Discussion: https://gist.github.com/PowerCreek/833cda14a6528f031fcc334305e56c63

Problem

After landing G1–G5 + #67 trio + devagentic#218 silo-v2 + #71 lazy-load, the poly-explorer worker spawned successfully but empties across all prompts that exercise MCP tools, even with #71's reduced preamble. Direct API probe with a single tool attached returns a perfect tool_call; hermes attaching all 33 tools per turn produces finish_reason=stop tool_calls=0 content="". Pure tool-paralysis from schema-attention overload.

This is the integration ceiling: the fused stack is architecturally complete but the per-call tool surface is too broad for current models to use without choking.

Root cause (static analysis)

• agent.tools populated once at session boot via agent/agent_init.py:818:

  agent.tools = _ra().get_tool_definitions(enabled_toolsets=…, disabled_toolsets=…, …)

• Same agent.tools attached to EVERY chat-completions request — three call sites at agent/conversation_loop.py:513 / 559 / 3471: tools=agent.tools or None. No per-turn filtering.
• Existing --enable-toolset / --disable-toolset flags only work at TOOLSET (plugin) granularity. Post-G1-G5 stack pulls 33 tools across 6 plugins; toolset disabling is too coarse for surgical narrowing (disabling canvas loses all canvas tools as a unit).

Proposed fix

Add HERMES_TOOLS_SUBSET env var — comma-separated allow-list of tool names. When set, agent.tools is filtered to only those tools at session boot. Composes with existing toolset flags (subset narrows what's already enabled; doesn't override).

Implementation (~15 LOC)

Insert immediately after agent.tools = _ra().get_tool_definitions(…) in agent_init.py:818:

_subset_raw = (os.environ.get("HERMES_TOOLS_SUBSET") or "").strip()
if _subset_raw and agent.tools:
    _wanted = {n.strip() for n in _subset_raw.split(",") if n.strip()}
    if _wanted:
        _before = len(agent.tools)
        agent.tools = [
            t for t in agent.tools
            if (t.get("function") or {}).get("name") in _wanted
        ]
        if not agent.quiet_mode:
            _kept = sorted({(t.get("function") or {}).get("name", "?")
                            for t in agent.tools})
            print(f"🎯 HERMES_TOOLS_SUBSET narrowed tool surface: "
                  f"{_before} → {len(agent.tools)} ({', '.join(_kept)})")

agent.valid_tool_names recomputation immediately below (agent.valid_tool_names = {tool["function"]["name"] for tool in agent.tools}) automatically reflects the filtered set — no extra change needed.

Operator usage

# Polynomial-explorer in observation mode (read-only):
HERMES_TOOLS_SUBSET=grafted_context_fetch,lane_h_list,lane_h_fetch,doc_search,silo_query hermes

# Polynomial-explorer in execution mode (read + confer):
HERMES_TOOLS_SUBSET=grafted_context_fetch,doc_search,doc_write,silo_query,confer_run hermes

# Default profile dev work (no narrowing needed):
hermes  # all 33 tools attached as today

Acceptance

• HERMES_TOOLS_SUBSET=A,B,C env in the hermes process narrows agent.tools to just A/B/C (when those names exist in the underlying registry).
• Empty / unset env preserves current behavior (no filtering).
• Names not present in the registry are silently ignored (plugins add/remove tools at runtime; pre-validation would over-warn).
• One INFO line at session start showing the narrowing effect (33 → 5 (grafted_context_fetch, …)).
• 5-6 unit tests cover: env unset → no change · env set with subset → filtered · env set with names not in registry → silent skip · env set to empty string → no change · composability with enabled_toolsets (toolset filter runs first, env subset narrows further).
• README addition in agent/ or docs/ documenting the env var + when to use it.

Why now

• Bridge to /retry and /undo gateway commands have no effect, /reset silently loses memories NousResearch/hermes-agent#210 R1/R2 (intent classifier → per-turn routing policy). The env var is operator-tunable today; R1/R2 will eventually replace it with dynamic per-turn selection. Code that filters agent.tools is the same hook point R1/R2 will use, just driven by a classifier instead of an env var.
• Unblocks polynomial-explorer immediately — operator picks a 5-7 tool subset, worker can actually do its mission while R1/R2 are designed.
• Composes with G1 preamble doesn't scale past ~10 grafts even with caps — needs carving / lazy-load / summarization #71 (lazy-load preamble — already shipped via Path traversal in skill_view allows reading arbitrary files including API keys NousResearch/hermes-agent#220/feat(mcp): grafted_context_fetch tool — extends devagentic-lane-h plugin (#71 PR2) #72/refactor(preamble): G1 renders graft INDEX only — workers fetch on demand (closes #71 PR3) #73): G1 preamble doesn't scale past ~10 grafts even with caps — needs carving / lazy-load / summarization #71 cuts graft-content from the request; HERMES_TOOLS_SUBSET cuts tool-schema from the request. Together they bring per-turn context well below model attention saturation.
• Composes with fix(conversation): structural-empty terminal recovery for stop+no-tool_calls+tools (closes #67) #69 (structural-empty synthetic recovery — already shipped): if a worker still empties despite narrowing, the synthetic recovery catches it. Defense in depth.

Out of scope

• Dynamic per-turn narrowing — that's /retry and /undo gateway commands have no effect, /reset silently loses memories NousResearch/hermes-agent#210 R1/R2 territory (intent classifier → routing policy). HERMES_TOOLS_SUBSET is a static-per-session knob; the long-term fix is dynamic.
• Per-tool capability metadata (e.g. "this tool is read-only, attach freely") — also belongs with /retry and /undo gateway commands have no effect, /reset silently loses memories NousResearch/hermes-agent#210's classifier work.
• Auto-discovery of "ideal" subset for a given vertical — operators tune by hand for now.

Severity

Quick unblock for polynomial-explorer + any future >5-tool vertical session. Small surgical change. Filing per orchestrator's direct ask.