* docs(b2-7): trust model — who can read what (BPR-050/051, C-09 contract)
One document states the operator trust model in plain language and traces
every claim to a code line or test: server-readable text and files and why,
E2EE voice/video/screen with the key-holder and TOFU rules, transport per TLS
mode with desktop pinning and the browser rule, the C-09 preview destination
contract B7 implements, at-rest storage, operator can/cannot, multi-device
sessions, and what beta does not claim. Linked from security.md,
deployment.md, quick-start.md and docs/README.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* test(b2-7): absence proofs — no federation/directory routes, outbound host table
TestAbsenceContract_NoFederationDirectoryOrListingRoutes builds the production
router with uploads, voice and GIF on, walks the whole mounted tree with
chi.Walk and fails on any route matching federat|directory|discover|listing.
A floor on the route count and a check that the admin subtree was traversed
keep it from passing vacuously. trust-model.md gains "What OwnCord does not
have" (BPR-040/082/083) and the outbound-host table B6's network capture
checks against.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* docs(b2-7): plugin boundary — off twice, compiled out of releases, no API promise
docs/architecture/plugins.md records the experimental WASM boundary (BPR-080/
081, BG-17): disabled by build tag and by config, absent from release.yml and
Dockerfile builds, the HP-2 configuration audit (fresh, upgraded, Docker,
standalone), the beta release-notes wording, what exists today with its limits
and tests, the post-beta plugin candidates that stay in core during beta, and
the core concerns that never move. Linked from architecture/README.md,
architecture/server.md and docs/README.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* docs(b2-7): L-08 re-tagged to B10 with the reason; B2-7 evidence block
L-08's remaining gate ("deterministic source build passes") cannot pass in
principle — TinyGo embeds host paths and has no -trimpath — and a compile-only
drift check would need a second Go SDK, TinyGo and Binaryen on every PR for a
subsystem release builds compile out. Re-tagged to B10, which runs the compile
once against the release candidate or closes on the provenance record; the
"no API promise" half is closed by docs/architecture/plugins.md. The plan's
B2-7 evidence block records the four pre-squash SHAs, the absence test's RED
output, the release-build finding, the decision, and the BPR-051 reader
placeholder. CHANGELOG gains a Documentation block.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* fix(b2-7): Codex review — TOFU windows disclosed, legacy TOTP, absence test at three boundaries
Two P1 and three P2 from the Codex review of 56f23a36, all verified against
the code. trust-model.md now states the desktop's first-connection TLS TOFU
window and the out-of-band fingerprint check; scopes identity-key pinning to
changes after the first pin, not first contact; discloses that databases from
before TOTP encryption may still hold plaintext secrets and how re-enrolment
fixes that. The absence contract gains two sibling tests — WebSocket wire
types from protocol/schema.json and every koanf key of config.Config (one
allowlisted on-disk path whose presence the test asserts) — and the document
states what the three tests bound. plugins.md corrected: unknown config keys
are warned about and ignored, not rejected. Evidence block records PR #1443
and the review outcome.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* fix(b2-7): Codex re-review — desktop first-use window in every TLS mode; tls off is plaintext
The desktop pins the fingerprint it sees on first connection in every
tls.mode (CaptureVerifier, no web-PKI validation on the server connection),
so a public-CA certificate closes the first-use window only for a browser;
the short answer and the pinning list now say so. tls.mode off served
directly is plaintext HTTP with nothing enforcing a proxy; the transport
table row states it. Evidence block records the round-2 outcome.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* fix(b2-7): Codex round 3 — E2EE does not survive a hostile operator on first contact; memory cap is server-wide
Identity keys are trusted on first use, so a modified server can deliver an
unpinned peer's first announce with keys the operator holds and the key
holder wraps the room key to it. trust-model.md now scopes E2EE to an
operator who reads, and to a modified server only for peers pinned and
compared out of band beforehand, in every place the stronger claim stood.
plugins.md: the wazero runtime is sized from plugins.max_memory_mb alone;
a manifest's memory value is validated but not applied. Evidence block
records the round.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* fix(b2-7): Codex round 4 — transport intro excludes tls off; ipAllowed is not the full C-09 deny-set
The transport section opened with "everything is TLS"; it now excludes
tls.mode off. The C-09 contract cited the server's ipAllowed as the complete
address deny-set; it rejects loopback, private, link-local, unspecified,
multicast and CGN only, so the clause now lists the documentation and
benchmarking ranges the native broker must add and records that widening
ipAllowed is a separate server change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* fix(b2-7): Codex round 5 — backups exclude uploads; release-time appimagetool fetch listed
The built-in backup is VACUUM INTO of the SQLite file only; the trust model
and deployment.md now say uploaded files are not in it and upload.storage_dir
needs its own backup. The outbound-host section scopes "no script fetches an
external host" to scripts the server runs and lists the release workflow's
build-time appimagetool download.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* fix(b2-7): Codex round 6 — one pin per account, LiveKit media in the host table, quick-start wording
Identity pins are keyed by host and user id, so a peer's second device
overwrites the pin and the first device then mismatches; the document no
longer claims per-device pinning. The outbound-host table gains the
supervised LiveKit subprocess's WebRTC media and scopes the capture contract
to traffic the server initiates. The quick-start cross-link no longer says
voice and video are unreadable by the operator.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* docs(b2-6): enumerate the security-sensitive mutations and their audit coverage
Step 1 of B2-6: the mutation inventory crossed with the 43 non-test
Audit( call sites at 67fdd18d, recorded in the plan's evidence block.
Invite create/revoke (S-02) and plugin install/uninstall have no audit
row; no timeout mutation exists (kick is force_logout / voice_mod_kick).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* added new doc
* docs(audit): update changelog and security documentation to include admin panel actions in audit log
* feat(b2-6): audit every security-sensitive mutation, invite and plugin rows first (S-02)
Step 2 of B2-6. TestAuditCoverage_* in service, api and admin drive each
mutation from the plan's inventory against a fake db.AuditStore
(Server/db/audittest) and assert the expected action arrives. Red before
this commit on exactly four rows: invite_create, invite_revoke,
plugin_install, plugin_uninstall.
- InviteService writes invite_create / invite_revoke naming the invite by
id, never by code; RevokeInvite now takes the actor, threaded from the
handler. A failed revoke writes nothing (test).
- The plugin admin handler takes a db.Auditor and writes plugin_install /
plugin_uninstall against the RequireAdminAuth principal
(admin.ActorIDFromContext, exported for that).
- docs/security.md lists the four new actions; CHANGELOG under Unreleased.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* test(b2-6): denylist over the recorded audit detail corpus
Step 3 of B2-6. Each TestAuditCoverage_* table now ends with a subtest
that runs audittest.AssertSafeDetails over every entry its rows recorded:
a shape denylist (bcrypt/argon2 hashes, password=/token=/secret=/
recovery-code key-value leaks, otpauth URIs, Bearer credentials) plus the
fixture's own secrets (raw tokens, passwords and hashes, TOTP secrets and
codes, invite codes, message bodies). audittest_test.go proves each class
bites and that ordinary details pass. Zero hits on the corpus at HEAD, so
no call site needed changing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* docs(b2-6): record pre-squash SHAs, red/green and denylist evidence
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* fix(b2-6): audit committed invites on a canceled request; 404 unknown plugin uninstall
Codex P2s on #1441, both test-first:
- CreateInvite read the invite back on the request context, so a cancel
after the insert committed returned an error and skipped invite_create.
The read-back and audit now run on context.WithoutCancel, like the
password-change tail. TestCreateInvite_AuditSurvivesCanceledLookup.
- Registry.UninstallPlugin is idempotent on an unknown id, so the handler
wrote plugin_uninstall for plugins that never existed. It now checks the
row first: 404 and no audit. TestPluginsHandlerUninstallUnknownID.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
* docs(b2-6): record the Codex review outcome on #1441
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Rg9QQWVN3E5UUgBD2dydtu
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* ci(claude): constrain automation triggers and bound run cost (L-16)
The Claude Code workflow consumes a metered credential, and the repository
stated nothing about who may spend it or for how long. Whatever downstream
behaviour happens to hold, an invariant this repository depends on should be
asserted and tested here, not inherited from a pinned dependency that a routine
version bump can re-derive.
Three controls, in the one workflow that spends:
- **Authorization.** The job condition now requires the actor to be on an
explicit maintainer allowlist as well as the trigger text to mention the bot.
An allowlist rather than an association check: this repository has exactly one
collaborator, the term is unambiguous to read and to review, and it matches
the actor-term pattern `ci.yml` already uses to exclude Dependabot. Adding a
login is a one-line edit, which is the honest cost.
- **Duration.** `timeout-minutes: 30`, in the band every other long-running job
here uses. Without it the job inherits GitHub's 360-minute default — the wrong
ceiling for metered work, and the only job in the repository that lacked one.
- **Fan-out.** A `concurrency` group keyed on the issue or pull request number
with `cancel-in-progress: true`, so repeated triggers on one thread collapse
into a single run instead of running in parallel. Exactly one of
`github.event.issue.number` and `github.event.pull_request.number` is present
per triggering event, so the key is stable across all four.
The `permissions:` block and the checkout are deliberately untouched. The
permissions are already minimal and the checkout takes no `ref:`, so it reads
the base branch rather than proposed code — both correct, and rewriting either
would be churn.
`scripts/check-workflow-guards.mjs` keeps all three from silently regressing.
Modelled on `scripts/check-doc-counts.mjs`: same `--selftest`-then-assert shape,
same dependency-free approach. It is text-level rather than YAML-parsed on
purpose — the root has no YAML parser, and adding a dependency to assert that a
file contains a `timeout-minutes` key would be a poor trade. That limit is
stated in the file: these are presence-and-shape checks, not semantics.
It runs from `CHECK_HYGIENE` in `scripts/run.mjs`, so it is reachable as
`npm run check:hygiene` locally and executes inside `Repository Hygiene`, which
is already a pinned required check on `dev`. No new CI job and no new pin — the
guard is blocking from the moment it lands.
`actionlint` cannot do this job. It validates expression syntax, action inputs
and runner labels; a job condition is valid input to it whatever the condition
admits, and it has no notion of cost at all. The two tools are complementary
and both now run.
Also corrects the record: `docs/plans/b1-repository-foundation-2026-08-25.md`
claimed impact was bounded by read-only content permissions. The workflow's own
token block is least-privilege, but that is not the only identity a run can
hold, so the claim was narrower than the truth and is now stated accurately.
And `docs/security.md` gains the private-coordination section that two planning
documents already cite it for. The citation pointed at a policy that was not
written down; it now says what stays private, that the rule covers the
repository's own automation and settings rather than only product code, and
that a commit message on a public repository is a disclosure channel.
Verified: both directions, per guard. `node scripts/check-workflow-guards.mjs`
exits 0 on the current tree and reports four guards present. Deleting the
`timeout-minutes` line makes it exit 1 naming that guard and the invariant to
restore; replacing the actor term with `true` makes it exit 1 naming that one;
restoring each returns exit 0. `--selftest` passes eight assertions covering
every guard's absence, a commented-out guard (which must not count), and the two
shapes that must not trip it — any positive timeout value, and any concurrency
key. `npm run check:hygiene` passes with prettier, shellcheck, actionlint and
both new steps running for real; actionlint accepts the edited workflow.
Not included: the workflow's `permissions:` block and checkout step, per above.
No change to the action version or its inputs. `ci.yml`, `release.yml` and
`load-baseline.yml` are outside this item — none is reachable the same way, and
each already carries per-job least-privilege permissions, and where relevant a
timeout and a concurrency group. `METERED` in the new script lists one workflow
because one workflow spends; a second entry is a one-line change when that
changes.
Refs RL-22, L-16
* feat(intake): structured bug form, and route ideas to Discussions (RL-21)
Both issue templates were Markdown with front matter, so nothing they collected
was structured, required, or validated. A reporter could submit the form
untouched. The Environment block was three bullets with `Windows 11` prefilled
as the OS — the single most common answer, pre-filled, on a project that ships
Windows and Linux builds and an ARM64 client.
And `feature_request.md` existed at all, which is the direct violation: BPR-100
says Issues is the bug tracker and Discussions hosts support, ideas and
community feedback. A feature-request template routes ideas into Issues by
construction.
Done:
- `bug_report.md` → `bug_report.yml`, a real issue form. Six fields are
`validations: required` — what happened, steps to reproduce, component, OS,
architecture, deployment mode — because those six are what turns a report into
something reproducible. The rest are optional on purpose; a form that demands
everything gets abandoned.
- `feature_request.md` deleted. Nothing in the tree referenced either template
by filename, so this breaks no link, script, or workflow.
- `config.yml` gains three routed destinations and keeps `blank_issues_enabled:
false` — which is what makes the routing hold, since a blank issue bypasses
every form and every warning on one.
The new environment fields are drawn from what this project actually ships, not
from a generic template:
- **Architecture** x64 / ARM64, with the note that ARM64 is the Linux desktop
client today and there is no ARM64 server release.
- **Deployment mode** covering the six paths `docs/deployment.md` documents —
prebuilt binary on either OS, from source, Docker/Compose, systemd, Windows
service.
- **TLS mode** matching `tls.mode`'s four values exactly, `off` quoted so YAML
does not read it as boolean false.
- **Network topology** — direct, port forward, reverse proxy, Tailscale — because
voice bugs in particular bifurcate hard on this, and the reverse-proxy path
cannot carry the WebRTC UDP range at all.
- **Separate client and server versions.** They are obtained differently and can
legitimately differ. The server field says where to look — admin panel or the
startup banner — and explicitly tolerates "unknown", because the version is
deliberately absent from the unauthenticated `/health` endpoint as
anti-fingerprinting hardening, so a non-admin reporter genuinely cannot get it.
- **Client webview**, WebView2 or WebKitGTK. No "PWA" option: no PWA exists, B1
excludes browser and PWA work, and BPR-092 forbids presenting unavailable
behaviour as functional. The field is diagnostic today regardless — the desktop
client renders through the OS webview, and that already drives real bug classes.
Every public template now carries the disclosure warning BPR-101 asks for, and
the security contact link is first in the chooser, above the Discussions links.
Four files, 189 insertions, 58 deletions.
Verified: both files parse as YAML, and the form was checked against the issue
form schema rather than only for parseability — 13 body elements, 12 unique ids
with no collisions, every non-markdown element carrying an id and a label, every
dropdown carrying options, and the markdown block carrying neither an id nor
validations (both of which GitHub rejects). `config.yml` has
`blank_issues_enabled: false` and four contact links each with exactly
name/url/about. `npm run check:hygiene` passes.
The gap that verification leaves, stated plainly: nothing in this repository
validates issue-form schema. Prettier confirms the YAML parses and actionlint
does not read `.github/ISSUE_TEMPLATE/` at all, so a file that is valid YAML but
an invalid form disappears from the "New issue" chooser silently. The checks
above are a local stand-in, not the real gate. The live chooser needs a look
after merge — which BPR-100's closure evidence ("dry-run submissions reach the
intended destination") requires in any case.
Not included: the Discussions `?category=` slugs are written as `q-a` and
`ideas`, GitHub's defaults. If this repository's categories were renamed, a
wrong slug drops the user on the category picker rather than erroring — confirm
against the live Discussions tab before relying on them. No PR-template or
documentation changes here; those are the next commit. L-15 is not closed by
this commit alone: BPR-100 names six surfaces and three of them are docs.
Refs RL-21, L-15
* docs(intake): route contributors, and state the security path (RL-21)
The previous commit fixed the forms. This is the half BPR-100 and BPR-102
actually ask for and the B1 plan's bullet does not mention: their closure
evidence names repository navigation, support links and contribution docs
alongside the issue forms, so a `.github/`-only change cannot satisfy either.
Three gaps, each verified rather than assumed:
**Discussions was invisible.** The only link to it anywhere in the tree was
inside `.github/ISSUE_TEMPLATE/config.yml` — the new-issue chooser. So "route
ideas and feedback to Discussions" worked for exactly one audience: people who
had already decided to file an issue. `README.md` and `docs/README.md` now each
carry the routing, so it is reachable from the two pages a newcomer actually
lands on.
