AGENT SKILL

seatkeeper

Keep MAIN MEMORY continuous across sessions — mount the fenced auto-memory, hash-diff it against the durable project seat, reseat (copy) stranded notes home, mirror offsite, and grow new notes by content-identity. Use for "reseat memory", "sync memory to the seat", "run the memory manifest", "seatdiff", "is my memory stranded/drifting", "my notes aren't persisting across sessions", or a session-close memory snapshot.

License 许可
GPL-3.0
所属仓库
inputloop/garden
最近更新
2026-10-14
Stars
★ 0
Forks
0

01 前言 Front Matter

name
seatkeeper
description
Keep MAIN MEMORY continuous across sessions — mount the fenced auto-memory, hash-diff it against the durable project seat, reseat (copy) stranded notes home, mirror offsite, and grow new notes by content-identity. Use for "reseat memory", "sync memory to the seat", "run the memory manifest", "seatdiff", "is my memory stranded/drifting", "my notes aren't persisting across sessions", or a session-close memory snapshot.
preferences
[ "skill.seatkeeper.<hash>" ]

02 资源文件 18

03 SKILL.md 正文

seatkeeper — keep MAIN MEMORY continuous across sessions

Corwin installs it his way (Settings → Capabilities, or via the polygram forge). It is delivered as a file, never auto-saved.

STATUS — pre-wear (define stage). Path: Model → LENS → define → build → test → cool. Lensed 2026-09-01. The lens PASSED the core (hash-identity · multi-copy · the mount workaround — all proven this session) and flagged three hardenings to build+test before wearing: (1) fail loud on an empty or absent mount — never report success on 0 copies (it silently no-op'd once); (2) a trigger / cadence (session-close) so it isn't a hand-run chore that lapses — the lapse is what made the mess; (3) the SAVANNAH carve-out exclusion + dedup awareness (hash-identity preserves everything, it does not consolidate meaning). Verdict: sound as an interim bridge toward the real destination — canon riding as an always-loaded skill/PI, growable notes in the seat. Hardenings built 2026-09-01 (fail-loud · trigger · carve-out/dedup) and a SAME-NAME-DIFF-preserve fold 2026-09-02; tested on synthetic fixtures 2026-09-02 (reseat / manifest-only / empty-mount-fail-loud all green; the preserve fold verified). Pending test on a live mount + your enable — the last step.

Cowork's built-in memory (the "auto-memory") lives under the packaged app-data root, which the desktop app fences: notes written there can strand when the data root migrates, and the folder is not reliably reachable from the shell. The durable home is the project seat — a mountable, version-controlled folder outside any app-data root — plus an offsite copy (e.g. Google Drive). Memory continuity is keeping these in sync, keyed by content-hash identity: a path is an address, a hash is an identity — key on the hash, carry the path as a hint, and motion costs only a hint refresh.

The stores

  • A — auto-memory (…/spaces/<id>/memory, under the packaged app root): the reflex write-target each session; fenced, per-session, migration-fragile.
  • B — the seat (a mountable non-root folder, e.g. a project's memory/): durable, versioned; the seat of record.
  • Offsite (a Google Drive folder via its MCP connector): a copy, never the only copy — also a second trust source.

No store is the single point of failure, and git-as-transport is assumed never to exist — the work must be durable in its own right and degrade gracefully.

The signals (semaphores)

SEAT (the authoritative home; only the keeper, or a successor system, moves it) · MANIFEST-MEMORY (the hash-keyed ledger) · STRANDED (reachable only in a fenced/retired root, not yet in the seat) · GROW (the accretion protocol below). A really-good item also gets a MODEL to go with its semaphore.

