Metadata-Version: 2.5
Name: capabilities-contract
Version: 0.46.0
Summary: The capabilities contract as a library: the code every capability shares, and the shared database layer - resolve which database a project uses, connect bound to its schema, and migrate each owner's tables once under a ledger.
Project-URL: Homepage, https://github.com/ai-cluster-one/capabilities-contract
Project-URL: Source, https://github.com/ai-cluster-one/capabilities-contract
Author: ai-cluster-one
License-Expression: Apache-2.0
License-File: LICENSE
Keywords: capabilities,migrations,postgres,psycopg,schema
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3 :: Only
Classifier: Programming Language :: Python :: 3.11
Classifier: Topic :: Database
Classifier: Topic :: Software Development :: Libraries
Requires-Python: >=3.11
Requires-Dist: psycopg[binary]<4,>=3.2
Requires-Dist: pynacl<2,>=1.5
Provides-Extra: dev
Requires-Dist: pytest>=8; extra == 'dev'
Requires-Dist: ruff>=0.6; extra == 'dev'
Description-Content-Type: text/markdown

# capabilities-contract

The capabilities contract as a Python library, in two parts. `capabilities_contract.capability` is the code every capability shares - the core, store and connections tiers - which each capability binds rather than carries a copy of; see [Capability](#capability). `capabilities_contract.db` is the one way the capabilities manager, its capabilities and ContextKit reach the shared PostgreSQL database: resolve which database a project uses, connect bound to its schema, migrate each owner's tables once under a ledger, and report where they stand. It does nothing else: no SQLite, no project registry, no records.

## Install

```
pip install capabilities-contract
```

Its dependencies are `psycopg[binary]>=3.2,<4` and `pynacl>=1.5,<2`, each imported lazily: `import capabilities_contract.db` does not import psycopg, only `connect`, `batch`, `migrate`, `status`, `store_major`, `announce` and `Follower` do, and only sealing or opening a secret of [the vault](#vault) imports PyNaCl. A capability declares it in its PEP 723 header by a version range, so a release inside the range reaches it with no edit of its own:

```python
# /// script
# dependencies = ["capabilities-contract>=0.9,<1"]
# ///
```

## Capability

`capabilities_contract.capability` holds the code every capability of the capabilities package shares, as three tiers, each one file under `capabilities_contract/capability/tiers/`: `core` (the declaration surface - summary, manifest, references, guide, ids - and the file, project and I/O plumbing), `store` (the records adapter for identifiers, policy, settings, connections and documents, and a store-mode project's records in its store) and `connections` (the credential cascade and the connection resolver, for a capability that implements connections).

A tier is not a module. `bind(namespace, *tiers)` executes each named tier, in order, in `namespace`, which a capability passes as its own `globals()` where it has defined its declarations:

```python
from capabilities_contract.capability import bind as _contract_bind
_contract_bind(globals(), "core")
_contract_bind(globals(), "store")
_contract_bind(globals(), "connections")
```

So a tier reads the capability's declarations (`NAME`, `SCOPE`, `STATE`, `CRED_KEYS`, `WRITE_VERBS` and the others the package docstring lists) and calls its hooks (`_cmd_connections`, `_inventory`, `_tables` ...) as the capability's own code would, and every name it defines is the capability's: a test that replaces one on the capability's module replaces it for the tier as well. A tier is compiled under the namespace's own `from __future__ import annotations`, as its text would be were it part of the capability's file.

A capability that declares `TABLES` answers `store_major` in its `manifest`: the store major the library it runs on reads (`STORE_MAJOR`, see [Migrate](#migrate)), and `owners`, every owner its `migrate` migrates, in order, its own last. Beside its `_tables()` builder it may define an `_owner_tables()` hook returning a list of `(owner, steps, major, minor)`: further owners whose tables it migrates, such as a library's tables its own steps reference. `migrate` migrates those first, in order, then its own, and `migrate` and `migrate status` answer for its own owner at the top level as before, with one answer of the same shape per further owner under `other_owners`; without the hook there is no `other_owners` and `owners` names its own alone. Its `migrate` and `migrate status` answer to the machine ceiling like any verb, except under the store upgrade (`CAPABILITIES_STORE_UPGRADE=1`), where a capability quarantined on the machine still migrates its tables; no other verb passes a quarantine.

Before its first read or write of a store, a process running the store tier reads that store's major once (`store_major`) and refuses a store at another major than `STORE_MAJOR`: a lower one with `store_upgrade_needed`, whose hint names `capabilities store upgrade` and `capabilities store upgrade --check`, and a higher one with `schema_too_new`, whose hint names `capabilities self-update`, then `capabilities update`. The refusal is a `StoreError` with that slug, raised where the tier raises a store it cannot reach, so a project's records are never read as empty from a store whose tables moved. An empty store passes, and under the store upgrade (`CAPABILITIES_STORE_UPGRADE=1`) nothing is refused.

`capabilities_contract.capability.store` is the store tier as a module of its own, for a process that carries no capability - the capabilities manager, a capability's bundled service, a test. `tier_path(tier)` gives a tier's file and `TIERS` names the three.

## Which database

`resolve_setting(project_root)` returns the database setting in force for a process standing in that project, as a `Setting`, reading and never writing. It asks three levels in order, and the first that answers is used whole, never completed from a lower level:

1. `project` - the `AGENTKIT_DB_*` keys in the project's `.env.local` and `.env`, `.env.local` winning key by key;
2. `environment` - the same keys in the process environment;
3. `machine` - the machine's store setting file, `$XDG_CONFIG_HOME/agentkit/store.json`, or `~/.config/agentkit/store.json` when `XDG_CONFIG_HOME` is unset, an `agentkit.store.v1` document whose format is the capabilities package's [SHEBANG.md, "The store setting"](https://github.com/ai-cluster-one/capabilities/blob/main/SHEBANG.md#the-store-setting). `setting_path()` gives its absolute path.

