Repository navigation
trail note: count a learning's cost from where its work began - #620
Conversation
🌱 graft blast radiusNothing outside this diff depends on it. 2 areas changed; no indexed dependents at depth 2. Who knows this code — 1 person across 2 areas
Ownership is git history over each area's own files, weighted towards recent work (120-day half-life). Merge commits and bots are dropped, and you are dropped from your own PR. A name with no Test signal per changed area — 1 ✓ · 1 ✗Reached = a node under a test path has a resolved edge into the changed symbol. It undercounts anything called indirectly — through a CLI, a spawned process or a dynamic import — so read a low ratio as “look here”, never as a coverage gate.
Open the interactive graph → — click an area to see its dependent symbols at file:line. |
Moves: none of the project's metrics directly. It fixes the number every credit line shows ("saved ~958k tokens and ~129 min"), which the 80k rule uses to choose between tokens and "3x faster to a PR".
learnings_committedis the closest metric (PostHog query indaily/project.md). I didn't check its value, because this session has no Metabase access.Why
trail notecharges a learning for the whole session since the last note. In a long session that did several things, the learning gets the cost of all of them. One learning saved today, about committing the hook wiring, claimed "129 min and ~958k tokens". That stretch also covered a Slack search, two PR merges and a demo setup, none of which it records.What changes
A learning's cost now counts from the turn where its work first came up:
dropHtmlCommentsorrepo_id. Phrases and plain words liketrail initorpublishare skipped, because they come up in any session about the product.trail notecall being saved never counts, since it names them all.--minutesand--tokensstill win. The files a note goes to are still taken from everything edited since the last note.Before and after
I re-ran the cost of every
trail note/graft notein my last 21 days of local Claude Code transcripts, old way and new way, with a script that isn't in the PR. These are 55 notes from one person's machine, so treat them as an example, not a measurement:Notes from short, focused sessions don't change. The big drops are in long mixed sessions:
The number is still labelled as an estimate wherever it's shown.
Tests
trail-notes:🤖 Generated with Claude Code