antfarm-crash-dedup-merge
Use when asked to dedupe AntFarm (D:/NightSmith) crash tickets or tidy the //psrecap crash share, when ticket counts grow or a ticket named archive appears after a merge, or when changing how duplicat
SKILL.mdskills/managed/antfarm-crash-dedup-merge/SKILL.md
GitHubAntFarm crash dedup and share merges
For D:/NightSmith. Reduces duplicate crash tickets and consolidates their folders on the shared crash drive. The share is other people's data — every step is dry-run-first and reversible.
The two traps that cost real tickets
1. A merge changes a crash's identity unless you pin it.
crashes._read_folder reads subdirectories sorted(..., reverse=True) — newest first — and
takes fingerprint/lua_stack/stack from the first instance with a manifest. The group key
is sha256 of that cause (ident = repr(min(cause))), and that key IS the ticket id. So
moving a newer bundle into a canonical folder changes its cause, its key, and its ticket id:
the next ingest files a brand new ticket for the crash you just tidied.
Fix already in the tree: crash_merge.apply writes .antfarm-crash-key (crashes.IDENTITY_PIN,
JSON {"key": "crash-..."}) into the canonical before any bundle moves, and
group_by_fingerprint prefers the pinned key over the evidence. If you add a code path that
moves bundles, it must write the pin too.
A folder merged before the pin existed keeps its post-merge id. Fix by hand, and do it before linking the stray as a duplicate — otherwise the link absorbs a ticket nothing points at:
from antfarm.crashes import IDENTITY_PIN, DERIVED_KEY_RE
from antfarm.tickets import crash_ticket_id
assert DERIVED_KEY_RE.match(key) and crash_ticket_id(key) == key
(folder / IDENTITY_PIN).write_text(json.dumps({"key": key, "note": "..."}) + "\n", encoding="utf-8")
2. archive/ was ingested as one giant fake crash.
scan() walked every top-level directory and _read_folder counts every subdirectory as an
occurrence, so retiring a crash re-filed it. scan now skips crashes.ARCHIVE_DIR. If a ticket
literally named archive appears, that skip has been lost.
Matching rules
dedupe.groups() matches on same crash site and masked detail (crash_family.family_of)
or the first 3 culpable frames (dedupe._culpable_frames). Culpable means after
dropping the funnel — every stack here starts ucrtbase → std::terminate → AbortImmediate
→ paInstallCrashHandler → ntdll, so raw top-frame matching matches everything. Exclusions
live in crash_title._is_handler_frame, crash_history.ABORT_PLUMBING and
dedupe.LUA_BRIDGE_PLUMBING (paluaerrorhandler.cpp, paluabindfunction.h,
paluaclassbinder.h, lj_api.c, paluabindinvoker.h).
Adding a frame to the plumbing list is the safe direction (fewer merges). Measure before and after on a copy of the DB:
con = sqlite3.connect("file:.antfarm/antfarm.db?mode=ro", uri=True)
con.backup(sqlite3.connect(tmp)) # never open the live DB for writing
before = dedupe.groups(Store(tmp, log), log)
Procedure
- Dry run —
.venv/Scripts/python.exe -m antfarm.queen_cli dedupe. Default writes nothing. Read the per-group plan: moves, empties,retained, info.md line delta. The dashboard's Ready-to-Fix panel andGET /api/duplicatesshow the same groups. - Stop the colony —
.venv/Scripts/python.exe -m antfarm.queen_cli stop, then confirm.antfarm/run.lockis gone.--applyrefuses while it runs (it re-ingests the share on a cycle) unless--force. - Record pre-state — top-level folder count per share, ticket counts by state,
count(*) where duplicate_of<>''. - Apply —
.venv/Scripts/python.exe -m antfarm.queen_cli dedupe --apply. It moves the timestamped occurrence bundles into the canonical folder, mergesinfo.md(combining the## Machinestable and updating*First seen:*) and removes the emptied duplicate folders. Variants:--tickets-only(no share writes),--group HANDBALL-176,--undo BATCH_ID. - Restart the colony detached, as in
antfarm-colony-ops(from Git Bash usecmd //c; single-slashcmd /cruns nothing) — and verify the churn is closed: total ticket count andmax(created_at)must be unchanged. A rise means the identity pin is not being written or read. - Expect
/api/duplicatesto report 0 groups.
Undo log is .antfarm/crash-merge-undo.jsonl — local on purpose, because the share may be down
when the undo is needed and a log on the share could sit inside a folder the merge removes.
Automatic linking at ingest
Automatic ticket linking is on by default (pipeline.link_duplicates: true in antfarm.yaml):
- Ingest only absorbs tickets created or reopened by that specific pass.
- Incumbent tickets already on the board always keep their write-ups and canonical status.
- Tickets in
needs_clarification,needs_file, or with open questions are never absorbed. - The automatic pass never moves share folders — share consolidation stays an explicit
dedupe --apply.
Verifying counts
Folder-count arithmetic that does not add up is usually a second project's share
(antfarm.yaml configures Handball 27 and Cricket 8). Check the rmdir paths' parents in the
undo log before assuming folders appeared or vanished:
top = [p for p in rmdirs if p.parent == share] # the rest are another project
Interpreter
.venv/Scripts/python.exe, never a bare python; why a bare one fails silently is in
antfarm-colony-ops.