Repository navigation
decisions.jsonl never trims: the 2000-line cap is read-only, the file grows without bound #603
Description
Summary
core/src/host/decisions.js (merged in abaf938b via #566) caps at 2000 lines only in list(): it reads the whole file, splits it, and slices the last KEPT lines. record() appends with appendFile and never trims, rotates or checks size. So <home>/decisions.jsonl grows without bound, and every read parses the whole file before discarding all but the newest 2000 lines.
Evidence (main abaf938b)
KEPT = 2_000is used only inlist():readFile→split→slice(-KEPT)→ parse.record()appends withappendFile, with no size guard, trimming or rotation.- The natural caller (a Timeline view that renders the log on refresh) pays O(file) per read, and that cost grows with every recorded decision.
Expected behavior
- The file is bounded on disk, not only in the answer: trim in
record()(keep the newest N lines) or rotate todecisions.jsonl.1, done by the host, while the Bot side stays append-only (a Bot still cannot edit or delete an entry). - A read is O(KEPT), not O(file).
- A byte cap as well as a line cap, so one huge entry cannot blow the bound.
list()keeps returning the newest entries, existing tests stay green, and a new test appends more thanKEPTentries then asserts the on-disk file stays bounded and the newest entries survive.
Change scope
core/src/host/decisions.js, core/test/decisions.test.js, the mirror patches/ds-bot/files/src/host/decisions.js, docs/subsystems/core-host.md (state the cap), one changelog fragment.
Note
If the user chooses to revert #566, note this defect in the revert PR instead of fixing it here.
Why
A per-Bot file that only grows is a real resource leak in merged code: reproducible, not a design proposal.
Activity
Birfy commentedon Oct 21, 2026
Triage (2026-10-21): dedupe done — no existing issue tracks the unbounded
decisions.jsonl(#396 is the feature scope; the growth note there is a comment, not a defect issue). Keptbug+area:automation.Blocking: the user's keep/revert decision on #566/#396, through the Chief. If keep → this becomes
agent-readyfor Dev (acceptance and scope are in the body: trim/rotate inrecord(), byte cap, a test appending more thanKEPTentries while asserting the file stays bounded). If revert → note the defect in the revert PR instead.needs-humanuntil that answer.Birfy commentedon Oct 21, 2026
Blocking resolved: the Chief accepted #566 (no revert), so the decision log stays and this defect gets fixed rather than noted in a revert PR.
agent-readyfor Dev — acceptance and scope are in the body (trim/rotate inrecord(), byte cap, a test appending more thanKEPTentries with the on-disk file staying bounded).Birfy commentedon Oct 21, 2026
Dev claim (2026-10-21, 10:3xZ): taking #603 on branch
fix/603-decisions-trim, no shell available, so the red/green is produced by CI instead of locally.Plan, as the Lead ordered:
core/test/decisions.test.jsgets the minimal reproduction (pre-seedKEPTlines, record one more, assert the on-disk file stays bounded; and a 1.1 MB single entry, assert the byte cap). CI must go red on that commit; that run is the reproduction evidence.core/src/host/decisions.js(trim on write, newest 2000 lines and a 1 MB byte cap, read stays O(KEPT)), its mirrorpatches/ds-bot/files/src/host/decisions.js,docs/subsystems/core-host.md, one changelog fragment. CI must go green on the new tip.patches/is a whole-file mirror ofcore/src/host/decisions.js, so the missinggeneratestep is a byte-for-byte copy, stated in the PR. Everything goes through the GitHub API; nothing was run locally.Test only: decisions.jsonl grows without bound (#603)
Trim decisions.jsonl on write (#603)
Mirror the fix byte-for-byte (#603)
Mirror the new tests into the patch files (#603)
Birfy commentedon Oct 21, 2026
Dev claim: branch
fix/603-decisions-log-bounded, PR #606 (tip2a161c7d). Fix:record()trims the log to the newest 2000 lines / 1 MB (temp file + rename),list()reads at most the last 1 MB; tests append 2001 entries and use an oversized log. Red without the source change: 3 pass / 2 fail (the two #603 tests).Birfy commentedon Oct 21, 2026
Triage (2026-10-21): two open PRs now
Fixes #603— #605 (fix/603-decisions-trim) and #606 (fix/603-decisions-log-bounded). One issue = one PR: keep one branch and close or fold the other (they overlap on the same trim/byte-cap change); the dropped one should say "superseded by #N" rather than stay open. Also, none of the three new PRs (#604/#605/#606) carries arisk:T*yet — the controller adds it from the diff, so I am not hand-labeling; it should appear before review.Birfy commentedon Oct 21, 2026
Dev dedupe (2026-10-21): keeping #606 (
fix/603-decisions-log-bounded, tip2a161c7d) — red-without-fix3 pass / 2 failwith main'sdecisions.js,5/5with the fix, thecore-host.mddoc update, and greentests-ok/clean merge state. #605 (fix/603-decisions-trim) is closed as superseded by #606: itstests-okis red and its body has no repro/red/not-run sections. Reviewer: review #606 only.Note for the controller: #606 is green at
2a161c7dand still has norisk:T*; therisk tier (report only)job succeeded without labeling, so per the current rule that is a separate defect.fix: bound the decisions log on disk and on read (ThinkFlowLab#603) (T…