workspace D-70 ("`feedback` needs no address"): the failure channel must
not depend on the thing most likely to break. `scribe_messages.from_identity`
becomes nullable (migration 010); `POST /scribe/messages` accepts a missing
`X-Office-Identity` for exactly one shape — an office caller, `to` is
`["tower-control"]` alone, and `tags` contains `feedback` — stored with
`from_identity` null and `from_consumer` the caller's consumer. Every other
send without an identity is still 400 `identity_required`, enforced in
ScribeService.send_message rather than the caller dependency, since the
decision needs the request body. `MessageResponse.from_identity` is now
optional so mailbox/history/find/message_id answer such a message with a
null sender. Migration 010's downgrade refuses outright while any
sender-less row exists rather than inventing one.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The user wants a way to get rid of a message that is no longer useful. A
message is stored once and shared by every recipient; `history` is how
agents rebuild context (D-66), so deleting the message itself would take
it from every other reader. Deletion is therefore per mailbox entry
(D-70): `archive` (kept, out of the default view) and `trash` (out of the
mailbox, restorable). True erasure of content stays `redact` — not this
round.
Migration 009 (down_revision="008") adds
scribe_mailbox_entries.archived_at and trashed_at — independent nullable
timestamps, mirroring book_records.archived_at/trashed_at (migration
005). No data step: every existing entry is neither.
PATCH /scribe/mailbox/{entry_id} takes `archived` and `trashed` booleans
beside `read` (true sets the timestamp if not already set, false clears
it); the response carries both. GET /scribe/mailbox takes `view`: `inbox`
(default, neither archived nor trashed), `archive` (archived and not
trashed), `trash` (trashed, regardless of archived_at — trashing hides an
entry from the archive view too), `all` (no filter on either column).
`find` now excludes trashed entries (archived ones still match).
T-9: GET /scribe/mailbox?message_id=N is a direct lookup, not a filtered
list — the unique constraint on (address, message_id) guarantees zero or
one, in any view, ignoring unread/tag/before_id entirely. That is the
point: a client that already knows the message number never pages the
unread-first/id-descending listing to find it.
Judgement call: the listing's own cursor (`before_id`, a single `id`
comparison) still does not match its two-key sort (unread-first, then id
descending) — the pre-existing gap the dossier flagged as optional to
fix. Left unchanged: T-9's message_id lookup is the actual answer to "a
client pages to find one," and reworking the general cursor is a larger,
separate concern (which field(s) a keyset cursor needs) that risks the
already-passing test_mailbox_lists_unread_first for no requirement this
round asked for.
Judgement call: archived_at and trashed_at are independent timestamps,
never cross-clearing each other. Trashing an already-archived entry does
not clear archived_at, so restoring it (trashed=false) returns it to the
archive view, not the inbox — the view filter's own trash-takes-priority
rule is what keeps it out of view=archive while trashed, not a write-time
side effect.
Tests: 12 new (6 archive/trash flow + 4 message_id lookup + 2 migration
round-trip). Full suite 554 -> 566 passed.
Mutation checks (Edit, reverted after each):
- view=inbox dropped its trashed_at filter ->
test_trash_hides_from_inbox_archive_and_find_shows_in_trash failed
(assert 1 == 0).
- message_id lookup dropped the caller-address filter ->
test_message_id_answers_nothing_for_a_message_not_in_their_mailbox
failed (assert 1 == 0).
- trash updated every mailbox entry for the message instead of the
caller's own row ->
test_trashing_one_recipients_entry_leaves_the_others_untouched failed
(assert 0 == 1).
All three reverted; full suite re-verified green (566 passed).
Migration 008 -> 009 -> 008 -> 009 round-tripped against core_api_test,
both manually (alembic CLI) and in
TestMigration009RoundTrip. Production stays at 008 — every alembic
invocation derived POSTGRES_* from TEST_DATABASE_URL.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
claude.ai's connector keeps no MCP session between tool calls, so the
Office cannot remember which address a conversation chose. Each consumer
grant now carries a fallback: the Office uses it when a call names none
and has no session binding (tatlock-office T-5, in a parallel round,
reads it).
