Session Receipts: what your coding agent did but didn't tell you

The basic plumbing is taken. Entire, Git AI, Memento and the Agent Trace spec already capture agent transcripts and tie them to commits. Entire is very well funded, already has a GitHub App, and can summarize the key decisions in a session. So "link the transcript to the PR" is not a moat. What I could not find shipped is the narrow, opinionated part: a PR comment that names only the handful of risky decisions (deleted or weakened tests, auth changes, new dependencies, migrations, CI or config edits) and backs each claim with a link to the exact transcript step and the exact diff hunk. The two sides of the market each miss half of this. Diff-only reviewers like Macroscope or CodeRabbit-style bots never see the transcript, so they can't tell you why the agent did something. Provenance tools store everything and leave the reviewer to dig through it. A new entrant could win on three things. First, precision and trust: every claim has evidence behind it and there is little noise. Second, cloud Codex background tasks, which the capture tools cover less well. Third, zero-setup install for small teams, consuming Entire or Agent Trace data rather than competing on capture. The risk is that Entire, or an AI review vendor like Propel, adds exactly this comment as a feature within months. The stop-the-run hook is weak as a standalone product. Claude Code's native hooks already offer allow, deny and ask, with Slack notification recipes. HumanLayer pioneered Slack approval gates, and Coder sells enterprise agent firewalls. It works best as a free add-on to the receipts bot, for example by writing paused or approved actions into the same receipt, rather than as its own business.

• Entire (Checkpoints CLI and entire.io)

  Thomas Dohmke's startup, which announced a $60M seed round. Its open-source CLI captures Claude Code, Codex-adjacent, Gemini, Cursor and Copilot CLI sessions (prompts, transcript, tool activity, changed files) and links them to commits. A GitHub App feeds a web viewer with Sessions and Changes tabs. The CLI can 'explain' and summarize a checkpoint, including the key decisions made, and the docs list viewing session details in PR reviews. This is the closest and best-funded overlap. I found no evidence that it posts a risk-ranked decision list as a PR comment.

• Git AI

  Open-source git extension. It links every AI-written line to the agent, model and transcript through git notes and supports 14+ agents, including Codex and Claude Code. Its PR bot that rolls this data up at the PR level is still in development. It covers attribution, not decision review.

• Agent Trace spec (Cursor)

  Open JSON spec that links code ranges to agent conversations. Reported backers include Cognition, Vercel, Cloudflare and Google Jules. This is a data layer a receipts bot could consume, not a competing reviewer.

• Memento / session-as-commit-message tooling

  Open-source git extension that attaches Codex or Claude Code transcripts to commits as git notes. It stores raw provenance and does not summarize it.

• Propel Code

  AI code review vendor that publicly argues for storing structured session provenance (intent, risk-based routing) in every PR. It is a likely adjacent entrant.

• Macroscope custom review agents

  Diff-based AI PR reviewer. It can gate risky migrations and check test coverage, but it does not read agent transcripts.

• Claude Code hooks (native PreToolUse / PermissionRequest)

  Built into Claude Code. Hooks can deny or 'ask' before force-pushes, DROP TABLE or rm -rf, and the Notification hook commonly pings Slack. Many free guides and scripts already do this, which commoditizes the stop-the-run idea.

• HumanLayer / CodeLayer

  YC-backed open-source SDK that blocks function calls until a human approves them over Slack or email. The team has since pivoted to CodeLayer, an IDE for orchestrating Claude Code sessions. This is a direct precedent for Slack approval gates.

• Coder Agent Firewall

  Paid AI Governance add-on. The Claude Code module uses it by default, giving enterprise policy control over agent actions.

• fuushyn armour

  Small community Claude Code plugin that acts as an agent firewall with policy and recovery controls. It is early stage.

Checked 2026-10-09