memhouse 0.18.12 → 0.20.1

package/AGENTS.md

@@ -1,149 +1,8 @@
-# Working on memhouse
+# Installing and deploying memhouse (for agents)
 
-Conventions for any coding agent in this repository — Claude Code, Codex, Gemini CLI,
-OpenCode, whatever comes next. `CLAUDE.md` points here; this file is the canonical copy.
-
-Constellation-wide rules (worktrees, branch naming, which remotes are frozen) live in the
-parent directory's [`CLAUDE.md`](../CLAUDE.md) and are not repeated here.
-
-
-## Release lines — read before tagging, publishing, or merging a release
-
-memhouse shipped two lines while the first production house ran the pre-0.18 layout. Since
-its conversion there is one line; the rules below are kept because a second line can happen again.
-
-| line | git branch | npm dist-tag | versions | who installs it |
-|---|---|---|---|---|
-| the 0.18 line | `release/0.18` (cut at v0.18.11; `main` stays the default branch) | `latest`, and `team` as an alias | 0.18.x | everyone — `npm install -g memhouse` (the production house was converted to the one layout with `memhouse convert`; `release/0.17` is retired) |
-
-Rules, each learned once:
-
-- **Tag on the line's branch, publish from the tag, from the primary checkout.** A 0.17.1 was
-  built on `master` (the retired default) and published from a worktree: the branch and the
-  registry disagreed for a day, and the package lost its `gitHead`. `misc/publish-guard.js`
-  runs as `prepublishOnly` and refuses a publish whose version, `--tag`, branch, or HEAD tag
-  disagree with this table. If you need a new line (0.19.x), edit the table **and** the
-  guard's `LINES` in the same commit.
-- **A fix lands on one line.** Porting to the other is a deliberate merge or cherry-pick with
-  its own CHANGELOG entry, never assumed. The one-layout (0.18) renamed every room to
-  `<member>_*`; anything touching room names ports with care (`house/HOUSE.md`).
-- **`memhouse update` follows the channel a machine was installed from** (`MEMHOUSE_CHANNEL`),
-  so a machine on the old line never receives a team build by accident, and vice versa.
-- **`master` is retired.** It was the pre-0.18 default; `main` is the default now. Do not
-  branch from it, tag on it, or push to it.
-- **A publish is live when `npm view` says so.** `npm publish` prints `+ memhouse@x.y.z`;
-  confirm with `npm view memhouse dist-tags --prefer-online`. A version that seems missing
-  right after a publish is registry lag or a stale local npm cache (`npm cache verify`),
-  not a "staged" publish. That misdiagnosis cost three releases.
-- **Which token.** The publish token is a secret reference, resolved at the moment of use
-  and lent to the one `npm publish` (a temporary userconfig); never print it.
-- **One owner per line at a time.** Before releasing, check for another
-  session working on the same repo and agree who cuts it; the release record lives in `CHANGELOG.md`
-  on the line's branch and in `drills/`.
-
-## What memhouse is
-
-A local shipper parses coding-agent sessions (16 adapters covering 18 apps) and writes them as typed rows
-into a ClickHouse database the user owns. A **house** is a database; its **rooms** are
-tables, one set per **member** and named for them (`mem.alice_messages`). A member holds
-one grant — `ON mem.alice_*` — and nothing else in the database; nobody is ever granted the
-database itself. `roomNames()` in `memhouse/house/house.js` is the only place a table name
-is produced, and `memhouse/provision.js` is the only description of what a member is
-granted — the live path executes it and `--print-sql` prints it.
-The ClickHouse is whichever one the user points at — local container, their own server,
-or ClickHouse Cloud; memhouse runs no service of its own and proxies nothing. There is no
-LLM anywhere in the write or read path.
-
-`memhouse/DESIGN.md` argues the design. `TERMINOLOGY.md` is the canon for names — if a
-thing has a name there, use it and do not invent another.
-
-## The three instruments, and what each cannot see
-
-| instrument | asks | blind to |
-|---|---|---|
-| `npm test` | is the logic right | anything needing a server |
-| `misc/invite-matrix.sh` | does it work against a real ClickHouse | anything a person has to read |
-| **a drill** | can someone actually use it | nothing — which is why it is slow |
-
-**[`drills/DRILLS.md`](drills/DRILLS.md) is the drill protocol**, borrowed from
-[`kernel`](https://github.com/agent-realm/kernel). A drill is memhouse being used by
-agents in real roles who have **not read this repository**. Four things it is easiest to
-get wrong:
-
-- **Write a plan first** — roles staffed, and the situation to create.
-- **A plan must contain no expected results**, or it is a test that knows the answer.
-- **Agents must run where this repo is unreachable**, or they read the source instead of
-  reporting the confusion. The installed package's `README.md` and the `/mem:*` skills
-  *are* in scope; a real user has those.
-- **Never fix memhouse mid-drill**, or the record describes a build that never existed.
-
-Run one after the matrix passes, and before any release that changes a surface a person
-touches.
-
-## Things this codebase has learned the hard way
-
-Each of these cost a real incident. They are in the commit log with the measurements.
-
-- **A probe that says yes to everyone is worse than no probe.** `invite` tested
-  `SELECT 1 FROM system.users`, which every member passes since they all hold `SHOW USERS`
-  — so it announced success and then failed at `CREATE DATABASE`.
-- **Scope is part of a privilege.** A member holds `CREATE DATABASE` inside their own
-  database and can mint nothing. Match the grant *and* its scope, never the name alone.
-- **A flag that is ignored rather than refused** is how a `--dry-run` typo becomes a real
-  migration. Every command declares its options in `memhouse/flags.js`.
-- **One `data` event is not one keystroke.** Reading a hidden password chunk-at-a-time
-  hung forever the moment anyone pasted.
-- **Row policies are OR'd, so a permissive catch-all fails open.** A second scoped share
-  saw everything. See `memhouse/share.js`.
-- **A skill that reasons about privileges in prose gets it wrong.** Put the logic in a
-  command and have the skill call it — that is why `memhouse whoami` and `memhouse share`
-  exist.
-- **Never print a credential.** memhouse ships this transcript into the house being
-  administered, and the archive is insert-only. Expand a shell variable; never a literal.
-
-## Testing before you claim it works
-
-```bash
-npm test                                            # syntax gate + unit checks
-misc/invite-matrix.sh <url> <admin-user> <password> # against a real ClickHouse
-misc/origin-matrix.sh …                             # the import/ship data guarantees
-memhouse nightly --out /tmp                         # a real installable tarball
-```
-
-`node bin/memhouse.js` from a checkout is **not** the product — the files are laid out
-differently and the design is sitting next to it. Anything about installation, first run,
-or what a user sees must be verified from a built package on a machine without a checkout.
-
-CI runs `npm test`, both matrices and a headless e2e across ClickHouse 25.11 and 26.7, on
-ubuntu only. macOS and Windows are untested there; say so rather than implying coverage.
-
-## Releasing
-
-Version lives in `package.json`, `package-lock.json` and the plugin manifest
-(`memhouse/delivery/plugin/.claude-plugin/plugin.json`). All three move together, along
-with the memhouse.io badges. `## Unreleased` in `CHANGELOG.md` becomes
-`## <version> — <date>`. Tag `v<version>`, push the tag, then publish. A new user-facing
-capability is a **minor** bump; so is a behaviour change like refusing a flag that used
-to be ignored.
-
-Every release also has:
-
-- **A release document**, `docs/releases/v<version>.md`: what changed, what a user must
-  do, the upgrade steps, deprecations with their removal version, and what was deferred.
-  The CHANGELOG entry links to it, and the GitHub release uses it as its notes
-  (`gh release create v<version> --notes-file docs/releases/v<version>.md`).
-- **The documentation standard:**
-  - a short README (what, why, how);
-  - `docs/`, with tutorials and guides for common operations;
-  - `examples/`, from smallest to full, each with its own README;
-  - the agent install section below.
-
-  A change to a command, a flag or a default updates the guide that describes it in the
-  same PR.
-- **A drill** when the release changes a surface a person touches (see above), recorded
-  in `drills/`.
-
-## Installing and deploying memhouse (for agents)
+This file ships with the npm package. It is for an agent installing memhouse for its
+principal. Working on memhouse itself? The conventions for this repository are in its
+`CONTRIBUTING.md`, which is not part of the package.
 