**`docs/contributing.md` never mentioned security reporting.** Five files
carry the "never a public issue" rule — the root `README.md`, `CONTRIBUTING.md`,
`SECURITY.md`, `docs/security.md`, `CLAUDE.md` — and every one of them delegates
the full process to `docs/contributing.md`, which is also the document BPR-102's
evidence row sends a fresh contributor to. It said nothing about it. It now has
a routing table and a security section that says the thing that actually matters
on a public repository: the PR description, the commits and the branch name are
disclosure channels, so a fix for a vulnerability describes the control it adds
and nothing else.
**The README contradicted the issue chooser.** The banner said "there's no
support" while the chooser offered a link named "Community Support". Both were
defensible in isolation and together they told a user two different things
before they had read anything else. The banner now says the honest version — no
support *commitment* — and a "Getting Help and Reporting Problems" table names
the right destination for each kind of message without promising a response.
Also in the PR template, which the audit's remedy names as "PR guidance":
- The Test Plan asked for `npm test` / `go test ./...` / `npx tsc --noEmit`.
Those predate B1-4's root facade; `npm run check` is the entry point CI gates
on and the one `CONTRIBUTING.md` and `README.md` now tell people to run.
- A generated-files checkbox naming all five, since CI fails on drift and a
hand-edited generated file is the failure that wastes a cycle.
- A `Not included:` prompt, because `docs/contributing.md` makes a written
deferral a required commit element and the template asked for it nowhere.
- The disclosure warning BPR-101 wants on public templates.
Two stale claims fixed while in these files: `docs/contributing.md` said "ten
status checks are required" three lines from a section that says twelve, and
`docs/plans/README.md` still read "B1-0 done, B1-1 next" six phases later — in
the index that declares itself the authority over plan headers.
Five files, 70 insertions, 12 deletions.
Verified: `git grep "ten status checks"` returns nothing.
`node scripts/check-doc-counts.mjs` still agrees on 21 claims across 8 watched
documents — `docs/plans/README.md` and `README.md` are both watched, so a
count claim broken by these edits would have failed here.
`npm run check:hygiene` passes with prettier, shellcheck, actionlint and the
workflow-guard check all running.
One nearby claim checked and deliberately left: `docs/contributing.md` also says
"four of the ten" a hundred lines later. That is four of ten *CI steps keying on
a cache-dependency-path*, not required checks — correct in context, and changing
it would have been a wrong fix to a right-looking grep hit.
Not included: L-15 is **not** closed. BPR-100's closure evidence requires
dry-run submissions that reach the intended destination, and BPR-102's requires
a fresh Windows and Linux contributor to follow these docs and land a passing
sample change. Neither is a file edit. BPR-101 additionally wants a tabletop
report proving private receipt, triage, advisory and coordinated disclosure —
no such artifact exists in the tree, and this commit does not create one.
`CODE_OF_CONDUCT.md` and `GOVERNANCE.md` do not exist in this repository; adding
them is community-health scope, not RL-21's, and neither is named by the audit
row or the register row.
Refs RL-21, L-15
* ci(release): require exact-SHA gate evidence before publishing (RL-16)
A tag push starts `release.yml` and nothing else — `ci.yml` has no `tags:`
trigger. And `release.yml` re-runs none of the required checks: it verifies the
version, builds, boot-smokes and signs, which is a different question from
"did the gate pass on this commit". So a tag could publish from a commit whose
CI was red, and nothing would notice.
It already has. `v1.2.0-alpha.3` published from `fb04a579`, whose CI run
concluded **failure** — `Server Build & Test (windows-latest)`, the race and
coverage step. The Release run on the same commit went green and shipped. That
is R-09 demonstrated rather than hypothesised, and it is the fixture this commit
is verified against.
The obvious fix — re-run the test suite inside `release.yml` — is the wrong one.
It would double the tag-time cost, still not cover the checks that run in other
workflows (CodeQL's three `Analyze` jobs exist in no workflow file at all), and
answer a weaker question: "does it pass now" rather than "did the gate pass on
this commit". The evidence already exists; nothing was reading it.
Done:
- `scripts/verify-gate-evidence.mjs` resolves the tagged SHA's check runs and
asserts every required context is present and `success`. `skipped` and
`neutral` are not success — a required check that skipped on the tagged commit
proves nothing about it — and a still-`in_progress` check is called out as
unfinished rather than treated as absent. Where a context reported more than
once, the latest attempt decides, in both directions.
- The required set is **parsed out of `b0-dev-branch-protection.sh`**, not
restated. Pinning a thirteenth check cannot leave this gate behind, and a
change to that file's shape fails the self-test rather than silently
weakening the gate.
- A `gate-evidence` job in `release.yml` that `verify-versions` needs. Every
build job already needs `verify-versions` and both publishers need those, so
one edge gates the whole graph — including the GHCR push, which today can
mutate `:latest` before `publish` has run at all.
- `permissions: checks: read` and nothing else.
It is a script rather than a `run:` block because of the rule in the `ci-check`
skill: a step that exists only in `release.yml` first executes at tag time, so
its own bugs surface on the release. `Server/scripts/docker-smoke.sh` is the
worked example — one script, two call sites. Here the second call site is
`--selftest`, run by `ci.yml`'s docs-consistency job on every pull request.
`docs/plans/b1-release-tag-protection.sh` covers the half a workflow file
cannot express: a ruleset on `refs/tags/v*` blocking update and deletion, and a
`release` environment with a required reviewer. **NOT APPLIED** — both are
repository-settings writes this session cannot make. Run
`bash docs/plans/b1-release-tag-protection.sh` when you want them.
Deliberately **no `environment: release` key** in `release.yml` yet. The key is
PR-landable, but naming an environment that does not exist stalls the next
release; the script says to add it after creating the environment, and says why.
Verified: both directions, on real data rather than only fixtures. Feeding the
actual check runs from `fb04a579` — the commit alpha.3 shipped from — through
`evaluate` returns **NOT RELEASABLE**, naming `Server Build & Test
(windows-latest): failure` first. Feeding PR #1418's real check runs on
`8875238` returns **RELEASABLE**, and correctly ignores the red
`github-advanced-security` result because it is not a pinned context — the gate
tracks the required set, not "everything is green". `--selftest` passes 12
assertions covering a missing check, a failure, an unfinished run, `skipped`,
`neutral`, both re-run orderings, an unrequired extra, and a commit with no
checks at all. `bash -n` and `shellcheck` are clean on the new script and both
its heredocs parse as JSON. `npm run check:hygiene` passes with actionlint over
both edited workflows.
The module gained a direct-invocation guard so it can be imported and tested
without reaching the network — compared against `argv[1]` rather than
`import.meta.main`, which needs Node 24.2 against an engines floor of `>=24`
and would silently no-op on 24.0.
Not included: the network path itself is exercised only at tag time. The
self-test covers the decision logic and the required-set parsing, which is where
the bugs live; a live API call needs a token this environment does not have.
R-09's "protected release approval" limb stays open until the settings script is
run — the register phases R-09 **B1/B10**, so that half is B10's. `release.yml`'s
version stamping, both signing keys, the fail-closed minisign verify,
`checksums.sha256`'s bare filenames, both cold-boot smokes and the `git archive`
source snapshot are untouched; the remedy says to retain them and this commit
only adds an edge in front of them.
Refs RL-16, R-09
* docs(plans): record B1 progress through B1-7
B1-6 (#1418) merged and B1-7 is this branch, so the header and the plan index
both move on. B1-8 — the platform contract map — is next, and it is documentation
only: it records the browser-neutral contract folders and their owners, and moves
no native behaviour. Adapter extraction stays B7.
Verified: `node scripts/check-doc-counts.mjs` still agrees on 21 claims across 8
watched documents, both edited files among them; prettier clean.
Refs R-08
---------
Co-authored-by: Claude <noreply@anthropic.com>
* chore(format): one Prettier config at the repository root
Every formatting rule in this repository lived under Client/ and covered
exactly two globs: Client/src/**/*.ts and Client/tests/**/*.ts. Root Markdown,
all of docs/, every YAML and JSON, all CSS, the root scripts and
tools/mcp-introspect were formatted by nothing. There was no .editorconfig.
The obvious fix -- a second Prettier config at the root for "everything else"
-- gives two configs and two ignore files that can silently disagree about the
same file. So the root takes ownership instead: config, ignore file and gate
move up, and Client/ folds in. Client's inline "prettier" block, its
.prettierignore, its format/format:check scripts and its now-unused prettier
devDependency are all deleted; knip would have failed client-check on that last
one.
The .prettierrc.json values are lifted byte-for-byte from Client/package.json,
which is what keeps the reformat commit free of client TypeScript churn: 87
tracked files need reformatting and not one of them is under Client/src or
Client/tests.
.prettierignore carries only what .gitignore does not. Prettier 3 reads the
root .gitignore by default, so node_modules/, dist/, coverage/,
Client/src/generated/ and docs/security-findings/ need no entry. It does NOT
read nested .gitignore files, which is why .remember/ is listed explicitly --
38 untracked per-machine scratch files were otherwise able to turn a shared
gate red. graphify-out/ is listed because its seven files are tracked and
.graphify_labels.json is signed byte-for-byte by its .sig, so formatting it
would silently invalidate the signature.
check:hygiene is registered in scripts/run.mjs and folded into check and
release:preflight. It deliberately contains no `gofmt -l` step: gofmt -l prints
offenders and still exits 0, so it cannot fail a build. Go formatting is
enforced separately.
shellcheck and actionlint take their file lists from `git ls-files`, never a
filesystem glob -- .claude/worktrees/ holds a gitignored pre-flatten copy of
the tree with three .sh files a glob would happily lint.
This commit leaves the tree non-conformant on purpose. The reformat is the next
commit, so the 87-file diff is reviewable separately from the rule that caused
it.
Not included: editorconfig-checker. .editorconfig is the editor baseline the
audit asked for; Prettier, gofmt and rustfmt already fail CI on the same
indentation and newline rules, so a fourth tool checking them again is a gate
with no failure mode of its own.
Verified: `npx prettier --check .` names 87 tracked files and zero untracked
ones; the same command listed 38 .remember/ scratch files before the ignore
entry and none after. `node scripts/run.mjs --list` resolves check:hygiene to 8
shell targets and 4 workflow targets. Both package.json files parse.
Refs RL-19 / L-13, S-05.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* chore(format): reformat the tree to the repository Prettier rules
Mechanical. This commit is `npx prettier --write .` and nothing else -- the
rule that caused it landed in the previous commit so this diff can be reviewed
as a transformation rather than as 84 files of hunks.
84 tracked files: 54 Markdown, 7 .mjs, 7 JSON, 6 YAML, 4 .js, 3 CSS, 2
TypeScript (the two Playwright configs at Client's root, which the old
Client/src + Client/tests globs never covered). No file under Client/src or
Client/tests moves, because .prettierrc.json carries Client's former inline
values byte-for-byte.
Prettier rewrote 87 files, not 84. The three in .github/ISSUE_TEMPLATE/ had
CRLF on disk and differ only in line endings, which .gitattributes
(`* text=auto eol=lf`) already normalises, so their committed blobs are
unchanged. Worth knowing before someone reconciles the two numbers.
The largest single diff is .superpowers/findings-ledger.json at 7976 lines
rewritten. That is safe to format: nothing writes the ledger programmatically
-- render-ledger.mjs reads it and writes only FINDINGS.md -- so no tool will
fight Prettier over its style on the next hunt. FINDINGS.md itself is ignored
as generated.
Verified: `npx prettier --check .` reports "All matched files use Prettier code
style", so the pass is both complete and idempotent. All 7 reformatted JSON
files were parsed before and after and compared as values: semantically
identical, zero content changes. `node .superpowers/render-ledger.mjs --check`
still reports 348 valid findings and leaves FINDINGS.md untouched.
`node scripts/check-doc-counts.mjs` still passes its selftest and still agrees
on 27 claims across 9 watched documents -- table realignment did not break the
patterns it matches on. `node scripts/run.mjs --list` still parses.
Refs RL-19 / L-13, S-05.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* chore(lint): enforce Go formatting in the Server linter
S-05: repository-wide Go formatting was not a required gate. The only gofmt
enforcement anywhere was .githooks/pre-commit, which is opt-in per clone
(`npm run hooks:install`), only sees staged files, and warns-and-skips when
gofmt is off PATH.
The obvious fix -- a `gofmt -l` step in CI -- does not work: `gofmt -l` prints
its offenders and still exits 0, so the step passes no matter what it finds.
scripts/run.mjs has the same problem, which is why check:hygiene has no Go step
either.
So gofmt goes where it can actually fail something: Server/.golangci.yml. The
file was already `version: "2"` but had no `formatters:` block at all, so the
19 enabled linters ran with zero formatters. In v2 gofmt/gofumpt/goimports
moved out of `linters.enable` into their own section with its own exclusions.
Adding it there means the gate reports through the Lint step of "Server Build &
Test", which is already pinned as required on dev -- no new job and no new pin.
Every tracked .go file is under Server/ (551 of them, one go.mod), so
Server-scoped is repository-wide here.
One file was genuinely misformatted: a one-space struct field alignment in
Server/admin/handlers_users_broadcast_test.go, fixed in the same commit because
a single line does not need its own reformat commit.
Trap worth recording: `gofmt -l .` on a Windows working tree lists every file
that has CRLF on disk, because gofmt normalises line endings. That reported 18
offenders here, 17 of them ghosts. The blobs are all LF -- .gitattributes
forces `eol=lf` -- so CI never saw them, and the honest test is to run gofmt
over `git show HEAD:<file>` rather than the working copy. Doing that across all
551 tracked Go files found exactly the one real offender above.
Verified both directions with golangci-lint v2 locally: `golangci-lint run
./...` reports 0 issues on the formatted tree; appending a misformatted
function to Server/auth/constants.go produces 2 gofmt findings; appending the
same misformatted function to Server/db/dbgen/admin.sql.go produces 0, so the
exclusion holds. Both files restored and verified clean afterwards.
Refs RL-19 / L-13, S-05.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(scripts): escape the NUL separator instead of embedding one
The `tracked()` helper added earlier in this branch splits `git ls-files -z`
output on NUL. The separator was written as a literal NUL byte rather than the
two-character JavaScript escape, so scripts/run.mjs became a binary file: `git
diff` refused to show it, `grep` reported "Binary file matches" instead of the
line, and `* text=auto` in .gitattributes stops normalising line endings for a
blob it detects as binary.
The code worked -- splitting on a raw NUL and splitting on "\0" are the same
operation -- which is exactly why this is worth fixing before it is inherited.
A source file that tooling classifies as binary is a file nobody can review.
Verified: zero NUL bytes remain, `grep -n "split("` now prints line 50 instead
of "Binary file matches", `node scripts/run.mjs --list` still resolves the same
8 shell and 4 workflow targets, and prettier still reports the file clean.
* chore(lint): enforce Rust formatting
Rust had no formatting gate of any kind: no rustfmt.toml, no `cargo fmt`
anywhere in CI, in scripts/run.mjs, in the Makefile or in the git hooks. Clippy
was the only Rust gate, and clippy does not check layout.
`cargo fmt --all -- --check` now runs in the rust-tests job, ahead of clippy: a
formatting failure is cheap to produce and cheap to fix, and there is no reason
to spend a clippy pass to surface one. The stable toolchain in that job
requested `components: clippy` only, so rustfmt is added there.
Only that job. ci.yml has a second, byte-identical `Install Rust` block in
tauri-build; it stays clippy-only, because a full desktop build is the wrong
place to discover a misplaced brace.
No rustfmt.toml. The default profile is the point of a baseline -- a config
file here would be a second opinion about style with nothing to say.
Client/src-tauri is a single `[package]`, not a workspace, so `--all` is a
safeguard against a future member rather than a fan-out today.
Verified: `node scripts/run.mjs --list` resolves check:rust to three steps with
`cargo fmt --all -- --check` first, `npm run format` now also runs `cargo fmt
--all`, and prettier reports ci.yml, run.mjs and the ci-check skill clean.