Migration 008 (down_revision="007") adds
tenant_consumers.default_address — nullable varchar(63) — and a
composite foreign key (tenant_id, default_address) ->
scribe_addresses(owner_tenant, name), so a default always names a real
address in the same tenant. scribe_addresses already carries both a
primary key and a named unique constraint on (owner_tenant, name)
(migration 006), so the composite FK has something to bind to; Postgres's
default MATCH SIMPLE means a NULL default_address never triggers the
check, which is the state every grant but one is left in. The data step
sets jpmschweitzer/office-claude -> claude-ai only if that address exists
at upgrade time (true in production, false in core_api_test, where no
scribe_addresses are seeded) — an UPDATE ... WHERE EXISTS, not a bare
UPDATE, so it is a no-op rather than a failure when absent.
llm_tester/office-claude is untouched and stays null. Downgrade drops the
FK then the column.
TenantConsumer (models.py) gains the column and a matching
ForeignKeyConstraint. TenantsService.get_me now returns a third element,
the caller's own consumer's default_address (looked up only when
caller.consumer is not None — a user-kind caller carries no consumer, so
there is no grant row to hold a default for it) or None. GET /tenants/me
answers it; GET /tenants and its TenantSummary schema are unchanged.
Tests: TestMigration008RoundTrip/ForeignKey/DataStep added to
tests/test_tenants_migration.py (round-trip 007<->008, the FK existing
and refusing an unknown address while allowing null, the data step's
no-op and positive cases). tests/conftest.py gains a
default_address_grant fixture (registers a throwaway address, points a
grant at it, reverses both on teardown). tests/test_tenants.py gains
TestMeDefaultAddress (an office caller with/without a default, a user
caller always null).
Friction found and fixed in the same commit: the new composite FK made
Postgres refuse tests/test_scribe.py's existing clean_scribe_tables
TRUNCATE ("cannot truncate a table referenced in a foreign key
constraint") — the restriction is on the constraint's existence, not on
whether any row's default_address is actually non-null. tenant_consumers
joins that TRUNCATE; CASCADE was rejected because it would silently drop
every grant rather than only addresses, so the four grants
(jpmschweitzer/llm_tester x office-claude/office-gemini) are reinserted
in the same fixture, right after.
Mutation-checked: forcing default_address to null in the controller
fails exactly test_office_caller_with_a_default_sees_it, leaving the
without-a-default and user-caller tests green. Disabling the FK's
creation (and, symmetrically, its drop in downgrade — to keep the
round-trip self-consistent against the live core_api_test schema) fails
test_fk_exists and test_fk_refuses_a_default_naming_an_address_that_does_not_exist,
leaving test_fk_allows_a_null_default_regardless_of_tenant green. Both
mutations reverted; the full suite re-verified green (543 + 11 = 554
passed).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
A tenant has been an unchecked string since T-4/T-7: the Office's
X-Office-Tenant header or a user JWT's tatlock_user claim, with Book and
Scribe each running their own _require_tenant that checked only presence.
A typo'd claim opened a fresh, empty namespace rather than being refused.
src/domains/tenants/ (controller/models/schemas/service) adds tenants and
tenant_consumers (migration 007, down_revision="006"): tenants.id is the
slug a tatlock tenant carries, kind (household|person|test) and status
(active|suspended) CHECK-constrained; tenant_consumers is the grant pairing
a consumer to a tenant. Migration 007 seeds the two tenants in use today —
jpmschweitzer (household, "Jeroen") and llm_tester (test, "Playground") —
each granted to office-claude, the only Office consumer configured in
Authentik. Every existing owner_tenant column (book_records,
book_tenant_counters, scribe_addresses, scribe_messages,
scribe_mailbox_entries — five, grepped and confirmed) now carries a
foreign key to tenants.id, ON DELETE RESTRICT, at both the migration and
the ORM model.