 How an agent installs memhouse for its principal, unattended. Every command takes `--yes`
 and most take `--json`. Follow the first route that applies, and never invent a house.
@@ -166,7 +25,7 @@ with-secret MEMHOUSE_ADMIN_PASSWORD=<reference> -- \
 memhouse deploy --local --yes
 
 # Then verify, and keep it running:
-memhouse status --json        # connection, rooms, last pass
+memhouse status --json        # connection, schema, tables, last pass
 memhouse whoami --json        # which credential, and what it may do
 memhouse service install      # survive reboots
 memhouse plugins install claude --yes
@@ -175,19 +34,31 @@ memhouse plugins install claude --yes
 `memhouse prompt --install` prints an install prompt rendered for the current machine,
 with its state and the one route that applies.
 
+Which layout an install lands on is not yours to choose: a new house is schema 3, a house
+on the 0.18 rooms keeps them, and an ambiguous house refuses. Moving a 0.18 house to
+schema 3 is an administrator's procedure, in order: `memhouse migrate --v3-prepare`,
+`--v3-convert`, `--v3-reconcile`, `--v3-shares` (as the admin: `MEMHOUSE_ADMIN_USER` and
+`MEMHOUSE_ADMIN_PASSWORD`, like `invite`), then `--v3-switch` on every machine, as its member
+([`docs/guides/upgrade-to-0.20.md`](docs/guides/upgrade-to-0.20.md)). `memhouse update`
+never does it.
+
 Rules:
 