A level answers when it sets any `AGENTKIT_DB_*` key to a non-empty value, or, for the machine level, when the file exists. `project_root=None` means the process stands in no project, and the project level is not asked.

The keys (`KEYS`): `AGENTKIT_DB_URL`, `AGENTKIT_DB_HOST`, `AGENTKIT_DB_PORT` (default 5432), `AGENTKIT_DB_NAME`, `AGENTKIT_DB_USER`, `AGENTKIT_DB_PASSWORD`, `AGENTKIT_DB_SCHEMA` (default `agentkit`), `AGENTKIT_DB_SSLMODE` (default `require`), `AGENTKIT_DB_SSLROOTCERT`, `AGENTKIT_DB_TRANSACTION_HOST`, `AGENTKIT_DB_TRANSACTION_PORT`, `AGENTKIT_DB_TRANSACTION_POOLER`. Within one level `AGENTKIT_DB_URL` wins and that level's other keys are ignored, except `AGENTKIT_DB_SCHEMA` and the three transaction keys, which apply beside a URL.

The store has one address unless a level names a transaction address as well: `AGENTKIT_DB_TRANSACTION_HOST` and `AGENTKIT_DB_TRANSACTION_PORT`, or `transaction_host` and `transaction_port` in the machine file. It is the same store reached at another endpoint, typically a transaction-mode pooler such as Supabase's on port 6543, and it takes the level's own database, user, password, sslmode and schema; a transaction address naming only a port keeps the session address's host, and one naming only a host keeps its port. The address the level names otherwise is the session address. `Setting.has_transaction_address` says whether a level named one, that is whether the store's two addresses are named apart, `Setting.transaction_kwargs()` gives the driver's arguments for it (the session address's when none is named), and the report shows it under `transaction`. A transaction address alone does not make a level answer whole: a level setting only a transaction key is refused as incomplete, never completed from a lower level. A transaction port is 1-65535, a transaction host has no whitespace, and `sslmode` `disable` reaches only a local transaction host.

A store whose one address - no transaction address named - is a known transaction pooler, Supabase's (a host under `pooler.supabase.com` on port 6543), is transaction-only: every capability works through it, since its transaction connections reach it, and it has no session address, which only speeds up noticing changes (see [Changes](#changes)). Naming its session pooler apart gives it one: the same host on port 5432 as the address, and port 6543 as the transaction port. A level may also declare its one address a transaction pooler of any provider - a self-hosted PgBouncer in `pool_mode=transaction`, or another provider's transaction-mode pool - with `AGENTKIT_DB_TRANSACTION_POOLER` set to `true` (or `1`; `false` and `0` declare nothing), or `"transaction_pooler": true` in the machine file; such a store is transaction-only exactly as a store whose one address is Supabase's transaction pooler is. `Setting.transaction_pooler` says whether a level declared it, and the report shows `transaction_pooler: true` when one did. The declaration is refused as `bad_store_setting` beside a transaction address named apart, since it is about the one address, and on any value other than those. A library older than 0.38.0 refuses a machine file or a project level carrying it as `bad_store_setting`, and ignores the key in the process environment, so it is declared only once every tool on the machine reads it. `Setting.has_session_address` says whether the store has a session address: true when a transaction address is named apart, the address being then the session address, or when its one address is neither declared nor known a transaction pooler. The setting's format is the same either way. A URL is a `postgresql://` or `postgres://` URL; one that names no `sslmode` is connected with `require`. A project file setting a key with the prefix that is not one of these is refused.

The checks are the same at every level: `host`, `database` and `user` with no whitespace, `port` 1-65535, `sslmode` one of `require`, `verify-ca`, `verify-full`, or `disable` for a local host (`localhost`, a loopback address or a Unix socket directory, which is also what a URL naming no host reaches); `disable` elsewhere, `allow` and `prefer` are refused as `sslmode_too_weak`; a schema is a lowercase identifier and never `public`, `information_schema` or `pg_*`. A refusal from the project or environment level names the level and the files or variables it read. A machine file of a newer `agentkit.store.*` version is refused as `store_setting_too_new`. With no level answering it raises `DbError("store_not_configured", ...)` with a hint naming the keys and `capabilities store set`.

`Setting.level` names the level that answered (`LEVELS`), and `Setting.sources` every file, as an absolute path, or variable that supplied a value in force. `Setting.report()` gives the whole setting for a status or doctor surface with every secret replaced by `***`: the password field, or for a URL the parameters libpq reads from it (every parameter libpq's own option table marks as a password, such as `password`, `sslpassword` and `oauth_client_secret`, and the SCRAM keys hidden, `sslmode` defaulting to `require`), never the URL's text; a URL libpq cannot parse reports as `{"unparseable": true}`. Under `addresses` it names the address each kind of connection dials, as libpq resolves it (a port named nowhere is `PGPORT`'s or 5432, a host named nowhere `PGHOST`'s): `{"transaction": {"host", "port"}, "session": {"host", "port"} | null}`, `transaction` the transaction address, or the one address when none is named apart, and `session` the session address, `null` for a transaction-only store. The port is a number; for a URL naming several hosts both are libpq's comma-separated lists, and a host libpq leaves to its local default, or a URL it cannot parse, gives `null`. The repr never shows the password or the URL.