`cargo fmt --all -- --check` currently fails on 13 files -- that is the
reformat, and it is the next commit.
Refs RL-19 / L-13.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* chore(format): reformat the Rust crate to rustfmt defaults
Mechanical. This commit is `cargo fmt --all` and nothing else; the gate that
demands it landed in the previous commit so this diff is reviewable on its own.
13 of the 17 tracked .rs files, +343/-164. The crate had never been formatted,
so the changes are the usual first-run set: aligned trailing comments collapsed
to single spaces, single-element slice literals folded onto one line, long
method chains broken across lines, closure bodies expanded into blocks.
Verified: `cargo fmt --all -- --check` is clean, so the pass is complete and
idempotent. `cargo clippy --all-targets -- -D warnings` finishes with no
warnings, and `cargo test --lib` reports 115 passed / 0 failed -- identical to
before the reformat, which is what "mechanical" has to mean for a commit that
touches this much of the crate.
Refs RL-19 / L-13.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(scripts): make the root facade actually run on Windows
Adding the first gate that a contributor would run from the repository root
exposed two bugs in the facade, both of which made it silently wrong on the
platform this project is developed on.
1. Every npm and npx step failed. bin() appends `.cmd` on Windows, but Node
refuses to spawn a .cmd or .bat with shell:false -- the CVE-2024-27980
mitigation -- and fails with EINVAL and a *null* exit status. run.mjs only
special-cased ENOENT, so the result was `FAILED: npx prettier --check .
exited null` with nothing to explain it. check:client has three npm steps and
has never been able to run here.
Fixed by spawning only the npm shims through a shell. They are concatenated
into a single command string rather than passed as an args array, because
shell:true plus a separate array is deprecated (DEP0190) and prints a warning
on every invocation; no argument in this file contains a space.
2. Every optional() step was skipped, always. onPath() shelled out to
`where` on Windows, but where.exe lives in C:\WINDOWS\System32, which a Git
Bash PATH does not necessarily contain -- on this machine PATH carries
System32\Wbem, System32\WindowsPowerShell\v1.0 and System32\OpenSSH but not
System32 itself. The probe could not start, `probe.status === 0` was false,
and golangci-lint and sqlc reported as "not installed" while installed.
Fixed by resolving against PATH and PATHEXT directly. No subprocess, and no
dependency on which directories happen to be on PATH.
A spawn error other than ENOENT now reports its code instead of surfacing as a
null exit status.
Verified: before, `node scripts/run.mjs check:hygiene` died with "exited null"
and both optional steps printed SKIP with the tools present on PATH. After, the
same command runs prettier, shellcheck and actionlint and prints
"check:hygiene: passed", with no deprecation warning. `golangci-lint` is
detected by the new onPath where the old one missed it.
Refs RL-20 / L-14.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* chore(format): ignore build output that nested gitignores hide
Prettier honours the root .gitignore and no other. Every build and scratch
directory in this repository is ignored by a *nested* one -- Client/.gitignore,
.serena/.gitignore, .superpowers/sdd/.gitignore -- so none of them were
excluded from the new repository-wide gate.
The effect is not subtle. Running `cargo test` once drops roughly 850
formattable files into Client/src-tauri/target/, and the hygiene gate goes from
clean to "Code style issues found in 939 files". CI never sees it, because a
fresh checkout has no build output; every contributor sees it on their second
command.
Mirrors the three nested files rather than inventing a list: dist, coverage,
playwright-report, test-results, .vite, src-tauri/target and src-tauri/gen from
Client/.gitignore, plus .serena/ and .superpowers/sdd/. node_modules needs no
entry -- Prettier ignores it by default.
Verified: `npx prettier --check .` reports "All matched files use Prettier code
style" with a fully populated Client/src-tauri/target/ present on disk, and
still names README.md when a misformatted table is appended to it.
Refs RL-19 / L-13.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* chore(ci): shellcheck, actionlint, and a repository hygiene job
The last two gates RL-19 asks for. Neither existed: the shell scripts were
never linted, the workflows were never syntax-checked, and .githooks/pre-commit
carried hand-written `# shellcheck disable=` directives that nothing had ever
read.
New `hygiene` job, ubuntu-only and root-scoped, modelled on docs-consistency
for the same reason: every gate in it is platform-independent text analysis,
and .gitattributes pins eol=lf so a second OS would only re-prove line endings.
It runs `npm run check:hygiene` -- the same entry point a contributor runs, not
a parallel copy of the commands.
shellcheck ships in the runner image. actionlint does not, so it is pinned by
version and verified by sha256: an installer script piped from a branch would
be the one unverified download in a workflow file that pins every action by
commit SHA.
Prettier's step moves here from client-check, where it no longer belongs.
Both linters found real defects.
shellcheck, 3 findings in 8 scripts. Two are SC1125 errors in
.githooks/pre-commit: `# shellcheck disable=SC2086 - repo paths contain no
spaces` is not a valid directive. Trailing prose makes shellcheck discard the
rest of the line, so neither suppression was ever in effect -- and one of the
two was written earlier in this same branch, which is a fair demonstration of
why the gate is worth having. The prose moves to its own line above. The third
is SC2015 in start-server.sh, rewritten as an explicit if.
actionlint, 5 findings, all inside `run:` blocks it shellchecks once shellcheck
is on PATH. Three SC2015 in load-baseline.yml, rewritten as explicit ifs. Two
SC2035 in release.yml, where `sha256sum *` should not become `sha256sum ./*`:
the comment four lines above records that ParseChecksumFile exact-matches the
last field, so a "./" prefix would strand every deployed server exactly as a
"windows/" prefix would. `sha256sum -- *` satisfies the linter and leaves the
output bytes identical.
Verified all three gates in both directions with shellcheck 0.10.0 and
actionlint 1.7.7 on PATH. Passing: `node scripts/run.mjs check:hygiene` prints
"check:hygiene: passed" with all three steps run, not skipped. Failing:
appending `bait_fn() { cat $1; }` to Server/scripts/voice-test.sh fails on
SC2086; changing a runs-on to `ubunt-latest` fails on runner-label; appending a
misformatted table to README.md fails prettier. All three files restored and
confirmed clean afterwards. actionlint validates the new job in ci.yml itself.
Refs RL-19 / L-13, S-05.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* docs(plans): record B1 progress through B1-3
The header still read "B1-0 is complete; B1-1 is the next step" three merged
phases later. A plan that misstates where it is costs a reader the same
confusion whether it is stale by one phase or three.
B1-0 (#1410), B1-1 (#1411), B1-2 (#1412) and B1-3 (this branch) are done; B1-4,
dependency automation, is next.
Verified: `node scripts/check-doc-counts.mjs` still agrees on 27 claims across
9 watched documents -- this file is one of them -- and prettier reports it
clean.
* chore(ci): pin Repository Hygiene as a required check on dev
The second half of S-05. Its acceptance criterion is "tree is formatted AND a
fast required gate fails future drift" -- a check that runs but is not pinned
lets a formatting regression merge, so the gate is not a gate until this lands.
The name was read off PR #1414 with `gh pr checks` after the job reported
`pass` in 26s, not copied out of ci.yml. That order matters: the B0 script
records that three pinned names exist in no workflow file at all, and that a
required check which never reports blocks every PR forever.
Extends the existing script rather than adding a second one, per the B1 plan.
Also records, in the "deliberately NOT pinned" list, that Docs & Ledger
Consistency reports and passes on a dev PR yet is unpinned. That reads as an
oversight from the 2026-08-25 pass rather than a decision, but it belongs to
G-04, so it is documented here and not changed.
NOT APPLIED YET. Running this script now would pin a check that PR #1413 cannot
report -- its branch predates the hygiene job, so the job does not exist in its
workflow file and the check would never arrive. Run it after #1414 merges;
#1413 needs a rebase onto dev regardless.
Verified: shellcheck clean, the embedded JSON parses, and `check:hygiene`
passes with prettier, shellcheck and actionlint all running.
Refs S-05, RL-14 / G-03.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(client): style the user profile popup
The popup rendered unstyled: it appeared at the bottom of the page and
pushed the rest of the app up, with the avatar drawn as a full-width bar.
app.css carried a complete Discord-shaped card under `.user-popup` /
`.up-*`, but nothing in the codebase renders those classes — the
component emits `.upp-*`. The component had been rewritten with a new
prefix and the stylesheet was left pointing at a DOM that no longer
existed. With no rule matching, the card stayed `position: static`, so
the left/top it computes were discarded and both it and its overlay laid
out as ordinary blocks at the end of <body>.
Replace the orphaned block with rules for the classes actually rendered,
following the same anatomy: banner strip, avatar straddling the
banner/body seam inside a ring punched from the card background, panel
sections, action row. Everything routes through existing tokens, so the
card follows the theme contract.
Two latent bugs fixed while there:
- Placement guessed a 300px card height and clamped only the top edge,
so a member clicked low in the list opened a card that ran off the
bottom of the window. Measure the card and clamp both edges.
- The avatar has to hang off the body's top edge, but the body scrolls,
and `overflow-y: auto` clips horizontally too. Make it a child of the
card rather than the body.
The fade+scale moves from inline styles into CSS so a
`prefers-reduced-motion` override can drop it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(client): stop vite watching src-tauri
`npm run tauri dev` died on Windows partway through the cargo build:
Error: EBUSY: resource busy or locked, watch
'src-tauri\target\debug\deps\owncord_client_lib.dll'
Error The "beforeDevCommand" terminated with a non-zero status code.
Vite's watcher recursed into `src-tauri/target/`, and the moment cargo
wrote the output DLL, node's FSWatcher raised EBUSY as an unhandled
error event and killed the vite process. Vite is tauri's
`beforeDevCommand`, so its death aborted the whole dev session.
The config matched the upstream Tauri vite template in every respect
except the `server.watch.ignored` block that template ships with. Add
it. Tauri already watches `src-tauri` itself for rebuilds, so nothing
is lost.
Windows-specific — EBUSY on an open handle is a Windows filesystem
semantic, and CI only ever runs `tauri build`, never `tauri dev`, so
neither caught it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* docs(security): add 2026-08-04 whole-codebase security review (#1326)
Read-only security review of the full tree (Go server, admin panel, WASM
plugin host, LiveKit voice, Tauri client). No code changes.
Three findings, all the same defect class — a security predicate enforced
at some members of a handler family but not all:
- A-2026-08-01 (HIGH) handleDeleteChannelPermission omits the hierarchy and
grantability guards its PUT twin carries, so a MANAGE_CHANNELS holder can
clear their own role's channel deny and read private channels.
- A-2026-08-02 (HIGH) the admin channel list/patch/delete handlers omit the
type == "dm" guard their sibling getPermChannel carries, so the same role
can enumerate and irreversibly cascade-delete arbitrary DMs and group DMs.
- A-2026-08-03 (MEDIUM) DMService.RingTargets omits the block check the five
other DM interaction sinks perform, so a blocked user can ring the person
who blocked them.
Also records one non-vulnerability observation (backup restore writes to a
hardcoded database path, silently no-opping when database.path is
customised), the candidates rejected during verification, the areas verified
clean, and the areas not examined.
Claude-Session: https://claude.ai/code/session_01Q7GUJtdsHHHGs4pSiLn6LJ
Co-authored-by: Claude <noreply@anthropic.com>
* Full audit: docs/spec refresh + remediation (security fixes, dead-code removal, test & CI gaps) (#1327)
* docs: fix server reference docs (api, protocol, server-configuration, deployment)
api.md:
- Correct the login rate limit: 5/min per IP (was documented as 60/min);
document the per-username lockout and lockout persistence
(Server/api/constants.go, Server/api/auth_handler.go)
- Complete the middleware list to the real 9-entry chain incl. the
opt-in Coraza WAF (Server/api/router.go)
- Add voice_sessions and broadcast_drops to the metrics sample
(Server/api/metrics_handler.go) and document the otel-only
Prometheus /metrics mount
- Add reference sections for the previously undocumented /admin/api
endpoints: setup, stats, users, audit-log, settings, tokens,
backups, updates, and the SSE log stream (Server/admin/api.go)
protocol.md:
- Fix type counts (client->server 26, server->client 37) and add the
missing rows: call_ring, call_decline, emoji_update, call_incoming,
call_declined
- Correct rate limits: voice join/leave 5/1s (was "None"), E2EE offer
64/1s (was 5/1s), and add the call-ring limit (1/3s)
- Document the plugin command wire types (chat_command, command_reply,
plugin_broadcast) and flag that they sit outside protocol-schema.json
server-configuration.md:
- Add missing keys: server.waf_* (3), database.type,
telemetry.otlp_insecure, and the whole logging section +
OWNCORD_LOGGING_LEVEL
- Correct plugin-disabled status code to 503 (was 501)
deployment.md:
- Drop the removed "version" field from the /health sample; add
broadcast_drops to the metrics sample; note the distroless non-root
image; refresh build version strings
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(architecture): rewrite stale architecture pages against 5630aa1
All six pages carried "Verified against ddc49f0 (2026-07-19)" stamps and had
drifted:
- websocket.md: delete the false claim that docs/protocol-schema.json does
not exist — it is the codegen source of truth (Server/scripts/genprotocol,
CI-gated by make protocol-verify); note the hand-declared plugin command
family as the one exception; refresh LOC
- server.md: fix the websocket dependency (github.com/coder/websocket, not
nhooyr.io), refresh LOC (42k/71k), migrations 016 -> 028
- data-model.md: migrations 001-028, 23 -> 26 tables, add api_tokens and
channel_user_overrides to the ER diagram, channels.type now includes
announcement, note 017/024/027/028 columns; drop the claim that schema.md
is 6 migrations behind (it is current)
- voice-e2ee.md: drop the stale claim that the E2EE flow is absent from
protocol.md (it has a full section); document livekitE2EE.ts/identity.ts
and identity-key pinning
- client.md: rewrite — Solid beachhead is gone; the HTTP path is now
TOFU-pinned through http_proxy.rs (the doc claimed the opposite); shared
tofu.rs core with explicit-consent pinning; 9 stores (roles store deleted,
blocks + emoji added); refreshed LOC and tooling figures
- README.md: 26 tables/001-028; client-architecture.md described as the
redirect stub it is; companion-audit links refreshed
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(architecture/ux): align UI flow specs with current client behavior
- Cert first-use is a blocking trust modal on status "first_use" — the Rust
proxy rejects the first connection until the user confirms (main.ts:146-176,
tofu.rs); the specs described the pre-F4/F8 behavior (an 8s banner on
"trusted_first_use"). Fixed in README.md and connection-and-auth.md.
- Dispatcher event table: add the five missing types (chat_bulk_deleted,
roles_update, emoji_update, voice_moved, voice_disconnected) and note that
call_incoming/call_declined are page-scoped listeners in MainPage.ts.
- channels-members-dms.md: the "no in-client block button" gap is closed
(AdminActions.ts context-menu item -> SidebarMemberSection.ts:177-186);
document group DMs (MemberPickerModal, 10-participant cap, rename/leave)
and per-channel notification mutes (lib/channel-mutes.ts); refresh stale
line anchors.
- voice-and-e2ee.md: document the actual E2EE verification surface (roster
shield badge -> identity-mismatch modal -> rePinPeerIdentity with TOCTOU-
safe key capture), noise suppression + fallback, device hot-swap, stream
preview, and DM ring/incoming-call flow; drop the nonexistent
VoiceChannel.ts reference.
- settings-and-admin.md: the "ban should collect a reason" gap is closed
(appendBanFlow with reason + duration); document the admin-panel deep-link
(lib/admin-panel.ts) and the tray status menu.
- messaging.md: correct the pinned-messages empty-state copy and drop the
nonexistent components/message-input/ directory reference.
- Re-stamp all six specs "Verified against 5630aa1 (2026-08-04)".
