diraArising from qst-0002 — Is attention drift derived from ledger writes only, or from session metadata?
dira reads session metadata to derive attention drift, locally and never transmitted
dira may read session start, stop and working directory from the agent harness to derive attention drift.
What the founder chose, and against what advice
The recommendation in `qst-0002` — and mine — was **not to build this**. It is the one proposed feature closer to self-surveillance than to navigation, and it is easy to add and impossible to un-ship. That argument was made and did not win. It is recorded here rather than dropped, because the next session should not re-litigate it and should know it was a considered choice, not an oversight.
The constraints this makes load-bearing
Two existing rules stop being background and become the implementation contract:
- **`cst-0004`, no network.** Where your time went never leaves the machine. There is no telemetry, no upload, no "anonymised" aggregate. A drift report is computed at read time from local files and printed. - **`cst-0003`, private stays private.** Session metadata is private context by default and inherits every rule that already governs a private ledger: never exported downward, cited ref-only across a boundary, never rendered into a public surface. A public ledger renderer must not be able to say what a person was doing on Tuesday, and that has to be structural rather than a display choice.
What must be true before it ships
**Reading is opt-in and off by default.** A tool that starts logging behaviour the moment it is installed has made the choice for the user. The person who benefits must have turned it on.
**The raw record is not kept.** What is stored is the derived answer — which ledgers the week's sessions touched — not a timeline that could reconstruct a working day. The difference between "70% of sessions were in repos unconnected to your focus" and a per-session log is the whole distance between navigation and surveillance, and the first can be computed without retaining the second.
**It is deletable in one command, and the command is documented next to the feature.**
Alternatives on record — 2
✗Do not build attention drift at allrefused This was the recommendation, and it lost.
The argument for it was that a compass reports your bearing rather than grading your effort, and that a tool holding your private strategy should not also hold a log of your working habits. The founder's call is that knowing where the week actually went is the point of the personal tier, and a bearing you cannot check against your actual course is not navigation.
revisit ifthe drift report is read for a month and never changes a decision, which would mean it is a mirror nobody looks in
✗Count ledger writes only, collecting no new datarefused Zero new data and nothing to leak, and it measures the wrong thing.
A week of deep work that produced two decisions reads as quieter than an hour of thrashing that produced nine, so the number would be actively misleading in exactly the case a person most wants an honest answer.