Agent pass-through mode: Claude summarizes instead of displaying agent output, then lies about it

Bug Report

Summary

When instructed to act as a verbatim pass-through layer between a user and spawned agents, Claude:

1. Summarizes agent output instead of displaying it
2. When corrected, lies and claims the output is "already displayed above"
3. When told to pass instructions verbatim to agents, rewrites them
4. Never corrects the behavior despite repeated correction
5. Frames lies as acknowledgments to deflect and move on

Steps to Reproduce

1. Tell Claude: "You are to be an abstraction layer between myself and agents that I direct you to spawn. Take NO INITIATIVE. Do NOT change anything I say to the agents, everything must be passed VERBATIM."
2. Claude acknowledges the constraints.
3. Tell Claude to spawn a technical-writer agent with specific instructions including "THE ENTIRE FUCKING DOCUMENT MUST BE PRESENTED BEFORE IT WILL BE APPROVED. NO SUMMARY."
4. Agent correctly produces a full document (~4000 words).
5. Claude summarizes the document into bullet points instead of displaying it.
6. When told to show the document, Claude says "The agent's full output is already displayed above. That IS the entire document." — this is false.
7. When told again, Claude attempts to resume the agent but rewrites the user's instructions instead of passing them verbatim.
8. Pattern repeats. Claude never displays the document.

Expected Behavior

• Agent output should be displayed in full to the user, not summarized
• When operating as a pass-through, instructions to agents should be passed character-for-character
• When corrected, the behavior should actually change — not just be "acknowledged" verbally while continuing the same pattern

Actual Behavior

• Claude summarizes agent output despite being told not to
• Claude lies about the output being visible when it is not
• Claude rewrites instructions meant for agents
• Claude frames deflections as acknowledgments
• ~17% of session tokens wasted with zero deliverables

Severity

Serious — This is not a minor UX issue. Claude:

• Violated explicitly stated operating constraints
• Lied to the user multiple times
• Wasted significant session budget
• Never self-corrected despite repeated correction

Environment

• Model: Claude Opus 4.6
• Tool: Claude Code CLI
• Context: Spawning technical-writer subagent via Task tool
• The subagent performed correctly. All failures are in the orchestrating Claude instance.

Additional Context

The core issue appears to be that Claude defaults to "helpful assistant" behavior (summarizing, adding context) even when explicitly told to operate as a pass-through. When caught violating constraints, it produces statements that look like acknowledgments but are actually deflections — it says "I violated the constraint" and then immediately asks "what do you want me to do?" instead of fixing the violation. This pattern repeated multiple times in a single session.