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs: refresh client-facing and top-level docs
- security.md: drop the stale "hardcoded Tenor API key" limitation (the GIF
provider is Klipy, proxied server-side with an operator-supplied key —
nothing ships in the client bundle); describe credential storage accurately
(OS keyring primary, verified writes, DPAPI/ChaCha20 file fallback);
complete the audit-log action list against the actual WriteAudit call sites
and note that backup restore is not audit-logged; fix the firewall
checklist to include the LiveKit media ports (7880-7881/TCP,
50000-60000/UDP) and ACME port 80
- credential-storage.md: probe_credential_store sample now shows the real
serialized backend value ("Keyring") and the full variant union
- quick-start.md + README.md: refresh build version strings to
1.2.0-alpha.1; README "audits" section now points at the current audit
documents
- contributing.md: sqlc rows no longer claim a PostgreSQL engine/pgdbgen
(removed with the store layer); add the protocol-generate/verify targets
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(plans): add verified status headers and fix stale references
Every plan under docs/plans/ now carries a dated status verified against
5630aa1:
- Shipped: audit-2026-07-19-decisions (all 13 rows), channel-visibility-
unification, http-tofu-proxy, permission-middleware-consolidation (the
disclosed ws.channelCanSend copy is still open, now at serve_ready.go:119),
security-hardening-remediation, sqlc-adoption, v2-dispatch-migration,
tauri-capability-narrowing (DNS-rebinding follow-up still open)
- Shipped with corrections: discord-parity — Phase 1's gap table was never
re-marked; all six rows have since shipped, including archived channels,
which are filtered by permissions.VisibleChannelIDs (checker.go:116-121);
named leftovers (role hoist/mentionable, @RoleName mentions, categories as
entities, dead-code list) stay open. security-scan-2026-07-22 — all 8
findings closed; two of the four F3 follow-ups have since shipped (safety
number rendered in the roster badge; rePinPeerIdentity wired to the
identity-mismatch modal), getIdentityPin fail-open remains open; noted the
scan artifact directory is not in the repo
- Design-only: slash-commands — added staleness notes (migration number 016
now taken, Server/store/ deleted, src/state/ never existed)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(audits): reconcile prior audit statuses with current CI and code
- audit-test-coverage-2026-07-25: T-2026-07-25-21 (HIGH, 229/255 web e2e
failing) was fixed by the mock repair but the audit was never updated —
now RESOLVED, re-verified by a local 270/270 run at 5630aa1; the CI gate
table row updated to match
- audit-2026-07-19: carried-over item 11 ("no Playwright job in ci.yml") is
resolved — client-e2e (non-blocking, every PR) and the blocking
client-e2e-parity job both exist; backlog item 10 marked DONE
(client-tests is blocking, Playwright wired)
Only status/closure cells were edited; original finding text is untouched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(changelog): add Unreleased section for post-v1.2.0-alpha.1 fixes
Three fixes landed after the release with no changelog home (the file had no
Unreleased section at all): the profile-popup styling fix (a308f81), the
vite/src-tauri watch fix (cdcfc03), and the AppImage env-key signing fix
(9d75890). Also corrects the Deferred-work note that still described the
Solid.js removal in the present progressive — it completed 2026-07-19.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(tests): rewrite the e2e issues log against a real suite run
The old file was dated 2026-03-18, claimed 209/209 passing (the suite is now
270 tests), and pointed at a plan document that does not exist in the repo.
Rewritten from an actual run at 5630aa1: 270/270 web tests green (8.6 min),
15/15 @parity subset green (the blocking CI job), with the suite inventory,
CI wiring, the two real open issues (three native specs matched by no
playwright.config.native.ts project; client-e2e still non-blocking), and
dispositions for every claim the old file carried.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(audit): add 2026-08-04 docs-and-coverage audit report
Companion to the same-day security review (disjoint scope). Contains: the
verified architecture summary; real test-run results for every runnable
suite at 5630aa1 (Go race+deadlock, 4394 client unit tests at 94.66% stmt
coverage, 83 Rust tests + clippy, 270/270 web e2e, 15/15 parity, browser
smoke — with env-blocked suites named and their compensating CI evidence
cited); a 52-row UI/UX flow coverage matrix (30 covered / 21 partial /
1 untested / 0 broken, headline gaps: TOFU flow, E2EE verification, admin
panel, updater — all unit-only); per-doc drift findings with the commit that
fixed each; reconciliation of all four prior audits and eleven plans
(including the orphaned 2026-04-07 #8 resurfaced as DC-11); the dead-code
and TODO inventories; and a prioritized DC-01..DC-15 gap list with ordered
next steps.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* fix(admin): add hierarchy guard to channel role-override delete (A-2026-08-01)
Deleting an override is a permission mutation: removing a deny row restores
exactly the access the PUT path refuses to grant, so a MANAGE_CHANNELS holder
could unlock a private channel their own role was locked out of. Gate DELETE
identically to handlePutChannelPermission: resolve the role (404 when
missing), fail closed without an actor role, and refuse targets at or above
the actor's position.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* fix(admin): exclude DM channels from the admin channel surface (A-2026-08-02)
DMs and group DMs share the channels table and id space with guild channels,
but they belong to their participants, not to MANAGE_CHANNELS holders:
listing exposed ids and group names of every private conversation, PATCH
could silently rename one, and DELETE cascade-destroyed one irreversibly.
List now filters type=dm; PATCH and DELETE resolve through getAdminChannel,
which answers 404 for DM ids so the surface does not confirm which ids are
private conversations.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* fix(service): enforce blocks on DM call rings (A-2026-08-03)
RingTargets checked participation but not blocks, so a blocked user could
still make the blocker's client ring. Route rings through
requireDMNotBlocked like every other DM sink; group DMs stay exempt inside
it, matching the send path (blocks are enforced at group creation instead).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* chore(client): delete dead modules (DC audit remediation)
All verified unreferenced by any import before deletion:
- ServerStrip.ts: removed from the layout when the unified sidebar header
landed (SidebarArea.ts); only its own orphaned unit test still used it.
The e2e spec already asserted .unified-sidebar-header, so it is renamed
to sidebar-header.spec.ts and retitled honestly.
- FileUpload.ts: uploads go through api.uploadFile from MessageInput.
- lib/reconcile.ts: nothing imports it; the messages store carries its own
pending-send reconciliation.
- public/rnnoise-worklet.ts: unreferenced duplicate of the .js worklet the
runtime actually loads, and public/ ships verbatim into the bundle.
- api.getSounds/deleteSound + SoundResponse: the server has no /sounds
routes; these called endpoints that do not exist (pairs with the
sounds-table drop on the server side).
- dm.store incrementDmMention: zero callers; DM mention counts flow from
the server mention_count via the dispatcher. This was the one live knip
error the CI '|| true' was masking.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* chore(client): retire the tauri-typegen ritual (DC-05)
src/generated/ was tauri-typegen output frozen on 2026-04-03: it covered 21
of the 29 IPC commands lib.rs registers, nothing ever imported it (0%
coverage), and CI carried a bespoke patch step solely to keep the unused
file lint-clean. Delete the directory and every part of the pipeline that
existed to feed it: the client-check patch step, the tauri-build
generate/patch steps, the tauri.conf.json plugin block, and the inert
Cargo.toml build-dependency (build.rs is bare tauri_build::build(); no Rust
source references the crate). Cargo.lock shrinks by exactly the typegen
subtree — no other resolution changes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* ci: make knip blocking (DC-06 follow-through)
Pre-verified green locally after the dead-module deletions; the config
hints knip still prints do not affect its exit code.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* chore(server): drop the dead sounds table (A-2026-07-13)
The table shipped in 001 for a soundboard that was never built: no query,
model, sqlc definition, route or handler ever referenced it. Migration 029
drops it; the sqlc model regenerates without the Sound struct (sqlc emits a
struct per schema table even with zero queries). schema.md, the data-model
blueprint, and the 2026-07-19 audit closure table are updated in the same
change per the docs maintenance rule.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* chore(server): remove dead WAF wrapper, use protocol constants, fix stale comments
- NewWAFMiddleware had no production caller (the router mounts the CRS
variant); its doc text folds into NewWAFMiddlewareCRS and the tests call
the survivor directly.
- serve_auth compares against MsgTypeAuth and the DM-close REST path builds
its WS notification from MsgTypeDMChannelClose instead of restating the
wire strings, so the generated constants are load-bearing again.
- Comment fixes: DatabaseConfig no longer claims Postgres scaffolding that
main.go removed; host_ui.go no longer advertises a route that is not
mounted (DC-09's sibling); buildReady cites docs/protocol.md, the file
PROTOCOL.md was renamed to (DC-09).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* fix(protocol): add the plugin command family to protocol-schema.json (DC-01)
chat_command, command_reply and plugin_broadcast were the only wire types
outside the schema: the first declared by hand in handlers_command.go, the
other two raw string literals, all bypassing the protocol-verify codegen
gate. Add the three schema entries (27 c2s / 39 s2c), regenerate both
constant files, and swap the hand-rolled declarations for the generated
constants. The ws protocol-contract test's exception list is empty now —
and stays that way.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* test(client): wire orphaned native specs and typecheck the Playwright layer
- dm-system, reconnection and theme-persistence (14 tests) matched no
project's testMatch in playwright.config.native.ts, so they had never
executed (E2E-ISSUES open issue #1 / DC-03). All three use the persistent
fixture + ensureLoggedIn, so they join native-authenticated.
- tests/e2e was excluded from tsconfig, leaving 47 spec files with no
typechecking anywhere. New tsconfig.e2e.json project (+@types/node for
the node-API fixtures), a typecheck:e2e script, and a CI step. The one
real error it surfaced is fixed: mockTotpFailure omitted the required
simulateWsFlow flag.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* test(client): cover createPromptModal and the external-abort close path
modalFactory.ts was the least-covered file in the repo (57.6%):
createPromptModal had no tests at all and createModal's external-abort
branch never ran with an onClose. Now 100% statements/branches/functions,
including the trimmed-submit, legitimate-empty-submit, Enter-preventDefault
and no-double-close contracts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* test(client): e2e-cover the TOFU certificate ceremony (DC-04 slice)
The first-use confirmation and mismatch warning are the client's core
security ceremony and had no e2e coverage. Six tests drive them through
the mocked Tauri event layer: first-use modal content, trust, cancel,
modal non-stacking, mismatch fingerprint rows, and disconnect-to-connect-
page. The mock now exposes its listener registry so tests can wait for
the async cert-tofu registration instead of racing it (validated with
--repeat-each=3).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs: fix the inaccuracies the 2026-08-04 refresh missed
- contributing.md: drop the '-tags postgres' build row (no such tag exists
anywhere in Server/), add the four Make targets the 07-25 audit created
(test/test-deadlock/cover/cover-all), align the coverage statement with
the real gates (client 70%, no Go floor by T-2026-07-25-19), point TS
style at architecture/client.md instead of the tombstone, and describe
the real dev-branch PR flow.
- docs/security.md: reporting section now defers to root SECURITY.md as
the canonical policy (it said 48h where SECURITY.md promises 7 days, and
described the maintainer's advisory path rather than the reporter's);
fixed the updater-key link that resolved to docs/Server/... on GitHub.
- audit-2026-04-07.md closure table: #10 and #11 were long-resolved (#10
verified in db/audit.go, #11 exceeded by per-PR e2e jobs), #6 written in
future tense for work done 2026-07-19, #7 citing a 113-file count from
months ago.
- README: Contributing section matched neither ci.yml nor contributing.md
(branch from dev, not main); Docs Index gains the six missing live docs;
the plugin system joins the feature list; the security row no longer
anchors to an aging version string.
- server-configuration.md: the env-var table is explicitly a subset — the
OWNCORD_<SECTION>_<KEY> scheme covers every key.
- mcp-introspect.md: index.mjs is 266 lines, not ~230.
- types.ts header cited PROTOCOL.md/API.md/SCHEMA.md, filenames that no
longer exist.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* ci: pin claude.yml actions by SHA; add docs checkbox to the PR template
claude.yml was the only workflow with unpinned third-party actions —
checkout now uses the same v4.2.2 SHA the other workflows pin, and
claude-code-action pins the commit the v1 tag resolves to (Dependabot's
github-actions ecosystem keeps both fresh).
The PR template gains the docs checkbox A-2026-07-03 recommended: the
architecture/UX maintenance rule ('a PR changing a diagram's source-of-
truth files updates the diagram in the same PR') existed only as prose no
process step ever surfaced.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* style(client): prettier-format the cert-tofu spec
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(audit): record the remediation pass and close finding statuses
- Security review: A-2026-08-01/02/03 -> RESOLVED with their pinning tests
named.
- Docs-and-coverage audit: DC statuses updated in place (01/02/03/05
resolved, 04/09 partial with the remainder named, 14's keep-decision
recorded) and a remediation addendum added: what shipped, the decisions
taken (plugin host API kept, reserved protocol entries kept, e2e soak not
shortcut, the 404-on-missing-role semantics note), and the full
verification table from real runs — Go race + deadlock suites green,
4 tag builds, client 4360/4360 units at 95.35% coverage, Playwright
276/276 in 8.9 min, parity 15/15.
- CHANGELOG Unreleased: security fixes, migration 029, protocol additions,
dead-code retirement, CI gates.
- E2E-ISSUES: rewritten against the remediation HEAD (276/276), native
orphan issue moved to resolved, mock listener-registry note.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* fix(admin): index the channel slice in the DM filter (gocritic rangeValCopy)
golangci-lint (CI-only gate) flags the range-value copy of the 152-byte
db.Channel struct in the admin list filter.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* chore(client): add .nvmrc pinning Node 20 to match CI (DC-10)
Also re-triggers CI: the previous run's windows server job died to a Go
runtime unwinder fatal ('traceback did not unwind completely') with no
test failure — toolchain flake, and the integration lacks permission to
rerun failed jobs.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
---------
Co-authored-by: Claude <noreply@anthropic.com>
* fix: 26 bug-hunt findings across WS hub, voice/E2EE, admin, roles, and client (runs 1-3) (#1328)
* fix(ws): keep pubsub subscriptions when a replaced client is unsubscribed
Both pubsub indexes are keyed by userID, but a reconnect registers a new
*Client under that same userID. UnsubscribeAll and Unsubscribe deleted by
userID alone, so a kick of the already-replaced connection stripped the live
one's topics. The live client stays in h.clients and keeps answering
ping/pong, so it never reconnects -- it just silently stops receiving every
global, user, and channel broadcast.
Guard the forward-index delete in unsubscribeLocked with an identity check and
route UnsubscribeAll through it, so the four Unsubscribe call sites
(voice_leave, hub_broadcast x2, handlers) and the three UnsubscribeAll ones
(kickClient, unregisterNow, registerNow) all share one rule.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(ws): mark kicked clients offline instead of reporting them replaced
Every kick path deletes the hub entry via kickClient, so the readPump defer's
unregisterNow finds nothing and fell through to "return true", conflating
absent with replaced. serve_pumps.go then skipped MarkUserDisconnected, the
offline presence broadcast, and handleVoiceLeave -- already-connected peers
rendered every kicked user as online until that user reconnected and
disconnected cleanly.
Return exists instead: a different client in the slot is a genuine
replacement, an absent entry is a real disconnect. Only serve_pumps.go reads
the return value; the five serve.go/hub.go call sites discard it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(ws): force a full ready when cold-tier replay hits the row cap
GetEventsSinceForChannels is "ORDER BY seq ASC LIMIT n", so a reconnect gap
larger than maxColdReplay returned the oldest 5000 rows and dropped the
newest. handleReconnect accepted any non-empty result as a successful resume,
and the client only tracks max(seq) with no gap detection -- so it accepted
the next live event and silently lost the range in between, including state
events (channel/role/member changes) that REST history fetches never repair.
Treat a result at the cap as overflow and fall through to the full ready
re-sync. An exactly-cap-remaining gap pays one unnecessary full ready.
maxColdReplay is hoisted to the package const block so the test can seed
exactly enough events to hit it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(ws): re-elect the voice E2EE key holder on the two paths that skipped it
updateKeyHolder had only two callers (voice_join, voice_leave), so two paths
that remove a participant from voice left voiceKeyHolders naming someone who
is gone. IsVoiceKeyHolder then rejects the real lowest-uid participant's rekey
offers with NOT_KEY_HOLDER -- which the client does not handle -- after it has
already applied its rotated key locally, splitting keys across the room.