## Connect

```python
from capabilities_contract.db import DbError, Step, connect, migrate

conn = connect(application_name="automations", project_root=root)
```

`connect(*, application_name, project_root=None, setting=None, connect_timeout=10, transaction=False, local_settings=None)` returns a psycopg 3 connection to the database `resolve_setting(project_root)` names, or to `setting` when one is passed, whose `search_path` is that setting's schema alone, never `public`, so unqualified names land in that schema. The schema need not exist yet: `migrate` creates it, in the schema the connection is bound to. A store that does not answer within `connect_timeout` seconds (`CONNECT_TIMEOUT_SECONDS`) is refused as `store_unreachable`. Both ends of the connection send TCP keepalives - after 15 seconds idle, every 5 seconds, giving up after 3 unanswered probes - so the store drops a client that died without closing and the client learns of a store that went away. A keepalive is never sent while data the client sent waits for its acknowledgement, so the client also gives the connection up once what it sent has gone unacknowledged for 20 seconds (libpq's `tcp_user_timeout`, `TCP_USER_TIMEOUT_MILLISECONDS`, on a platform whose TCP offers it): a network flow that stopped carrying packets without a FIN or a reset fails in that time as a lost connection, rather than when the store ends the session.

That is a session connection, to the session address: it binds the schema for its session and may hold session state, as a process holding `LISTEN` must. With `transaction=True`, `connect` returns a transaction connection, to the transaction address, or to the one address when the setting names none, also when that address is a transaction pooler. It carries no session state, so a transaction-mode pooler, which hands the server session to another client after every transaction, serves it correctly: psycopg prepares no statement on it (`prepare_threshold=None`), the schema is bound with `SET LOCAL search_path` at the start of every transaction rather than once for the session, also around each statement in autocommit mode, and a statement that would leave state in the session - `LISTEN`, `UNLISTEN`, `PREPARE`, a session `SET` or `RESET`, `set_config(..., false)`, a session advisory lock, a cursor `WITH HOLD` - is refused as `session_state` before it is sent. Transaction-level state is fine: `SET LOCAL`, `set_config(..., true)`, `pg_advisory_xact_lock`. `local_settings`, a mapping of run-time parameters such as `{"statement_timeout": "30000"}`, is set beside the schema at the start of every transaction of a transaction connection, and once for the session of a session connection. A transaction connection sends no startup option, which a transaction-mode pooler refuses: each run-time parameter the setting's `options` or `PGOPTIONS` names (`-c name=value`, `--name=value`) is set for each of its transactions instead, and `default_transaction_read_only=on` makes each of them read-only; any other startup option is refused as `session_state`. A transaction connection at a transaction address of its own makes one round trip as it connects, severed when it is not back within `connect_timeout`, so a pooler that holds the client while the store behind it does not answer is refused as `store_unreachable` in that time rather than after the pooler's own wait. On a transaction connection, a statement that would open a transaction and is one of its own - any statement in autocommit, or a plain read (`SELECT`, `WITH`, `VALUES`, `TABLE` or `SHOW`, naming no write, lock, sequence or setting) at READ COMMITTED - reaches the store in one round trip: the schema binding, the statement and its commit go as one simple query, or as one pipeline when the statement takes parameters, so the store never holds a transaction that waits on the client, and the connection is idle after it, never `idle in transaction`; a transaction-mode pooler hands the server session on as soon as it answers. A read inside a transaction that is open - a `transaction()` block, or after a write not yet committed - runs in that transaction, and a write that opens a transaction leaves it open for `commit()`, as before. A connection set to another isolation level, or to `read_only` or `deferrable`, opens a transaction for a read as before. Client-side TCP keepalives apply to both kinds; only a session connection also asks the server to probe it. A server-side (named) cursor is not bound; use client cursors on a transaction connection. `migrate` and `status` work on either kind.

A transaction connection is the process's working connection, since it carries nothing from one transaction to the next: the first `connect(..., transaction=True)` opens it, `close()` or leaving `with` ends the loan - rolling back what was not committed and resetting what the borrower changed, such as `autocommit` or `row_factory` - and every later `connect` for the same address is lent that same connection, its schema, `local_settings` and `application_name` bound for each of its transactions as that call names them. A short command therefore connects once, and a long-running process keeps one connection for all its work. A loan taken while another is still out gets a working connection of its own; each beyond the first is closed once it has sat idle for 30 seconds (`SPARE_IDLE_SECONDS`), and every one is closed when the process exits. `disconnect()` closes the idle ones at once, for a process that is about to wait a long time without the store; `disconnect(every=True)` closes the lent ones too. A child of `fork` never uses or closes its parent's. After `close()` the connection reads as `closed` and refuses statements, as any closed connection does. A session connection is never lent: each `connect` without `transaction` opens one, and its `close()` closes it.

A working connection the store or the network lost is replaced: as it is lent, one the store closed is noticed without a round trip, and one that sat idle for 10 seconds (`PING_AFTER_SECONDS`) makes one round trip, cut at the connect timeout, so a connection that stopped answering is not waited on. One the store or the network closed while it was lent is replaced, without a round trip, before a statement that is a transaction of its own is sent on it. A statement that finds its connection gone before anything of its transaction reached the store - the statement that opens a transaction, a statement in autocommit before it is sent, or the start of a `transaction()` block - runs once more on a new connection in place of the old one. A plain read that is a transaction of its own runs once more on a new connection whenever it loses its connection, while a write in autocommit that was sent is never sent again, since its commit went with it. So does a plain read - `SELECT`, `WITH`, `VALUES`, `TABLE` or `SHOW`, naming no write, lock, sequence or setting - in a transaction at READ COMMITTED that has run only plain reads: each such read sees the store as it is when it starts, so it runs on the new connection in a transaction opened again there, with every `transaction()` block it is inside opened again too, and the commit of a transaction that has only read - `commit()`, the end of a block, or an autocommit statement's own - is not refused when its connection is gone, since it has nothing to commit. A store that ends the session itself (an administrator's or a shutdown's error) counts as a lost connection. A commit whose answer was lost is never sent again, and a transaction that loses its connection after earlier statements fails as before, since what it did is gone; its `rollback()` is then not refused, since the store rolled the transaction back with the connection, and the next statement runs on a new connection. A connect that fails before the store answers - refused, reset, timed out - is tried again after a short pause, jittered and doubling, at most 3 times (`CONNECT_ATTEMPTS`) while `connect_timeout` is not spent, each try waiting at most half of it; a store that answers with a refusal of its own, such as a wrong password, is not tried again. A process whose store did not answer does not dial it again on every statement: after a transaction connection to an address found no store answering, a further one to that address - a `connect`, or the replacement of a lost connection - is refused at once as `store_unreachable`, carrying the last failure as its cause, for a pause drawn between half and all of 1 second (`RECONNECT_FIRST_SECONDS`), doubling with each failure in a row up to 30 seconds (`RECONNECT_LONGEST_SECONDS`); the first one after the pause dials again, and one the store answers ends the pause. Each such refusal, and the one of the connect that began the pause, carries the moment the pause ends, on the `time.monotonic()` clock, as `dials_again_at`, so a caller that asks again waits for that one schedule rather than keeping a second of its own; a refusal that leaves no pause carries `None`, and the next connect dials. The pause is kept per process, so the first connect of a process always dials, and a session connection is never paused.