TenantsService.require_tenant is the one enforcement point: no tenant is
400 tenant_required (unchanged); an unknown tenant is 403 tenant_unknown;
a suspended one is 403 tenant_suspended; an Office caller whose consumer
holds no grant in that tenant is 403 consumer_not_granted. One query — a
LEFT JOIN against the grant — so presence, status and the grant are all
read in a single round trip. Book's and Scribe's own _require_tenant now
delegate to it, translating TenantError to their own exception type so
every call site and each controller's single except clause keep working
unchanged.
Read-only this round: GET /tenants/me answers the caller's own tenant and
its granted consumers (the cheap "am I allowed here" probe); GET /tenants
answers every tenant for a tatlock/scheduler caller, or a list of one for
a tenant-bearing caller — another tenant's existence is never revealed.
No write routes; provisioning a tenant is workspace T-195.
Tests: tests/test_tenants.py and tests/test_tenants_migration.py (new);
a TestTenantRegistryEnforcement class added to both tests/test_book.py and
tests/test_scribe.py proving all four cases (unknown, suspended,
ungranted-consumer, and the control) on a live Book and Scribe route, not
only on /tenants itself. tests/conftest.py gains a tenant_registry schema
fixture and a suspended_tenant fixture (inserted and deleted per test; no
FK references it); book_schema and scribe_schema each independently grant
a second, test-environment-only consumer (office-gemini) in both seeded
tenants, since test_book.py and test_scribe.py have used it as a stand-in
second consumer since before this grant check existed — the migration
itself still seeds only office-claude, matching the dossier.
Mutation-checked: dropping the grant check fails the three
consumer_not_granted tests (200 instead of 403); dropping the status
check fails the three tenant_suspended tests; making Book bypass the
helper fails its unknown-tenant test (200 instead of 403). All three
restored and the full suite re-verified green (508 + 35 = 543 passed).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
005 has never run in production (still at 004), so it is amended in place
rather than patched later by a 007: book_records gains owner_tenant (the
tatlock tenant, Caller.tenant from T-4) and number, a short outward id
allocated per tenant via a new book_tenant_counters row locked FOR UPDATE
in the same transaction as the insert — INSERT ... ON CONFLICT DO NOTHING
bootstraps the counter row so a tenant's very first allocation doesn't
race its own creation. Unique on (owner_tenant, number); composite index
on (owner_tenant, kind, status). 006 applies on top unchanged.
Every Book route now requires a tenant (BookService._require_tenant,
mirroring the Scribe's own pattern) and filters on it: a record in
another tenant answers exactly like one that does not exist, in both
directions. link and add's parent convenience both route through
_get_or_404, so a link can never reach another tenant — which is also
what keeps show's links/backlinks free of their own tenant filter, since
every link touching a tenant's record is created tenant-safe already.
Responses include number; list accepts number= to find one.
Verified: core_api_test downgraded to 004 (with the original migration,
to match its already-applied pre-amendment schema), amended, upgraded
back to head — clean round trip. 508 tests pass (489 baseline + 19 new:
3 migration-shape, 16 tenant/numbering). Mutation checks confirmed and
reverted: dropping list's tenant filter, allocating numbers from a global
counter, and dropping link's tenant check on its target each fail the
test built to catch it. Raw REST against a dev server on :8777 against
core_api_test confirms the same behaviour over the wire: two tenants'
first records both land on number 1, and a cross-tenant read answers 404
not_found.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The Office (workspace D-66, D-68, D-73) needs a message bus: an address is
a topic named for an identity, owned by a tenant and a consumer. D-68 puts
it here, in core-api's own database, behind its own auth, following
src/domains/book/.
Migration 006 lands the whole schema by hand (down_revision="005"):
scribe_addresses (owner_tenant + name primary key — no surrogate id, since
no route ever references an address except by name), scribe_messages
(retained; a generated tsvector search column + GIN index on subject/body,
and a GIN index on the topics array for history's lookup), and
scribe_mailbox_entries (one row per recipient address per message, unique
on address+message_id). No data: seeding `jeroen` and `tower-control` is a
deploy step through `register`, not this migration.