 - **Never put a secret in the conversation, in argv, or in a file you write.** memhouse
   ships transcripts into the house, and the archive is insert-only. Take passwords by
   reference (`with-secret`, `MEMHOUSE_ADMIN_PASSWORD`, `--admin-password-file -`). Never
-  print `~/.memhouse/env`. If a secret reaches a transcript, say so and recommend rotation.
+  print `~/.memhouse/env`, nor the dashboard's `~/.memhouse/dashboard-token` and
+  `~/.memhouse/dashboard.html` (`memhouse dashboard` opens the dashboard without showing the
+  token). If a secret reaches a transcript, say so and recommend rotation.
 - **Never guess a house.** With no URL from the principal and no invite file, stop and ask.
   Do not point at `localhost:8123` because something answers there.
 - **Destructive commands need the principal's explicit yes, for that object:**
   `memhouse reset`, `deploy --down` (deletes the local house's data),
-  `uninstall --full-removal`, and any `DROP`/`DELETE` in SQL. Look first, report the
-  counts, then ask.
-- **Verify from the outside.** `status --json` and a row count in `mem.<member>_messages`
-  are the evidence that it works. A started daemon is not.
+  `uninstall --full-removal`, the `migrate --v3-*` steps (they fence writers, rename a
+  member's rooms and move its grants), and any `DROP`/`DELETE` in SQL. Run the `--dry-run`
+  first, report the counts, then ask.
+- **Verify from the outside.** `status --json` and a row count in
+  `mem.<member>_sessions FINAL` (schema 3) or `mem.<member>_messages` (the 0.18 rooms) are
+  the evidence that it works. A started daemon is not.
 
 The user-facing version of all this is in [`docs/guides/install.md`](docs/guides/install.md).

package/CHANGELOG.md

@@ -4,18 +4,252 @@ Versions before 0.8.0 were beta-only. Beta installs of 0.7.x and earlier should
 uninstall and reinstall — the 0.8.0 layout is new, and an in-place `update` +
 migration path is planned work, not a promise the old versions can cash.
 