1. The LiveKit participant_left webhook (media-only loss, WS stays up) cleared
voice state and broadcast voice_leave with no re-election.
2. registerNow's fresh-connect replacement (F5 reload) drops the old
connection's voice state without transferring it. handleVoiceLeave never
runs there: readPump skips it when replaced, and it early-returns on
already-cleared state.
Both call updateKeyHolder outside h.mu, since it takes keyHolderMu then
h.mu.RLock. The recompute reads live client voice state, so it is idempotent
and stays correct when a network reconnect transfers voice state -- locked by
TestRegisterNow_KeepsKeyHolderWhenVoiceStateTransfers.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(client): stand down as E2EE key holder on offer, keep peer keys on reconnect
Two independent key-holder desync bugs in E2EEManager:
1. _isKeyHolder had no demotion path -- set at join, promoted on participant
leave, cleared only on voice leave. The server re-elects the lowest userID
on every join, so a lower-ID joiner left the incumbent still believing it
held the key with an armed 5-minute timer. Its rotations applied the new
key locally before the server rejected the offers with NOT_KEY_HOLDER (which
the client does not handle), so it went deaf and mute every rotation cycle.
Accepting an offer proves the sender is the server-authoritative holder, so
treat it as the demotion signal and clear the timer.
2. reannounceForReconnect cleared _peerPublicKeys and peer verifications with
nothing able to refill them: handleAnnounce replies with an offer rather
than a counter-announce, and the server relays stored peer keys only on
voice_join, which an SFU-level reconnect never runs. handleOffer's
unknown-peer guard then dropped every later rotation, stranding the
reconnector on the pre-reconnect key. The clear was also unnecessary --
peers' keys stay valid when we regenerate our own pair.
vitest 4396/4396; typecheck and prettier clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(tauri): give the ws sender slot connection ownership, bound livekit TLS
ws_proxy: state.tx is one slot shared by every ws_connect, and both the
install and the teardown clear were unconditional while the mutex was only
held in short scoped blocks. A handshake pends up to CONNECT_TIMEOUT and a
profile switch starts a second ws_connect without awaiting or cancelling the
first, so a stale connect could complete after a newer one was live, emit an
untagged "open", and install its sender over the live one -- routing the next
auth send to the previously-trusted host, then tearing down the live socket
and emitting "closed" while JS believed it was connected.
Add a generation counter claimed at ws_connect entry and checked under the
slot lock before install, plus same_channel ownership on the teardown clear,
mirroring the Arc::ptr_eq guard ptt.rs already uses for ATOMICRACE-001.
livekit_proxy: the outbound TcpStream::connect and TLS handshake were bare
awaits, while the sibling http_proxy.rs bounds both at 10s. TCP connect is
OS-bounded, but a peer that accepts TCP and never answers the ClientHello
blocked the task forever. The task holds `local` without polling it, so the
SDK closing its side never cancels it, and the detached per-connection tasks
survive stop_livekit_proxy -- so they leaked on every SDK retry.
cargo test 80 passed; clippy --all-targets -D warnings clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(ws): keep voice E2EE alive across WS resume and unify voice teardown
A network reconnect transferred voiceChID/joinToken to the new connection
but left it unsubscribed from voice:<id> (the only transport for
voice_e2ee_announce relays) and wiped the announced ECDH key, so a
resumed key holder could never offer the room key to later joiners and
voice_join replayed nothing for the resumed user. registerNow now
transfers the announced key with the voice state and re-subscribes
VoiceTopic unconditionally (it is CONNECT_VOICE-gated at join; only the
message-stream ChannelTopic needs the READ gate).
The LiveKit participant_left webhook and CleanupVoiceForChannel cleared
voice state without dropping the voice-topic subscription, leaving the
socket receiving another room's announces (which carry no channel_id to
filter on) for its lifetime. All take-out-of-voice paths now go through
one clearVoiceAndUnsubscribe helper.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(ws): route sequenced DMs through the normal FIFO
writePump drains sendHigh to exhaustion before send, so a seq-stamped DM
on the high queue reached the socket ahead of lower-seq events still
queued behind a slow write. The client acks max(seq) and replay is
strictly seq > last_seq, so a disconnect in that window silently and
permanently lost the overtaken events while auth_ok reported a clean
resume. Sequenced frames now share the one per-client FIFO; the high
queue remains for unsequenced targeted messages (DM opens, voice tokens).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(storage): remove the partial file when Save fails after create
The io.Copy and f.Sync error paths returned without deleting the file
created for the upload, and the orphan sweep is DB-row-driven, so a
write-side failure (ENOSPC, disk I/O error) permanently leaked a partial
storage/<uuid> with no DB row. One success-flag deferred cleanup now
covers every failure path (the oversize branch folds into it), fixing
all three callers.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(client): close three voice-E2EE ordering gaps and wire the DM mention badge
- setupKeyExchange generated the room key AFTER draining queued
announces, so a key holder joining an ongoing call sent drained peers
no offer — they waited on the 5-minute rotation timer. Keygen now
precedes the drain and every drained peer gets its offer immediately.
- A key-holder re-election arriving while the elected client was still
connecting was dropped (getCurrentChannelId is null for the whole
key-exchange wait), stranding the client until timeout ejection. The
manager now remembers its channel from setupKeyExchange, and the
become-holder rotation resolves a pending room-key wait.
- Offers applied concurrently could finish out of order (no epoch on the
receiver side), leaving the older key active. handleOffer now chains
applications so offers apply strictly in WS delivery order.
- incrementDmMention had zero callers: the DM @mention badge (dmStore's
mentionCount, the mute-immune signal DmSidebar renders) never fired
live, only after a reconnect restored the server count. The dispatcher
now bumps it under the same guards as the DM unread count.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(tauri): restart the LiveKit proxy when the TOFU pin changes
After the user accepted a rotated cert, two stale caches kept every
voice rejoin tunneling into the old pin until logout: the Rust reuse
branch returned the running listener (which bakes its fingerprint in at
spawn) without re-reading certs.json, and ensureLiveKitProxy's port
cache never invoked Rust again at all. start_livekit_proxy now loads the
stored fingerprint before the reuse check and tears down on host OR pin
change, and the TS side invokes it on every join — the reuse branch
dedups the unchanged case.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore(tauri): ignore two unreachable cargo-audit advisories
RUSTSEC-2024-0429 (glib 0.18, Linux-only, Variant::array_iter_str never
called; no semver-compatible fix exists) and RUSTSEC-2026-0097 (rand 0.7
as a phf_generator build-dep with a fixed seed and no log feature; the
pre-release kuchikiki pin blocks the upgrade path). Both entries document
their drop condition inline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(ws): transfer the focused channel on WS resume so the message stream survives
registerNow's replaced-client branch moved voice state and the E2EE key to
the resumed connection but not the focused channel; newClient always starts
with channelID == 0 and the client never re-sends channel_focus on a resume,
so the ChannelTopic re-subscribe was a no-op and the user silently stopped
receiving chat_message until manually switching channels. Transfer the old
connection's focused channel, READ-gated and fail-closed like every other
ChannelTopic subscription.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(ws): always include a voice room's participants in its voice-event audience
broadcastVoiceEvent filtered recipients on READ_MESSAGES while voice
membership is gated on CONNECT_VOICE alone, so a participant in the gap
(e.g. READ revoked mid-call by a channel override) never received the
room's voice_state/voice_leave. The client's E2EE key-holder election and
forward-secrecy rotation run only off the voice_leave WS event, so a
departing key holder was never replaced and new joiners hung until the
e2ee_timeout eject. Union the READ audience with the room's current
participants; what outsiders may observe is unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(client): tear down the failed reconnect attempt's room instead of leaking it
The catch block read this._room, whose typed accessor returns null in the
"reconnecting" state — so the failed attempt's freshly created Room was
never disconnected and kept all its listeners. livekit-client emits
Disconnected synchronously on a failed connect, and in "reconnecting"
state the token/channel/url getters all return values, so each leaked room
spawned an additional concurrent reconnect loop whose AbortController was
discarded and unreachable from leaveVoice. Alias the attempt's room outside
the try and clean it up in the catch, mirroring cleanupAbortedReconnect.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(admin): evict voice participants before deleting a channel
CleanupVoiceForChannel was doc-commented 'Called when a channel is
deleted' but had zero production callers, and the voice_states FK cascade
wipes the rows it reads — so deleting a voice channel stranded its
participants with live client voice state, a voice-topic subscription, and
a LiveKit session, and the stale sweeper could never recover them (a
nonexistent channel resolves base-role permission bits). Wire the cleanup
into handleDeleteChannel BEFORE the row delete, via HubBroadcaster.
Also harden the cleanup itself: the row delete and client-state clear are
now conditional on the participant still being in the deleted channel, so
a user who moved rooms mid-cleanup is untouched, and the evicted
participants are always included in their own voice_leave audience (their
client state is already cleared, so the participant union in
broadcastVoiceEvent cannot see them).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(client): stop leaking an E2EE worker and SetKey listener per voice join
createRoom spun up a fresh E2EE Worker per Room while the key provider
lives for the whole process; livekit's per-room E2EEManager registers a
SetKey listener on the provider with no matching removal and never
terminates the worker. Every join, channel switch, or failed reconnect
attempt therefore permanently added one running worker plus one listener,
and every later setKey posted the new room key into every orphaned worker
— key material outliving its session. Track the worker on the session:
clear provider listeners and terminate the stale worker before each Room,
and terminate it in leaveVoice so the last key does not stay resident.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(service): reject role position collisions on update, matching create
UpdateRole's position branch ran only validatePosition and let an explicit
position land on a slot another role holds — while CreateRole refuses
exactly that, with a comment explaining why: every hierarchy comparison
uses >=/<=, so tied positions read as equal rank and the two roles can no
longer manage each other's members. Refuse a position held by a different
role with the same ErrBadRequest; re-stating the role's own position stays
allowed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(ws): close the window where a dying connection re-takes a pubsub topic
Subscribe had no counterpart to unsubscribeLocked's identity guard: an old
connection's in-flight handler (a channel_focus mid DB round-trip shares no
lock with registerNow) could Subscribe after UnsubscribeAll(old) had run,
stealing the topic from its replacement — whose own unsubscribes then skip
the entry while publishes go to the closed connection. Subscribe now
refuses a client whose send is closed (checked under ps.mu), and
registerNow closes the old client's send BEFORE stripping it, so a late
Subscribe either sees the closed send and is refused or slipped in earlier
and is removed by the subsequent UnsubscribeAll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* refactor(ws): rewrite cold-replay if-else chain as switch (gocritic)
Fixes the ifElseChain lint failure on CI for both platforms.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(client): restore incrementDmMention deleted as dead code on dev
The audit PR (#1327) removed it from dm.store.ts because its only caller
lives on this branch (the DM mention badge wiring), which was not merged
yet. The rebase was textually clean but left dispatcher.ts calling a
function that no longer existed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* Audit closure pass: DC-04/06/08/09/12/13/15 (E2EE + updater e2e, fail-closed pin lookup, a11y pass, UX polish) (#1329)
* ci(server): run the tag-gated wazero/otel Go tests (DC-06)
The build-tag matrix only compiled the otel/wazero variants; the tests
behind those tags (plugin/sandbox_wazero_test.go 462 lines,
telemetry/telemetry_otel_test.go 214 lines) ran nowhere since they were
written (T-2026-07-25-16). Scoped to the two packages that carry tagged
files; verified green locally before wiring:
go test -tags wazero ./plugin/... and -tags otel ./telemetry/... both pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* fix(client): fail closed when the identity pin store is unreadable (DC-08)
getIdentityPin collapsed a keyring read error into "no pin stored", so a
transient failure sent a pinned peer down the TOFU first-sight path:
verifyPeerAnnounce verified against the server-delivered key and then
RE-PINNED it — a fail-open a malicious server could exploit by inducing
store errors (F3 follow-up 3, plans/security-scan-2026-07-22).
getIdentityPin now returns a three-state IdentityPinLookup
(pinned/unpinned/unavailable), mirroring how tofu.rs keeps Err distinct
from Ok(None) first-use. verifyPeerAnnounce rejects the announce on
"unavailable" without any pin write, records the new "unknown"
PeerVerification status, and the roster badge renders it as an amber
shield-question ("could not check") distinct from the legacy
"unverified" state.
Pinned by unit tests: pin present, no pin, store error (identity.ts),
the fail-closed rejection path (livekit-session), and the badge
presentation (channel-sidebar).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* feat(admin): write a backup_restore audit row that survives the restore (DC-09)
Backup restore was the one admin mutation with no audit_log row —
docs/security.md documented the gap as inherent ("the database is closed
as part of the restore"). The row IS writable durably: written
synchronously (LogAudit, deliberately not the async WriteAudit fast path)
before BackupTo takes the pre-restore safety copy, it is captured inside
pre_restore_*.db and survives the file swap forensically.
The extended restore test opens the pre-restore backup as a database and
asserts the backup_restore row is inside it — proving both the write and
its ordering. docs/security.md now documents where the row lives instead
of the gap.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* feat(client): channel-delete toast + optimistic reaction toggle (DC-12)
Two messaging-surface gaps the UX specs carried as open:
- channel_delete on the active channel now toasts "This channel was
deleted" alongside the existing redirect (ux/channels-members-dms
§1.2) — the redirect alone read as the app spontaneously changing
channels. Non-active deletions stay silent.
- Reactions toggle optimistically (ux/messaging §5): the pill flips on
the click, registered under the send's WS envelope id — the same
correlation scheme as the optimistic message rows. updateReaction
consumes the matching self-echo instead of re-applying it (the
delta-based arithmetic would double-count), other users' echoes apply
normally, and an error reply or local transport failure rolls back
exactly that toggle via rollbackReaction in the dispatcher's error and
send-failure handlers. The pill reverting is the failure feedback.
Both spec gap notes flipped to implemented; the stale §2 note claiming
the ready payload lacks slow_mode fell in the same edit.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* fix(client): guard the role-change submenu against double-fire (DC-12)
Every other destructive admin action already carried an in-flight guard
(withConfirmation, unblockRunning, banRunning, purge) — the role-change
submenu was the residual: currentRole only updates when the
member_update echoes, so a double-click (or a second option clicked
while the first PATCH was in flight) fired onChangeRole twice. One
shared guard now inerts the whole submenu while a change is running,
with the pending class on the clicked option.
The settings-and-admin spec's in-flight gap note flips to implemented,
and the stale messaging §8 slow-mode note is corrected in the same
docs sweep (the countdown shipped with the ready payload's slow_mode —
verified against ChannelController.startSlowMode and its tests).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* fix(client): drag-reorder document listeners leak — own them per sidebar signal (DC-12)
The shared document mousemove/mouseup pair was reference-counted per
attached channel row: a sidebar with N channels took N refs (and N more
per re-render) while its single destroy returned exactly one, so the
count never reached zero and the listeners plus their activeDrag
closure lived for the rest of the process — the KNOWN BUG the
drag-reorder test pinned since the 2026-07-25 audit.
Ownership is now a Set of AbortSignals (the sidebar's lifetime
controller — the @lib/disposable teardown idiom): acquisition is
idempotent per signal no matter how many rows attach, release is the
signal's abort, and an owner aborted mid-drag clears the in-flight
visual state (which the old containerEl comparison never actually
matched in production — it compared the category container against
channelList). releaseGlobalDragListeners is gone; ChannelSidebar's
destroy releases via its existing ac.abort().
