### Preflight Checklist - [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) 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? SECTION 1 — WHAT'S WRONG: Session timeout reduction breaks scientific verification workflows built on Claude Code Recent reductions to Claude Code sessi

[BUG] Execution timeout regression silently kills scientific scripts

Resolved 💬 4 comments Opened Mar 22, 2026 by WPOpenClaw Closed Apr 20, 2026

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?

SECTION 1 — WHAT'S WRONG:

Session timeout reduction breaks scientific verification workflows built on Claude Code

Recent reductions to Claude Code session timeouts have broken a working scientific verification pipeline that was built specifically around Claude Code's capabilities over 236 commits and 15+ sessions across multiple weeks.

The project — BRT Explorer (https://github.com/WPOpenClaw/BRT) — is an open-source physics framework with 223 self-contained Python scripts and 2373+ binary PASS/FAIL tests. It can no longer complete its full verification chain within a single Claude Code session. This defeats the core methodology: AI-substrate verification with cryptographic provenance.

The project consists of:

223 gap scripts — self-contained Python verification scripts, each deriving physical constants from geometry with zero free parameters
2373+ binary PASS/FAIL tests — every claim has a script, every script has tests
SHA256 solace chains — tamper-evident provenance showing the AI verified the complete derivation chain
A 9-step installer (installer.sh) that sets up Python, Node.js, Julia, and sl2cfoam-next from scratch
A tiered pipeline (run_pipeline.py) with quick (~1 min), light (~10 min), moderate (~22 min), and heavy (~96 min) tiers
Three heavy tier scripts that previously completed in Claude Code sessions now get killed:

gap5_ms_bounce.py — Mukhanov-Sasaki bounce solver (~45 min)
gap7_pantheon_full.py — Full Pantheon+ covariance MCMC, 1617 supernovae (~25 min)
gap19_subset_validation.py — MCMC subset validation, jackknife (~26 min)
This breaks: (1) the complete verification chain, (2) the SHA256 solace provenance chains which require every link, and (3) the AI-substrate verification claim — you can't honestly say "an AI verified the complete chain" when the AI can't run the complete chain.

Beyond timeouts, sessions have repeatedly disconnected, lost context, and the AI has rewritten working physics scripts causing damage that took days to recover from. 41 Python files were rewritten across 30 commits and had to be fully reverted (commit 599e228). The session sidebar shows 15+ sessions across multiple weeks — the majority spent on recovery, not science.

This is exactly the kind of unconventional, AI-assisted scientific work that Anthropic's mission supports: an independent researcher using AI to do open, reproducible, falsifiable science with built-in safeguards against hallucination. I am dyslexic — AI tools like Claude Code are not a convenience for me, they are how I do my best work. I came to Anthropic specifically because of the company's ethos around beneficial AI use and making AI accessible. Having it broken by the same company whose values drew me here is disheartening.

What Should Happen?

Claude Code sessions should support long-running scientific computations that previously worked. Scripts requiring 25-45 minutes each (MCMC sampling, numerical solvers) completed successfully in earlier sessions. The platform should either maintain previous timeout limits, offer a configurable timeout for scientific workloads, or at minimum provide advance notice when session limits change so users can adapt their workflows before they break.

Specifically:

Awareness — this is a real project with real scientific output that was built around Claude Code's previous capabilities. Reducing session timeouts without warning broke a working workflow.
Consideration for long-running scientific workloads — not every use case is "write a function and test it." Some projects need extended compute time for MCMC sampling, numerical solvers, and large-dataset validation.
Transparency — if session limits change, users building long-running workflows need advance notice.
Session stability — disconnections mid-computation lose hours of work and corrupt state.

Error Messages/Logs

TIMELINE OF ISSUES ACROSS 15+ SESSIONS (multiple weeks, Feb-March 2026):

PLATFORM FAILURES:
- "Claude Code is unavailable — There was a problem loading your account
  data" — full platform outage mid-session (screenshot: Mar 20, 18:19)
- "Request timed out" — session killed while writing gap192 (Mar 21, 11:50)
  Multiple "[Request interrupted by user]" — user trying /bug, /debug,
  /doctor commands to recover. Nothing works.
- Julia process at 1476% CPU for 2.7 hours computing sl2cfoam vertex
  data — session dies underneath it. Computation lost.

