[BUG]

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

The built-in /deep-research skill (invoked via the Workflow tool) spawned 112 subagents that all failed
identically with:

agent({schema}): subagent completed without calling StructuredOutput (after 2 in-conversation nudges)

The workflow burned 4,299,207 tokens across 338 tool uses over 264 seconds and returned zero usable result.
Every parallel[N] slot failed the same way, retry after retry.

Because I'm on Claude Max 5x, this consumed roughly 85% of my current 5-hour session budget with no output to show
for it. The failure was silent until completion — there is no early-fail circuit when the first N
structured-output calls fail, so the workflow drains its full budget before erroring out.

Task ID: wgu23qy1f
Run ID: wf_d99364f0-f2b

What Should Happen?

Either:

1. Every subagent invoked with a schema: option should successfully call the StructuredOutput tool, OR
2. The Workflow harness should fast-fail after the first 3–5 subagent completed without calling StructuredOutput

errors (assume harness-level bug, stop spawning more agents), AND

1. Rate-limit / session quota consumed by a workflow that returns zero successful agent() results should be

creditable via support.

Users should also get progress signals during long-running workflows so they can TaskStop before full-budget
drain when things are going wrong.

Error Messages/Logs

Dynamic workflow "Deep research harness — fan-out web searches, fetch sources, adversarially verify claims,
  synthesize a cited report." failed: Error: agent({schema}): subagent completed without calling StructuredOutput
  (after 2 in-conversation nudges)
      at R (B:/~BUN/root/src/entrypoints/cli.js:3623:2694)
      at processTicksAndRejections (native:7:39)

  Failures (pattern repeated across all 5 phases):
  parallel[0] failed: agent({schema}): subagent completed without calling StructuredOutput (after 2 in-conversation
  nudges)
  parallel[1] failed: agent({schema}): subagent completed without calling StructuredOutput (after 2 in-conversation
  nudges)
  parallel[2] failed: agent({schema}): subagent completed without calling StructuredOutput (after 2 in-conversation
  nudges)
  ... (31+ identical rounds)

  Usage:
    agent_count:   112
    total_tokens:  4,299,207
    tool_uses:     338
    duration_ms:   264,385

  Task ID: wgu23qy1f
  Run ID:  wf_d99364f0-f2b

Steps to Reproduce

Steps to Reproduce

1. Start Claude Code on Windows with Opus 4.7 (claude --model claude-opus-4-7).
2. Invoke the built-in deep-research skill with any substantive research question, e.g.:

/deep-research How does a solo operator get maximum leverage out of Claude Code CLI in 2026?

1. The skill invokes Workflow({name: "deep-research", args: "<question>"}) internally.
2. The workflow proceeds through phases: Scope → Search → Fetch → Verify → Synthesize.
3. Every agent() call with a schema: option fails after 2 retries with:

agent({schema}): subagent completed without calling StructuredOutput

1. The workflow continues spawning agents (never fast-fails) until ~4 minutes and 100+ failed agents in.
2. Workflow errors out. No research output is produced. Session token budget is fully consumed.

Claude Model

Opus

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.153 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

### Impact / Severity

For Claude Max and Pro users, a single failed run of this workflow can consume the majority of a 5-hour rate-limit
window. Because the failure mode returns nothing usable, users have no way to know the workflow is broken before
invoking it, and no way to abort mid-run before their budget is gone.

### Suggested fixes (in priority order)

1. Root cause: Investigate why subagents ignore the StructuredOutput tool-call instruction. Possible causes:

• System-prompt truncation when the parent has many MCP servers loaded (this user has ~28 MCP servers

configured, contributing significant tool-descriptor overhead)

• Prompt-cache poisoning across schema-mode agent calls
• Tool definition not reaching the subagent's system prompt
• Model-version mismatch between the parent Opus 4.7 and the subagent runtime

1. Circuit breaker: If the first N (e.g., 5) agent({schema}) calls fail with `subagent completed without

calling StructuredOutput`, abort the workflow rather than continuing to spawn more agents. This is a clear
harness-level failure signal, not a per-agent flake.

1. Progress visibility: Emit a <task-notification> when structured-output failure rate exceeds a threshold

so users can TaskStop before full budget drain. Currently /workflows shows live progress but the token counter
runs silently.

1. Automatic credits: Consider automatic session-quota refund for Max/Pro users when a workflow spawns >10

agents and returns zero successful agent() results. The failure signal is unambiguous and detectable
server-side.

### Reproduction environment

• Windows 11 Home 10.0.26200
• PowerShell 5.1
• Claude Max 5x plan (default_claude_max_5x)
• ~28 MCP servers configured (github, playwright, cloudflare, discord, r0ice-server2 SSH bash, etc.)
• No custom modifications to the built-in deep-research workflow script
• Question passed was substantive but well-formed (multi-paragraph research prompt about Claude Code leverage

tactics)