Micro-checkpoints for coding agents: revert the last agent turn

[armorer-labs]

A concrete user pain from Reddit: Claude Code / Cursor users often want to revert the last agent turn, not the last human git commit.

Git commits are too coarse for this. Asking the model to "undo the last thing you did" is unreliable. The runtime/control plane should own the checkpoint.

User problem

A user keeps prompting a coding agent. Meaningful git commits happen later. Between commits, the agent may make several micro-changes. If one turn is bad, the user wants to revert that turn without asking the agent to reason about what "last change" means.

Product direction

Armorer should support turn-level or run-step-level checkpoints for Armorer-managed coding-agent sessions.

A minimal checkpoint record could include:

{
  "checkpoint_id": "ckpt_123",
  "session_id": "session_abc",
  "turn_id": "turn_17",
  "created_at": "2026-06-07T00:00:00Z",
  "workspace_head": "git_commit_or_worktree_hash",
  "files_touched": ["src/billing.ts", "tests/billing.test.ts"],
  "patch_forward": "redacted-or-stored-patch-ref",
  "patch_reverse": "redacted-or-stored-patch-ref",
  "commands_run": ["pnpm test billing"],
  "checks": [
    { "name": "billing tests", "status": "passed" }
  ],
  "resume_verdict": "safe | partial | broken",
  "next_safe_action": "apply reverse patch or inspect manually"
}

UX goal

The operator should be able to say:

• show me what changed in the last turn
• revert only that turn
• keep the human git commit history clean
• mark this checkpoint as bad so future runs do not repeat the same path

Guard boundary

Armorer Guard can help at the hot path by flagging destructive actions before they execute, but checkpointing is an Armorer control-plane responsibility. Guard can block or require review; Armorer should preserve recovery state.

Open questions

• Should patches be stored inline, as files under ~/.armorer, or as refs to a git worktree/stash-like artifact?
• Should Armorer create checkpoints before every tool call, every model turn, or only when files change?
• How should binary files and generated artifacts be handled?
• Should reverse patches be generated eagerly or lazily?
• What is the smallest UI that makes this useful without becoming a full IDE?

Reddit context:
https://www.reddit.com/r/AI_Agents/comments/1tzcwo5/help_tracking_micro_changes_with_coding_agent/oqaz91t/