-## 0.18.12 — 2026-10-29
-
-Release notes: [`docs/releases/v0.18.12.md`](docs/releases/v0.18.12.md).
-
-- **Skills:** each `/mem` skill's `allowed-tools` is restricted to the read-only memhouse
-  commands it teaches, instead of `Bash` as a whole. Any other command goes through the
-  normal permission prompt. `memhouse update` refreshes installed skills; then
-  `/reload-plugins`.
-- `misc/skill-allowlist-test.js` and `misc/env-guard-test.js` join `npm test`. Test tooling
-  reads its deploy targets from an untracked `deploy/targets.env`, and fixtures use stand-ins.
-- The 0.18 line is cut from `release/0.18`.
-- No drill was run for this patch release.
+## 0.20.1 — 2026-11-02
+
+Release notes: [`docs/releases/v0.20.1.md`](docs/releases/v0.20.1.md).
+
+- **Dashboard: a browser that refuses sessionStorage still opens it.** The page keeps the token
+  in memory for that load when sessionStorage is refused; it used to show the "open it with
+  `memhouse dashboard`" page instead.
+- **doctor names the channel when it says a release is the latest.** An unpriced model's hint now
+  says "the latest release on '<channel>'", so an install pinned to another line is not told it
+  is on the newest release overall. Offline, the hint no longer claims the install is current.
+- **/mem skills: the raw-SQL recipe runs `memhouse room-sql` as written.** The reference's
+  schema-2 recipe wrapped it in a `q()` shell function, so every call asked for approval despite
+  the skills' pre-approval. It now runs `memhouse room-sql <<'SQL'` directly.
+- **install, invite and `deploy --local` name a refused credential up front.** When the server
+  refuses the credential (ClickHouse code 516, or 194 for a missing password), they now stop
+  before writing anything with "could not authenticate as '<user>' at <url>", instead of reading
+  the house as new and failing at the next step. Code 194 counts as an authentication failure
+  everywhere 516 does.
+- **`memhouse <command> --help` prints that command's help.** It used to print the whole page
+  for every command. A command with no entry of its own still gets the whole page, and
+  `memhouse q --help` lists the named queries (the same as `memhouse q list`).
+- **`memhouse share --list` also lists the shares you received.** Beside who reads your
+  sessions, it names each member whose sessions this credential can read, with the command
+  that reads them (`memhouse q recent --member <them>`); `--json` adds `received`. An
+  administrator is told it reads every member's sessions instead.
+- **`memhouse discover` flags an unreachable house only when it is this install's.** The
+  default local address (`http://localhost:8123`) is probed as a discovery; when nothing
+  answers there and nothing is configured to use it, it is a plain line, not a ✗. `--json`
+  marks each endpoint `configured`.
+- **`memhouse update` on a machine with no member install checks no house.** An
+  administrator's machine (only `MEMHOUSE_URL` and the admin variables) used to run the schema
+  heal and the migration check as the default member and print its authentication failures as
+  the house's state. It now says there is no member install here and skips those steps.
+- **`memhouse members` tells a share from a member's own tables.** Each account's line names its
+  own tables, then each member whose sessions it reads ("reads alice's sessions (a share,
+  scoped)" or "full", where the row policies can be listed), then any other grant, which
+  memhouse does not make itself. A scoped share used to read as a plain list of another
+  member's tables. `--json` adds `own`, `shares` and `other`.
+- **A new minor version's first run says so.** The first CLI call of a newer major.minor on a
+  machine prints one line to stderr, naming the version it replaces and the shipped release
+  notes to read. Patch releases, downgrades and fresh installs say nothing; a stamp in the
+  memhouse home (`version-seen`) records the last version that ran. An install from before the
+  stamp is read from its home: a 0.18 install moving to 0.20.1 is told "it ran 0.18 or
+  earlier".
+- **`memhouse plugins install codex`.** Codex gets a short, marker-delimited block in its
+  `AGENTS.md` (never read or print the env file or a `MEMHOUSE_*` value; `MEMHOUSE_URL` is the
+  house, `MEMHOUSE_PORT` the dashboard; past sessions go to the skill) and a `memhouse` skill with
+  the read recipes, under `CODEX_HOME` or `~/.codex`. `plugins remove codex` takes out exactly
+  those: the rest of `AGENTS.md` stays byte-identical, and a `skills/memhouse` that memhouse did
+  not install is left alone. Opt-in: nothing is written there unless it ran; the first run of a
+  new version refreshes a copy it finds installed. The `/mem` skills stay Claude Code's.
+- **An invite's password expires.** `memhouse invite` sets the new member's password to stop
+  working after 7 days (`--expires <days>`, 1 to 365), by the server's clock. The invite's
+  output, the invite file and `MEMHOUSE-INVITATION.md` say when. `install --env` lifts the
+  expiry when it changes the password, and when `--keep-password` keeps it. An expired invite
+  is refused at install with "this invite expired at … ask whoever invited you for a new one",
+  before anything is contacted or written. Inviting the same name again says the unused
+  account must be removed first (`DROP USER <name>`); its tables are kept and reused.
+  `invite --print-sql` prints the same expiry, with a placeholder for the date.
+
+## 0.20.0 — 2026-10-29
+
+Release notes: [`docs/releases/v0.20.0.md`](docs/releases/v0.20.0.md). Read its "Read this
+before upgrading" first: `memhouse update` installs 0.20 and does not convert the house to schema 3;
+converting a 0.18 house is an admin step, and a house in daily use is converted on its own
+schedule.
+
+- **Schema 3.** A session is keyed by its store (machine, editor, the editor's data folder or,
+  for OpenCode, its database file, and the session's id). Each member has the six conversation
+  tables of the harness-sessions-sql profile (`<m>_session_parts`, `_session_usage`, `_session_links`,
+  `_session_lifecycle`, `_calls`, `_sessions`), plus a capture archive in `<db>_archive`
+  (`<m>_session_units`, `_session_sources`) holding every byte as captured. A new house is
+  schema 3; joining a house follows its layout.
+- **Ten editors through normalisers** (agy, Copilot CLI, Cursor, cursor-agent, Devin, Gemini
+  CLI, Goose, OpenCode, VS Code, Zed); Claude Code and Codex natively; the rest through the
+  0.18 readers.
+- **The host** is sessions-model/host/v0. The 0.18 id is kept as this machine's alias.
