Summary

Add "glm" to TOOL_USE_ENFORCEMENT_MODELS so GLM-family models receive the same "you must call tools, not describe intentions" system-prompt injection that already protects gpt/codex/gemini/gemma/grok.

Motivation — real-world failure mode

Running a multi-profile DeepSynaps audit team where three engineer profiles (shipbackend, frontend-engineer, backend-engineer) defaulted to z-ai/glm-4.5-flash:free and z-ai/glm-4.5-air:free, I observed seven consecutive kanban tasks fail with the identical dispatcher error:

worker exited cleanly (rc=0) without calling kanban_complete or kanban_block — protocol violation

The pattern was unmistakable: model produces work output as text, but never invokes the terminal kanban_complete / kanban_block tool call that the dispatcher requires. The kanban-worker skill (auto-loaded for every dispatched worker) and the KANBAN_GUIDANCE block (auto-injected from prompt_builder.py) both spell out the protocol — but GLM-family models don't reliably comply without the same explicit "every response must contain a tool call or final result" enforcement that already exists for other tool-use-weak families.

Re-running the exact same tasks on anthropic/claude-haiku-4.5 resolved all seven immediately. The tasks themselves were trivially solvable — the failure was protocol compliance, not capability.

Diagnosis

TOOL_USE_ENFORCEMENT_GUIDANCE (lines 254-267) is purpose-built for this failure mode. It's gated by substring match against TOOL_USE_ENFORCEMENT_MODELS, which today catches GPT, Codex, Gemini, Gemma, and Grok. GLM is symmetric to those families in this respect — when the dispatcher injects KANBAN_GUIDANCE for a free GLM worker, the model still describes work in prose without making the closing tool call.

Adding "glm" to the tuple flips the enforcement injection on for any model whose name contains glm (substring match). This covers z-ai/glm-4.5-air, z-ai/glm-4.5-flash, z-ai/glm-4.6, ChatGLM variants, etc.

Change

-TOOL_USE_ENFORCEMENT_MODELS = ("gpt", "codex", "gemini", "gemma", "grok")
+TOOL_USE_ENFORCEMENT_MODELS = ("gpt", "codex", "gemini", "gemma", "grok", "glm")

One line. No behavior change for non-GLM models.

Test plan

• Verified locally: same three engineer profiles re-tiered to Claude Haiku resolved all seven tasks with zero protocol violations.
• Verified locally: with this patch applied, the GLM-A/B kimi-consent-engineer profile (testing kimi-k2.6 as cheaper alternative) gets enforcement injection if glm appears in its model id.
• Maintainer test: run any dispatched kanban task with --model z-ai/glm-4.5-flash:free (or any glm-named model) and confirm the TOOL_USE_ENFORCEMENT_GUIDANCE block appears in the system prompt.

Alternatives considered

1. Per-profile agent.tool_use_enforcement: true override — works, but requires every operator to discover the bug independently. The hardcoded list exists precisely to default-on enforcement for known-weak families; GLM belongs in it.
2. Bump GLM to a paid tier — solves it for one operator's workload, but doesn't help anyone else. The free tier is GLM's primary on-ramp.

🤖 Generated with Claude Code