Summary

ContextEngine.assemble should fire per LLM call, not per user prompt

Current behavior

User prompt
→ assembleAttemptContextEngine() ← fires ONCE
→ activeSession.prompt() ← agentic loop
LLM call 1 ← ✅ assemble ran before this
→ tool result
→ LLM call 2 ← ❌ no assemble before this
→ tool result
→ LLM call 3 ← ❌ no assemble before this
→ final response

In attempt.ts:1904, assembleAttemptContextEngine() is called before activeSession.prompt(), which contains the agentic loop. Once the loop starts executing multiple LLM calls, assemble is never invoked again.

Why installContextEngineLoopHook is insufficient

transformContext's call frequency does not guarantee invocation before every LLM call — it fires at specific internal transition points.

Related

• PR feat(plugins): add transform_context hook for per-LLM-call message re… #57585 (closed)

Problem to solve

ContextEngine.assemble() currently fires once per user prompt (per attempt), not per LLM call within the agentic loop. Context engines — especially those doing context compression — cannot make informed decisions about the message array before subsequent LLM calls in a multi-turn tool loop.

Proposed solution

Add a transform_context hook that fires before each LLM call within the agentic loop. Two approaches:

1. New transform_context hook type — conservative, doesn't change existing assemble semantics
2. Make assemble fire at each LLM call boundary — simpler but may have backward-compatibility implications
   Option 1 is recommended as the safer path.

Alternatives considered

No response

Impact

Context overflow on long agentic loops, stale context in subsequent LLM calls, inability to implement smart trimming strategies.

Evidence/examples

No response

Additional information

No response