…#124) (#125)

PR #122 / #121 added raw-response fallback for mistral-shaped
tool_calls (type field absent → SDK strips entry → empty
``msg.tool_calls``). The fallback was gated on
``finish_reason in {"tool_calls", "function_call"}`` — assuming the
SDK preserved finish_reason while dropping the tool_call entry.

Field-tested on duplex sandbox container: the SDK actually
normalizes finish_reason to ``stop`` when it strips the type-less
tool_call. The recovery NEVER FIRED on the actual mistral shape;
the structural-empty recovery (#69 / #67) fired instead and
surfaced the noisy "Your previous response was empty" message.

## Fix

Drop the finish_reason gate. Whenever the SDK gives empty
``msg.tool_calls`` but the raw response shape carries tool_calls,
recover them — wire-level evidence is authoritative. Additionally,
rewrite the normalized ``finish_reason`` to ``"tool_calls"`` so
downstream consumers (``conversation_loop.py:3180`` tool branch,
the structural-empty + finish_reason-tools guards) see a
consistent state.

## Tests

- Updated 1 test in
  ``tests/agent/test_tool_call_type_default.py``:
  ``test_normalize_recovers_when_sdk_normalized_finish_reason_to_stop``
  (previously asserted recovery DIDN'T fire on stop, now asserts
  it DOES — matches field-observed SDK behavior + the bug
  symptom). Also asserts finish_reason gets rewritten to
  ``tool_calls`` for downstream consistency.
- 49 total green across affected suites — no regression on the
  SDK-happy path or the #67/#99/#108/#118 recovery families.

## Composition

- #121 / PR #122 — original fix, gated finish_reason too narrowly
- **this PR** — drops the gate; recovers whenever raw has tool_calls

…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

…MPTY_RECOVERY (refs #143) (#147)

T2 of the #143 thin-client refactor scope. Devagentic-side cascade
(NousResearch#324) + runaway detector (NousResearch#345-348) + exec-terminus (NousResearch#349-354)
now cover the empty-content recovery layer with full intent /
role / dispatch-trace context. Hermes-side recovery layered on top
caused 3×4 dispatch stacking (#118) + mode confusion + invisible
swallow points (#133 debug funnel).

This patch default-flips: the four hermes-side recovery paths are
short-circuited unless ``HERMES_LEGACY_EMPTY_RECOVERY`` is set to
a truthy value. Legacy users opt-in to keep the pre-T2 behavior.

## Gated paths (all skipped when env unset)

| Path | Source | Loc |
|---|---|---|
| ``_finish_wants_tools`` synthetic recovery | PR #108 / #99 | line ~3618 |
| ``_post_tool_empty_retried`` nudge | pre-existing NousResearch#9400-class | line ~3749 |
| ``_structural_empty`` synthetic recovery | PR #69 / #67 | line ~3880 |
| 3-retry empty-content loop | pre-#67 | line ~3927 |

When env unset (default), empty responses fall through to either
the fallback-chain provider switch (if configured) or the clean
``(empty)`` terminal with ``_empty_terminal_sentinel=True``.

## Preserved through T2 (verified by test_t2_legacy_empty_recovery_gate.py)

- **PR #119 cascade_exhausted short-circuit** — hermes correctly
  deferring to devagentic's sentinel; NOT recovery. Source-level
  test asserts the ``if _cascade_err:`` block is NOT prefixed by
  ``_legacy_recovery_on``.
- **PR #122/#125 raw tool_calls fallback** — pre-recovery wire
  parsing in transports/chat_completions.py. Source-level test
  asserts the helper name doesn't appear in conversation_loop
  (lives elsewhere; untouched).
- **PR #131/#136/#138/#141 diagnostics** — env-gated via
  ``HERMES_DIAG_RAW_CAPTURE`` (#140), independent of this env.

## Operator deploy

Default behavior changes: empty responses surface cleanly (no
synthetic re-prompt). To preserve pre-T2 behavior:

```bash
export HERMES_LEGACY_EMPTY_RECOVERY=1
```

The legacy escape hatch is intended as a temporary safety net
while operators validate the thin-client architecture. Once the
devagentic-side cascade is universally deployed + observed to
cover all empty-content cases, the legacy gate can be removed in
a follow-up (the env var stays as a no-op for backward compat).

## Tests

- 23 new tests in
  ``tests/agent/test_t2_legacy_empty_recovery_gate.py``:
  resolver default-false / empty-false / 6 truthy / 6 falsy-or-
  unknown; source-level gate assertions for each of the four
  recovery branches; preserved-path assertions for cascade_exhausted
  + raw tool_calls fallback location; default-off + opt-in
  resolver round-trip.
- 163 total green across affected suites (T2 + existing empty-
  terminal mirror tests + finish_reason_tools_recovery +
  cascade_exhausted + internal_marker_stripping +
  tool_call_type_default + tool_use_enforcement + T3 + T1 persona
  + doctor persona probe).

## Composition

Per #143 sequencing:
- T3 / PR #145 (clarify default-out) — merged ✓
- T1 / PR #146 (persona default-flip) — merged ✓
- **T2 / this PR (empty-content recovery removal, env-gated)**
- Later: T4-T6 (tool list / iteration cap / summary fallback)

The legacy escape hatch design choice (vs. full removal) lets
operators roll back per-deployment if a previously-recovered edge
case surfaces in the field. Diagnostic from #140
(HERMES_DIAG_RAW_CAPTURE) remains the observability tool to spot
any uncovered case.