+- **Upgrade commands:** `memhouse migrate --v3-prepare`, `--v3-convert`, `--v3-reconcile`,
+  `--v3-shares`, `--v3-switch`; each has `--dry-run` and refuses before changing anything.
+  `status` calls a shipper that started before the switch or the install stale, repeats
+  `memhouse stop && memhouse start`, and exits 1 while its passes fail; `instance` says when an
+  install is stranded on the 0.18 rooms; `q schema` says `none` with no house configured.
+- **Share on schema 3** is `SELECT` on exactly the six conversation tables, full or scoped. It
+  never reaches meta, events or the archive. The share commands fail closed.
+- **`MEMHOUSE_URL` is the house, `MEMHOUSE_PORT` the dashboard.** `--help`, the env file and the
+  configuration guide say so; the house's port is part of `MEMHOUSE_URL`. The `/mem` skills
+  install into Claude Code only, and `--help` and the install guide say how other agents read
+  the house (`memhouse q`, `memhouse room-sql`).
+- **Security: member names that collide are refused.** `mem.bob_*` also matches `bob_smith`'s
+  tables (read and write; true on 0.18 too). install and invite refuse such a name, doctor
+  names an existing pair. Admins: check your members.
+- **Security:** The dashboard now requires a per-user token; a loopback connection alone is no
+  longer enough. Open it with `memhouse dashboard`.
+- When the dashboard's port is in use, possibly by another account on this machine, the
+  dashboard takes the next free port up to 10 above and says which (`memhouse status` and
+  `memhouse dashboard` use it); a dashboard that cannot start is a warning from
+  `memhouse start`, not a failure.
+- **Reading an editor's SQLite store never writes into it** (0.18 could leave `-wal`/`-shm`
+  files in an editor's folder).
+- **`memhouse import claude-ai`** stores a claude.ai data export in the house: the `.zip`, the
+  unpacked export, or an archive folder filed by project. Each conversation becomes a
+  `claude-ai` session with origin `import`, in one batch per run.
+  - **A dry run unless `--write`.** It prints counts only.
+  - **A later export adds only conversations the house does not hold.** `--update` also
+    replaces the ones changed since.
+  - **Batches.** Each batch is recorded in the member's meta. `--list` shows them, and
+    `--undo <batch> --write` tombstones one.
+  - **Changed at the same `updated_at`.** A conversation whose content differs from what the
+    house holds, at the same `updated_at` (a summary claude.ai added later, a rename), counts as
+    changed. An export older than the house's copy never replaces it.
+  - **The export's part zips** (`conversations-000.zip`, `design_chats-000.zip`, …) are read as
+    given, or from their folder; the older single zip still works.
+  - **On the profile tables** (`<m>_session_parts`, `_calls`, `_sessions`) a conversation is a
+    session in the service's store: host `s-` plus 32 hex (sessions-model's service host for
+    `claude.ai`), `harness_path` `https://claude.ai`, adapter `memhouse-import`. The keys are the
+    parser's, as on layout 3.2, and so are the digest and `--update`. A development house's 3.2
+    tables are still written; a member with both is written on the profile.
+  - **A run that stops part way is finished by the next.** Sessions whose totals were not
+    filled are refreshed by the next `--write`, and a batch whose rows all went in then reads
+    `applied`. An undo records `undoing` before it starts, so a stopped undo is never taken for a
+    stopped write; running it again finishes it.
+  - Guide: [`docs/guides/import-claude-ai.md`](docs/guides/import-claude-ai.md).
+- `memhouse q` named queries; `q removed`; the dashboard and `/mem` skills read schema 3.
+- The `/mem` skills write every command as plain `memhouse …`, the form their pre-approved read
+  commands match, so those run without a prompt; a share you received is found with `SHOW TABLES`
+  and read with `memhouse q … --member <them>`, not looked for with `SHOW DATABASES`.
+- **Deleted sessions on the profile layout** (Claude Code and Codex in 0.20) are recorded as gone,
+  their records kept, only when the editor's reader answers that they are not there: a session
+  a short list leaves out but the reader still finds is captured as usual. Only a store whose list
+  names none of the sessions the house holds of it is refused whole.
+  - **The shipper log and doctor say so.** A store the pass refused prints a `vanish refused:`
+    line after the pass line and fails `memhouse doctor` until a later pass is clean; a listed
+    session's source the list left out that still reads is a `vanish warning:`, shown by doctor
+    without failing.
+- **doctor: an unpriced model is a warning.** It no longer fails `memhouse doctor`, and its hint
+  names `memhouse update` only when a newer release exists; on the latest release it says the
+  model's cost stays unknown until a release prices it.
+  - **`memhouse doctor --confirm-empty-store <editor>:<store>`** confirms a store whose every
+    session was deleted (an empty list is refused, since an unreadable store gives one too), for
+    the sessions its last pass refused.
+  - **The archive write** of a pass carries one `captured_at` set by the capture, and writes its
+    units before its read manifests, so a replay from the archive alone sees the pass it was.
+  - Guide: [`docs/guides/daemons.md`](docs/guides/daemons.md) (Logs).
+- Fixed in the release work: `invite` follows the house's layout; `deploy --local` builds the
+  schema-3 tables; a machine without Claude Code or Codex ships its other editors; example 04
+  backs up the archive; Gemini CLI tokens use the house split (cached tokens were charged
+  twice and thoughts not at all), through the usage flags `input_includes_cache` and
+  `output_includes_reasoning` (sessions-model §2.4); `--v3-convert` takes the cached tokens out
+  of 0.18's Gemini CLI input, where 0.18 counted them twice (0.18 never stored Gemini's
+  thoughts, which cannot be recovered).
+- The invitation page `invite` writes (`MEMHOUSE-INVITATION.md`) lists the tables the invite
+  created, by the house's layout, and names the env file the invite wrote (`--out` included)
+  instead of `ls invite-*.env`.
+- Fixed in the normalisers: Zed usage is the thread's `cumulative_token_usage`, one per-session
+  series kept as appended increments (it took only each turn's last request); Copilot CLI keeps
+  a capped reply's usage record and a compaction's own tokens; an unknown Zed message keeps its
+  place, so the anchored ids after it do not shift; a Copilot CLI fork ignores the markers it
+  copied from its origin. Both normalisers' `VERSION` rose, so a machine that ran a pre-release