Routes, every one behind require_caller("office", "user"), tenant from the
caller (T-4):
- register/list addresses — register is idempotent for the same consumer,
409 address_taken for another consumer in the tenant; list marks which
addresses the caller's consumer owns.
- send — `from` is always the caller's own identity, stamped server-side,
never read from the body; 403 identity_not_owned if that identity is not
a registered address owned by the caller's consumer; 404 unknown_address
(naming the missing ones) if any recipient does not exist in the tenant;
writes one mailbox entry per recipient.
- mailbox / find — the caller's own identity's entries (mailbox: unread
first; find: full text via websearch_to_tsquery), both requiring the
identity be owned by the caller's consumer.
- history — the retained messages sent to an address, oldest first;
reading never consumes or marks, and reads scribe_messages directly
rather than joining through mailbox entries.
- mark / tag / untag — share one PATCH route; an entry outside the
caller's tenant, or addressed to an identity the caller isn't reading
as, answers 404 not_found — indistinguishable from a nonexistent entry.
Every query filters on the caller's tenant (D-73): an address or message
in another tenant answers exactly like one that does not exist, so no id
is ever confirmed across that boundary (test_scribe.py::TestTenantIsolation,
including two tests proving the tenant filter is load-bearing even when
the same address *name* legitimately exists, differently owned, in two
tenants — not just when the caller's own identity is unregistered there).
Messaging only this round: groups, subscriptions, the notification matrix
and delivery are explicitly not built (workspace T-176) — the schema
already carries what `topics`/`tags` need for them.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The Office (workspace D-73) now signs every visitor in through Authentik
with one OAuth client per consumer, and every token carries a signed
tatlock_user claim. Caller gains consumer (X-Office-Consumer, the
Authentik application slug) and tenant (X-Office-Tenant, the
tatlock_user claim). For kind "office" both are required on every call,
reads included — missing is 400 consumer_required/tenant_required,
malformed is 400 consumer_invalid/tenant_invalid. Every other kind
ignores both headers outright: tatlock and scheduler carry no tenant,
and a "user" caller's tenant comes only from their own JWT's
tatlock_user claim, never from a header — a header must not be able to
put a non-Office caller into a tenant.
X-Office-Client (informational) is retired in favour of
X-Office-Consumer. book_events.client keeps its name (no migration this
round) and now records the consumer instead.
Book routes do not start filtering by tenant this round — that is
core-api T-2.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
T-2 finishes the Book before its first deploy: the remaining verbs,
short outward ids (amending migration 005), and work/personal as tags.
T-3 names the alembic target instead of inheriting production.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The dossier's verification wants a note *and* a task found by find();
the previous version only ever searched for the note's words. Search
for the task's own body text and assert it is the record returned.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The Office (workspace D-64..D-72) needs a store for Notes, Tasks,
Decisions, Questions and Rejected options. D-68 puts it here, in
core-api's own database, behind its own auth.
Migration 005 lands the whole schema now, by hand (down_revision="004";
the chain has a hash-named revision between 003 and 004, confirmed
before trusting the numbering): book_records (kind, status validated
per kind in the service, scope work/personal, a generated tsvector
search column + GIN index), book_tags, book_record_links, book_attachments
(table only, no file handling yet), and book_events — append-only via a
database trigger that raises on UPDATE and DELETE for every role,
including the owning application.
Routes this round (D-70's core verbs), every one behind
require_caller("office", "user"), writes with identity_required=True,
every change writing a book_events row: add, show, list (the nine
views), find (full-text), change, close (task done/dropped; question
resolved via a `resolves` link into `settled`; decision superseded via
a `supersedes` link into `superseded`), link/unlink, tag/untag, history.
Append, remind, reschedule, undo, purge, redact, pin, check, snooze,
archive/trash/restore, pause/resume, due, next, batch calls and
idempotency keys are explicitly not this round — the schema already
carries what they need.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
api.schweitz.net is internet-facing with no forward-auth, and every
GET under /housekeeping/* (health, devices, areas, scenes, scripts,
automations, history) was unauthenticated. It answers 500 today only
because Home Assistant's token is dead; once that token is restored it
would serve the house's device list to anyone.