The pinning test changes with the fix, as its own comment instructed;
the lifecycle block now pins the fixed contract, including
re-registration after full teardown.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* test(e2e): E2EE identity-verification and updater journeys (DC-04)
The two remaining client-side headline coverage gaps from the 2026-08-04
audit's flow matrix (rows 38 and 49), both driven through the web mock
harness:
voice-e2ee-verify.spec.ts (6 tests, ux/voice-and-e2ee §7): the peer's
announce is real crypto — an ECDSA P-256 identity key signing an ECDH
ephemeral key exactly as e2eeCrypto does — so the badge states come out
of the production verification path. Covers the verified badge with
safety number (+ TOFU pin on first sight), the legacy unverified badge,
the mismatch block, the mismatch modal's reject path (peer stays
blocked, nothing pinned) and Trust New Key (re-pins the displayed key),
and the DC-08 fail-closed 'could not check' badge when the pin store is
unreadable. The harness gap that kept this untestable is closed by a
voice_join handler that grants a key-holder voice_token plus a WebSocket
shim that parks LiveKit's room.connect forever, holding the session
stably in 'securing'.
updater.spec.ts (4 tests, ux/settings-and-admin §5): no-update silence,
banner + Later dismissal, the full banner → download progress (% and MB
fallback via real update-progress events) → automatic relaunch journey,
and the failure state with Dismiss. There is no restart prompt by design
— the applied state IS the relaunch, asserted via the recorded
plugin:process|restart invoke.
Harness: buildTauriMockScript gains per-test identity-pin config
(identityPins / identityPinError) and a window.__invokeLog recording
every IPC call so tests can assert side effects with no DOM footprint.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* style(client): prettier-format the drag-reorder module and new e2e specs
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* feat(client): accessibility pass over the modal/overlay stack (DC-13)
The repo had exactly one focus-trapped dialog (UserProfilePopup), one
aria-live region, and no role="dialog" anywhere else. This pass
generalizes that one good implementation into lib/a11y.ts
(applyDialogSemantics, trapFocus, focusDialog, setRovingTabindex,
enableRovingNavigation) and applies it across the stack — all additive:
no DOM classes, testids, structure, or visible text changed.
- modalFactory: every factory modal now carries role=dialog + aria-modal,
moves focus in on open, restores it on every close path, and Tab-cycles
inside; createPromptModal is labelled by its title.
- Hand-rolled modals (CertMismatch/CertFirstUse/IdentityMismatch,
Create/Edit/DeleteChannel, InviteManager): dialog semantics labelled by
their existing headings, focus trap + restore, aria-label on icon-only
close buttons, and Escape mapped to each modal's SAFE action (reject on
the trust prompts — dismissal must never grant trust; cancel on the
channel modals — never the destructive/submit callback).
- SettingsOverlay: dialog on the panel, focus in on open/restore on
close; the sidebar is a vertical role=tablist with roving tabindex and
ArrowUp/Down/Home/End activate-on-focus; the content pane is a
tabpanel labelled by the active tab.
- QuickSwitcher: dialog + combobox/listbox/option wiring with
aria-activedescendant tracking the active row. QuickSwitchOverlay:
dialog + keyboard-operable rows (the inert current-server row stays
unfocusable on purpose).
- EmojiPicker/GifPicker: listbox/option cells with a roving tabindex
(Arrow/Home/End move the single Tab stop, Enter/Space activate through
the click path). inline-autocomplete: option ids + combobox attrs and
aria-activedescendant on the composer textarea — deliberately NOT
roving tabindex, since moving DOM focus out of the textarea would
break typing (the combobox pattern).
- Toast and TypingIndicator are polite live regions (role=status).
Tests: +71 unit cases across 18 files (4474 total, all green) pinning
roles, traps, restores, Escape safety, and roving behavior; plus an
axe-style e2e smoke (a11y-smoke.spec.ts, 5 tests) proving the wiring in
the running app — settings tablist + focus restore to the opener, quick
switcher combobox, member-picker Tab containment, live regions, and the
cert first-use dialog where Escape rejects without trusting.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(ux): replace file:line anchors with symbol references (DC-15)
The UX specs cited code as file:line anchors, and a three-week-old
snapshot already had 15 of 58 pointing at entirely wrong code (the audit
measured 200-700 lines of drift). All 55 remaining anchors across the
six spec files now reference the owning symbol instead
("validateForm() in pages/connect-page/LoginForm.ts"), each target
verified to exist before rewriting; the stale ones were re-aimed at the
correct symbol, not just de-numbered. Zero file:line references remain
under docs/architecture/ux/.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(audit): record the 2026-08-05 closure pass; stamp DC statuses in place
- audit-2026-08-04-docs-and-coverage.md: DC-06/08/12/13/15 marked
resolved, DC-04 and DC-09 further-resolved (admin-panel journey and the
handleApplyUpdate TODO are the remainders), matrix rows 38/49 flipped
from headline gaps to covered, the §4 UX-problem bullets closed, and a
§12 closure addendum records what shipped and the verification runs.
- CHANGELOG Unreleased: operator-facing entries for the DC-08 fail-closed
fix, the a11y pass, the UX polish, the backup_restore audit row, the
tag-gated CI tests, and the new e2e journeys.
- E2E-ISSUES.md: fresh full-suite run recorded at this HEAD — 291/291
passed in 9.2 min with zero flaky retries (276 baseline + 6 E2EE + 4
updater + 5 a11y smoke), @parity 15/15; suite inventory now 40 files.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* test(client): fix install-settle race in the updater e2e spec
The install settle handles are created only when the app's
download_and_install_update invoke reaches the mock wrapper, but the spec
called them right after asserting the banner text — which flips
synchronously on click, before the invoke's microtask runs. Local runners
won that race; CI lost it three attempts in a row
(window.__rejectInstall is not a function). Both settle sites now wait
for the handles, same pattern as the listener waits the file already uses.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
---------
Co-authored-by: Claude <noreply@anthropic.com>
* Final audit closure: admin-panel e2e, container-safe updates, blocking e2e gate, dependency policy (#1330)
* ci(server): run the tag-gated wazero/otel Go tests (DC-06)
The build-tag matrix only compiled the otel/wazero variants; the tests
behind those tags (plugin/sandbox_wazero_test.go 462 lines,
telemetry/telemetry_otel_test.go 214 lines) ran nowhere since they were
written (T-2026-07-25-16). Scoped to the two packages that carry tagged
files; verified green locally before wiring:
go test -tags wazero ./plugin/... and -tags otel ./telemetry/... both pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* fix(client): fail closed when the identity pin store is unreadable (DC-08)
getIdentityPin collapsed a keyring read error into "no pin stored", so a
transient failure sent a pinned peer down the TOFU first-sight path:
verifyPeerAnnounce verified against the server-delivered key and then
RE-PINNED it — a fail-open a malicious server could exploit by inducing
store errors (F3 follow-up 3, plans/security-scan-2026-07-22).
getIdentityPin now returns a three-state IdentityPinLookup
(pinned/unpinned/unavailable), mirroring how tofu.rs keeps Err distinct
from Ok(None) first-use. verifyPeerAnnounce rejects the announce on
"unavailable" without any pin write, records the new "unknown"
PeerVerification status, and the roster badge renders it as an amber
shield-question ("could not check") distinct from the legacy
"unverified" state.
Pinned by unit tests: pin present, no pin, store error (identity.ts),
the fail-closed rejection path (livekit-session), and the badge
presentation (channel-sidebar).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* feat(admin): write a backup_restore audit row that survives the restore (DC-09)
Backup restore was the one admin mutation with no audit_log row —
docs/security.md documented the gap as inherent ("the database is closed
as part of the restore"). The row IS writable durably: written
synchronously (LogAudit, deliberately not the async WriteAudit fast path)
before BackupTo takes the pre-restore safety copy, it is captured inside
pre_restore_*.db and survives the file swap forensically.
The extended restore test opens the pre-restore backup as a database and
asserts the backup_restore row is inside it — proving both the write and
its ordering. docs/security.md now documents where the row lives instead
of the gap.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* feat(client): channel-delete toast + optimistic reaction toggle (DC-12)
Two messaging-surface gaps the UX specs carried as open:
- channel_delete on the active channel now toasts "This channel was
deleted" alongside the existing redirect (ux/channels-members-dms
§1.2) — the redirect alone read as the app spontaneously changing
channels. Non-active deletions stay silent.
- Reactions toggle optimistically (ux/messaging §5): the pill flips on
the click, registered under the send's WS envelope id — the same
correlation scheme as the optimistic message rows. updateReaction
consumes the matching self-echo instead of re-applying it (the
delta-based arithmetic would double-count), other users' echoes apply
normally, and an error reply or local transport failure rolls back
exactly that toggle via rollbackReaction in the dispatcher's error and
send-failure handlers. The pill reverting is the failure feedback.
Both spec gap notes flipped to implemented; the stale §2 note claiming
the ready payload lacks slow_mode fell in the same edit.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* fix(client): guard the role-change submenu against double-fire (DC-12)
Every other destructive admin action already carried an in-flight guard
(withConfirmation, unblockRunning, banRunning, purge) — the role-change
submenu was the residual: currentRole only updates when the
member_update echoes, so a double-click (or a second option clicked
while the first PATCH was in flight) fired onChangeRole twice. One
shared guard now inerts the whole submenu while a change is running,
with the pending class on the clicked option.
The settings-and-admin spec's in-flight gap note flips to implemented,
and the stale messaging §8 slow-mode note is corrected in the same
docs sweep (the countdown shipped with the ready payload's slow_mode —
verified against ChannelController.startSlowMode and its tests).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* fix(client): drag-reorder document listeners leak — own them per sidebar signal (DC-12)
The shared document mousemove/mouseup pair was reference-counted per
attached channel row: a sidebar with N channels took N refs (and N more
per re-render) while its single destroy returned exactly one, so the
count never reached zero and the listeners plus their activeDrag
closure lived for the rest of the process — the KNOWN BUG the
drag-reorder test pinned since the 2026-07-25 audit.
Ownership is now a Set of AbortSignals (the sidebar's lifetime
controller — the @lib/disposable teardown idiom): acquisition is
idempotent per signal no matter how many rows attach, release is the
signal's abort, and an owner aborted mid-drag clears the in-flight
visual state (which the old containerEl comparison never actually
matched in production — it compared the category container against
channelList). releaseGlobalDragListeners is gone; ChannelSidebar's
destroy releases via its existing ac.abort().
The pinning test changes with the fix, as its own comment instructed;
the lifecycle block now pins the fixed contract, including
re-registration after full teardown.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* test(e2e): E2EE identity-verification and updater journeys (DC-04)
The two remaining client-side headline coverage gaps from the 2026-08-04
audit's flow matrix (rows 38 and 49), both driven through the web mock
harness:
voice-e2ee-verify.spec.ts (6 tests, ux/voice-and-e2ee §7): the peer's
announce is real crypto — an ECDSA P-256 identity key signing an ECDH
ephemeral key exactly as e2eeCrypto does — so the badge states come out
of the production verification path. Covers the verified badge with
safety number (+ TOFU pin on first sight), the legacy unverified badge,
the mismatch block, the mismatch modal's reject path (peer stays
blocked, nothing pinned) and Trust New Key (re-pins the displayed key),
and the DC-08 fail-closed 'could not check' badge when the pin store is
unreadable. The harness gap that kept this untestable is closed by a
voice_join handler that grants a key-holder voice_token plus a WebSocket
shim that parks LiveKit's room.connect forever, holding the session
stably in 'securing'.
updater.spec.ts (4 tests, ux/settings-and-admin §5): no-update silence,
banner + Later dismissal, the full banner → download progress (% and MB
fallback via real update-progress events) → automatic relaunch journey,
and the failure state with Dismiss. There is no restart prompt by design
— the applied state IS the relaunch, asserted via the recorded
plugin:process|restart invoke.
Harness: buildTauriMockScript gains per-test identity-pin config
(identityPins / identityPinError) and a window.__invokeLog recording
every IPC call so tests can assert side effects with no DOM footprint.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* style(client): prettier-format the drag-reorder module and new e2e specs
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* feat(client): accessibility pass over the modal/overlay stack (DC-13)
The repo had exactly one focus-trapped dialog (UserProfilePopup), one
aria-live region, and no role="dialog" anywhere else. This pass
generalizes that one good implementation into lib/a11y.ts
(applyDialogSemantics, trapFocus, focusDialog, setRovingTabindex,
enableRovingNavigation) and applies it across the stack — all additive:
no DOM classes, testids, structure, or visible text changed.
- modalFactory: every factory modal now carries role=dialog + aria-modal,
moves focus in on open, restores it on every close path, and Tab-cycles
inside; createPromptModal is labelled by its title.
- Hand-rolled modals (CertMismatch/CertFirstUse/IdentityMismatch,
Create/Edit/DeleteChannel, InviteManager): dialog semantics labelled by
their existing headings, focus trap + restore, aria-label on icon-only
close buttons, and Escape mapped to each modal's SAFE action (reject on
the trust prompts — dismissal must never grant trust; cancel on the
channel modals — never the destructive/submit callback).
- SettingsOverlay: dialog on the panel, focus in on open/restore on
close; the sidebar is a vertical role=tablist with roving tabindex and
ArrowUp/Down/Home/End activate-on-focus; the content pane is a
tabpanel labelled by the active tab.
- QuickSwitcher: dialog + combobox/listbox/option wiring with
aria-activedescendant tracking the active row. QuickSwitchOverlay:
dialog + keyboard-operable rows (the inert current-server row stays
unfocusable on purpose).
- EmojiPicker/GifPicker: listbox/option cells with a roving tabindex
(Arrow/Home/End move the single Tab stop, Enter/Space activate through
the click path). inline-autocomplete: option ids + combobox attrs and
aria-activedescendant on the composer textarea — deliberately NOT
roving tabindex, since moving DOM focus out of the textarea would
break typing (the combobox pattern).
- Toast and TypingIndicator are polite live regions (role=status).
Tests: +71 unit cases across 18 files (4474 total, all green) pinning
roles, traps, restores, Escape safety, and roving behavior; plus an
axe-style e2e smoke (a11y-smoke.spec.ts, 5 tests) proving the wiring in
the running app — settings tablist + focus restore to the opener, quick
switcher combobox, member-picker Tab containment, live regions, and the
cert first-use dialog where Escape rejects without trusting.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(ux): replace file:line anchors with symbol references (DC-15)
The UX specs cited code as file:line anchors, and a three-week-old
snapshot already had 15 of 58 pointing at entirely wrong code (the audit
measured 200-700 lines of drift). All 55 remaining anchors across the
six spec files now reference the owning symbol instead
("validateForm() in pages/connect-page/LoginForm.ts"), each target
verified to exist before rewriting; the stale ones were re-aimed at the
correct symbol, not just de-numbered. Zero file:line references remain
under docs/architecture/ux/.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(audit): record the 2026-08-05 closure pass; stamp DC statuses in place
- audit-2026-08-04-docs-and-coverage.md: DC-06/08/12/13/15 marked
resolved, DC-04 and DC-09 further-resolved (admin-panel journey and the
handleApplyUpdate TODO are the remainders), matrix rows 38/49 flipped
from headline gaps to covered, the §4 UX-problem bullets closed, and a
§12 closure addendum records what shipped and the verification runs.
- CHANGELOG Unreleased: operator-facing entries for the DC-08 fail-closed
fix, the a11y pass, the UX polish, the backup_restore audit row, the
tag-gated CI tests, and the new e2e journeys.