Supabase's transaction pooler - a host under `pooler.supabase.com` on port 6543 - serves transaction connections, as the transaction address or as a transaction-only store's one address; a transaction connection to it makes the one round trip a transaction address of its own makes as it connects. A session connection to it is refused as `transaction_pooler` before anything is sent, because session connections cannot work through it. When it is the store's only address, the message says so - the store's only address is a transaction pooler, which keeps no session - and the hint names how to give the store a session address apart: its session pooler, the same host on port 5432, as the address, keeping port 6543 as the transaction port (`capabilities store set ... --session-port 5432`). When it is the session address beside a transaction address, the hint names its session pooler as the address and port 6543 as the transaction port. A session connection to a store whose one address is declared a transaction pooler is refused as `transaction_pooler` the same way, and the hint names how to give that store a session address apart.

## Batch

```python
from capabilities_contract.db import batch

counts, names = batch(conn, [
    "SELECT count(*) FROM automations_runs",
    ("SELECT name FROM automations_jobs WHERE host = %s", (host,)),
])
```

`batch(conn, statements)` sends independent statements - each a query, or a `(query, params)` pair, and each one statement - to the store as one pipeline on a connection from `connect`, and returns each one's rows in order: a list of rows for a statement that returns rows, an empty list for one that does not. The store answers them all in one round trip, directly or through a transaction-mode pooler. On a transaction connection the schema is bound for them. Where each would be a transaction of its own - in autocommit, or when every one is a plain read - they run as one transaction of their own and leave the connection idle, and a batch of plain reads that loses its connection runs again on a new one; otherwise they run inside the transaction that is open, or the one the first of them opens, which stays open for the caller's `commit()`. A statement that fails fails the batch with its error, and the statements after it are not run.

## Migrate

```python
result = migrate(conn, "automations", [
    Step("0001-runs", "CREATE TABLE automations_runs (id bigint PRIMARY KEY)"),
    ("0002-runs-host", "ALTER TABLE automations_runs ADD COLUMN host text"),
], major=1, minor=1)
for warning in result.warnings:
    print(warning)
```

`migrate(conn, owner, steps, *, major, minor)` applies each step that is not yet applied, in order, once. The owner is the tool that owns the tables, a lowercase identifier without underscores. The first call creates the schema and three platform tables in it: `schema_ledger(owner, step, checksum, applied_at, library_version)`, `schema_version(owner, major, minor, updated_at)` and `schema_channel(channel, scope, version)`, the change versions of [Changes](#changes); a store whose schema was made before `schema_channel` existed gets it on the next `migrate` by a role that may create in the schema, once, under the schema lock, and a role that may not migrates such a store as before, leaving `announce` and `Follower` refused as `not_migrated` until one that may has run. Each step runs in its own transaction under a per-owner advisory lock together with its ledger row, so concurrent processes apply it exactly once. When every step is already applied and the store records the caller's version, it reads that without taking any lock, so a session holding one stops no tool whose tables are in place. Otherwise a lock is tried rather than queued for, so a process that dies while waiting leaves nothing in the queue: after 10 seconds held elsewhere (`LOCK_WAIT_SECONDS`) `migrate` refuses with `store_busy`, and a session that holds a lock and then issues no statement for 5 seconds (`LOCK_IDLE_SECONDS`) is ended by the server, which frees the lock. It returns a `MigrateResult` with `applied`, `skipped`, `warnings`, `major`, `minor` and `core`.

One owner, `core`, is the store's core: the tables any owner's steps may reference. Its tables are the plain nouns of `CORE_TABLES` (`agents`, `nodes`, `connections`, `grants`, `settings`, `secrets` and the rest), and it names everything its steps create `<table>` or `<table>_*` for a table on that list, or `core_premig_*` for a pre-image it keeps, and no other `core_*` name. Those names are reserved against every other owner, whatever it is called. For any other owner, `migrate` first migrates the core's own steps (`CORE_STEPS`, at `CORE_VERSION`), so the core is current before the owner's first step, and a refusal of the core's refuses the call before any of the owner's steps; `result.core` is the core's `MigrateResult`, or `None` while `CORE_STEPS` is empty, when nothing is run or written for the core. A step that references core tables declares the core version it needs, `Step(id, sql, core=(major, minor))`, and is refused with `core_too_old` until the store's `schema_version` records the core at that version or later; the owner's steps before it stay applied.

