Русский | English

What the state projection costs, and why every kind stays

The tausik/ Markdown tree is 68.6% of this repository's tracked paths. This page says what that buys, what it costs, and what was decided about each kind — measured, because the story that asked the question had guessed wrong twice.

Re-measure at any time:

python scripts/projection_census.py --db .tausik/tausik.db

The measurement

Kind	Files	Bytes	Avg	Share of tree bytes
tasks	1744	11,933,248	6,842	39.1%
memory	765	1,465,331	1,915	4.8%
decisions	401	655,171	1,633	2.1%
stories	297	102,467	345	0.3%
epics	109	36,786	337	0.1%
graph-snapshots	1	157,751	157,751	0.5%
projection	3317	14,350,754		47.1%
tracked tree	4832	30,484,187		100%

Two more numbers from the same tree:

• Journals are 47.8% of task-file bytes (5.7 MB). The rest is front matter, goal and acceptance criteria — a row's shape. The journal is the part a writing habit moves.
• Over the whole history, the projection is 48.7% of file-touches but 14.8% of line churn. It is rewritten constantly and changes little each time.

What the guess got wrong

The hygiene story planned on soft-archiving old done tasks to cut a projection "occupying 69% of the tree". Both halves were wrong.

1. Archival does not touch the projection. state_export selects FROM tasks with no archived_at filter. Archiving hides a task from task list; the file stays.
2. The reachable ceiling is 19%, not 69%. 917 task files match "done and older than 90 days" — 2.7 MB, 19.0% of tracked paths. Even a filter that removed all of them would leave the projection at half the tree.

Why removing them is not on the table

.tausik/tausik.db is gitignored. The Markdown tree is the only carrier of state between machines: a row dropped from the tree is a row a fresh clone never sees. Hiding a task from a listing and deleting its record are different acts, and archival was built for the first. tests/test_projection_census.py::TestTheTreeIsTheOnlyCarrier pins both facts — if the database ever starts travelling, this decision has to be retaken rather than inherited.

The verdict, per kind

Kind	Verdict	Why
tasks	stays	The tree is the only carrier; dropping archived rows would lose 917 tasks on a fresh clone. The journal is the evidence a closure happened, which is the record SENAR asks for.
memory	stays	Same carrier argument. It already filters archived_at IS NULL on export, so what is hidden is hidden and what is kept travels.
decisions	stays	2.1% of the tree and the thing most often read by a fresh agent. Nothing to win.
stories	stays	0.3% of bytes. Removing it would cost more in review than it saves in checkout.
epics	stays	0.1% of bytes. Same.
graph-snapshots	stays	One gzipped file per release, kept on purpose as a before-picture. It is the largest single file in the projection and still half a percent of the tree.

The lever that does exist

Growth is bounded at the point of writing, not by deletion: the journal budget advises on every task log entry, and journals are where 47.8% of the task bytes are. That is a habit with a brake on it, which is a different mechanism from a threshold on the total.

There is deliberately no ratchet on the projection's share. Every closed task adds a file, so a threshold on the total would be crossed by ordinary work — and a threshold ordinary work crosses is one somebody switches off, taking the measurement with it. The census reports; it never fails a build.

See also

• Task archive spec — what archival does and does not do, and the one command back.
• Architecture — where the exporter sits.