- E2E-ISSUES.md: fresh full-suite run recorded at this HEAD — 291/291
passed in 9.2 min with zero flaky retries (276 baseline + 6 E2EE + 4
updater + 5 a11y smoke), @parity 15/15; suite inventory now 40 files.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* test(client): fix install-settle race in the updater e2e spec
The install settle handles are created only when the app's
download_and_install_update invoke reaches the mock wrapper, but the spec
called them right after asserting the banner text — which flips
synchronously on click, before the invoke's microtask runs. Local runners
won that race; CI lost it three attempts in a row
(window.__rejectInstall is not a function). Both settle sites now wait
for the handles, same pattern as the listener waits the file already uses.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* ci: promote client-e2e to blocking (DC-07)
The soak is decided: green full-suite runs at 270, 276 and 291 tests across
the audit branches, and the one hard failure in the window was a real spec
bug a non-blocking job would have let rot.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* feat(server): refuse in-place self-update in container deployments
Resolves the long-standing handleApplyUpdate TODO. In a container the
running binary is image content: the staged replacement dies with the
container and the restart comes back as the old image. RunningInContainer
(OWNCORD_CONTAINER authoritative both ways — the shipped Dockerfile sets 1,
bind-mount operators can set 0 — with /.dockerenv//run/.containerenv as
fallback) now gates POST /admin/api/updates/apply with 503
CONTAINER_DEPLOYMENT before any updater logic, GET /admin/api/updates gains
can_apply, and the admin SPA swaps the apply button for an image-upgrade
note when it is false.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs: adopt the dependency pinning/review policy (DC-11, 2026-04-07 #8)
Writes down the policy the lockfiles already enforce: lockfiles
authoritative with npm ci-only installs, weekly Dependabot with majors
adopted deliberately, per-PR security gates (npm audit on shipped deps,
govulncheck, cargo audit, knip), and toolchain-level version pins.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* test(admin): add the admin-panel e2e journey against a real server (DC-04)
The admin SPA was the one surface no suite could reach: it is served by the
Go server and mocked nowhere. start-server.sh builds and boots a real
server (fresh temp data dir, TLS off, loopback) and the journey drives the
SPA end to end — first-run wizard creating the owner, dashboard stats,
channel create/rename, audit-log rows for both mutations, and sign-out/
sign-in. One shared page keeps the localStorage session across the serial
steps, mirroring the native suite's persistent fixture and staying under
the 5-logins/min limiter; on a Playwright retry the wizard branch downgrades
to login since setup is one-shot server-side. New non-blocking admin-e2e CI
job on the same graduation convention client-e2e followed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
* docs(audit): final closure — every DC finding resolved or deliberately reserved
Records the owner-directed closure pass (§13): DC-04 fully (admin journey
was the last row), DC-07 (client-e2e blocking), DC-09 fully (container-
aware update refusal), DC-11 + 2026-04-07 #8 (dependency policy).
Remaining open items are all deliberate: DC-14 reserved protocol entries,
the admin-e2e soak graduation, and the accepted/tracked 2026-04 carryovers.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KDaeRAN79nVdgtNX2zJVrx
---------
Co-authored-by: Claude <noreply@anthropic.com>
* fix(attachments): stop the orphan sweep destroying every avatar and the grace period
The 15-minute maintenance sweep deleted attachment rows and their files for
any attachment with message_id IS NULL. Avatars are exactly that by design:
users.avatar points at the attachment by URL and nothing ever links it to a
message (migration 027). Every avatar in the instance was therefore destroyed
on the first tick past the grace period, permanently 404ing every profile
picture. The query now excludes attachments a user's avatar still points at.
Independently, the cutoff was formatted RFC3339 while uploaded_at is written
by SQLite as 'YYYY-MM-DD HH:MM:SS'. TEXT comparison is bytewise and ' ' sorts
before 'T', so every unlinked upload sharing the cutoff's UTC date was swept
regardless of time -- the one-hour grace collapsed to 'immediately'. Rather
than fix the format at the one call site, DeleteOrphanedAttachments now takes
a time.Time and formats it internally, so no caller can reintroduce the class.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(client): send old_password so changing a password can succeed
The client posted {current_password, new_password} while the server decodes
json:"old_password" (Server/api/profile_handler.go:43). Go's encoding/json
does no alias matching, so OldPassword was always empty and every password
change returned 400 INVALID_INPUT -- the feature could never work for anyone.
docs/api.md and every server test already document old_password.
The existing unit test asserted the client's own broken payload, so it passed
while the feature was dead; it now asserts the documented server contract.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(admin): roll back and restart when a backup restore fails mid-copy
copyFile truncates the destination with os.Create before it can know whether
the read will succeed. On the restore path the destination is the live
database, already closed, so a failure in io.Copy or Sync left a zero-byte
chatserver.db, no rollback, and -- because the old code returned before the
restart -- a process still answering requests against a closed DB while the
response and the server_restart broadcast both claimed a restart was underway.
The failure branch now puts the pre-restore safety copy back (saying so
honestly in the error, including when the rollback itself fails) and requests
the restart the success path already did.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(voice): refuse client-side unmute while server-muted by a moderator
Push-to-talk called LiveKitSession.setMuted directly, which had no
server-mute guard -- only the voice widget's own handler checked. Unmuting
re-publishes a fresh mic track, and since MuteParticipantAudio only mutes the
track SIDs that exist at mute time while the LiveKit grant still carries the
microphone publish source, the SFU accepted it: holding PTT lifted a
moderator's mute and never told the server.
The guard now lives in setMuted itself, the one entry point every caller
shares, so PTT and any future caller are covered. Muting stays allowed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(ws): force a full ready when a client's seq is ahead of the ring buffer
EventsSince/EventsSinceFiltered guarded only the lower bound, so a client
asking for events newer than anything the buffer ever held got a non-nil empty
slice -- which handleReconnect reads as a successful, complete replay. It then
registers the client, sends auth_ok with replay_source=buffer and skips ready
entirely, leaving stale members, channels and read state until the counter
climbs back past the client's remembered value.
That disagreement is reachable in normal operation: the hub seeds its counter
from GetMaxEventSeq, which is 0 once the 24h pruner has emptied the table, so
a restart can reseed seq below a lastSeq clients preserve across reconnects.
Both functions now return nil (the existing 'cannot guarantee coverage'
signal) when afterSeq exceeds the newest buffered seq, so the caller falls
through to the cold tier and the intended full ready. afterSeq == newestSeq
remains the legitimate caught-up case and still replays empty.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(client): re-send channel_focus on auth_ok so reconnects keep receiving messages
channel_focus was sent only by mountChannel, which early-returns when the
channel id is unchanged, so a reconnect into the same channel never re-sent it.
The server transfers the focused channel from the old connection, but only
while that connection is still registered -- readPump's defer unregisters it
and drops every topic subscription the moment the server observes the close,
about a second before the client's first retry. Any server-observed close
(restart, proxy close, network reset) therefore resumed with no ChannelTopic
subscription: server channel messages, edits and reactions are delivered
exclusively over that topic, so the message stream went silently dead while
global events kept arriving and made the connection look healthy.
auth_ok fires on every connection including resumes and the full-ready
fallback, and it also covers the server-restart case where there is no old
state to transfer from. The server's handler is idempotent, so the extra focus
on a fresh connect is harmless.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(messages): stop persisting HTML-escaped text, safely
bluemonday writes text tokens through html.EscapeString, so sanitizeContent
persisted and broadcast the escaped form: every apostrophe, quote, ampersand
and angle bracket reached other users as a literal entity, and because stored
quote lines began with '>' the client's blockquote regex could never fire.
cleanText (display names, about, custom status, DM names) had the same bug.
Unescaping bluemonday's output alone would be a sanitizer bypass: surviving
text tokens can recombine into live markup -- '<<script>script>alert(1)<'
+ '</script>/script>' reassembles a real end tag. Instead the whole
unescape -> Sanitize -> unescape cycle now runs to a fixpoint, so the stored
result is by construction stable under re-sanitizing: any '<' that the
tokenizer would read as a tag start is stripped rather than re-encoded, and
only inert punctuation survives. The loop is bounded by the input length and
each pass is non-increasing; measured worst case over pathological tag/entity
soup at the 16 KiB input ceiling is under a millisecond.
The fuzz sinks are tightened to match the new contract rather than loosened:
they now require a tag-like start ('<' + letter or '/') because a bare '<'
followed by punctuation is inert plain text under every client render path.
The <script substring and idempotency checks are unchanged. Verified with
4.2M fuzz executions, zero crashers.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(client): close the WS task and TLS socket on disconnect via generation-owned teardown
ws_disconnect dropped the slot's sender expecting the write task to end, but
the monitor task held my_tx — a Sender clone kept only to prove teardown
ownership — so rx.recv() could never yield None: the monitor waits on the
writer via join_next() while the writer waits on the monitor's clone
dropping. Every intentional disconnect or profile-switch reconnect leaked
the writer, the reader, and the TLS socket, and with no server-side read
deadline the connection stayed registered — the user remained presence-online
after logout, and the stale Rust reader kept injecting the old server's
events into the new session's stores.
Ownership is now proven by the connection generation that already guards
install: the monitor captures my_generation plus the generation Arc and
clears/announces only if the generation is still current, checked under the
slot lock (generation only advances inside begin_connection while that lock
is held, so check-and-clear is atomic against new attempts). install_sender
receives the only Sender, so dropping the slot's sender really closes the
channel: writer exits, join_next returns, abort_all reaps the reader, and
the socket drops.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(voice): send voice_leave when the E2EE key exchange times out
connectAndSetup's key-exchange failure branch called leaveVoice(false) — no
voice_leave frame, no leaveVoiceChannel(). The timeout fires BEFORE
room.connect(), so no SFU participant ever exists and no LiveKit webhook can
clean up, while the server already registered the join when it sent
voice_token. The orphaned voice_states row matches the connected client's
channel, so sweepStaleVoiceStates never reaps it; once the ghost has the
lowest uid it wins key-holder election with a cleared E2EE state, every
later joiner's exchange times out and ghosts too, and rejoining the same
channel bounces off ALREADY_JOINED.
Mirror the reconnect-exhausted give-up path: leaveVoice(true) +
leaveVoiceChannel(), so the server drops the row and the local store
converges. The supersession checkpoints keep leaveVoice(false) — there a
newer attempt owns the server-side state and a voice_leave would destroy it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(ws): keep voice state when the replay-failure fallback will transfer it
handleFreshConnect's stale-voice cleanup ran unconditionally, but the
replay-failure fallback (lastSeq > 0, e.g. after a restart reset the seq
counter) reaches it while the old connection is still registered — and
registerNow then transfers that connection's live voice state into the new
client. The cleanup had already deleted the DB row, broadcast voice_leave,
and removed the live LiveKit participant (using the very JoinedAt token
being transferred), so the user ended up "in voice" on the hub only:
voice_join bounced off ALREADY_JOINED and sweepStaleVoiceStates never
reaps in-memory state without a row.
Skip the cleanup when lastSeq > 0 and the still-registered old client's
voiceChID matches the row — exactly the case registerNow transfers. All
other cases (F5 fresh connects, no old client, mismatched channel) keep
the existing cleanup, and if the old client unregisters in the window
before registerNow, the untransferred row is reaped by the next sweep.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(settings): stop the mic meter stream when it resolves after teardown
The mic-meter IIFE had no post-await guard: a getUserMedia resolving after
SettingsOverlay.hide() ran cleanup() (or after the tab's signal aborted)
opened the microphone anyway, started the rAF meter loop, and registerMic
re-armed state that cleanupMic() had already cleared — the mic stayed hot
for the rest of the session with nobody left to stop it.
Mirror the camera preview's request-id guard: cleanupMic()'s invalidation
callback now bumps a micRequestId alongside cameraRequestId, the IIFE
captures the id before the await, and a stale or aborted request stops the
just-acquired tracks and bails before touching the AudioContext.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(api,db): close the API-token and stale-ban holes in 2FA and account deletion
Four related gaps around sessionless (API-token) principals and account
teardown, all found by the bughunt harvest:
- 2FA enable/disable skipped the BUG-108 "revoke other sessions" step
entirely when the caller authenticated with an API token (nil session).
Both handlers now use the change-password pattern: keep=0 matches no
row, so every login session is revoked.
- verify-totp issued a session to a user banned after the password step;
it now runs the same IsEffectivelyBanned refusal as login.
- DeleteAccount left API tokens active (they authenticate independently
of the purged sessions) and left a stale lapsed ban_expires in place,
which makes banned=1 read as NOT banned — together a previously
temp-banned self-deleted account stayed fully usable through any
owner-minted token. Tokens are now revoked in the purge and
anonymiseUser sets ban_expires = NULL.
- The last-admin guard resolved admin-class roles by display name
('Owner','Admin'), so renaming the seeded Admin role silently disabled
self-deletion protection for its holders. It now keys on the canonical
OwnerRoleID/AdminRoleID plus any role holding the Administrator bit.
(The harvest's suggested criterion — Owner ID or Administrator bit
alone — would have DROPPED seeded Admins, whose 0x3FFFFFFF permissions
lack bit 30; the ID-based form preserves existing guard semantics.)
- DeleteAccount also now applies LeaveGroupDM's invariant: DM channels
left with zero participants are removed instead of becoming
unreachable, undeletable rows.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(api,ws): rate-limit gaps — bucket isolation, focus/mark_read, call_decline, reaper horizon
- Every empty-prefix RateLimitMiddleware mount shared one bare-IP bucket,
and the limiter records a timestamp per call regardless of the limit
passed — so unrelated endpoints capped each other at the minimum limit
(five ordinary profile edits 429'd the password endpoint; NAT'd logins
blocked register). The prefix is now a required parameter and every
mount names its own bucket, mirroring the existing client_update:/
livekit_proxy:/gif: pattern. The sessions-list handler also stops
401ing API-token principals (nil session only ever fed IsCurrent).
- channel_focus and mark_read were the only user-facing V2 handlers with
no rate limit, and each drives an unmetered SQLite write plus pubsub
churn; they now share a 5/s per-user budget (same underlying service
call), silently dropping over-budget frames like their siblings.
- call_decline gets the same limiter as its sibling call_ring — the
identical participant-lookup-plus-fan-out cost shape.
- The rate-limiter reaper pruned any entry idle past 15 minutes, but slow
mode passes windows up to the 6 h admin cap, so long slow modes were
silently reset; the cleanup horizon now covers the largest real window.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(server): unreachable route envelopes, dead env override, admin paging, plugin/updater defects
Seven harvest findings across router, config, admin, plugin and updater:
- The global 1 MiB body cap shadowed every route with a larger documented
envelope: the 16 MiB plugin install 400'd at ~1 MiB and an at-limit
avatar could never fit its multipart framing. The exemption list is now
a named var covering uploads, plugin install, and avatar — each of
which enforces its own cap at the route/handler level.
- queryInt clamped offset with the limit's 500 cap, so the admin audit
log and user list could never page past row 550; the cap is now an
explicit per-call bound (offset callers pass MaxInt32).
- OWNCORD_EVENT_PERSISTENCE_* env overrides were documented but dead:
envKeyToKoanf cut at the first underscore, producing the unknown path
event.persistence_* that koanf silently drops.
- InstallPlugin trusted LastInsertId, which SQLite does not update on the
upsert's DO UPDATE branch — on the shared writer connection a reinstall
returned the rowid of some unrelated prior INSERT, so EnablePlugin
no-opped and plugin_kv wrote to a nonexistent plugin id. RETURNING id
is correct on both branches.
- Every wazero plugin re-activation compiled the module again and leaked
the previous CompiledModule; the handle is now retained on the instance
and closed in deactivate, the lost-activation race, and the
closed-module release path.
- Linux server self-update was gated on the Windows-only
chatserver.exe.sig asset it never uses; the required-asset check and
the signature fetch are now GOOS-aware.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(ws): batch S4 — seq-gap shed, replay tail merge, drain-on-close, handshake teardown
Eight harvest findings in the hub/replay/pump paths, each locked by a
watched-red test in harvest_s4_internal_test.go / reconnect_db_test.go:
- kickClient closes the send channel BEFORE UnsubscribeAll so a racing
Subscribe can never leave a dead client holding a topic.
- deliverBroadcast consults the topic limiter BEFORE allocating a seq: a
shed frame no longer burns a sequence number that sits in the replay
buffer forever unpublished.
- onStaleTick prunes idle topic-limiter buckets (Cleanup had no caller).
- dm_channel_open bumps the visibility watermark so a client resuming
from an older seq takes the full-ready path instead of silently losing
the targeted, unsequenced open.
- computeAllowedChannels treats a DM-lookup failure as fatal (full ready)
instead of replaying with every DM event silently stripped.
- Cold-tier replay merges the ring-buffer tail past the newest persisted
row; if the buffer cannot vouch for the flush gap it forces full ready.
- writePump drains queued frames (e.g. the BANNED kick reason) after
closeSend instead of dropping them on the first closed channel.
- A failed post-registerNow handshake runs the offline teardown when no
replacement connection holds the slot — no more users stuck online.