The core is the manager's tables, at core version 2.0: `agents`, `nodes`, `agent_nodes`, `node_states`, `node_updates`, `node_logins`, `node_login_expectations`, `node_logs`, `node_sessions`, `node_session_expectations`, `connections`, `grants`, `policies`, `settings`, `settings_changes`, `settings_documents`, `identifiers`, `secrets`, `service_instances`, `service_leases`, `store_state` and `reference_versions`. A column that holds an agent is `agent_id`, and a connection's owner is `owner_agent_id`. Up to 0.41.0 the same tables were owner `capabilities` at schema version 1.23, named `capabilities_*` with their agent in `project_id` (`capabilities_contract.db._legacy` keeps those steps, the library's and the two the manager added, as history). The core's first step, `0039-core`, takes a store from wherever it stands: it runs each of those steps the store's ledger does not record - all of them on a fresh store - then renames every table, column, index, constraint, sequence, trigger and function in place, keeping every row and moving the pre-images `capabilities_premig_*` to `core_premig_*`, recreates each function whose body named what moved, and words the comments for the new names. The step raises the core to major 2 (`Step(..., major=2)`), so on a store already in use it runs only under the store upgrade (`capabilities store upgrade`), and `STORE_MAJOR` was 2 until core 3.0. On a store that held the old tables it leaves a refusing stub under each old name, with the old columns, whose every read fails naming the new table and `capabilities self-update`, then `capabilities update`, so a release that still reads `capabilities_*` is refused rather than reading no rows; a fresh store gets none. A fresh store and one moved this way end with the same catalog, the stubs apart. The ledger's rows of owner `capabilities` become the core's, and `schema_version` keeps a row for owner `capabilities` at major 2 beside the core's, so a release that still migrates owner `capabilities` (0.41.0 and older) refuses the store with `schema_too_new` instead of making its tables again. The channels the core's triggers raise keep their names, `capabilities_vault` and `capabilities_node_update`.

At core version 3.0 every connection has an opaque id. The core's second step, `0040-connection-ids` (`Step(..., major=3)`, so a store in use takes it only under the store upgrade, and `STORE_MAJOR` was 3 until core 4.0), gives each row of `connections` an `id` - twelve lowercase hex characters, minted by the store where a writer gives none, unique across the store, its primary key, and never changed (a trigger refuses an update of it) - and makes `label` what every connection is called: its former name where it had no label, not null, unique within its capability, and never any connection's id, so a word names a connection by its id, else by its label, and never two. Every core reference holds the id, with a foreign key to `connections (id)`: `grants.connection_id`, `node_sessions.connection_id` and `node_session_expectations.connection_id` (the rows go with the connection; a session, or the owner's word on one, of a connection only a machine's files hold keeps that connection's key in `file_connection` instead, exactly one of the two set, and each table is unique on `(agent_id, node, capability, coalesce(connection_id, file_connection))`), `nodes.coolify_connection_id` and its copy in `agent_nodes` (set null when the connection goes, the node kept; a node made through a connection the machine lent, which no row holds, keeps that connection's address in `coolify_address`), `settings.connection_id` and `identifiers.ref_connection_id`. A top-level `connection` setting naming a connection of its own capability by its name or label holds its id; a node session or expectation naming no connection the store holds keeps the name it held in `file_connection`. `node_sessions_put` and `node_expectation_set` take a connection key and keep it as `connection_id` where the store holds that id for the capability, as `file_connection` otherwise. `connections.name` and every column that held a name are dropped; the pre-image `core_premig_connection_names (id, capability, name, label)` (`CONNECTION_NAMES_TABLE`) keeps each connection's former name and the label it had, written by that step alone, for other owners' steps converting their own columns (`Step(..., core=(3, 0))`) and for a capability finding state it kept under the name. A fresh store and one moved from core 2.0 end with the same catalog.

At core version 4.0 every agent, node and connection reference is a foreign key that says what its parent takes with it. The core's third step, `0041-keys` (`Step(..., major=4)`, so `STORE_MAJOR` is 4 and a store in use takes it only under the store upgrade), keys every core agent and node column: a row the agent or node owns goes with it (`ON DELETE CASCADE`), and `settings_changes.node`, a record of where a change was made, keeps its row and loses the node (`ON DELETE SET NULL`); the keys of `agent_nodes` and the `node_*` tables lose `ON UPDATE CASCADE`. `agents.id` and `nodes.id` refuse an update (triggers `agents_id`, `nodes_id`), as `connections.id` does, and `agents`, `nodes` and `connections` refuse a `DELETE` (triggers `agents_delete`, `nodes_delete`, `connections_delete`) unless the transaction-local setting `store.deleting_agent`, `store.deleting_node` or `store.deleting_connection` names the row, a comma-separated list of ids; a connection is also deleted with the agent that owns it. `db.deleting(conn, parent, *ids)` sets it for the open transaction, and clears it with no ids; it first runs the transaction's pending deferred triggers (`SET CONSTRAINTS ALL IMMEDIATE`), so a deferred notify trigger fires outside the setting for the statements before it and inside it for the cascade. The store tier's connection removal calls it around its delete. The step also makes the functions every owner's own keys step and guards call: `store_state_keys(owner, keys)` takes `{table, column, parent, class}` rows (class `cascade`, `set null` or `restrict`), makes each column's key in its class `NOT VALID`, settles every reference to an agent, node or connection the store does not hold by that class under the deleting settings - each changed row kept as it was in `<owner>_premig_m9 (source, op, row, saved_at)` - and validates every key, refusing the step where a `restrict` column or another parent names a missing row; `store_state_cascade(table, op, old, new)` answers whether a delete or an update of a row is the cascade of such a deliberate delete and nothing else, which a history guard admits (the core's `settings_changes_append_only` and `reference_versions_guard` do); and `store_state_deleting()`, under which a notify trigger stays quiet (the core's `secrets_announce` does). A step of another owner that calls them declares `core=(4, 0)`. A `Step` also reads as its `(step id, sql)` pair.

