Git Already Logs the Why. This Developer Wants AI Agents to Read It.

Veselin Dimitrov had tried everything. Mega-prompts. Slash commands. Rules files. MCP tools. Session transcripts. After more than a year working daily with AI coding agents, he kept hitting the same failure at the same moment: the end of a session, when he was tired, when he'd just shipped something, and when the last thing he wanted to do was write documentation.

"Every approach I tried required a separate documentation step," he said. "And that step always happened at exactly the wrong time — when you're done, when your cognitive load is highest, when you just want to close the laptop."

So agents like Claude Code started each new session with no record of prior decisions. They could read the codebase and reconstruct patterns and conventions. What they couldn't recover was the reasoning: why a particular approach had been chosen over another, what alternatives had been tried and discarded, what constraints had emerged mid-work and never made it into the code itself.

Dimitrov's solution, published this month as 'Contextual Commits', is almost aggressively low-tech. It extends the Conventional Commits specification by defining five action types — intent, decision, rejected, constraint, and learned — that developers record in the Git commit body. Not in a separate file. Not in a wiki. In the same message they were already going to write.

"Git's commit log is already append-only, already conflict-free, already at every session boundary," he said. "We've been treating it as a changelog when it could be a memory system."

That framing has resonated with developers dealing with the same frustration. "The context decay problem is real and it's been quietly annoying everyone who works with these agents regularly," said one senior engineer at a developer tools company, who wasn't authorised to speak publicly. "Whether a commit message is the right container is worth debating, but the instinct to use existing developer behaviour rather than adding a new ritual is sound."

Dimitrov tested the alternatives before landing on commit messages. Hooks — both Git and Claude's own — broke down in team settings because they depended on consistent environment setup. Rules files accumulated stale entries. Extracting session artifacts was still a manual step at exactly the wrong moment. The commit message survived because developers already have the discipline to write one, and because it stays attached to the change that prompted it.

The standard ships with a companion 'recall' utility skill that lets agents query contextual history by scope, action type, or branch. Dimitrov says he noticed Claude Code invoking the utility without being prompted once it was available — the agent apparently sensing when context was thin and pulling the structured history on its own.

The v0.1.0 spec is MIT-licensed on GitHub under the 'berserkdisruptors' organisation. It requires no new infrastructure and no changes to existing tooling. Dimitrov's longer aim is native adoption by agent harnesses rather than skills-layer integration, which would produce more consistent behaviour across sessions. That shift is not imminent.