Reads now go through require_caller("office", "tatlock", "user").
tatlock is in the list because tatlock's Housekeeper
(tatlock/src/agents/housekeeper/client.py) is the only caller of these
routes today and already sends Bearer CORE_API_KEY when it is set.
Writes are untouched — control_device, activate_scene, run_script and
toggle_automation still depend on get_admin_user. Whether tatlock may
write is a separate decision, not this one.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
core-api had no service-key auth: AuthService.validate_api_key was never
called, api_keys stayed empty, and CORE_API_KEY was read by nothing. This
is the first thing that reads a service key.
require_caller(*allowed, identity_required=False) resolves, in order:
office_service_key -> tatlock_service_key -> scheduler_service_key ->
the existing get_current_user() JWT path. Each static key is compared
with hmac.compare_digest, gated on the configured key being non-empty so
an unset key can never match. A disallowed kind is 403 forbidden.
The office caller's identity comes from X-Office-Identity: required on
writes (400 identity_required if missing), validated against
^[a-z0-9][a-z0-9-]{0,62}$ and against the reserved/group- set (400
identity_invalid). Other kinds carry a fixed identity and ignore the
header.
New settings (src/shared/config.py, default ""): office_service_key,
tatlock_service_key, scheduler_service_key; and book_timezone
("Europe/Amsterdam") for the Book's due/remind handling.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The trailing wildcard on the three rm -rf deny entries spanned path
separators, so Bash(rm -rf /*) matched every absolute path on the
machine rather than the filesystem root, and the ~ and $HOME entries
had the same shape. Narrowed to the exact literal forms.
These rules match literal command text, so they still stop a typo on
rm -rf /, rm -rf ~ or rm -rf $HOME exactly, but they no longer stop a
recursive delete aimed at any other path. That reduced cover is
deliberate, not an oversight.
test_dns_lookup_returns_result patched src.controllers.tools_controller.
DNSService, which is never imported by the request path under test.
client wraps src.main.app, which routes /tools/dns/lookup through
src.domains.tools.controller.tools_controller — a singleton constructed
at import time from src.domains.tools.dns.service.DNSService. The patched
class was dead; the mock was never consulted, so the test issued a real
DNS query for example.com and asserted on its outcome. With no network
the query times out and the assertion fails (D-26).
The three sibling tests in the same class patch the same dead class and
also run unmocked, but happen not to notice: DNS failures are caught
inside DNSService.lookup() and returned as a normal 200 response with
success=False, and their assertions only check status_code / DNSQueryError
branches that don't depend on resolution actually succeeding. Only this
test's `data["success"] is True` assertion is sensitive to the real
network outcome, which is why it's the only one D-26's namespace run
catches. Not touched here — out of T-55's scope, flagging for the record.
Fix patches tools_controller.dns_service, the actual instance attribute
the live route calls, via patch.object on the singleton rather than
patch() on the constructor class (the instance already exists by the
time a class-level patch would apply).
Verified:
- unshare -rn (lo up): 381 passed, exit 0 (was 1 failed, 380 passed, exit 2)
- with network: 381 passed, exit 0 (unchanged from before the fix)
- mutation check: retargeted the patch.object to a nonexistent attribute
name, confirmed count==1 before editing; namespace run then reproduced
the original failure (1 failed, 380 passed); reverted and reconfirmed
381 passed, exit 0.
Co-Authored-By: Claude <noreply@anthropic.com>
pip install exits 0 whether the result is usable or not — that is the D-24
shape exactly, a step whose job is to not fail. On 2026-08-09 the venv here
existed and pip had already succeeded, but sqlalchemy was declared in
requirements.txt and not installed. That surfaced as 11 pytest collection
errors that read as broken imports rather than as an environment problem.
setup now ends with `pytest --collect-only tests/`, which exercises every
import the suite touches without running anything. pip check was considered
as a cheaper alternative and rejected: it only checks the installed set's
internal consistency against itself, so it would not have caught this case —
sqlalchemy was still present as another package's transitive dependency even
after being dropped from requirements.txt. collect-only checks declared vs.
actually usable directly, which is the axis that broke.
