…vances #89 Direction A) (#98)

Operator-supplied intent override (Option A3 from #97). When
``HERMES_INTENT_OVERRIDE=code`` is set, the system prompt's
``stable`` layer narrows for tool-call-heavy traffic — addresses
#89's prompt-saturation symptom on mid-tier coding models.

## What narrows under code intent

| Block | Action | Why |
|---|---|---|
| SOUL.md | Skip | Largest single contributor; falls back to short DEFAULT_AGENT_IDENTITY floor |
| HERMES_AGENT_HELP_GUIDANCE | Skip | Off-topic for tool-call traffic |
| SKILLS_GUIDANCE | Skip | Per-tool block, off-topic for code |
| KANBAN_GUIDANCE | Skip | Worker-lifecycle, off-topic for code |
| SESSION_SEARCH_GUIDANCE | Skip | Off-topic for code |
| skills_prompt (the big one) | Skip | Biggest contributor when many skills loaded |
| MEMORY_GUIDANCE | **Keep** | Small + sometimes useful even for code |
| TOOL_USE_ENFORCEMENT_GUIDANCE | **Keep** | Critical for tool emission |
| Per-model operational guidance | **Keep** | Model-quality-specific |
| Env / platform hints | **Keep** | Execution-environment essentials |
| nous-subscription + computer-use + alibaba | **Keep** | Operational invariants |
| ``context`` + ``volatile`` layers | **Untouched** | Out of scope per #97 |

Other intents (``confer`` / ``planning`` / ``exploration`` /
``refinement`` / ``generic``) are recognized as valid but pass
through without narrowing in v1 (keeps the door open for per-
intent shape later).

## Intent vocabulary

Matches devagentic#240's ``intent_classifier`` 6-key enum exactly,
so the same operator-side classifier that's wired into devagentic's
R5 dispatch hook can also drive hermes-side prompt narrowing
without a second vocabulary.

## Doctor probe

New ``_check_intent_override_env`` probe surfaces the active
override at ``hermes doctor`` time — silent when unset, check_ok
when valid (with a narrowing-active note for ``code``), check_warn
with the full valid-keys list when typo'd. Mirrors the silent-
when-irrelevant pattern from PR #95 / #96.

## Tests

- 22 new prompt-narrowing tests in
  ``tests/agent/test_system_prompt_intent_override.py``: resolver
  enum + normalization (5), per-section drops under code (7),
  pass-through for non-code intents (5), typo falls back (1),
  byte-count regression (1), default-still-includes counter-case (1),
  case-insensitive (1), runtime-vs-doctor-config sanity (1).
- 6 new doctor-probe tests in
  ``tests/hermes_cli/test_doctor_intent_override_probe.py``:
  silent-when-unset / silent-when-empty / code-ok-with-narrowing-note /
  non-code-valid-pass-through / typo-warn-with-valid-sample /
  case-insensitive.
- 258 total green across affected suites (system-prompt + prompt-
  builder + restore + doctor + provider-env + tools-subset). No
  regression in the existing prompt-shape pins.

## Composition note

Option A1 (port classifier) + A2 (devagentic GraphQL surface) are
deferred per the #97 sequencing — A3 unblocks deployment-specific
narrowing immediately; A1/A2 only matter when dynamic per-turn
classification is needed on the hermes side. The classifier output
on the devagentic side (NousResearch#240) drives R5 dispatch decisions there.

#115) (#116)

Companion to devagentic#315 (initiative preamble). When operator
sets ``HERMES_TOOL_USE_ENFORCEMENT=required``, the chat_completions
transport injects ``tool_choice: "required"`` on every dispatch
where tools are attached — the model-layer enforcement that closes
the gap devagentic#315's soft-signal preamble leaves open.

## Behavior

- Unset / empty / unknown value → default behavior unchanged (no
  ``tool_choice`` injected by hermes)
- ``HERMES_TOOL_USE_ENFORCEMENT=required`` + tools attached →
  ``tool_choice: "required"`` set on the API kwargs
- Tools NOT attached → no injection (sending ``tool_choice=required``
  with empty tools is a 400 on most providers)
- Caller-supplied ``tool_choice`` already on kwargs → no override
  (the dispatcher-tier signal wins; env is a session-tier default)

Per devagentic#203 §1.3 — hermes owns model-call-shape decisions
(per-call enforcement). Devagentic's models.json
``default_tool_choice`` is the dispatcher-tier default; this env is
the session-tier override.

## Where it fires

Both build_kwargs paths in ``chat_completions.py``:
- Legacy fallback path (unregistered providers)
- Provider-profile path (known providers via providers/ registry)

Shared helper ``_maybe_inject_required_tool_choice(api_kwargs,
tools)`` keeps the two sites in sync.

## Doctor probe

New ``_check_tool_use_enforcement_env`` surfaces the active setting
— silent when unset, ``check_ok`` on ``required``, ``check_warn``
with valid-values hint on typos. Mirrors the silent-when-irrelevant
pattern from #95 / #96 / persona-deferred.

## Tests

- 18 new tests in tests/agent/test_tool_use_enforcement.py:
  resolver returns None/required/case-insensitive/unknown (8
  parametrized), injection happy path (1), no-inject-when-unset (1),
  no-inject-when-no-tools (1 covering both None and empty list),
  does-not-clobber-existing-tool_choice (1), no-inject-on-unknown
  (1), doctor silent-when-unset (1), doctor check_ok on required
  (1), doctor check_warn on unknown (1).
- 128 total green across affected suites (new + doctor + provider/
  intent/persona/tools-subset probes). No regression.

## Sequencing per #115 body

The issue says "Land after devagentic#315 Phase 1 has deployed +
been observed. If the preamble alone closes the reliability gap to
operator satisfaction, this issue may not need to ship."

This PR ships the env-knob in opt-in OFF-by-default mode, so:
- Operators can enable it the moment they observe NousResearch#315's preamble
  is insufficient (no further hermes-side dev cycle needed)
- Default behavior unchanged → zero risk to non-client-tier sessions
- Doctor probe surfaces the active state so operators can confirm
  enablement at boot

Saves the round-trip of waiting + then dev'ing once the signal
arrives.

After v0.18.4's tool_call recovery (#124) landed, the next-level
bug surfaced in sandbox field-test: model calls a tool name
hermes' worker didn't register, the invalid_tool_call retry path
fires, but its verbose-only print is invisible in default runs.
Combined with model hallucination ("the file has been created..."
narration on the NEXT turn), the mismatch becomes invisible —
operators see model narration, not the underlying tool-name
mismatch.

## Fix

Upgrade conversation_loop.py:3219's verbose-only print to:

1. ``logger.warning`` with the invented name + count + first 10
   registered names + model + provider for cross-system log
   correlation
2. ``agent._emit_status`` surfacing the mismatch in the user-
   facing stream

Operator immediately sees:
- WHICH name the model invented
- HOW MANY tools the worker has registered
- WHICH tools (sample) ARE registered
- Across which retry of 3

No behavior change — existing invalid_tool_call retry semantics
unchanged. Pure observability boost.

## Tests

- 3 new source-level tests in
  tests/agent/test_loud_invalid_tool_call.py: patch-landed,
  emit_status template includes name + count, warning includes
  model + provider for correlation.
- 20 total green across affected suites — no regression.

## Composition

Same observability family as the #95 / #96 doctor probes. Helps
operators distinguish "hermes ate the tool_call" from "sandbox
toolset doesn't expose what the model is calling".