+  re-derives its Zed and Copilot CLI sessions.
+- Known issues, retired when memhouse reads editors through the harness-session-adapters package: Devin
+  token totals exclude abandoned branches in 0.20; seven low-impact normaliser differences
+  (Copilot CLI effort, orphan tool results and encrypted reasoning; a VS Code synthetic call id;
+  an OpenCode redaction marker and v2 usage basis; optional fields left empty). See the release
+  notes.
+- Known issue: a session whose capture state (its `files` row and the machine's capture cache)
+  is lost after its log was rewritten repeats its time-less records in `turns`; its totals are
+  not affected. See the release notes.
+- Known issue: `memhouse import claude-ai` takes no lock. Run one import or undo per member at a
+  time; two at once can split a conversation between two batches or leave a batch's recorded
+  status wrong. See the release notes.
+- **Apache-2.0** from this release (below).
+
+
+- **A fork's copied prefix is billed once.** Forks copy their origin's messages, usage included.
+  The house used to bill them in both sessions; now the copies are verified on full content and
+  billed only in the origin.
+  - Explicit links: agy, Copilot CLI. On the profile tables memhouse resolves each into one
+    engine edge on the fork (`session_links`, owner `engine`), verifies it, and the rollup leaves
+    the fork's copied usage out of its totals. A fork captured before its origin is verified when
+    the origin arrives; a rewound origin gives the copies back.
+  - Forks without a pointer, found by content (OpenCode, VS Code, cursor-agent, Cursor, Goose):
+    a session whose first reply is an earlier session's in the same store is nominated, verified
+    the same way, and recorded as an inferred edge; it ends when the verification no longer
+    holds. A pass verifies at most 100 nominated forks and owes the rest to the next.
+  - The copies stay messages of the fork.
+  - A link an editor states (sub-agent, fork, rewind copy, clear, continuation) whose two sessions
+    are both in the same store is resolved into one engine edge on the profile tables
+    (harness-sessions-sql §4.5), stated on the child's side, the origin's or both. Only forks and rewind
+    copies are verified and billed once; the others are lineage. A link to another store gets no
+    edge, nor does one whose child cannot be named in the store (not stored by the editor, or an
+    id with two lives); one whose origin cannot be resolved gets an unverified edge.
+- **Removed content in view.** `memhouse doctor` reports what is kept as removed versions (a
+  revert, a rewind, a message deleted at the source), and the forks billed once.
+  `memhouse q removed` lists those messages.
+
+- **Read manifests (harness-sessions-sql §4.1).** Every read that opens a generation, or moves a cursor,
+  archives a manifest row beside its records. It lists the units the read delivered, so the
+  archive alone replays an emptied or shrunk source correctly: before this, a rewrite to empty
+  left no row in its generation, and a replay took the previous one. `replayArchive` is that
+  replay.
+  - A session file the harness deleted stays in the house, and is recorded once as
+    `source_gone` with `"policy": "kept"`.
+  - A file that comes back is read again from the start.
+  - Details: `docs/design/sessions-engine.md` (repository) item 8.
+- **A pass that stops and is run again archives its read once.** A source the stopped pass read
+  from scratch (no cursor yet, or a reset) was read again from scratch and archived under a new
+  generation, so the archive held the same units twice. A re-read that delivers exactly what the
+  generation the stopped pass opened holds now stays in it and moves only its position; other
+  content, or a source that vanished in the stopped pass, still opens the next generation.
+  Streams, snapshots and empty delta sources were affected.
+- **The profile tables follow harness-sessions-sql v1's types.** `adapter_ver` is the version string an
+  adapter states (`0.2.0`), not a number; a link's `child_stored` is null when its statement does
+  not say; an archive unit with no original bytes has a null `orig_sha256`. The schema step brings
+  an older table to these types (`MODIFY COLUMN`, these three columns only, and the old defaults
+  removed). A member retypes its conversation tables itself; the two archive tables need an admin
+  (`memhouse install --schema 3 --admin-user …`), and until then the capture writes them in the
+  old types and `--ensure-schema` says so. The capture reads an old numeric `adapter_ver` back as
+  the version it encodes. Versions have no order, so a row another version of the same adapter
+  wrote is an upgrade's, whichever is newer.
+- **An adapter upgrade re-derives a session from its archive, removed versions included.** The
+  first pass under a new adapter version replays the session's archived passes one by one under
+  that version (harness-sessions-sql §8, "Re-derive from the archive"):
+  - every record the new version makes is written again above the house's revs, at the new
+    version, whether it is current or a removed version;
+  - every key the new version no longer makes is deleted, a removed version included (before
+    this, removed versions kept the old version and the old keys, and an old removed key was never
+    deleted);
+  - a pass that also reads a change replays up to each cursor, then applies the change.
+
+- **memhouse is relicensed from MIT to Apache-2.0, from 0.19.0.** That covers the whole
+  repository: CLI, shipper, dashboard, tools and docs. Releases up to and including 0.18.11
+  remain MIT; that text is kept in `LICENSES/memhouse-MIT-until-0.18.11.txt`.
+  - `LICENSE` is the Apache License 2.0, verbatim from apache.org.
+  - `NOTICE` credits:
+    - agentlytics (Fatih Kadir Akın, MIT as declared in its README;
+      `LICENSES/agentlytics-MIT.txt`).
+      Every commit by an author other than Ramazan Polat predates the fork at `4cba3d8`
+      (2026-06-21);
+    - models.dev, the source of `pricing.json` (MIT; `LICENSES/models.dev-MIT.txt`);
+    - the dashboard's bundled packages, via `public/THIRD_PARTY_NOTICES.txt`.
+  - `package.json` and `ui/package.json` declare `Apache-2.0`. The npm package now also
+    ships `NOTICE` and `LICENSES/`, and CI's tarball audit checks for both.
+
+- **Security reports go to security@memhouse.io.** `SECURITY.md` and the troubleshooting
+  guide name it, and `package.json`'s `bugs` field carries it. `package.json` no longer names
+  a repository while the repository is private.
 
