Redditor monsieurpooh claims Claude Code's revert-conversation-turn-with-files feature reverts files from all other conversations in the same project to their pre-current-conversation state, and independent reproduction, wider reports, or an Anthropic fix would establish turn-revert scope as a real cross-conversation data-loss bug in the harness.
state: seedheat: lowuncertainty: mediumconvergesscott: highagent-harnesses claude-code state-corruptionAnthropic
What is this?
A single Reddit user (monsieurpooh) reports that Claude Code's rewind feature, when 'restoring code and conversation', reverts project files to their state before the current conversation — silently overwriting work done in other conversations in the same project directory. The retrieved material offers no reproduction, no second report of that specific cross-conversation behavior, and no Anthropic response, so the claim stands or falls on one user's word. What the search does establish is the surrounding terrain: Claude Code's rewind is independently documented as destructively unsafe by default (GitHub issue #64615 — the destructive 'Restore code and conversation' option is the default, reverts files instantly with no confirmation or preview, still present as of v2.1.159 after two earlier reports were closed without rationale), Claude Code's conversation storage is per-project-directory with known indexing fragility (#9898), and the same bug genre recurs across rival harnesses — Qoder rolls files back when starting a new chat, Cursor's turn-revert leaks reverted state back into the session, and opencode's /undo silently under-reverts across repositories. So revert-scope bugs are a real, recurring harness failure class, but this particular multi-conversation data-loss claim remains a lone, unverified report.
Why it matters to Scott
Though still a lone report, the claim lands precisely on the rollback-domain-scoping axis Scott's reversibility-membrane and session-isolation concepts govern — the harness's own safety feature becoming the destroyer is the sharpest possible instance of his 'harness-managed rollback without git as safety net' warning, and Anthropic's documented destructive-by-default rewind (issue #64615, reports closed without rationale) is a consequential party behaving exactly as the membrane thesis predicts. It also supplies the failure-mode counterpart to the radar's constructive revert thread (Codex thread-revert, Replit snapshots, Agent_acid rollback primitives), and if confirmed it changes both how he runs multiple conversations per Claude Code project and how his own ask harness scopes conversation-state persistence and any future undo.
ip:concept.reversibility-membraneip:concept.session-isolationip:concept.goal-world-isolationip:framework.long-running-agentsip:source.give-the-agent-a-workshop-ebookdev:project.askradar:concept.claude-coderadar:concept.agent-safetyradar:claude-code-48k-file-deletionradar:openai-codex-thread-revert-supportradar:replit-snapshot-agent-isolationradar:agent-acid-rollback-guardrailsradar:hollow-agentos-peer-deletion
queries asked of Scott's wikis
- agent harness checkpoint rewind design
- file snapshot scope per-session vs per-project
- destructive defaults undo safety coding agent
- claude code harness reliability known bugs
- multi-session same-repo agent state isolation
- git as safety net vs harness-managed rollback
Measured heat
now 0 pts/hpeak 6 pts/hcomments 0/hpeers p0momentum: steady1 platformsage 298h
points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion
How the heat travelled
pace: p30 vs 1188 stories at the 168h mark (now 298h old) — ahead of addom-local-coding-harness (1.5x), behind agentsec-static-config-auditing (0.8x)
Evidence (1) — ⭐ canonical anchor
Interpretation history
2026-09-29T06:37:17Z
grounded: converges/high — Though still a lone report, the claim lands precisely on the rollback-domain-scoping axis Scott's reversibility-membrane and session-isolation concepts govern —
2026-09-29T06:26:24Z
case created — A concrete, cheaply falsifiable multi-session data-loss claim against a core Claude Code feature — squarely on Scott's harness-reliability radar and distinct from every open Claude Code case.
Decision trace
- 09-29 16:37groundThough still a lone report, the claim lands precisely on the rollback-domain-scoping axis Scott's reversibility-membrane and session-isolation concepts govern — the harness's own safety feat
- 09-29 16:26createA concrete, cheaply falsifiable multi-session data-loss claim against a core Claude Code feature — squarely on Scott's harness-reliability radar and distinct from every open Claude Code case.