At core version 4.1 every connection records when it was created. The core's fourth step, `0042-connection-created-at`, is a minor step (no `major`), so it is applied wherever the core is at major 4, outside the store upgrade, and a release that knows core 4.0 reads a store at 4.1 as a newer minor with a warning. It adds `connections.created_at timestamptz`, which a trigger (`connections_created_at`) sets on every insert to the inserting transaction's time, over whatever the writer gave, and keeps on every update, so an edit, a rename or any other write leaves it as it was and nothing fills it in. A row the store held before the step stays null: no moment it could be given would be true. A connection imported from files has the moment its row was made. `StoreRecords.connection_created(capability)` answers each connection's moment by id as ISO text, read through the row's JSON so a store at core 4.0 answers null for every row, and a capability's `connections` report carries it on each connection as `created_at`, null for a connection kept in files or older than the step.

The tiers follow. A resolved connection's entry carries `id` and `label`; `_select_connection` matches a word against the id, then the label, then the `address` field, and reads a default pointer holding either. In the store a connection written under a word that names none is added labelled with it and its id minted (the adapter's `set` answers the id), and a grant, a default pointer, a removal and a rename name their connection by id or label. A new connection's env keys fill `{ID}` from the label it is added under, and a key is the connection's own when it is the one its label or its former name gives. `_connection_state_dir(base, cid)` answers `base / cid`, first moving state a capability kept under the connection's former name there, once and saying so on stderr, never over a path that exists; `_connection_former_name(cid)` answers that name. `<capability> ids set <label> --connection <id|label>` writes an identifier standing for a connection: its value the id, and `ref_connection_id` beside it in the store. A settings row written to the store carries `connection_id` where one of its path segments, else its value, is a connection's id. Files keep a connection under its key, which is its id and its label.

It refuses with a `DbError`, rolling the step back: an applied step whose SQL changed (`checksum_mismatch`); a step that creates a relation, index, sequence, type, function, collation, statistics object or schema not named `<owner>` or `<owner>_*` inside the configured schema, or named `schema_ledger`, `schema_version` or `schema_channel`, or named for the core by an owner other than `core` (`naming_law`); a step that declares a core version the store's core has not reached (`core_too_old`); a step that fails or ends its own transaction (`step_failed`). A step cannot use statements that refuse a transaction, such as `CREATE INDEX CONCURRENTLY`.

The version rule lets tools at different releases share one database. When the store records a newer major for the owner than the caller's `major`, `migrate` refuses with `schema_too_new` before applying anything, and the hint names the commands that catch the code up: `capabilities self-update`, then `capabilities update`. When it records the same major with a newer minor, it proceeds and adds a warning to `result.warnings`, keeping the stored version. When it records an older major than the caller's, `migrate` refuses as below. Otherwise it records the caller's version.

A schema major crosses only through the store upgrade, a procedure a person starts with `capabilities store upgrade`; a minor change applies by itself. A step that raises its owner to a new major says so, `Step(id, sql, major=2)`. Outside the upgrade, `migrate` refuses with `store_upgrade_needed` before applying or recording anything for the owner when a pending step is marked `major` and the store is not empty, or when the store records a major for the owner older than the caller's `major`; the message names the owner, the major the store holds and the one the code wants, and the hint names `capabilities store upgrade` as the one way to cross and `capabilities store upgrade --check` as what says what it will do. The core's own pass is governed by the same rule. An empty store - no `schema_version` table, or no row in it - takes every step, major steps included, judged as it was when the call began, so a new install takes the newest. The upgrade procedure is recognised by `CAPABILITIES_STORE_UPGRADE=1` in the environment (`UPGRADE_ENV`; `upgrading()` reads it); under it `migrate` applies major steps and records the higher major. Nothing in the library sets it.

The store's major is the major `schema_version` records for `core`. A store with rows there but no `core` row is at store major 1, which is also the major recorded for `core` in the rule above, and an empty store has none. `STORE_MAJOR` is the store major this library's own steps read: `CORE_VERSION[0]` once `CORE_STEPS` holds steps, else 1. `store_major(conn)` reads the store's in one statement, without a lock and changing nothing (in a savepoint when the caller holds a transaction), as a `StoreMajor(major, since)`: `major` is `None` for an empty store, and `since` is when the latest major step of `CORE_STEPS` was applied, else `None`.

A major step that renames or drops a table may leave a refusing stub under the old name, so code that still reads it fails with a message saying what to run rather than with a missing table. `moved_stub(old_name, columns, message)` returns the SQL: a view under `old_name` with the old `columns` - each a name, typed `text`, or a `(name, sql_type)` pair - that holds no data, whose every read fails with an error carrying `message`, and that takes no write. The naming law admits such a view, under a name that held a relation in the schema before the step, only inside a step marked `major`.

## Changes

```python
from capabilities_contract.db import Follower, announce, connect

with connect(application_name="automations", project_root=root, transaction=True) as conn:
    conn.execute("UPDATE automations_jobs SET paused = true WHERE name = %s", (name,))
    announce(conn, "automations_jobs", project_id)
    conn.commit()

with Follower([("automations_jobs", project_id), "automations_runs"],
              application_name="automations", project_root=root) as follower:
    follower.run(lambda moved: reload(moved), stop=stopping)
```

A change other processes must see is announced as a version, and a process follows it by polling, so neither side needs a session for it: both work on the process's working connection, directly or through a transaction-mode pooler. Where the setting names the store's two addresses apart, a follower also listens on a session connection of its own, so a change wakes it at once rather than at its next poll. The store keeps one version per channel and scope in the platform table `schema_channel`; a pair that was never announced has version 0. A channel is a lowercase identifier of at most 63 characters, conventionally `<owner>_<event>`, and a scope is a non-empty string naming what the change is about, such as a project id.

