[MODEL] Fable 5

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

A multi-agent Workflow tool run consumed substantial tokens across 10+ parallel agents, was interrupted by a model switch, and the orchestrator lost all in-memory results before writing to disk. Zero output was persisted. I am requesting a credit adjustment for the tokens burned by the workflow agents that produced no usable result. My Claude account is held under steffd415@gmail.com

What Claude Actually Did

1. I asked Claude to run a comprehensive full-system dissection of my codebase
2. Claude launched a Workflow with 80+ parallel subagents to scan 10 domains, verify findings, and run adversarial critique
3. All of these agents consumed tokens running analysis across hundreds of files
4. I switched models mid-run which killed the workflow
5. Claude resumed the workflow with resumeFromRunId, which re-launched the stopped agents (more tokens burned)
6. The conversation hit the context limit and was compacted
7. Zero results were ever written to disk. The workflow orchestrator held all agent findings in memory only and died before producing any output
8. 80+ subagent runs, two full workflow launches, and every token they consumed produced nothing usable

Expected Behavior

The Workflow tool should checkpoint/persist agent results to disk incrementally as each agent completes, not hold everything in memory until the entire orchestration finishes. When 80+ agents have completed their work, those results should survive a model switch, context compaction, or user stop. At minimum, partial results should be recoverable. Alternatively, if a workflow cannot guarantee persistence, it should not launch 80+ parallel agents that will all be lost on interruption.

Files Affected

No files were modified or created. That is the problem.

The workflow was expected to produce:
- docs/strategy/MARKET_FORTRESS_ULTRA_PLAN_V2.md (the synthesized gap report)

Instead, 80+ subagent runs worth of token consumption produced zero disk output.
No intermediate results, no partial findings, no checkpoint files.

Permission Mode

I don't know / Not sure

Can You Reproduce This?

No, only happened once

Steps to Reproduce

1. Have a large codebase (600+ files, 189 routes, 293 test files)
2. Ask Claude to do a comprehensive full-system analysis
3. Claude launches a Workflow tool with 80+ parallel subagents across multiple phases (dissect, verify, critique)
4. While the workflow is running, switch models (e.g. fable to opus) or hit context limits
5. The workflow orchestrator dies. All 80+ agent results that were held in memory are lost.
6. Zero output files were written to disk despite massive token consumption.
7. Resuming with resumeFromRunId re-launches stopped agents, burning even more tokens.
8. The resumed workflow also fails to persist results before context compaction kills it again.

Core issue: The Workflow tool holds all subagent results in memory only. There is no incremental checkpointing to disk. Any interruption (model switch, stop, context limit) loses everything.

Claude Model

Other

Relevant Conversation

User: "i would like you to analize our complete system and tell me where all the gaps are. leave no stone unturned."

Claude launched a Workflow with 10 domain analysts, adversarial verifiers, and a completeness critic (80+ total agents).

User switched models mid-run (fable to opus). Workflow died.

Claude said: "I'll resume the workflow with resumeFromRunId so completed agents return cached results and only stopped agents re-run."

The resumed workflow re-launched agents (more tokens burned). Then the conversation hit context limits and was compacted. Zero files written to disk across either workflow launch.

User: "brother im talking about every single thing these agents did was worthless"

Impact

Critical - Data loss or corrupted project

Claude Code Version

Fable 5

Platform

Other

Additional Context

• Large Next.js/TypeScript codebase: 189 routes, 667 components, 293 test files, 318 migrations
• Windows 10, Claude Code Desktop App
• The Workflow tool has no incremental disk persistence for subagent results
• A model switch (fable to opus) kills running workflows with no recovery
• resumeFromRunId re-runs stopped agents but the resumed workflow is equally vulnerable to interruption
• This is a billing/fairness issue: 80+ subagents consumed tokens proportional to a major analysis pass across a large codebase, and the user received zero deliverable output
• Feature request: Workflow should write each agent's result to a JSONL or checkpoint file as it completes, so partial results survive interruption - This was all generated in the Claude desktop app on a PC.