PR #1916 rule: a refusal names the field or file, never a config value
or a host path. readJsonFile's three refusals and parseJsonFileText now
name the file by its `what` label only. Callers whose label alone is
ambiguous add a name that already passed a strict pattern (business id,
role name, org id, turn id, project key) or a fixed path under the data
root. Exit codes, refusal codes and check order are unchanged.
organization: util, access, project, config-command, library-seats,
org-service deleteOrg receipt, agent-levels regex; T1 files role.mjs and
organization-config.mjs change only the readJsonFile label.
cli: access lock, accounts owner, host state/lock, notifier journal,
notify-config, run-evidence, dev-fixture, git-socket, org-bus, org-console.
discord: turn record label names the turn id.
Sentinel tests (sk-test-NOT-A-SECRET-1234 as a path component) cover each
changed refusal and fail on BASE; old path pins updated.
The row299 root-side pair (mos-1409) spans supervisor start-up and
clean_session, which scans every /proc entry on the host. With about
1,850 processes that adds 0.6 to 1.6 s, so the unchanged inner probe read
2.5 to 3.9 s and failed the [1.25, 2.75) band, while the 1 s inner-timer
mutant read 2.06 to 2.38 s and passed. Widening the bound alone would
have kept that mutant alive.
A second observer pair, also confined to the private test copy, brackets
only the inner timer: begin just before the supervisor starts its sleep,
end at the start of finish(). It writes only for the SO_PEERCRED peer and
at most one pair. The unchanged band now applies to that interval
(2.01 s on base, 1.00 s and 3.00 s for the 1 s and 3 s mutants, so both
fail). The root pair remains required and must contain the timer pair;
probe-inner-timeout already places it before the outer bound.
test-case2.sh failed 8 of 21 steps on base from S3b on. Bisect names
f23a998ee (#797, "close project archive admission races") as the first
bad commit: it made every launch call request.admission and updated each
in-process launcher fixture except scripts/case2/launch.mjs. The fixture
host refused the op with invalid-host-request, so no launch started; the
S6/S7 "unauthenticated" and S8/S10 reds followed from the missing launch
capability, not from a separate session change.
The fixture now spreads requestHostOps from packages/cli/tests, the shared
map every other fixture uses, so it tracks the production protocol.
mosaic accounts init refuses accounts-unsafe since #1919 when a folder above the
config directory is writable by another user. Step 3 now gives the read-only check.
assertAccountsOwner refused a symlink or non-directory on the config path,
but checked owner and mode only on the leaf. An ancestor owned by another
user, or writable by others without the sticky bit, let that user rename the
directory away and put their own accounts.json in its place. Since M3 that
file decides root authority on a deployed Console.
One shared helper, config-dir-trust.mjs, now holds A2's rule: every ancestor
belongs to root or this uid and is group or other writable only with the
sticky bit; configDir itself is this uid's and not group or other writable.
The accounts store (load, init, lock, locked write, mosaic accounts) and the
root agent file's writer both call it, each with its own refusal code. A2's
messages are unchanged.
Accounts refusals no longer quote a path: the config directory and accounts
file uid messages name the rule and the uid only (redaction rule, #1916).
A missing config directory now refuses accounts-unsafe instead of a raw
ENOENT.
The rehearsal provider matches routes on the parsed pathname, so a
query string no longer turns authorize or token into a 404. Authorize
and token answer 403 to any method before the GET check, and only
discovery and the JWKS answer 405 to another method. R20 now sends a
POST to token and a GET to authorize with a query, and expects 403 for
both.
The upstream A2 agent-levels write test stubbed an owner session
without rootOwner. M3 sets the hook's owner flag from rootOwner, as
sso.mjs grants it, so the stub now carries rootOwner: true, as
system-types-http does. No check changes.
