phased:

• Phase 1: tier enum + routing + SYSTEM_PROMPT_ACT_SMALL authoring + tests
• Phase 2: settings override + context-window coupling

---

Derived from maintainer engineering TODO §1. Happy to implement if assigned.

[Ayush7614]

cc: @esokullu

[esokullu]

@Ayush7614 great points,

The three-tier enum already exists (compact/mid/full), so the structural part is done.

The genuinely new work is the small tier — but I'm skeptical that expanding the prompt helps 7B models. My intuition says more text = more attention dilution, not less hallucination.

Do you already have traces or data showing the expanded prompt reduces errors vs. the compact one on qwen3.6-27b or similar? If yes, share them and I'll merge on the evidence.

If not, I'm building a frozen small-model harness under test/llm/ that will run the same scenarios across compact vs small tiers (measuring error rate / re-download loops / scratchpad adherence). I'd rather we wait for that and evaluate the small prompt against it before shipping 2K extra tokens to every small-model request on a hunch. The harness itself will land first regardless — even if small stays gated behind a flag — so we have the measurement infrastructure either way.

If the data shows small wins, great, we ship it. If it doesn't, we've at least answered the question empirically.

[esokullu]

@Ayush7614 https://github.com/webbrain-one/webbrain/blob/main/docs/webbrain-tool-tiers.xlsx

[esokullu]

I'm taking this over. I wll be committing something soon.

[esokullu]

esokullu#73

will merge only after the benchmark results

[esokullu]

now live.