The moves

  1. Mount A — and know what each route can prove. A sits under the packaged app root, so the routes differ and so do the words each one earns. (a) An explicit-path request is refused. request_cowork_directory on the …/spaces/<id>/memory path comes back "Cannot mount a folder that is, or is inside, Claude's own application data." That is a policy refusal, not a fickle mount, so retrying the same path in the same session buys nothing (receipt: SEATDIFF-2026-09-04b.md §"What the tooling did and didn't do", row 1 mount A). Still attempt it once per session before calling it closed — which folders are mounted is a per-session runtime fact, not a property of the project ([[verify-mount-access-fresh-per-session]]), and the morning account of the same day recorded A "mounted via request, read-only" (SEATDIFF-2026-09-04.md), so either the fence moved or the picker was used. One attempt dates the line. If a mount does appear, it can relocate or drop between shell calls — verify it (ls .auto-memory/*.md | wc -l) before trusting it. (b) The picker shows the keeper no purpose and returns whatever he clicked. request_cowork_directory with no path opens the native folder picker; the dialog names no reason and the call hands back the folder that was clicked — once (on the author's machine) F:\claude\AS4CE\AS4CE0002\New folder, which did not exist and was not A (same table, row 1 mount A, picker). So say in chat which folder and why before popping a dialog on him (SEATDIFF-2026-09-04b.md CARD 1, [[dialog-tools-need-visible-line]]), and check the returned path is actually A: a picker result is a claim to verify, not a mount. (c) The file-tool fallback works but cannot hash the source — its receipt word is "copied, re-read", never "verified". Glob + Read over A read all of it (103 notes enumerated on the evening the mount was refused, row 1 fallback), but the sandbox cannot run sha256sum against A, so the content manifest is NOT ATTEMPTED and same-name pairs stay uncompared (row 2 manifest by content: 96 pairs). A copy made this way has its destination hashed and its source unhashable: write "copied, re-read". Keep "copied, verified" for the case where both sides were hashed, as the morning run could (SEATDIFF-2026-09-04b.md §"Two piles, closed" (2)). Print the weaker word and name the check you could not run. (d) The session-close trigger must be a hook, not a memory. Move 7's cadence only exists if something outside this skill fires it: skills have no event hooks, they fire on conversational triggers (skills/skilltank/SKILL.md:27, skills/ethicsminions/SKILL.md:39). A session-close run carried as a note to remember is the hand-run chore that lapsed and made the mess (STATUS above, hardening 2). Until the keeper wires a real close-time trigger, report the cadence as unwired and never count a run that happened only because someone remembered. Fail loud: if the mount is empty or absent, STOP and say so — never report success on 0 copies; a silent no-op is worse than an error (proven 2026-09-01).
  2. Manifest. Hash every note in A and B (sha256) and classify each A-note vs the seat: TRACED (same content already in the seat) · SAME-NAME-DIFF (name in seat, content differs) · ORPHAN (only in A). Write MANIFEST-MEMORY.tsv to the seat; ORPHAN = a bring-in candidate.
    A=.auto-memory; SEAT=<seat>/memory
    for f in "$SEAT"/*.md; do sha256sum "$f"; done | awk '{print $1}' | sort -u > /tmp/bhash
    for f in "$A"/*.md; do h=$(sha256sum "$f" | awk '{print $1}'); grep -qxF "$h" /tmp/bhash \
      && echo "TRACED $f" || echo "ORPHAN $f"; done
    
  3. Reseat. Copy the ORPHANs A→seat, name-collision-safe (never overwrite an existing name) and verify by hash afterward. Nothing is deleted; the seat wins ties; extra copies are fine (bytes are cheap). On a SAME-NAME-DIFF, never drop the divergent copy — that is the one thing a continuity tool must not do. Update the seat in place only when the auto copy is a strictly newer edit of the same note; otherwise PRESERVE the auto copy beside the seat original under a non-clobbering name (<stem>.auto-<short-sha>.md), leave the seat original untouched, and FLAG the pair for the keeper to reconcile. Two divergent notes ride as two until the keeper merges them. Carve-out: never reseat a SEAT@SAVANNAH (enclave / embargoed) note into the general seat. Dedup awareness: hash-identity preserves every note but does not consolidate meaning — two notes on one topic ride as two; flag topical dupes for the keeper, never silently merge.
    for f in "$A"/*.md; do bn=$(basename "$f"); [ "$bn" = MEMORY.md ] && continue
      [ -e "$SEAT/$bn" ] || cp "$f" "$SEAT/$bn"; done
    