AI DAMAGE INCIDENTS:

1. Mar 5 — AI session applied "code cleanup" assumptions to physics scripts:
   - Downgraded DERIVED status to DATA-DRIVEN on verified constants
   - Added "fits WORSE" editorial commentary
   - Inserted "DISCONNECTION" warnings between domain scales
   - Added hedging language to proven geometric results
   Recovery: commit e82af23 (full revert + added symbolic integrity rules)

2. Mar 21 — Random audit discovered 56 missing .anchor.solace files:
   - Broken sessions (AWS timeouts, context limits) terminated after
     writing JSON but before generating tamper-evident proofs
   - Chain of trust broken at Level 1
   Recovery: commit be70881 (regenerated all 56 chains)

3. Mar 22 — Discovered 41 Python files rewritten across 30 commits:
   - Working physics derivations overwritten by AI "improvements"
   Recovery: commit 599e228 (restore all to known-good state)
   Then:    commit 485eba2 (re-anchor all files, 55/55 PASS)

4. Mar 22 — Heavy tier scripts killed by session timeout:
   - gap5_ms_bounce.py (~45 min) — killed
   - gap7_pantheon_full.py (~25 min) — killed
   - gap19_subset_validation.py (~26 min) — killed
   - No error message — scripts silently terminated

DOCUMENTED AI FAILURE MODES (from docs/NEW_SESSION_PROMPT.md):
   "Previous sessions have:
   - Merged all docs into a single file, losing cross-references
   - Deleted gap scripts they deemed 'redundant'
   - Flattened src/ subdirectories into root
   - Renamed files to 'improve' naming, breaking import chains
   - Downgraded DERIVED status on scripts with no free parameters"

SESSION SIDEBAR EVIDENCE (screenshot attached):
   15+ sessions across multiple weeks, including:
   - "Read BRT documentation and setup" x 8+ (AI losing context each time)
   - "Read BRT documentation and symbolic integrity rules" x 2+
   - "Debug connection drops in AI physics system"
   - "Fix timeout issues with verification scripts"
   - Plus more scrolled off the bottom
   The majority were NOT science — they were recovering from previous
   session damage or reconnecting after drops.

PROTECTIVE MEASURES NOW REQUIRED:
   - 7 rule files in .claude/rules/ to prevent AI destruction
   - 137-line NEW_SESSION_PROMPT mandatory for every new session
   - verify_symbolic_integrity.py gate before every commit
   - Automatic backups in public/backups/
   - Weeks spent on recovery instead of science

CURRENT STATE AFTER RECOVERY:
   Quick tier:    118/118 PASS (59s)
   Light tier:    211/211 PASS (591s)
   Moderate tier:   6/6   PASS (1316s)
   Heavy tier:    CANNOT COMPLETE — session timeout kills scripts
   Solace chain:  BROKEN — cannot produce complete provenance

PROJECT: https://github.com/WPOpenClaw/BRT
COMMITS: 236 (March 2026)
TESTS: 2373+ binary PASS/FAIL, 100% pass rate

Steps to Reproduce

Open a Claude Code web session on any project with long-running

Python scripts (25+ minutes each).

Run a script that takes more than ~10 minutes:

Example from this project:
python gap7_pantheon_full.py # ~25 min, Pantheon+ MCMC

The script will be silently killed before completion.
No error message is shown. It just stops.

This previously worked. The same scripts completed successfully

in earlier Claude Code sessions (Feb-March 2026). The session
limits were reduced without notice.

The project repository demonstrating this:

https://github.com/WPOpenClaw/BRT

Pipeline tiers that still complete in-session:

Quick: 118 scripts, 59 seconds
Light: 211 scripts, 10 minutes
Moderate: 6 scripts, 22 minutes

Pipeline tier that cannot complete:

Heavy: 3 scripts needing 25-45 minutes each (~96 min total)

The SHA256 provenance chain (--solace flag) requires ALL tiers

to run sequentially. If heavy tier cannot complete, the
cryptographic chain is broken and cannot be produced.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

https://github.com/WPOpenClaw/BRT/releases/tag/beta

Claude Code Version

Platform: Claude Code web (browser), Windows 11 Model: Claude Opus 4.6 (1M context)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

View original on GitHub ↗

This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