…iority

Authored by manuelschipper. Adds GLM-4.7 and GLM-5 context lengths (202752)
to model_metadata.py. The key priority fix (prefer OPENAI_API_KEY for
non-OpenRouter endpoints) was already applied in PR #295; merged the Z.ai
mention into the comment.

…NAI_API_KEY in all code paths

Authored by 0xbyt4. Fixes NousResearch#289.

…g API key priority

Authored by manuelschipper. Adds GLM-4.7 and GLM-5 context lengths (202752)
to model_metadata.py. The key priority fix (prefer OPENAI_API_KEY for
non-OpenRouter endpoints) was already applied in PR NousResearch#295; merged the Z.ai
mention into the comment.

…g API key priority

Authored by manuelschipper. Adds GLM-4.7 and GLM-5 context lengths (202752)
to model_metadata.py. The key priority fix (prefer OPENAI_API_KEY for
non-OpenRouter endpoints) was already applied in PR NousResearch#295; merged the Z.ai
mention into the comment.

…NAI_API_KEY in all code paths

Authored by 0xbyt4. Fixes NousResearch#289.

…g API key priority

Authored by manuelschipper. Adds GLM-4.7 and GLM-5 context lengths (202752)
to model_metadata.py. The key priority fix (prefer OPENAI_API_KEY for
non-OpenRouter endpoints) was already applied in PR NousResearch#295; merged the Z.ai
mention into the comment.

…loses #99) (#108)

When the model emits a tool-call-specific finish reason
(``tool_calls`` / ``function_call``) but the normalized
``assistant_message.tool_calls`` reaches the response handler empty
(SDK normalization gap, upstream shape issue, etc.), the previous
behavior:

1. Line 3180 ``if assistant_message.tool_calls:`` evaluates falsy
2. Falls into the ``else: # No tool calls`` branch
3. ``final_response = ""`` (empty content)
4. ``_truly_empty`` becomes True; ``finish_reason="tool_calls"``
   is NOT ``"stop"`` so the #67 structural-empty recovery doesn't
   apply
5. ``_empty_content_retries`` burns 3 retries returning the same
   empty shape
6. Falls through to ``"(empty)"`` terminal

This produces the user-visible "returned no content" + retry loop
documented in hermes-agent#99. Devagentic-side workaround NousResearch#295
strips tools at the request boundary to avoid the model emitting
the problematic finish-reason — once that workaround lifts, this
branch catches any remaining edge cases.

## Fix

In the else branch at conversation_loop.py:3493, before the
empty-retry logic, check:

  _finish_wants_tools = finish_reason in {"tool_calls", "function_call"}
  _tool_calls_absent = not (assistant_message.tool_calls or [])

When both hold (and ``_finish_reason_tools_handled`` flag isn't
set — one-shot semantics like the #67 path), surface synthetic
recovery instead of the retry loop:

- Emit WARNING log with the diagnostic shape (finish_reason,
  model, provider, response_id) so operators see the structural
  mismatch
- Emit user-visible status (``⚠️ Model emitted
  finish_reason=tool_calls but no tool_calls parsed — surfacing
  recovery prompt``)
- Append synthetic assistant message + user-level recovery prompt
  that tells the model the shape was broken + how to retry

Composes with #67's structural-empty recovery (different
finish_reason gates them).

## Tests

- 9 new tests in
  ``tests/agent/test_finish_reason_tools_recovery.py``: positive
  for ``tool_calls`` + ``function_call`` finish reasons with None
  / empty-list tool_calls (3); negative for ``stop`` / ``length``
  / ``content_filter`` finish reasons (3); negative when
  tool_calls populated (1); already-handled one-shot (1);
  source-level patch-landed check (1).
- 28 total green across affected suites (new + #67 terminal +
  empty-response-diagnostic). No regression.

## Composition

Unblocks hermes-agent#100 (R12 worker tool surface for
``run_pipeline`` / ``propose_pipeline``) — that issue's
"Recommendation: Stay Backlog until hermes#99 lands" gate
clears with this PR.

Devagentic#295 (tools-strip at boundary) can be lifted once this
PR + a redeploy is in place; the client-side fix is now the
authoritative path.

…NAI_API_KEY in all code paths

Authored by 0xbyt4. Fixes NousResearch#289.

…g API key priority

Authored by manuelschipper. Adds GLM-4.7 and GLM-5 context lengths (202752)
to model_metadata.py. The key priority fix (prefer OPENAI_API_KEY for
non-OpenRouter endpoints) was already applied in PR NousResearch#295; merged the Z.ai
mention into the comment.