 ## 0.18.11 — 2026-10-10
 
@@ -84,7 +318,7 @@ subagents once.
   update for this install" (a tarball is on no registry tag), exited 1, and left the old
   shipper running the old code in memory. `--no-install` no longer needs a channel: the
   files are already in place, and the restart, migrations and schema check run. Found while
-  upgrading testbed from 0.18.9 to a nightly of this release.
+  upgrading a test VM from 0.18.9 to a nightly of this release.
 
 - **A subagent caught before its first line is written holds its session for one pass.**
   A subagent with no timestamped line ranked last. When its first line then landed with an
@@ -120,7 +354,7 @@ subagents once.
   - An admin given for this command (`--admin-user` or an exported
     `MEMHOUSE_ADMIN_USER`, with its password from a file, stdin, the environment or a
     prompt) still provisions at `--url`.
-  - `misc/invite-matrix.sh` points `--url` at a spy server and asserts it receives
+  - `misc/invite-matrix.sh` (repository) points `--url` at a spy server and asserts it receives
     nothing but a ping.
 - **The dashboard no longer goes empty on a quiet house (G11).** The stats tables carry a
   one-day TTL and were refreshed only by a pass that shipped something. So a house whose
@@ -141,7 +375,7 @@ subagents once.
     by `seq` arithmetic, which a subagent's block numbering defeats.
 - **`HOUSE.md`'s subagent queries keep to the current parse (A3).** They had no epoch
   filter, so a re-parsed session doubled each subagent's turns.
-- `misc/skill-recipe.sh` now runs these queries verbatim, extracted from the docs, against
+- `misc/skill-recipe.sh` (repository) now runs these queries verbatim, extracted from the docs, against
   fixture rows with a superseded parse and a subagent block.
 
 ## 0.18.10 — 2026-10-08
@@ -167,7 +401,7 @@ changed, the upgrade steps, and the status of the 0.18.10 pre-tag list.
     it. They also print the env-file lines to hand over, with `MEMHOUSE_INVITE=1` from
     `invite`, so the member's `install --env` offers rotation. A `--member-password` given
     with `--print-sql` is not echoed.
-  - `misc/invite-matrix.sh` checks the output of every install, invite and passwd path
+  - `misc/invite-matrix.sh` (repository) checks the output of every install, invite and passwd path
     against the real passwords. A unit test fails on any print call that interpolates a
     password variable.
 - **Docs:** `memhouse passwd --member` does not exist. `--member` was accepted and ignored
@@ -295,7 +529,7 @@ The first release on **one line**: published to `latest` and `team` alike. The f
 - **A member's env repairs itself after the move.** `memhouse update` and the shipper's
   startup both notice that `<db>.messages` is gone and `mem.<name>_messages` exists, and
   rewrite `MEMHOUSE_DB=mem`; `doctor` names it. So the member's whole upgrade is one
-  command. Rehearsed by `misc/convert-matrix.sh`: a real 0.17.1-built house on 26.7,
+  command. Rehearsed by `misc/convert-matrix.sh` (repository): a real 0.17.1-built house on 26.7,
   converted, shares intact, isolation intact, the member's old env adopted, 24/24.
 
 ## 0.18.4 — 2026-09-17
@@ -385,7 +619,7 @@ the drill taught:
   — restart it: memhouse stop && memhouse start`, and a `✗` in doctor.
 - The invitation guide says what to do when an agent's `npm install -g` is blocked by a
   permission classifier: the person runs step 1, the agent continues from step 2.
-- `drills/DRILLBOOK-team-arrival-2026-09-10.md` and the drill record.
+- `drills/DRILLBOOK-team-arrival-2026-09-10.md` and the drill record (repository).
 
 ## 0.18.1 — 2026-09-08
 
@@ -591,7 +825,7 @@ possible.
   is left alone so a shared room still reads.
 
   Written with `[[:space:]]` classes, not `\b`: **BSD `sed` does not support `\b`**, and
-  the first version matched nothing on macOS. `misc/prefix-skill-recipe.sh` extracts
+  the first version matched nothing on macOS. `misc/prefix-skill-recipe.sh` (repository) extracts
   `q()` from the reference and runs it against both layouts — the doc is executable
   because it is load-bearing.
 
@@ -634,7 +868,7 @@ possible.
   the skill most likely to produce a number someone acts on, never mentioned it. Schema,
   connection, the three traps and the known measurement gaps now live in one file every
   skill points at. 11,388 words became 5,854.
-- **`/mem:recall` restores a name the canon already carried** (`TERMINOLOGY.md`: `memorecall`
+- **`/mem:recall` restores a name the canon already carried** (`TERMINOLOGY.md`, repository: `memorecall`
   retired in favour of `/mem:recall`, 2026-07-29). Its description carries every trigger
   phrase the three skills it replaces had, so auto-triggering does not narrow.
 - **The CLI no longer advertises a directory as a skill.** `plugins` listed
@@ -847,7 +1081,7 @@ possible.
 
