…loses #86) (#87)

Post-#82/#83/#85 the autowire surfaces ~36 MCP tools from
hermes-internal, pushing fresh workers to 52 total tools and into the
tool-paralysis ceiling (text responses with affirmation pattern, no
tool_call emission for verticals that should be one-shot).

#75 (HERMES_TOOLS_SUBSET) was supposed to let operators narrow the
surface per worker, but it had three gaps for MCP tools:

1. Subset filtering only ran in agent_init.py:838 (built-in tool path)
2. cli.py:9790 (/reload-mcp + auto-reload on config change) re-assigns
   agent.tools without re-applying the filter — regression risk where
   any post-init MCP server reload nukes the subset
3. Even when filtering at agent.tools, the registry still carried the
   unwanted MCP tools, costing schema-conversion + collision-check
   work per tool per boot

Fix: apply the allow-list at MCP tool registration in
tools/mcp_tool.py::_register_server_tools. Excluded tools never enter
the registry, so both initial discovery AND /reload-mcp paths honor
the subset uniformly + the registry stays clean.

Three edits:

1. hermes_cli/tool_subset.py (new) — shared helpers `get_subset_allow()`
   + `is_tool_allowed()`. Single source of truth for env parsing +
   allow-list semantics so the two call sites (agent_init + mcp_tool)
   can't drift in casing/whitespace/empty-vs-missing handling.

2. tools/mcp_tool.py::_register_server_tools — read the subset once
   per server registration (O(1) check per tool, not O(N) env parse),
   apply inside both the per-tool loop AND the utility-tools loop
   (list_resources/read_resource/list_prompts/get_prompt).

3. agent/agent_init.py — refactored the inline #75 filter to use the
   shared helper. Behavior unchanged; dedupes parsing logic.

Subset compares against the **prefixed** MCP name
(`mcp_<server>_<tool>`) — exact match, predictable behavior. Fuzzy /
unprefixed matching is a separate feature request.

Tests (tests/test_mcp_subset_filter.py — 14 cases, all pass):
- get_subset_allow: unset/empty/whitespace-only/single/multi/padded/
  MCP-prefixed parse cases
- is_tool_allowed: None passes through, exact match, exclusion, no
  substring matching, MCP prefixed vs unprefixed (contract test)
- Source-level: _register_server_tools imports the shared helper +
  calls it inside main loop
- Source-level: same call inside the utility-tools loop (regression
  catcher for "fix the main loop, forget utility tools")
- Source-level: agent_init.py uses shared helper, no stale inline
  parser

37 total green (14 new + 20 from #83 + 3 from #85).

Final layer of the post-#67 cascade. After this lands + container
rebuild, workers can be configured with HERMES_TOOLS_SUBSET to a 5-tool
surface that includes their specific MCP needs (e.g.
"doc_view,mcp_hermes-internal_grafted_context_fetch,..."), out of
tool-paralysis range.

#96)

Operators who narrow the tool surface via HERMES_TOOLS_SUBSET can
now confirm at ``hermes doctor`` time exactly which tools the
filter parsed to. Catches two failure modes that previously
required a separate ``hermes mcp list`` diff:

1. Operator typoed a tool name → still in the parsed list (no
   cross-check), but the diff against ``hermes mcp list`` is now
   trivial.
2. Operator forgot the ``mcp_<server>_<tool>`` prefix for MCP
   tools → no entry uses ``mcp_`` prefix but entries look
   structured → info reminder fires.

Silent when env var is unset/empty (silent-when-irrelevant pattern
from the #88/#53/#54 doctor probes). When set, surfaces:

  * check_ok with the count + a sample of names (first 6, then
    ``+N more`` suffix to keep the row readable);
  * check_info reminder when zero entries use the mcp_ prefix but
    some look structured (the most common parse-correctly-but-
    filter-nothing failure mode).

Cross-check against the live MCP registry was considered + rejected
for this PR — it would require spinning up ``create_mcp_server()``
at probe time. Operators can ``hermes mcp list`` separately if they
want the full diff. Filing as opportunistic follow-up if demand
shows up.

## Tests

- 8 new tests in tests/hermes_cli/test_doctor_tools_subset_probe.py:
  silent-when-unset / silent-when-empty / silent-when-whitespace /
  count-and-sample-shown / long-list-truncated / mcp-prefix-reminder
  / no-reminder-when-mcp-present / no-reminder-when-only-simple-bare-
  names.
- 30 total green across affected suites (probe + provider-env-probe
  + mcp_subset_filter). No regression.

…143) (#145)

T3 of the #143 thin-client refactor scope. When the active provider
is ``devagentic-local``, augment ``disabled_toolsets`` with
``"clarify"`` before the ``get_tool_definitions`` call.

## Rationale

Per devagentic#203 §1.3 + the #143 scope: devagentic-side intent
classifier knows when clarification is actually needed and can
surface it via an OpenAI-shaped assistant message. Hermes' modal
TUI clarify-tool was a layered opinion fighting devagentic-side
classification — both rotation's debug evidence + sandbox UX
showed the dual-source as confusing (clarify modal popped even
with --yolo, ignoring devagentic-side intent signals).

This is the smallest of the T1-T3 sequence and the cleanest
revert path — purely a tool-registry adjustment for one provider.

## Behavior

| Setting | Before | After |
|---|---|---|
| provider=devagentic-local, no --enable-toolset | clarify enabled | clarify implicitly disabled |
| provider=devagentic-local, --enable-toolset clarify | clarify enabled | clarify enabled (explicit override) |
| provider=other | unchanged | unchanged |

A boot-line print informs operators of the implicit disable + how
to re-enable for legacy workflows. Composes naturally with the
existing ``HERMES_TOOLS_SUBSET`` narrowing (#75/#87) — disable
happens first, then subset narrows further if set.

## Tests

- 4 source-level tests in
  ``tests/agent/test_t3_clarify_default_out.py``: patch-landed,
  explicit-enable-overrides-implicit-disable, re-enable hint
  visible in print message, strict-equality on provider name
  (no prefix/alias matching to avoid surprise on related
  providers).
- 21 total green across affected suites (T3 + diag-env-gate).

## Composition

Per #143 sequencing (T3 → T1 → T2-gated → T2-default-flip):
- This PR: T3 (clarify default-out)
- Next: T1 (HERMES_DEFER_PERSONA default-flip for devagentic-local)
- Then: T2 (empty-content recovery removal, env-gated then default)
- Later: T4-T6 (tool list / iteration cap / summary fallback)

## Preserved through the refactor

- PR #119 (cascade_exhausted short-circuit) — hermes correctly
  deferring to devagentic; NOT recovery
- PR #122/#125 (raw tool_calls fallback) — pre-recovery wire
  parsing; belt-and-suspenders against future streaming-chunker
  regressions
- PR #131/#136/#138/#141 diagnostics — env-gated via
  HERMES_DIAG_RAW_CAPTURE; no-op when off