[Bug] Warp terminal AI layer causes agent cognitive degradation and parallel agent isolation failures

Problem

When running Claude Code inside Warp terminal (free tier), two distinct failures occur
that do NOT reproduce in plain terminals (Alacritty, kitty, native Terminal):

1. Parallel agents violate file boundaries — subagents edit files owned by

other parallel agents and race to commit, ignoring worktree/directory isolation

1. Severe cognitive degradation ("降智") — the model performs nonsensical actions

such as creating frontend source code inside docs/architecture_decisions/,
then rationalizing it by claiming it intends to migrate README files into web/.
The same model + same CLAUDE.md on a clean terminal does not exhibit this.

Root cause hypothesis

Warp's architecture introduces multiple I/O interception layers on the same PTY stream
that Claude Code reads:

• Shell integration hooks: DCS escape sequences injected into PTY for command

block segmentation (\eP$f{"hook": "SourcedRcFileForWarp", ...}\x9c)

• Proactive AI: On the free tier, this cannot be disabled. It parses all stdout

in real-time, potentially injecting AI suggestion metadata back into the stream

• Command block parser: Real-time ANSI parsing that segments output — may interleave

or truncate tool results from concurrent subagent subprocesses

The net effect: Claude Code's tool results and system prompt context get polluted with
Warp-injected metadata, causing:

• Premature autocompact (CLAUDE.md rules lost → model forgets project structure)
• Cross-contamination between parallel agent stdout streams
• Model receiving truncated or corrupted tool results, forced to hallucinate

Evidence

Incident 1: Frontend code in docs/ directory

• Branch: adr/0206_ia_atomization (architecture documentation branch)
• Model: Claude Opus 4.7 1M (OAuth, user's own API key)
• Terminal: Warp (free tier, macOS/WSL2)
• Behavior: Agent created Svelte component files inside docs/01_architecture_decisions/
• When questioned, rationalized: "planning to migrate README.toml and markdown into web/"
• Same session context in Alacritty: does not reproduce

Incident 2: Parallel agent commit racing

• Multiple subagents launched via Agent tool with isolation: "worktree"
• Agents edited files in each other's worktrees
• Agents raced to git commit, causing lock conflicts and incorrect staged files
• Consistent with #47266, #45815, #48927 but Warp worsens the failure mode

Related issues

• #48028 — Context loss at <40% utilization (same symptom, possibly same root cause)
• #43038 — Warp + worktree causes pane split and agent stall
• #45815 — Sessions ignore worktree isolation rules
• #47266 — Parallel worktree agents hit git lock race condition
• #10375 — Terminal escape sequences leak into Claude Code input
• #34332 — Autocompact fires prematurely on 1M Opus

Warp-side issues:

• warpdotdev/Warp#8409, #8805 — 100% CPU hang during Claude Code sessions
• warpdotdev/Warp#9365 — Rendering corruption with Claude Code's alt-screen

Steps to reproduce

1. Open Warp terminal (free tier, Proactive AI enabled — cannot be disabled)
2. Run claude with a large CLAUDE.md (10K+ tokens of project rules)
3. Launch 2+ parallel subagents via Agent tool
4. Observe: agents cross file boundaries, commit races, and/or nonsensical actions
5. Repeat in Alacritty/kitty with identical CLAUDE.md — issue does not reproduce

Expected behavior

• Claude Code should be resilient to terminal-layer I/O interception
• Parallel agent isolation should hold regardless of terminal emulator
• Tool results should not be corrupted by terminal escape sequences or AI overlays

Suggested mitigations

1. Detect Warp ($WARP_IS_LOCAL_SHELL_SESSION == 1) and warn users about

known compatibility issues

1. Sanitize tool results — strip non-standard DCS/escape sequences before

feeding into model context

1. Document terminal compatibility — recommend plain terminals for agent-heavy

workflows in official docs

Environment

• Claude Code CLI (latest)
• Warp terminal (free tier)
• Platform: Linux (WSL2) / macOS
• Model: Claude Opus 4.7 1M (OAuth)
• CLAUDE.md size: ~15K tokens (heavy rule set)