[MODEL]

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

Claude modified files I didn't ask it to modify

What You Asked Claude to Do

Claude was forgetful, slow, and repeatedly failed to execute the task it had already understood. It spent roughly 36 hours re-investigating state, creating overlapping specs and worktrees, chasing packet-ID collisions, and checking unrelated dispatches instead of delivering the requested Live Activity. It acknowledged the two central gaps—the missing Unified Claude row and missing last-6-hours segment—but left them as unapproved planning documents rather than starting the build.

The routing behavior was especially costly: although I explicitly asked for headless x5 work because x20 was near its weekly limit, the interactive ORCH session continued doing substantial work on x20. Its attempted x5 dispatch failed, and it did not recover deterministically or fall back to Codex. The orchestration ledger was also confusing, showing work as “running” even when notes said it had finished.
Claude additionally tried to rationalize an obviously incomplete result as constrained by the phone layout instead of matching the approved mockup as closely as the platform permits. It produced a sparse three-row pill without the Unified row, the 6-hour data, the three-part bars, or the fuller information hierarchy. Codex implemented the real data path across Token Monitor, the runner, and iOS, then rendered and verified the four-row design on an actual simulator lock screen in one working session.

The main improvements needed are: retain the user’s explicit visual acceptance criteria; dispatch immediately when told to use headless capacity; recover from a failed dispatch without reverting to expensive interactive work; keep orchestration status truthful; stop multiplying specs when implementation is authorized; and treat “deploy and verify” as the completion condition, not “write another packet.”

What Claude Actually Did

Didn't complete my project. Took 36 hours to get to this.

<img width="534" height="221" alt="Image" src="https://github.com/user-attachments/assets/f2e1955d-f89b-44d8-a7cb-1086871fa748" />

Expected Behavior

<img width="433" height="272" alt="Image" src="https://github.com/user-attachments/assets/96a9dfcc-838a-4083-b19c-46a1c98d93f8" />

Codex built in 10 minutes.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

No, only happened once

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.269 (Claude Code)

Platform

Anthropic API

Additional Context

_No response_