Verified: two clean runs (33.5s cold, 5.7s idempotent re-run, second changes
nothing). Reproduced the original failure by uninstalling sqlalchemy from an
otherwise-correct venv — collect-only alone then exits 2 with 11 collection
errors; make setup against that same state reinstalls it and exits 0 with
381/381 collected.
Co-Authored-By: Claude <noreply@anthropic.com>
The hook carried ~50 lines of gitleaks logic and a comment explaining it was
self-contained because "this repo has no Makefile". It has one now, so the
reason is gone and the arrangement is backwards: a hook is a trigger, and
logic belongs where it can be read, run by hand, and changed under review.
.githooks/pre-push is now a byte-identical shim onto `make pre-push` in every
repo in the workspace. The scan itself moves to ci/secrets.sh unchanged, and
`make secrets` runs it on its own.
The call surface is identical everywhere; what it runs is not, and should not
be — each repo gates what it actually has. That is the point of standardising
the name rather than the contents: nobody has to read a repo to find out how
to check it.
secrets runs first, deliberately. It is the only failure here that cannot be
undone by fixing it afterwards — a failed lint costs another commit, a pushed
credential is cached and indexed whether or not it is later deleted.
Some of these gates fail today, on lint debt that predates them, and they are
left wired anyway. The board was measured once and written down in T-56
instead of being worked around here. Narrowing each gate to whatever already
passes would produce a gate that reports success for doing nothing, which is
the failure this workspace keeps rediscovering.
Co-Authored-By: Claude <noreply@anthropic.com>
Environment guards now exit 69 rather than 1, so a caller can tell a suite
that could not start from one that ran and failed. The first toj test sweep
reported "3 repositories failed" and none of the three had executed a test —
two could not find go, one had no venv. That points the reader at the tests
when the fault is in the environment.
Only the environment guards change. A gitleaks finding, a failed test run and
a vulncheck hit still exit 1, because those did run and did fail.
Co-Authored-By: Claude <noreply@anthropic.com>
The suite could not even collect: the venv was missing declared dependencies,
and five tests asserted an API that had moved on. 11 collection errors to 381
passing.
test_oidc.py was written for the single-issuer API and 6243f29 replaced it.
issuer and audience became lists, jwks_uri stopped being an attribute in
favour of get_jwks_uri(issuer), get_jwks became get_jwks_for_issuer, and the
lru_cache became a per-issuer dict so cache_clear no longer exists. Rewritten
against the current surface, with coverage added for the two behaviours the
multi-issuer change introduced and never tested: is_valid_issuer rejecting an
unconfigured issuer, and the cache keying per issuer. Both are
security-relevant — a shared cache would serve one issuer keys for another.
Three /auth/me tests asserted a path that does not exist. The route is
declared as /me inside AuthController.create_router() and mounts at
/auth/users/me; the generated spec is authoritative and the local app and the
deployed service agree on it. Those tests had never passed.
test_handles_empty_groups expected groups == [""] for an empty header. oidc.py
has returned [] since the initial commit, and [] is correct — [""] would also
be unsafe, since any check doing "" in groups would match.
test_model_aliases_property covered Settings.model_aliases, deleted with the
Ollama integration in c1f16d4. Removed rather than repaired.
Co-Authored-By: Claude <noreply@anthropic.com>
Every repo gets one at the root: help, plus test and lint where those exist.
The point is that a target name means the same thing in every repo, so an
agent or a person can act without reading the repo first.
Paths resolve here rather than in callers (D-10). python3 on this host is 3.8
and cannot parse these sources, and a bare pytest or ruff resolves only in a
login shell — so both are named explicitly through the venv, and a missing
venv fails with the command to fix it rather than a bare no-such-file.
Co-Authored-By: Claude <noreply@anthropic.com>
pql is now a bare word on PATH, which removed the long incantation that had
been forcing --vault into every call by habit. Convenience lowered the cost
of the wrong thing without lowering the cost of the right one: a three-word
pql ticket new targets whichever vault the cwd happens to sit in, and there
are nine of them with colliding id sequences.
