experimental.chat.system.transform hook silently discards plugin mutations

[tstanton1-ds]

Bug: experimental.chat.system.transform hook mutations silently discarded by runtime

Observed Behavior

The experimental.chat.system.transform hook fires correctly, and mutations to output.system execute without error within the plugin. However, the Go runtime silently discards all mutations before the system prompt reaches the LLM.

System prompt size remains constant across all hook invocations regardless of what the plugin appends or modifies — confirming the mutations are ignored at the platform level, not failing at the plugin level.

Expected Behavior

Mutations to output.system within the experimental.chat.system.transform hook should be reflected in the system prompt delivered to the LLM, per the type contract in @opencode-ai/plugin.

Reproduction

Two independent plugins were built to verify this behavior:

1. Probe plugin — Minimal plugin that appends a marker string to output.system. The marker never appears in the system prompt delivered to the model.
2. Skill injector plugin — Full implementation that injects skill content into output.system. Content is never delivered to the model.

Both plugins:

• Register correctly and load without error
• Hook fires on every qualifying event
• Mutations execute without throwing
• System prompt size is unchanged after mutation

The type contract (@opencode-ai/plugin v1.2.20) specifies a void return with mutation-based side effects. The plugins follow this contract exactly.

Evidence

• 37 unit tests pass on the plugin side — the plugin correctly mutates the object it receives
• output.system.length logged before and after mutation shows the mutation takes effect in-process
• System prompt measured at the LLM shows no change — the runtime does not propagate the mutations

Environment

• OpenCode version: Latest (as of March 2026)
• Plugin SDK: @opencode-ai/plugin v1.2.20
• Runtime: Go
• Model: Claude (Anthropic)

Impact

The experimental.chat.system.transform hook is unusable for its intended purpose — dynamically modifying system prompts based on session context. This blocks plugin-driven skill injection, context-aware system prompt modification, and any use case that relies on this hook.

Possibly Related

The tool.execute.after hook may exhibit similar behavior (mutations to tool output not propagated). We have not independently verified this but observed references to similar patterns.

[github-actions]

This issue doesn't fully meet our contributing guidelines.

What needs to be fixed:

• No issue template was used. Please use one of the three required templates: Bug Report, Feature Request, or Question.
• The issue appears to be AI-generated (highly polished prose, structured evidence sections, formal impact statement). AI-generated walls of text are not acceptable per our guidelines — please rewrite in your own words with a concise description.

Please edit this issue to address the above within 2 hours, or it will be automatically closed.

---

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

• Feature: Add sessionID to experimental.chat.system.transform hook for tracing integration #6142: References the same experimental.chat.system.transform hook and the same code path (packages/opencode/src/session/llm.ts). While that issue is a feature request (add sessionID), it exposes the same hook behavior and may be related to the root cause here.

If you believe this was flagged incorrectly, please let a maintainer know.

[tstanton1-ds]

Closed and confirmed this is due to interference from another plugin on my side

Thanks OpenCode is the absolute best!