Follow-up scoping issue for #89 Direction A. devagentic#240 (R1 classifier) has landed, so the upstream primitive is ready — this issue scopes the hermes-side narrowing work itself.

Background

#89 documented that the default hermes system prompt overwhelms tool-emission on mid-tier coding models (coding-groq, coding-gpt54). Direction A proposed per-intent narrowing — strip identity/SOUL/capabilities/behavior fragments from the system prompt when the classified intent is a tool-call-heavy case where the affected models hit saturation.

devagentic#240 ships a 6-key heuristic classifier ({code, confer, planning, exploration, refinement, generic}) — already merged + deployed. The classifier output is wired into devagentic's R5 dispatch hook but not surfaced to the hermes-side prompt builder. Direction A needs the same classification on the hermes side.

Composition options

Option A1 — port the classifier

Duplicate the heuristic into hermes_cli/intent_classifier.py mirroring devagentic's intent_classifier.py. Cheapest engineering; risks drift between the two copies.

Option A2 — devagentic-side surface

Expose classifyIntent(messages, role) as a devagentic GraphQL query; hermes calls it before assembling the system prompt. Single source of truth; adds a network round-trip per llm_call (latency on the hot path).

Option A3 — operator-supplied intent

Add HERMES_INTENT_OVERRIDE env var the operator sets per worker (e.g., HERMES_INTENT_OVERRIDE=code for a coding-only worker). No classifier on hermes side; narrowing decisions are static per deployment. Cheapest, least general; useful for the polynomial-explorer-style verticals where intent is known a priori.

Recommended order: ship A3 first (single env-knob, no classifier coupling, immediately unblocks deployment-specific narrowing). Defer A1/A2 until a use case shows up that needs dynamic per-turn classification on the hermes side.

Narrowing target — agent/system_prompt.py

The system prompt is assembled into 3 layers (see the module docstring at agent/system_prompt.py:10-20):

Specific candidates within stable to narrow when intent=code:

• SOUL.md / DEFAULT_AGENT_IDENTITY block (largest single contributor)
• HERMES_AGENT_HELP_GUIDANCE
• SKILLS_GUIDANCE (large if many skills loaded)
• KANBAN_GUIDANCE / SESSION_SEARCH_GUIDANCE (off-topic for code-only)
• MEMORY_GUIDANCE (small; could keep)
• Keep: TOOL_USE_ENFORCEMENT_GUIDANCE (critical for tool-emission)
• Keep: per-model operational guidance (e.g., GOOGLE_MODEL_OPERATIONAL_GUIDANCE)

Concrete first-PR scope (Option A3)

1. New env var HERMES_INTENT_OVERRIDE — operator-side, valid values per devagentic#240's enum: code / confer / planning / exploration / refinement / generic.
2. build_system_prompt_parts(agent, system_message) opt-in narrowing: when env is set to a known intent, swap to a narrowed stable layer (drop SOUL.md + SKILLS_GUIDANCE + KANBAN + SESSION_SEARCH + MEMORY guidance; keep identity-as-fallback + TOOL_USE_ENFORCEMENT + per-model operational guidance).
3. Doctor probe: surface the active intent override (silent when unset, check_ok when set+valid, check_warn when set+unknown — mirroring my doctor: validate HERMES_DEFAULT_PROVIDER + HERMES_INFERENCE_PROVIDER env vars #95/doctor: surface active HERMES_TOOLS_SUBSET at boot (#75/#87 follow-up) #96 doctor-probe pattern).
4. Composition test: with HERMES_INTENT_OVERRIDE=code, the system prompt is < N tokens (set the bar based on baseline).

Acceptance

• A3 ships; deployment-specific narrowing works (operator sets env, system prompt narrows).
• coding-groq / coding-gpt54 empirically emit tool calls with the narrowed prompt (the Research: default Hermes system prompt overwhelms tool-emission on mid-tier coding models #89 symptom).
• Doctor probe confirms the active override.
• A1/A2 filed as separate follow-ups when dynamic classification is needed.

Out of scope

• Per-turn classification (defer to A1/A2 when needed).
• Tool-surface narrowing (already handled by HERMES_TOOLS_SUBSET feat(env): HERMES_TOOLS_SUBSET — operator-side tool surface narrowing (closes #74) #75/feat(mcp): extend HERMES_TOOLS_SUBSET to filter MCP tools (closes #86) #87).
• Per-vertical preamble interplay (G1 INDEX-only already bounds that contributor; refactor(preamble): G1 renders graft INDEX only — workers fetch on demand (closes #71 PR3) #73).

References

• Research: default Hermes system prompt overwhelms tool-emission on mid-tier coding models #89 (parent, prompt-saturation research)
• devagentic#237 / PR Autonomous Skill-Management Ecosystem (AI for AI Framework NousResearch/hermes-agent#240 (R1 classifier on devagentic side)
• gist 8f238424 directive 2c (R1 classifier rollout)
• gist 8f238424 comment 6166859 (cross-team cascade-glue anchor)