fix(agent): inject OPENAI_MODEL_EXECUTION_GUIDANCE for GLM/Qwen/DeepSeek

[tauros1983]

Summary

Telegram sessions on weaker tool-use-format-trained models (GLM-5.2 in particular) stalled after a single assistant turn: the model emitted its tool intent as a plain-text [TOOL_CALL]...[/TOOL_CALL] markdown block instead of a structured tool_calls JSON array, so the runtime saw tool_calls=None and finished with finish_reason=stop after one turn.

Two Telegram sessions from 2026-06-27 reproduced this (IDs 20260627_144930_7531d0bd and 20260627_145132_3c5abc52) — 3 messages each, 0 tool calls each, then silence.

Root cause

build_system_prompt_parts injected OPENAI_MODEL_EXECUTION_GUIDANCE (the strong tool-persistence + mandatory_tool_use + anti-fabrication block) only when the model name contained gpt, codex, or grok. GLM/Qwen/DeepSeek matched TOOL_USE_ENFORCEMENT_MODELS so they got the lighter TOOL_USE_ENFORCEMENT_GUIDANCE block — but not the execution-discipline block that specifically steers the model away from emitting plans instead of executing. On a weaker tool_use format-trained model that lighter block alone is not enough to stop the model from emitting its tool intent as natural-language markdown.

The block's own comment already noted "body is family-agnostic; the OPENAI_ prefix reflects origin, not exclusivity" — the gate just wasn't widened to match the comment.

Fix

agent/system_prompt.py:257 — expand the gate to the non-Google slice of TOOL_USE_ENFORCEMENT_MODELS (i.e. gpt, codex, grok, glm, qwen, deepseek). Google models are excluded to avoid duplicating the parallel-tool-call steer that GOOGLE_MODEL_OPERATIONAL_GUIDANCE already provides.

19 lines added, 4 removed. Two files:

• agent/system_prompt.py — gate expansion + comment update
• tests/agent/test_system_prompt.py — 8 new regression tests

Tests

8 new regression tests in TestOpenAIExecutionGuidanceInjection:

• test_glm_5_2_receives_openai_execution_guidance (the exact model from the stall)
• test_deepseek_receives_openai_execution_guidance
• test_qwen_receives_openai_execution_guidance
• test_gpt_still_receives_openai_execution_guidance (no regression)
• test_grok_still_receives_openai_execution_guidance (no regression)
• test_anthropic_opus_does_not_receive_openai_execution_guidance (no false positive)
• test_google_gemini_does_not_receive_openai_execution_guidance (no duplicate steer)
• test_disabled_when_enforcement_off

169/169 tests pass in tests/agent/test_prompt_builder.py + tests/agent/test_system_prompt.py. Pre-existing test_anthropic_adapter.py::TestRunOauthSetupToken::test_returns_token_from_credential_files failure is unrelated (MagicMock JSONDecode error on main HEAD without this patch; reproduced with git stash). ruff + ty both clean.

Reproduction evidence

-- sqlite3 ~/.hermes/state.db
SELECT message_count, tool_call_count, end_reason FROM sessions
WHERE id IN ('20260627_144930_7531d0bd','20260627_145132_3c5abc52');
-- 3|0||   (both sessions)

SELECT substr(content, 1, 250) FROM messages
WHERE session_id='20260627_144930_7531d0bd' AND role='assistant';
-- 'Ik ga die 3 commits lezen. Eerst kijken waar ik sta en wat de context is.
--  [TOOL_CALL]
--  run_command
--  command=cd /Users/guidolassally/esmi-dashboard 2>/dev/null && ...'
-- finish_reason='stop', tool_calls=NULL

System prompt size also showed the cause: Telegram sessions received 19,600 chars of system prompt vs 38,666 chars for CLI sessions — the lighter guidance was the only one Telegram got.

Live verification (after patch)

Running build_system_prompt_parts with a mocked agent for each model name now produces:

model	tool_enforce	openai_exec (was)	openai_exec (now)	google_ops
z-ai/glm-5.2	yes	no	yes	no
openai/gpt-5.5	yes	yes	yes	no
xai/grok-4	yes	yes	yes	no
deepseek/deepseek-v4-pro	yes	no	yes	no
google/gemini-2.5-pro	yes	no	no	yes
anthropic/claude-opus-4.8	yes	no	no	no

E2E confirmation requires a fresh Telegram session with GLM-5.2 after merge — that's a downstream verification step for the user, not something the unit tests can cover.

[tauros1983]

Superseded by upstream PR NousResearch#53847. Same branch, same commit, same diff. The fork PR is redundant now that the upstream PR is open.