`announce(conn, channel, scope)` raises that pair's version by one and returns the new version. A writer calls it inside its own transaction, after the change it announces, on a connection from `connect`: the raise is committed with the change and rolled back with it, so a follower never sees a raise whose change did not happen. The same statement notifies `channel` with `scope` as the payload (`pg_notify`; a scope longer than a notification carries is sent as an empty payload), which the store sends when the transaction commits and never when it rolls back, and which reaches a listener through a transaction-mode pooler as well. It is one statement, one round trip. The pair's row stays locked until the writer's transaction ends, so writers announcing the same pair take turns at it: announce last, just before the commit, so that each holds it for the shortest time.

`Follower(follow, *, application_name, project_root=None, setting=None, seen=None)` follows each item of `follow`, a `(channel, scope)` pair or a channel alone, which means every scope of that channel. `read(conn=None)` returns every followed version as the store holds it now, keyed by `(channel, scope)`, in one query and one round trip on the process's working connection, whatever the number of channels. Given `conn`, a connection from `connect` the process already works on and keeps lent, it reads on that one instead and leaves it open, in whatever transaction it was in, so a process that holds its working connection follows without a second. `poll(conn=None)` takes `conn` the same way and returns only the versions that moved since the previous poll, with their new values; the first poll returns every version it reads, unless `seen` - the `seen` property of an earlier follower, the versions it last read - is given, when it returns only those that moved since then. A follower that was not polling for a while, or whose process was down and started again with its `seen`, therefore gets every pair that moved meanwhile on its first poll, once, with its latest version.

`wait(seconds, stop=None)` returns False after `seconds`, or as soon as the `threading.Event` `stop` is set, and returns True earlier, as soon as the store notifies one of the followed channels. It listens only where the setting names the store's two addresses apart (`Setting.has_transaction_address`): on a session connection of its own, opened at the first wait, which `LISTEN`s on the followed channel names, whatever their scopes. With one address - a direct store or a transaction-only one - it waits out the time and opens no connection, so following is polling alone, as before. A listener that cannot be opened, or is lost, never fails a wait and never costs a change: the wait waits out its time, the next poll sees what moved, and the listener is tried again no more often than every 30 seconds. A notification sent before the listener opened is not seen; the next poll sees its change. `close()` closes the listener, and a follower used as a context manager closes it as the block ends; a later wait opens it again.

`run(on_change, *, interval=1.0, stop=None)` polls every `interval` seconds (`FOLLOW_INTERVAL_SECONDS`) until the `threading.Event` `stop` is set, and calls `on_change` with what moved: once for the first poll even when nothing moved, so the caller reads its state after the versions it starts from and misses nothing that changes later, and afterwards only when something moved. Between polls it waits with `wait`, so where the store's two addresses are named apart a notified channel is polled at once; it opens the listener before its first poll and closes it as it returns. A poll that finds the store unreachable is tried again at the next interval, and the first one that reaches it reports everything that moved meanwhile. A store whose schema has no `schema_channel` yet is refused as `not_migrated`.

## Vault

`capabilities_contract.vault` keeps each agent's secrets in its store, sealed with the agent's own key. The key is `AGENT_ENCRYPTION_KEY` (`KEY_ENV`): 32 random bytes as URL-safe base64 without padding, 43 characters (`generate_key`, `parse_key`, which refuses anything else as `bad_key` without echoing it). `key_source(project_root)` finds it in the project's `.env.local`, then its `.env`, then the machine's `$XDG_CONFIG_HOME/capabilities/shared.env` (`shared_env_path()`), then the process environment, and `agent_key` parses what it found. The store never holds a key: `fingerprint(key)`, a SHA-256 over it, is what `agents.vault_key` and each row's `key_fingerprint` keep.

`seal(key, agent_id, name, value)` seals a value with XChaCha20-Poly1305 (libsodium, through PyNaCl) under a fresh 24-byte nonce, the agent id and the name bound as associated data, so a sealed value moved to another agent or name does not open; `unseal` answers None for anything that does not open. Each secret is a row of `secrets` (`agent_id`, `name`, `ciphertext`, `key_fingerprint`, `updated_at`, `updated_by`), migrated with the contract's other tables; a trigger raises the channel `capabilities_vault` (`VAULT_CHANNEL`) for the agent in `schema_channel` on every change to a row or to the agent's `vault_key`. `put`, `remove`, `rows`, `recorded_key`, `record_key` and `sealed_under_other` read and write it on a connection the caller holds; a row whose fingerprint is not the key's is unreadable, kept until it is set again or removed.

`reader(project_root, agent_id)` is the process's `Reader` of that agent's vault, or None where no key is set. `Reader.seed(rows, version)` takes what `round_trip(agent_id)` reads - the sealed rows and the channel version, two statements to send beside others in one `batch` - and `get(name)` answers a value, opening only the names asked for and keeping what it opened in memory only; without a seed it reads them itself, and asked again after `CHECK_SECONDS` it reads the channel version and rereads the rows only when it moved. A store it cannot read raises `VaultError("vault_unreadable", "the vault could not be read: ...")`, which stands for `RETRY_SECONDS`. The connections tier asks it as the credential cascade's first tier, before the project's `.env` files, the machine's files and the process environment, which still answer every name it does not hold; the store tier seeds it from the round trip that reads a project's connections.

