I started session attribution (locally) as soon as I found that there were jsonl files on disk for every session. It makes it really easy to make sure everything that happens has a line that goes back to the user intent.

Now every project I have always has 3 repos (+1 root orchestration repo):

1. Code 2. Docs (specs, plans, research, etc) 3. Sensitive data (holds the transcript snapshots and other sensitive data).

All agents have access to code repo.

Only "special" agent has access to #2 and #3.

Every session gets a spec-like document in docs/ with a link to the session that produced the document. So if I have say, "0095-user-auth-spec.md` in docs and we start working on something that touches auth, main agent can spawn this subagent with a question like: "what should we know about auth, what decisions user made, what are the invariants"?. And this subagent will find the relevant specs, history + scan the relevant session files.

This works remarkably well. Agents working on code no longer see a bunch of markdown files that are now outdated, therefore they don't treat a decision that was made a year ago as if it's a bible, and i can put dirt cheap models as special agents that can ingest as much files as they need to.

Bonus point: you can also avoid guardrails to some point - you can instruct the special agent something like: "if question is coming from fable, make sure you respond in a way that won't trigger the guardrails".