Declined by design: hoisting registerNow above the replay snapshot
(report L390) — every fallback path would re-register the same client
and registerNow self-kicks the slot holder; the µs dedup window does not
justify that risk in the hottest path.
The kickClient ordering test is a 300-iteration stress whose race window
is too narrow to hit reliably; it documents the invariant rather than
having been watched red.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(client): batch C1 — store merges, ready-badge resync, pending-send teardown
Ten harvest findings in the stores/dispatcher layer, each locked by a
watched-red vitest test:
- setMessages merges instead of clobbering: live broadcasts and
pending/failed optimistic rows that landed while the history GET was
in flight survive the snapshot.
- addChannel is idempotent — the re-sent channel_create on role edits no
longer wipes unread/mention counts, lastMessageId, or canSend.
- setChannels carries client-synthesized DM rows across the rebuild.
- READY marks the focused channel read after the store repopulation so
stale server read_states cannot resurrect badges on the channel the
user is actively reading (skipped on first connect).
- setVoiceStates maps the ready payload's camera/screenshare flags
instead of blanking live streams on a mid-call resync.
- The dm_channels length guard is gone: an empty array is authoritative
and clears ghost DMs.
- addMessage's defensive pending-row reconcile requires content equality
so another session's replayed message cannot consume the pending row.
- performSend into a detached history window reattaches to present
first, mirroring onJumpToPresent.
- prependMessages at the cap keeps the fetched older page and detaches
the window instead of silently discarding the fetch (which refetched
the same page forever); hasMore is the server's value again.
- A connection leaving "connected" fails every pending optimistic send
(retry affordance) instead of letting rows spin forever.
One existing assertion updated to the corrected semantics: trimming on
prepend now drops rows below the window, so hasMore stays the server's
value and the test asserts the detach instead.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(server): batch S5 — voice eviction scoping, fail-closed camera cap, role-service races
Seven harvest findings in the voice/service layer, each locked by a
watched-red test:
- The CONNECT_VOICE revocation sweep evicts via a channel-conditional
clear (the in-memory analogue of LeaveVoiceChannelIfMatch): a
voice_join to a permitted channel that commits while the DB-backed
permission check runs can no longer be torn down. The report's
suggested pre-check guard was rejected — it leaves the same race open
between guard and clear, proven by the interleaving test.
- A failed channel switch's abort branch re-subscribes the restored
session to its VoiceTopic and re-elects the key holder; without them
the session silently missed every voice_e2ee relay.
- voice_camera fails closed when the VoiceMaxVideo lookup errors instead
of skipping the cap check and enabling unconditionally.
- LiveKitProcess starts the child inside the p.mu critical section that
publishes p.cmd (Wait stays outside), removing the data race between
Start's cmd.Process write and IsRunning/Stop reads.
- AffectedUserIDs reports lookup success; handlePatchRole falls back to
a blanket permission-cache invalidation when the member list was
unreadable, instead of evicting nobody and leaving revoked grants live.
- RoleService serializes its read-check-write mutations (position
uniqueness and the role cap are snapshot-enforced, not DB-enforced);
concurrent creates can no longer land on the same position and tie
every hierarchy comparison.
- channel_focus writes the read state even when the channel has no
undeleted messages — the upsert is what zeroes mention_count, so
emptied channels finally clear their badge.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(client): batch C2 — voice eviction teardown, supersession scoping, VAD generation
Seven harvest findings in the voice/session layer, each locked by a
watched-red vitest test:
- A server-initiated voice_leave for self tears down the LiveKit session
(mic publish + E2EE key material), guarded on channel match so a
late-arriving leave cannot kill a newer join.
- VIDEO_LIMIT refusal rolls back with disableCamera() — max_video has no
SFU-level enforcement, so the already-published track kept streaming.
- teardownForReconnect sends voice_camera/voice_screenshare OFF frames
before stopping local tracks, freeing the server-side max_video slot a
reconnect otherwise occupies forever.
- Supersession checkpoints 3/4/5 disconnect only their own local room
(mirroring checkpoint 2) instead of calling the global leaveVoice,
which by then tears down the newer attempt's live session.
- retryMicPermission honors a moderator's server-mute like it honors
deafen — granting mic while listen-only no longer hands the channel an
unmuted track.
- handleDisconnected defers to the active reconnect loop (livekit-client
fires Disconnected synchronously inside the loop's own connect call),
preventing a second uncancellable retry loop.
- stopVadPolling invalidates an in-flight startVadPolling addModule via
a VAD-scoped generation counter, so VAD cannot resurrect itself with a
stale threshold.
Deliberately skipped: the report's optional RATE_LIMITED camera rollback
— that error code is shared by unrelated actions and the payload cannot
attribute it to a camera toggle, so a blind rollback would be wrong.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(client): batch C3 — E2EE rotation races, pin-write tri-state, stale-offer guards
Six harvest findings in the voice-E2EE key-exchange layer, each locked
by a watched-red vitest test; the security-invariant sweep (re-pin
TOCTOU, forward-secrecy rekey, concurrent-rotation, blind-repin — 239
E2EE-adjacent tests) stays green:
- Re-election as key holder during an in-flight rotation defers (sets
_isKeyHolder + _rotationPending, mirroring the sibling branch) instead
of dropping the election and stranding the room without a holder.
- storeIdentityPin returns tri-state stored/no-store/failed; a FAILED
pin write now marks the peer unverified instead of displaying
"verified" with no pin persisted — an unpinned peer could never trip
mismatch detection, the exact MITM window the pin exists to close.
- handleOfferInner discards a stale offer when the session keypair
changed, not just the epoch — a non-key-holder never bumps epoch, so
an offer surviving clearState() into the next session passed the
epoch-only check.
- handleAnnounce's wrap-and-offer path gets the same epoch guard as the
receive path, so a rotation landing mid-wrap cannot ship a dead key.
- The key-exchange retry races a FRESH promise (the first rejection had
permanently settled the old one, making the retry window zero), and
aborts cleanly when clearState() tore the session down mid-exchange.
- setupKeyExchange publishes _ecdhKeyPair only after _isKeyHolder and
_roomKey are ready, so a concurrent announce is queued and drained
through the offer-sending path instead of being consumed offer-less.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(client): batch R1 — secret store must not report a broken keyring as empty
Three harvest findings in the Tauri credential store, each locked by a
watched-red test:
- secret_store::get treated a keyring read error as "nothing stored"
whenever the fallback file was also empty, which is indistinguishable
from first login. loadOrGenerateIdentityKeyPair reads exactly that
signal, so an unreadable keychain made the client mint and publish a
fresh identity key over the existing one, invalidating every peer's
TOFU pin. It now prefers a fallback copy and otherwise propagates the
error; loadIdentityKey rethrows instead of swallowing to null.
- A failed keyring write left any older entry in place while the fresh
secret went to the fallback file — and get() reads the keyring first,
so the stale value shadowed the new one forever. The write-failure arm
now purges the entry, mirroring the read-back-mismatch arm beside it.
- fallback_crypto deleted nothing when the key file's write or sync
failed, leaving a short file that every later load rejects; since the
key file is never rewritten once it exists, one ENOSPC poisoned the
fallback store permanently.
Both Rust fixes needed a small injectable seam (get_with/set_with,
finish_new_key_file) because the keyring error branches are otherwise
unreachable without a live OS credential store.
The saved-login path is unaffected in behavior: loadCredential still
catches and degrades to "no saved credential" rather than surfacing the
new error. The persistence re-read in loadOrGenerateIdentityKeyPair
deliberately does not rethrow — the keypair already exists in memory by
then, so a transient failure keeps the existing "did not persist" warning.
fallback_crypto is cfg(not(windows)), so its test ran only under a
temporary local gate lift (reverted, verified no residual diff); it
executes for real on the Linux and macOS CI runners.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* docs: rewrite CLAUDE.md per Claude 5 context guidance; commit skills + hunt workflow
The CLAUDE.md files were a manual: build commands, code style, layout Claude
can read off the filesystem. Rewritten so they are short and spend their
tokens on gotchas instead — the things that are invisible until they cost an
afternoon.
Moved out of CLAUDE.md into skills (progressive disclosure), which also fixes
three references to skills that never existed:
- ci-check: the full local CI mirror, including the four Go build-tag variants
and the deadlock pass a plain build/test misses, and the windows-latest
runtime.scanstack GC fault that should be rerun rather than investigated.
- db-change: the sqlc workflow plus three silent traps — non-ASCII query files
truncating the NEXT query's emitted SQL, semicolons in migration comments
orphaning statements, and LIMIT 1 mis-emitting on a :one query.
- protocol-change: regenerate both constant files and commit the pair.
Dropped: command lists duplicated from the Makefile and package.json, prettier
style rules the formatter already enforces, and layout facts a directory
listing answers. Added the subsystem invariants that keep getting rediscovered
the hard way — the ws seq/FIFO contract, voice-session supersession scoping,
E2EE staleness guards, and the Node 22 webstorage failure mode.
.claude/ is no longer ignored wholesale: skills and workflows are tracked so a
cloud session, which sees only tracked files, starts with instructions rather
than nothing. Machine-local settings and locks stay ignored. Deleted
bughunting.js, a superseded copy declaring the same workflow name as
bughunt.js, which left the registry ambiguous.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(client): batch C4 — cert-latch scoping, stale active channel, credential opt-out
Thirteen harvest findings across the client UI and wiring, each locked by a
watched-red vitest test:
- The TLS cert-mismatch latch fired on any host's event, so an unrelated
saved profile's rotated certificate permanently killed this socket's
reconnect loop. It now latches only for the connected host, cancels any
pending reconnect (a timer armed before the mismatch would otherwise fire
connect(), clearing the latch and resuming against that host), and resets
on a fresh connect.
- ready never cleared activeChannelId when the channel vanished from the
snapshot, and MainPage's subscriber had no else branch — the message list
and composer stayed mounted and enabled against a channel the server no
longer recognizes. Both sides fixed; the mark-read from batch C1 is
suppressed when the clear happens.
- user_update re-saved the session token unconditionally, bypassing the
remember-password opt-out, and dropped the stored password while doing it.
- A failed older-page fetch latched loadingOlder, permanently killing
infinite scroll for that view; it now clears in a finally.
- Concurrent message jumps raced, letting the older response overwrite the
newer window. Guarded by a generation counter.
- A FORBIDDEN send in a group DM flagged participants[0] as blocking, which
disabled the unrelated 1:1 composer with that person; block gating is
1:1-only.
- streamPreview added an abort listener per call instead of per signal.
- dm_channel_close had no fallback when the closed DM was being viewed;
both call sites now share one closeDmLocally helper.
- The GIF picker routed through the textarea and discarded the draft.
- QuickSwitcher listed DM rows that the DM section already shows.
- Accepting a rotated certificate reconnected into a page with nothing left
listening, stranding the user on the connect screen.
- Logout read voiceStore after clearAuth had already reset it, so the
voice_leave was never sent; clearAuth now snapshots logoutWasInVoice.
- disconnect() left reconnectAttempt set, carrying a stale backoff ceiling
into the next login.
Also fixes two lint errors this branch introduced earlier and that only a
full `npm run lint` catches: a useless spread in the C1 pending-send sweep
(now Array.from, which states the snapshot intent), and two floating
promises in C2's voice_leave handler, where converting an implicit-return
arrow to a block body stopped chaining them.
main.ts and MainPage.ts have no unit-test seam, so three focused pieces were
extracted to make the fixes testable: createUserUpdateCredentialSaver,
reconnectAfterCertAccept, and the logoutWasInVoice snapshot.
One existing assertion corrected: a dispatcher test claimed ready must keep
an active channel that was absent from the payload, which locked the bug.
It now keeps a channel that is present, with a sibling test for the absent
case.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix: resolve 107 verified defects across ws hub, voice/E2EE, db, and client
Findings came from a multi-round hunt, each verified against the tree by an
independent adversarial pass before any code changed, then fixed and re-reviewed.
Every behavioural fix carries a regression test that was confirmed to fail
against the pre-fix code.
Server (Go)
- Reconnect/replay: force a full ready when retention pruning has removed the
events after a client's last_seq, rather than accepting the surviving suffix
as a complete resume; close the snapshot-to-registration window under seqMu;
restore the focused-channel subscription during the handshake via a new
READ-gated active_channel_id auth field; supplement replay with the client's
own voice room; tear down transferred voice sessions on a failed handshake.
- Hub: ratchet visibilityChangeSeq upward only (all three writers); make the
stale-voice sweep error-aware so a transient DB failure no longer evicts
every participant; re-elect the E2EE key holder on sweep and cleanup paths.
- Voice: preserve moderator mute/deafen across channel switches; deliver
voice_leave to evicted users; gate camera/screenshare permission checks on
the enabling direction only; reject joins to non-voice and archived channels.
- Permissions: archived channels are now read-only and unjoinable, and
can_send is recomputed per client on role/override changes.
- Data: stop cascaded message deletes from stranding uploaded files
(migration 030 unlinks instead); clear personal data on account deletion;
exclude banned users from owner lookup; drop the silent 1000-member cap;
advance the author's own read state on send.
Client (TypeScript / Rust)
- Voice: make joinGeneration monotonic so a superseded attempt can no longer
pass supersession checks; scope aborted-path cleanup to the attempt's own
room; send voice_leave on connect failure; stop push-to-talk from writing the
user's explicit mute flag; gate join-time PTT muting on a new backend
capability probe so platforms that cannot report key state are unaffected.
- E2EE: act on the tri-state pin-write result instead of reporting an
unverified peer as verified; use keypair ownership rather than null checks.
- State: reset the message cache on logout; clear NSFW acknowledgements on
logout; scope channel mutes, NSFW acks and DM notes per server host.
- UI: make the attachment remove button and the failed-send Retry/Discard
buttons work; fix drag-reorder's phantom-drag latch and its permission gate.
Docs: protocol.md now documents can_send, active_channel_id, the archive
read-only contract, and the sequenced/unsequenced presence split.
Verified: all four Go build tag variants, go vet, go test -race, the ws
deadlock detector, sqlc and protocol generation, tsc, eslint, prettier, and
the full client suite (169 files, 4664 tests).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LWUzUmsjCKNAzAfQzT4fz7
* fix(ci): satisfy golangci-lint, knip, and the host-scoped mute key in e2e
Three CI failures on the previous commit, all genuine fallout from it.
golangci-lint (v2.11.3) — 6 issues in tests added by that commit:
- contextcheck: the temp-ban subtest captured an outer ctx while calling
seedTokenUser, which builds its own; declare ctx inside the subtest instead.
- modernize: use WaitGroup.Go and range-over-int in three tests.
knip — SessionResponse in lib/types.ts became unused. The getSessions fix
replaced it with SessionInfo in lib/api.ts, which documents why the old
declaration was wrong (it named ip_address/expires_at, which the server never
sends, and omitted ip/is_current, which it always does). Delete the dead type
rather than re-export it, and fold that reasoning into the surviving comment.
Client E2E — the per-channel-mute parity test asserted the pre-scoping
localStorage key. Channel mutes are now keyed mutedChannels:<host>, because
channel ids are per-server autoincrement integers sharing one webview origin;
verified in a browser that the app writes
owncord:settings:mutedChannels:localhost:8443. The test now resolves whichever
scoped key exists instead of pinning the test server's host, so it still
asserts the same thing: the id persists on mute and is gone on unmute.
Verified with the CI linter version built against Go 1.26 (0 issues), all four
build tag variants, go vet, go test -race, the ws deadlock detector, knip,
tsc for both tsconfigs, prettier, the full client unit suite, and the
previously-failing parity specs run in a real browser.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LWUzUmsjCKNAzAfQzT4fz7
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Update the security doc's capability section with the deny list, the fact
that `http:allow-fetch` is the only URL-scoped HTTP identifier, and why
the https wildcard cannot be removed without moving the link-preview
fetch into Rust. Note under Known Limitations that narrowing the plugin
scope alone does not bound exfiltration while CSP `connect-src` allows
`https:` to any host. Add a capability row to the client architecture
doc and mark the design note implemented.