An installed unit must equal the render but for the copy id, so a
re-pin that adds --sso refused render-refused. case2-unit gains --check
(every render check, nothing written) and --check-new (a first install's
checks). stack-upgrade runs both checks before anything stops, and
--next-args reads root's <instance>.args.next: equal before -- but for
--copy, different after it. After the stage it moves the installed unit
files into a new <instance>.replaced.XXXXXX directory and renders a
first install. A failure moves them back unchanged and restarts the old
copy, and confirm swaps the args file. The run sheet gives the commands,
the rollback past the migration included.
A stamped rehearsal now starts an isolated provider on 127.0.0.1 port 0
with a certificate and keys made for the run. It serves discovery and a
JWKS for the live provider ids and answers 403 to authorize and token.
Only the scratch Console trusts its certificate. The gate requires each
provider ready (sso-ready); provider-discovery requires discovery and
JWKS fetches and no code or token request. The provider stops with the
run, and by itself once the run's shell is gone.
The system-path test helper asserts the log's from copy again. The
pinned-copy sign-in requests are bounded, and cleanup waits for each
Console to exit.
stack-release refuses a user-unit activate, rollback or pm-on when the
instance shows any accounts sign (stamped target, accounts.json,
console.sso, .deployed), before anything stops. Root renders the system
units with case2-unit.sh, and the op runs with --units external.
case2-unit renders a stamped copy's units only with --sso, console.sso,
an accounts.json that loads and a valid .deployed, and names each
missing step in run-sheet order before any write. Per Jason's ruling of
2026-11-10, a stamped library unit takes --sso and refuses proxy-preview
admission and its proof. The unstamped and dev paths are unchanged.
rehearse runs a stamped copy against rehearsal.invalid with scratch
slot placeholders, a copy of the live accounts.json and the copy's own
.deployed. It refuses sso-not-prepared and accounts-not-initialized
before any rehearsal directory exists, and sso-isolated before the host
starts.
The tests that turned SSO on through the user unit now pin the refusal
and reach the same end state through the system path. The copy tests
start the Console on real scratch prerequisites and pin each refusal.
The run sheet, TOOLS.md and the accounts design record the ruling and
the migration order.
- A checkout an earlier launch left for a project the instance may no longer
work in is covered by an empty tmpfs inside the sandbox; the host copy stays.
- The launcher rechecks every prepared project against the reloaded document
after preparation, and the bus's final admission takes the instance and
checks the broker's current unions.
- A project define whose vars changed since boot refuses its copy until the
next start.
- Instance-name collisions are recomputed at each reload from the frozen root
claims and the current project defines; every claimant refuses.
Review fix round for slice A3.
- Every project's union is recomputed from the current organization file
at each load. A refused union refuses the project's own copies and every
inherited root or org agent there: no mount at launch (checkout-skipped,
project-refused), no thread and no room, each naming the project and why.
- A project exclude removes that project's mount from the excluded agent;
a narrow reaches the instance's resolved limits, so its real bundle is
narrowed. A narrow in any project narrows the instance for every launch,
since one launch mounts every project it works in.
- Two copies whose instance names collide both refuse, naming each
claimant; the view lists the claimants. An organization instance of the
same name replaces a root or project copy, at boot and after it, and the
copy is gone rather than refused.
- Rooms hold only the thread instances that work in the room's project.
- The grid finds an agent's copy by name and project from the bus view,
never by rebuilding the instance name.
- Real-boot regressions: inherited agents in a refused union, exclude and
narrow in the bundle, an org override after boot, two organizations each
running a copy, a root file change while running, and a grid fed by a
booted bus.
Each root agent an organization does not exclude, and each agent a project
defines, becomes a copy when the organization's bus boots. The root file is
read once, through loadRootAgents, at the cli boot step; the snapshot is fixed
until the next boot. A bound copy joins the organization document's roles and
goes through validateBusinessDocument and checkRoles with the organization's
own credential map. A copy whose role needs a credential service is unbound
and refuses at launch by name and service (credential-unbound); a copy the
organization refuses is copy-refused; every agent in a project whose union
refuses is project-refused, naming the project. A project agent launches
only with its own project's mounts and takes a thread only in its project
(project-agent-elsewhere). Refusal events carry the code alone. The bus's
instances read carries the copies with their running runs, and the Agents
grid shows each copy's real status and reason.
channel-proof re-pins webui bus.mjs for four STATUS entries; no proof or
transport change. packages/business gets the agent-copies.mjs re-export.
