Our Postgres setup code had a line that capped shared buffers: if the VM has more than 16 GB, use 8 GB. The comment said "fix at 8 gb". Everyone knew it was decided. Nobody could say why, or why 8 and 16 specifically. A kernel limit on an old host? A crash we hit once? So nobody touched it, for years.
An AI agent hits that line with worse options than we had. It can delete it as arbitrary, or preserve it forever without knowing why. That one line is the whole problem I'm working on: the decision survived, the reasoning did not.
The version of this that made me quit arguing and start building was watching it hit machine speed. Ten engineers, all using coding agents, all prompting in their own words. The same task produced different code depending on who asked. One person's agent settled a question and another person's agent re-decided it the opposite way a week later. Ten small companies inside one ten-person company. We tried pinning conventions in a skills.md file. It went stale immediately, because nobody maintains a file that has no consequences for being wrong.
Decispher is a system of record for engineering decisions. Three parts, and I'll be specific because the interesting parts are mechanical:
Capture. Connectors on Slack, GitHub, GitLab, and Jira watch for decision-shaped moments, not everything. Separately, Branch Story records AI coding sessions themselves, with hooks for Claude Code, Codex CLI, and Grok. When your agent discovers a constraint at 2am, or you steer it off a dead end, that currently dies with the session. It shouldn't. Extraction produces seven unit types: decisions, conventions, constraints, rationale, ownership, history (what was tried and why it failed), and plans.
Fusion. The reasoning behind one decision is genuinely fragmented. Part of it is a Slack argument, part is a PR review, part is a Jira comment. No single tool sees all of it. Calibrating this on real conversations taught me something I did not expect: entity overlap beats semantic similarity for cross-source linking. Two messages about the same decision often share almost no vocabulary, but they share the same repo, service, and people.
Serving. This is where I broke from what everyone else does, and it's the part I'd most like this crowd to argue with. The industry answer is static context dumping: a fat CLAUDE.md or rules file that goes stale the day it's committed and burns tokens on every single request. Instead each agent's shell file (CLAUDE.md, AGENTS.md, .cursorrules, and six others) is a thin auto-regenerated index telling the agent what context exists and how to fetch what it needs, live, over MCP. When the record changes, the shells regenerate and agents pull fresh. Measured on our own repos: about 35% fewer tokens on a well-documented repo with real ADRs, about 65% on a rough one. The gap is the point. The savings come from the agent no longer rediscovering things that were written down somewhere it couldn't reach.
What I want. A small number of design partners. Concretely: teams of roughly 10 to 60 engineers who already run coding agents daily and have already had the "we decided this six months ago" moment. You'd get direct access to me and your priorities on the roadmap. I'd want your real workspace rather than a sandbox, and blunt feedback about where it's wrong.
Product: https://decispher.com/
Top comments (0)