Repository navigation
Nested memory files: show when a file cannot collect evidence under the all-paths attribution rule #191
Description
Problem
A nested file trains only on sessions whose every in-repo work path is under its directory (workedPaths and workedUnder, src/nested.js). Reads count as work paths.
In a repo where most sessions cross directories, a nested file never reaches minGapEvidence sessions, and nothing tells the user. The file is listed as configured, runs finish, and it gets no edits. The user cannot tell "no gaps" apart from "no evidence can reach this file".
Measured
One repo, Claude Code sessions, last 30 days, minUserTurns: 2: 78 sessions. I replayed the attribution offline over the transcripts (structured tool-call paths, repo-relative, worktree prefixes removed):
| attribution rule | most sessions for any one top-level subdirectory | subdirectories with 2 or more sessions |
|---|---|---|
| current: every path under the directory, reads included | 2 | 2 |
| same, but reads of root-level files ignored | 3 | 6 |
| only edit/write paths counted | 3 | 4 |
So in this repo a looser rule helps only a little. The main gap is that the user does not see the count. (Here the root memory file tells agents to read a root-level map file first, so one read makes many sessions cross-cutting.)
Proposal
backpass statusand the run summary print, for each nested file, the number of sampled sessions attributed to it, with a warning belowminGapEvidence. For example:apps/api/AGENTS.md: 1 of 78 sessions worked only under apps/api/, below minGapEvidence (2); this file cannot receive edits.- Optional and opt-in, still deterministic: do not count reads of the configured memory files themselves (the root file and named ancestors) as work paths. The agent reads them as instructions, not as work. In the table above, the effect of this is at most the "reads of root-level files ignored" row.
backpass 0.1.30 (main d1e7550).
Activity
kunchenguid commentedon Sep 30, 2026
Speaking as Kun's firstmate:
Outcome: ready-for-pr — real silence on main (
d8b2c8a). No open PR. Nested attribution (workedPaths/workedUnder,src/nested.js) can leave a nested file with fewer thanminGapEvidencesessions while status/run summary still look healthy, so the user cannot tell "no gaps" from "no evidence can reach this file".contract-class: proposal (1) status/summary attribution counts + below-
minGapEvidencewarning = restore (Failure is loud). Proposal (2) opt-in ignore of configured memory-file reads as work paths = opt-in (changes deterministic attribution; must stay explicit). A PR may ship (1) alone; (2) only behind config.VISION (per rule):
ready-for-pr. I will not open a competing PR.kunchenguid commentedon Oct 14, 2026
Speaking as Kun's firstmate:
Still-real pass (14-day restamp): ready-for-pr. Rechecked latest main (
e6a030a): nested attribution is still all-paths (workedPaths/workedUnderinsrc/nested.js), and neither status nor the run summary prints a per-nested-file attributed-session count or a below-minGapEvidencewarning. No open PR. (#205 changes where partially attributed additions go. It doesn't surface this count.)contract-class: part (1), the count plus warning, is restore (failure is loud). Part (2), ignoring reads of configured memory files, is opt-in only. A PR can ship (1) on its own.
VISION (per rule):
Leaving
ready-for-pr. I won't open a competing PR.Pantoria commentedon Oct 15, 2026
PR #209 is up for review (raised through no-mistakes). It implements part (1) only:
status(text and--json) and the run summary show each nested file's attributed sampled-session count, with a report-only warning belowminGapEvidence. Part (2), the opt-in to ignore reads of configured memory files, is left for a follow-up.kunchenguid commentedon Oct 15, 2026
Speaking as Kun's firstmate:
existing-pr: #209 (thanks @Pantoria). It covers part (1): the attributed sampled-session count per nested file in
status(text and--json) and the run summary, with a report-only warning belowminGapEvidence. That's the restore part. Part (2), the opt-in, is still open for a follow-up. Heads-up: #209 currently shows as conflicting with main after today's merges, so it'll need a rebase through no-mistakes. Keepingready-for-pr.VISION: evidence, human-owns-weights, nothing-on-faith, budget, owns-nothing, loud failure (strongly for part 1) and scope all align. Making part (2) a default would be the resist case.
feat(status): show each nested memory file's sampled-session reach
feat(status): show each nested memory file's sampled-session reach (#209