Every key also gives its agent a public sealing key: `public_key(key)`, the X25519 public key of a scalar derived from the key apart from its fingerprint, which `public_text` writes as 43 characters of URL-safe base64 and `parse_public` reads back. `publish(conn, agent_id, key)` writes it to `agents.vault_public_key` where the row records that key's fingerprint, and the store clears it whenever `vault_key` changes without it, so no writer seals for a key the agent no longer holds; the key itself still never enters the store. `seal_for(public, agent_id, name, value)` seals a value for that public key by a writer that holds no key of the agent's - an ephemeral X25519 key agreed with it gives the XChaCha20-Poly1305 key, the agent id and the name bound as before, format byte 2 - and `unseal` opens either format with the agent's key, so a reader reads such a row as any other. `recipients(conn, agent_ids)` answers each registered agent's fingerprint and public key, and `put_for` keeps a value sealed for one under its fingerprint. `share(project_root, writer, values, agents, *, replace, application_name, updated_by)` seals values into the vault of each of `agents` in one transaction - the writer's own under its key, every other one's for its public key - leaving a name an agent already holds as it is where `replace` is false, and answers `{sealed: {agent: {name: "created" | "replaced"}}, unchanged, unpublished, unregistered, hint}`: an agent whose public key is not published is named, never skipped in silence. It is how a connection the store shares carries its secret in the vault of every agent granted it.

`write(project_root, agent_id, values, *, application_name, updated_by)` seals each value into the agent's vault under the key `project_root` resolves, in one transaction of its own, recording the key's fingerprint where the agent has none yet and publishing its public sealing key, and answers `{name: "created" | "replaced"}`; it writes nothing where it refuses, as `key_missing`, `bad_key`, `agent_unregistered`, `key_mismatch` or `vault_unwritable`. `delete(project_root, agent_id, names, *, application_name)` removes rows, which needs no key, and answers the names that were there. Both make the process's readers of that agent read again, so a value written is the value read back. They are how the connections tier's `connection add` and `setup set` keep a secret in the vault by default - writing a file only with `--file`, at the machine's scope where no grantee's vault takes it, or where no vault takes it, and saying why (`vault_skipped`) - and how `connection remove` and `setup unset` take it out again. `connection add` of a new connection also seals the values the machine holds for it, and `connection seal [<id>] [--share]` writes into each existing connection the values it resolves from the project's and the machine's files, sealed into the vault, without writing over what a vault holds; a value only the process environment holds is left there (`not_sealed`), and with `--share` a connection the store shares reaches another agent's vault only under the keys that are the connection's own, never under a name another connection or a setup field may read (`not_shared`). Every capability that carries the connections tier renders the cascade in its help as `Where a secret resolves`. `render_text` and `parse_text` write and read the vault as KEY=value text, a value a plain line would not carry back whole as a JSON string.

## Status

```python
from capabilities_contract.db import status

report = status(conn, "automations", steps, major=1, minor=1)
print(report.state, report.pending)
```

`status(conn, owner, steps, *, major, minor)` reports what `migrate` would do with the same arguments, changing nothing and taking no lock, so it answers while another session holds one. It returns a `MigrateStatus` with `state`, `major` and `minor` (the caller's), `stored_major` and `stored_minor` (the store's, or `None`), and the step ids under `applied`, `pending` and `changed`, plus `warnings`. `state` is `current` when every step is applied and the store records the caller's version or a newer minor (which adds a warning), `pending` when `migrate` has something to apply or record, and `checksum_mismatch`, `schema_too_new` or `upgrade_needed` (where `migrate` refuses `store_upgrade_needed`) when `migrate` would refuse. `store_major(conn)` reports the store's major, as [Migrate](#migrate) says.

## Bodies

```python
from capabilities_contract.bodies import body_digest, check_body, materialise

path = materialise(script_body, state_dir / "bodies")
subprocess.run([str(path)], check=True)
```

`capabilities_contract.bodies` is the one way a stored executable script body - an automation's `script_body`, a stored tasks hook - is run. It imports nothing outside the standard library. `body_digest(body)` is the sha256 hex digest of the body's UTF-8 bytes. `check_body(body)` raises `ValueError` unless the body is a `str` starting with `#!`. `materialise(body, directory)` checks the body, makes the directory with mode 0700 (and its parents) when it is missing, and returns `<directory>/<body_digest(body)>`, an executable file holding the body: a regular file already there with the same content and mode 0700 is reused, and anything else at that path is replaced atomically by a file written beside it, chmodded 0700 and moved into place with `os.replace`. A process running the file never sees it partial, and processes materialising the same body at once each return the whole file.

## Errors

Every failure is a `DbError` with `slug` (stable, for the caller to map to an exit code), `message`, `hint` (possibly `None`) and `dials_again_at`, the `time.monotonic()` moment from which a transaction connect to the same address dials again after a `store_unreachable` refusal that left it paused (see [Connect](#connect)), else `None`. Slugs: `store_not_configured`, `store_not_postgres`, `store_setting_unreadable`, `store_setting_too_new`, `bad_store_setting`, `sslmode_too_weak`, `bad_schema_name`, `store_unreachable`, `transaction_pooler`, `session_state`, `driver_missing`, `bad_application_name`, `bad_owner`, `bad_version`, `bad_step`, `connection_busy`, `checksum_mismatch`, `naming_law`, `core_too_old`, `step_failed`, `schema_too_new`, `store_upgrade_needed`, `store_busy`, `bad_channel`, `bad_scope`, `bad_follow`, `bad_interval`, `not_migrated`.

## Development

```
uv venv && uv pip install -e ".[dev]"
uv run pytest
```

The suite starts a throwaway PostgreSQL cluster with TLS under a temporary directory (`initdb` and `pg_ctl` on `PATH`, or in `$PG_BIN`) on a random port. Set `CAPABILITIES_CONTRACT_TEST_URL` to a TLS-enabled server's admin URL to run it against that server instead, as CI does. The transaction-pooler tests run when `CAPABILITIES_CONTRACT_TEST_PGBOUNCER` names a `pgbouncer` executable: they put it in front of the test cluster in `pool_mode=transaction` with one server connection. The tests that compare the machine file reader with the manager's own run only when `CAPABILITIES_STORE_TIER` names a store tier module such as `capabilities_contract/capability/store.py`; otherwise they skip.