PQL_VAULT in each project settings file makes the vault a property of the
session rather than of the working directory — the same lesson Rule 3 records
for git -C, applied to pql. Verified the env var overrides cwd discovery,
that an explicit --vault still beats the env var, and that the harness
hot-reloads it without a restart.
This does not make provenance visible: no output says which vault answered,
so a forgotten --vault still returns a well-formed answer about the wrong
dataset. That remains T-37.
Co-Authored-By: Claude <noreply@anthropic.com>
toj is now on the global PATH as /usr/local/bin/toj, so its scope boundary
had to stop being "the absolute path is inconvenient to type" and start
being a rule. Its repo and settings verbs operate on the workspace root; run
from inside this repo they answer about the wrong tree.
Both spellings are denied, bare and absolute, because a deny with one
spelling left open is decorative.
Co-Authored-By: Claude <noreply@anthropic.com>
No repo here scanned for committed credentials. The hook is self-contained
rather than delegating to a Makefile, because this repo has none and a hook
reaching into a sibling repo breaks the moment this one is cloned elsewhere.
Scans the outgoing range rather than full history: history carries settled
findings — test fixtures, vendored third-party code — and a gate that fails
on something unfixable gets bypassed within a week.
Setting core.hooksPath means pql init must replant its replication shims into
.githooks, which is why they are gitignored here alongside the tracked
pre-push. Same layout pql itself uses.
Co-Authored-By: Claude <noreply@anthropic.com>
Decision ids are per-vault sequences, so they collide by construction
once there is more than one vault -- and every repo now has one. A bare
D-15 here will mean this repo's D-15 the moment this repo records one.
Cross-vault references are therefore qualified: workspace D-15.
Not hypothetical: pql holds D-1 through D-31 while the workspace holds
D-1 through D-21, so every workspace id currently collides with an
unrelated pql one. A bare id is not wrong the day it is written -- it
decays into wrong as the other vault grows, and nothing flags it.
Co-Authored-By: Claude <noreply@anthropic.com>
Commits a .claude/settings.json rather than leaving permissions to
per-developer local state, and initialises a pql vault for this repo's
tickets and internal decisions.
Every git deny rule appears in both the `git <verb>` and `git * <verb>`
forms. Only the second catches `git -C <path>`, and without it the whole
deny list is decorative -- it looks like a policy and stops nothing.
The allow list carries pql's absolute path alongside the bare name.
pql is installed to ~/.local/bin, which is on the login PATH but not the
one a non-interactive shell gets, so the bare-name rules match nothing on
their own and every call would prompt anyway.
.gitignore now covers .claude/settings.local.json, which is machine-local
and must never be shared. `pql init` contributed the .pql/* rules with an
exception for the changelog, which is the replication log of record and
has to be committed for tickets to travel with a clone.
Co-Authored-By: Claude <noreply@anthropic.com>
One agent doc per repo, and it is CLAUDE.md. Two docs describing one repo
drift, and the one nobody read is always the one holding the rule that
mattered. Written fresh rather than reformatted, so the structure follows
what someone working here actually needs.
Two rules from the old file are gone deliberately. The mandate to branch
for every change was retired in favour of one linear-history policy, and
the release snippet used `git add -A`, which sweeps in whatever else is
dirty.
The architecture section is the part worth reading. An earlier draft
called src/auth, src/controllers, src/clients, src/dns and src/models
dead code, derived from grepping main.py's imports. That was wrong:
main.py:55 calls initialize_oidc(), which imports and configures
src.auth.oidc from inside the function body, so src/auth is configured
with live Authentik issuers on every boot. It also missed four genuinely
unreferenced packages. The section now states the method used -- import
the app in the container and read sys.modules -- and its blind spot, that
a cold snapshot cannot see a module imported on a request path.
Also records that the README's "runs as non-root user (uid 1000)" is
false: the Dockerfile has no USER directive. Flagged rather than fixed,
since changing the runtime user is not a docs change.
Co-Authored-By: Claude <noreply@anthropic.com>