The CLI agent verbs validate the instance config through loadSystem before
any read or write, and print a fixed line for errors they do not own. The
root writer refuses a config directory with an ancestor another user can
write without the sticky bit. The project HTTP route checks agent names
with the organization ID_PATTERN and answers 400 naming the key.
The Settings level forms take their own class, so the role-instance
handlers no longer catch them, and keep both drafts in module state with
the revision each began at. The organization writer tests now reach the
file's own uid check and test the assistant protection with an assistant
in the fixture.
Root writer for <configDir>/agents.json (owner only, revision-checked, atomic, A1-validated,
receipts name agents and keys only), the project.agents writer verb on the organization
writer, `mosaic agents root` and `mosaic project agents` verbs, and owner-only Settings >
Agents edit forms on GET/POST hooks gated like /api/agents/levels.
The mark from the previous commit kept the wait off the old page, but
the wait could still start on it and be killed by the reload
("Inspected target navigated or closed", once in ten runs). navigate()
waits for the load; a query makes the same URL load again.
Review fix. dialog.close() clears `open` at once but queues the close
event that empties the dialog. A conflict read that resumed in between
still found the old form connected and reopened the dialog, and the
queued event then emptied it: an open modal with nothing in it. The
check is now the dialog's own `open` state and that it still holds the
form. A browser test closes the dialog and releases the read's body in
one task; it fails on the previous commit.
The wait after location.reload() could pass on the board that was
unloading, and the next read then found no board (seen once in ten
runs under load). A mark set before the reload goes with the old page,
so the wait now passes only on the reloaded one.
A product race, not only a test race. On task-conflict the dialog wrote
the refusal into the old form, turned its buttons back on and only then
read the current task. For the length of that read the form showed the
refusal beside the old task and a Save went out against the old digest.
The conflict test read the title in that window and failed under load.
The old form now stays disabled with "Reading the current task…" until
the read answers; the refusal and the reopened form appear together. A
read a poll overtook no longer reopens on the board's old read; the
dialog reads the task itself. If the task cannot be read the form keeps
only Cancel and takes no save. A dialog closed during the read stays
closed, and a form whose Save is disabled refuses a submit sent past it.
The requirements page drew project-request-select, project-north-star
and project-prd with fixed ids, so the same project in both panes put
each id in the document twice. mountContent takes the pane's idPrefix
(pane-a-, pane-b-) and every id and id lookup on the page follows it;
the full view passes none and keeps its ids.
The dashboard was already prefixed and the Agents page draws no ids.
A new browser test opens the same dashboard, Agents page and project
pages in both panes and checks for duplicate ids and for labels, ARIA
references and in-page links that leave their pane.
Review finding 1: step 7 compares messaging.receiveFrom with the chief's as
well as sendTo, since the broker checks the recipient's list too.
Review finding 2: checkComposition reads the role and the chief once, as a JSON
copy; a getter, proxy, toJSON, cycle or BigInt that throws refuses with a fixed
message, so the thrown text never reaches the caller.
Roles round 2, slice P1. checkComposition(role, registry, chief, { assertClean })
runs the design's seven steps in order against a registry from buildRegistry
(core vocabulary plus a caller-passed plugin set, empty by default) and returns
the flagged lines: every decision.resolve.* action and everything above the
chief's own resolved role. Refusals name the field and position, quote only
registered names, and never quote instructions or personality text.
COMPOSITION_BOUNDS holds the length bounds the design leaves open. Nothing
calls checkComposition yet; HIRE_ALLOWED and unhireableAction stay until A4.
jason.woltje