 - **`memhouse relocate --to <url>` moves a whole house to a new ClickHouse.** The copy is
   a server-to-server `remoteSecure()` INSERT SELECT — the destination pulls each room
-  directly from the source over the native protocol, so the pilot's laptop is never in the
+  directly from the source over the native protocol, so the user's laptop is never in the
   data path and, crucially, **the shipper never re-ingests**: once the new host holds a
   faithful copy, the shipper's skip predicate sees every old session already present and
   ships only genuinely new work. Provenance is carried, not restamped
@@ -858,12 +1092,12 @@ possible.
   SOURCE is only ever read — a failed run leaves the old house intact — and a hard
   row-count gate must pass before the local config is repointed (the previous env is kept
   as `env.pre-relocate`). Flags: `--to-user/--to-password/--to-db`, `--from-native-host`
-  (when the destination's route to the source differs from the pilot's URL),
+  (when the destination's route to the source differs from the user's URL),
   `--from-native-port` (default 9440 TLS), `--insecure-native` (`remote()` + 9000),
   `--keep-shipper`, `--dry-run`, `--yes`. The source password reaches the destination's
   `query_log` (never this transcript) — rotate it after if those logs are not yours.
 - This is the host-to-host data-copy piece named in
-  `docs/design/host-repoint-reconciliation.md`; the reconciliation gate (detecting an
+  `docs/design/host-repoint-reconciliation.md` (repository); the reconciliation gate (detecting an
   *accidental* repoint at an empty host) remains deferred.
 
 ## 0.11.3 — 2026-08-20
@@ -954,7 +1188,7 @@ possible.
   it detects what the house needs: `--migrate` (or `--yes`) runs them unasked,
   an interactive session is asked once (the inner confirm is not repeated), and a
   non-interactive run only names them and prints the command — a cron must never start
-  a house-wide copy on its own. `update --no-install` skips the npm/git step for pilots
+  a house-wide copy on its own. `update --no-install` skips the npm/git step for users
   who already upgraded by hand and want the half a bare `npm i -g` leaves undone.
 - **`memhouse uninstall` asks first.** It used to start removing the service the moment
   it was typed. It now prints exactly what the chosen tier removes and keeps (the house
@@ -979,8 +1213,8 @@ possible.
   a pre-0.10 machine still writing.
 - **The plugin is `mem` now, and it answers questions.** Skills install as `/mem:ask`,
   `/mem:search`, `/mem:sessions`, `/mem:sql` (the `mem` short name belonged to the
-  retired memory-house and moves to the living product). `/mem:ask` is new: retrieve the
-  relevant past sessions, read the transcripts, answer with citations — the synthesis is
+  retired memory-house, the predecessor tool of memhouse, and moves to the living product).
+  `/mem:ask` is new: retrieve the relevant past sessions, read the transcripts, answer with citations — the synthesis is
   the agent's; memhouse itself still runs no LLM anywhere. The installer removes a
   pre-0.10 `skills/memhouse/` copy (only when it is provably ours) so the old and new
   namespaces never load side by side.
@@ -1005,7 +1239,7 @@ possible.
   looks for `origin`), but its delete-before-reinsert reaches every retained parse of a
   session it re-ships. `memhouse migrate` warns about this at migration time.
 - **`memhouse migrate-rooms`** rebuilds rooms whose sorting key predates this version:
-  copy, one atomic `RENAME`, old room kept as `<room>_pre_epoch` for the pilot to drop.
+  copy, one atomic `RENAME`, old room kept as `<room>_pre_epoch` for the user to drop.
   Nothing is deleted, and `user_id` is carried across explicitly rather than restamped —
   it is `MATERIALIZED currentUser()`, so a plain `INSERT SELECT` would reassign every row
   in a shared house to whoever ran the migration. Rows a running shipper writes during the
@@ -1031,7 +1265,7 @@ now the whole install.
 - **`--allow-scripts=better-sqlite3` is retired.** It was required on every install and
   every upgrade, npm never remembered it, and omitting it left cursor, zed, opencode,
   goose and antigravity reading zero sessions — indistinguishable from editors the
-  pilot does not have. That failure class no longer exists, so the probe that reported
+  user does not have. That failure class no longer exists, so the probe that reported
   it (`MISSING_BINDING`, `probeSqlite`) and the `missingBinding` flag on every
   `getAdapterErrors()` entry are removed. Per-store failures — locking, permissions,
   corruption, schema drift — are reported exactly as before.
@@ -1100,7 +1334,7 @@ now the whole install.
   messages/tool_calls sorting keys and bound in the shipper's clear. The bug this fixed
   destroyed 27,948 imported messages in a single pass.
 - Local house delivery (`deploy --local`, docker/podman), OS service integration,
-  the herdr-driven acceptance suite, `memhouse.io`.
+  the herdr-driven acceptance suite (repository), `memhouse.io`.
 
 ## 0.3.x — 2026-07-22 → 2026-07-31