  4. Multi-copy. Mirror the model + key notes to the offsite store by content (same hash = same note). The seat is canonical-current and wins ties.
  5. Grow (per new note). on-ramp → pin (sha256) → semaphore (+ a MODEL if the item is really-good) → copy (multi-store) → link ([[wikilink]]) → reflect. Grow from durable sources, never from random bias.
  6. Index. Consolidate the seat's MEMORY.md. If it is signed (a sibling .asc), the re-sign is the keeper's keystroke — do the content, hand back the signature; a stale .asc is a raised flag to diagnose, not tampering.
  7. Trigger. Run on his word — "reseat", "rotate", "consolidate the seat" — and never on your own read that a session is closing: there is no "closing a session" in C4; a session is the assistant's bookkeeping, not the user's (his rule, 2026-09-23), and closing is pended to him, never proposed. A continuity tool that lapses recreates the very drift it was built to cure, so when he says the word, do all of it.

Two disciplines that ride along

  • Resolve, don't punt. Persona and definition questions ("is this the node or the persona?") are resolved by a reflection pass — a stance and its inverse held against a baseline — not bounced back to the keeper. Surface to the keeper only on a genuine divergence, not for every definition.
  • Ground every "should". A hold or a rule must cite the keeper's actual words, a superseding statement, or a hard floor (ToS / safety) — never a self-authored rule cited back as the keeper's. Budget the compute and stop when the work is done or at a bound. Never touch entrusted systems, family or third-party material, or anything irreversible on your own initiative.

How this skill can fail (falsifiers)

  • If a session opens with A mounted read-write by an explicit-path request_cowork_directory alone, (a) in move 1 is stale — re-run the request, paste the result, and date the line.
  • If sha256sum returns a digest for a file under A from the sandbox shell, (c) is stale — the content manifest is available again and the receipt word goes back to "copied, verified".
  • If a reseat reports copies but ls <seat>/*.md | wc -l is the same before and after, the run did nothing and the fail-loud rule did not fire — treat the report as false and re-run by hand.
  • If the session prints its memory directory as the seat path (auto-memory relocated onto the seat), this skill should not have fired at all: there is one store, and moves 1–4 are ceremony over a copy of itself — say so and stop.
  • Off-switch: say "don't reseat" (or "memory is one store now") and it stops for that session; remove the skill, or the seat's memory/ folder, and it has nothing to run against. It never runs unasked — a skill fires on a conversational trigger, not on its own.

A remembered "no" to seating a folder (extension edition, 2026-10-10)

When someone declines to seat a folder (st1dt, the seating detector, proposes it and asks once), the "no" is kept as a preference this skill reads, so the folder is not asked about again:

  • Key: skill.seatkeeper.<hash>, where <hash> is the first 12 hex digits of sha256 over the folder's path, with backslashes made / and one trailing / dropped (printf %s "$path" | sha256sum | cut -c1-12).
  • Value: no: <the folder's path>, so the path is there in words. The server refuses a value whose path does not hash to the key.
  • Write it with pref_set on the person's own turn (turn = their name), only after they said no.
  • Read it before proposing a seat: pref_list lists each one under declined and says "you said no to seating <path> on <date>". When one is in force for the folder, say so and ask nothing.
  • Take it back with pref_unset skill.seatkeeper.<hash>; the earlier rows stay as the record.
  • When usage capture is on, the "no" is also mirrored as one pref row in the frame (FRAME.md); only the preference acts, and the mirror is only counted.

<!-- footer: minted by bin/ft1mt from conventions/house.env and conventions/FOOTERS.org; edit those, not this --> Version 1.1.2: edition revised from 1.1.0 on 2026-10-10; spine: 2026-09-01.

Copyright 2026 Corwin Brust corwin@bru.st and Contributors; CC-BY-SA4: https://creativecommons.org/licenses/by-sa/4.0/ (scripts GPL-3.0-or-later); Mentat (for Corwin; claude-opus-5-5) <!-- /footer -->

04 同仓库更多 Skills