[BUG] GLM-5 model pollutes context window with broken thinking output

[tayroneoliveira2023]

Description

When using the GLM-5 model (Z.AI), OpenCode's context window gets polluted with broken/malformed thinking output. The terminal displays ? characters scattered throughout the interface, the rendered content becomes garbled, and the TUI layout breaks completely. The context window shows raw thinking tokens that should be hidden or properly parsed.

The GLM 4.7 model works correctly without any of these rendering issues under the same conditions.

Steps to Reproduce

1. Open OpenCode on Windows (cmd or PowerShell)
2. Select the GLM-5 (Z.AI Coding Plan) model
3. Run any task that triggers the model's thinking/reasoning mode (e.g., a build or code generation task)
4. Observe the terminal output becoming corrupted

Expected Behavior

The context window should display clean, properly formatted output — thinking tokens should be hidden or correctly rendered, just like the GLM 4.7 model does.

Actual Behavior

• Multiple ? characters appear throughout the terminal (top, bottom, status bar)
• The thinking/reasoning output leaks into the visible context window in a broken/malformed state
• The TUI layout breaks — code snippets, markdown, and status indicators become garbled
• The status bar at the bottom shows ???■■■■■ instead of proper indicators
• The overall context window becomes unusable and filled with noise

Specific observations from the screenshot

• The title bar and content area show ? characters on the left margin where they shouldn't be
• Code content (Java — EvidenciaRunner.java) is partially rendered but interspersed with broken formatting
• The bottom status bar shows: ? Build · glm-5 followed by more ? lines, then ? Build GLM-5 Z.AI Coding Plan
• The very bottom shows ???■■■■■ esc interrupt — the progress/status indicators are corrupted

Comparison

Model	Behavior
GLM 4.7	Clean output, no artifacts, thinking is properly handled
GLM-5	Broken thinking tokens leak into context window, ? characters everywhere, TUI breaks

Environment

• OS: Windows 10/11
• Terminal: cmd / PowerShell
• OpenCode version: Latest
• Model: GLM-5 (Z.AI Coding Plan)

Screenshot

---

• I have verified that this issue has not already been requested
• I have read the contributing guidelines

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Last token cut off in thinking (windows) #11194: Last token cut off in thinking (windows) — similar symptom of thinking tokens not rendering properly on Windows
• Error "reasoning part reasoning-0 not found" from tool calls using GLM 4.5 Air #6486: Error "reasoning part reasoning-0 not found" from tool calls using GLM 4.5 Air — related GLM reasoning/thinking token handling issue

If your issue is distinct from these, feel free to keep it open and provide any additional context that differentiates it.

[tayroneoliveira2023]

Related issues and root cause

After investigating, this issue appears to be directly related to the known problem with <think>/<thinking> tag handling in OpenCode:

• Support parsing <think> and <thinking> tags as reasoning blocks #11439 — Support parsing <think> and <thinking> tags as reasoning blocks (main tracking issue)
• Thinking models: raw <think> tag content displayed instead of being collapsed #15380 — Thinking models: raw <think> tag content displayed instead of being collapsed (closed as duplicate of Support parsing <think> and <thinking> tags as reasoning blocks #11439)
• thinking stop when use GLM-5 #15782 — thinking stop when use GLM-5
• GLM 4.7 on Zai coding plan puts tool calls inside the thinking/reasoning tag. #6708 — GLM 4.7 puts tool calls inside thinking/reasoning tags
• [bug] GLM 5 keeps screwing up the json parsing of read tool calling #13982 — GLM-5 keeps screwing up JSON parsing of read tool calling

Root cause

The extractReasoningMiddleware was intentionally removed in PR #11270 (v1.1.46+), which means <think> tags are now rendered as plain text instead of being collapsed into reasoning blocks. This causes the context window pollution seen in the screenshot — the GLM-5 thinking output leaks into the visible UI as garbled ? characters and broken formatting.

GLM 4.7 is unaffected because it handles reasoning differently and doesn't produce the same volume of raw <think> tag output.

Potential fix

PR #15382 by @erwinh22 implements the rendering-layer approach: tags are preserved in storage for multi-turn LLM context (respecting the decision in PR #11270), but parsed and rendered as reasoning blocks at display time in TUI, Web UI, and CLI. This PR appears to not have been merged yet.