History was selected by msg_to, so whoever next held a released nick, or
created a deleted channel again under its old name, could read the
earlier holder's messages. Each message row now records the session a
DM or INVITE was sent to (to_session_id) and the channel a channel
message was sent to (to_channel_id), and GET /api/v1/history selects by
the one that applies, using its index. A session's DM history now also
follows it across nick changes. A DM or channel message whose recipient
is deleted before it is stored gets the unknown nick or channel error.
Model: opus-5-5
The IRC relay registers a broker wait channel on each pass of its
loop, and when the 100 ms ticker fired instead of a notification it
looped without removing that channel. An idle IRC listener connection
added ten channels a second to the broker, and all but the last stayed
after the connection closed. The ticker case now removes its channel,
as the HTTP long-poll does after its timer.
Broker.ListenerCount returns how many wait channels a user holds, so
the test can check that an idle connection holds at most one, and
none after it closes.
Model: opus-5-5
Co-authored-by: clawbot <sneak+clawbot@sneak.cloud>
The nick rule moves from internal/handlers to internal/service as
ValidNick, and the HTTP API and both IRC listener NICK paths call it.
The IRC listener used to cut a nick to 32 bytes and accept anything
else, so a nick with spaces such as "bob PRIVMSG #inj :hi" made the
lines built from it read as another user's PRIVMSG. It now answers a
nick the rule refuses with ERR_ERRONEUSNICKNAME (432), before and
after registration, and keeps the nick as it was.
Model: opus-5-5
README.md rules that nicks are case-sensitive, but the IRC listener's
relay and user MODE on both transports compared nicks ignoring case:
an IRC listener user alice never saw a PRIVMSG, NOTICE, JOIN, PART,
NICK or QUIT from Alice, and MODE Alice from alice was taken as her
own nick. These checks now use ==.
Tests cover alice receiving each of these from Alice, MODE Alice from
alice getting 502 on both transports, and Alice registering while
alice exists. README.md drops the text that described the old
behaviour.
Model: opus-5-5
The IRC listener's JOIN now passes the channel key on, INVITE records
the invite, LIST and WHOIS hide +s channels from non-members with the
queries the HTTP API uses, and PRIVMSG to a +H channel gets 404.
The INVITE checks and the channel hashcash check move from the HTTP
handlers into the service, so both transports run the same code. An IRC
client cannot attach a stamp, so its PRIVMSG to a +H channel always
fails that check. The unfiltered channel list query, now unused, is
removed.
Model: opus-5-5
Co-authored-by: clawbot <sneak+clawbot@sneak.cloud>
The IRC listener relayed only the first line of a PRIVMSG or NOTICE
body; it now sends one line per body line. A relayed MODE change now
carries its key, limit or mask, each value its own parameter, also in
the echo to the user who made it. CR, LF and NUL are removed from each
parameter and from each whole line sent to an IRC listener connection,
so no stored value can start a line of its own. The channel name check
moved into service.JoinChannel, so the IRC listener's JOIN refuses the
same names with 403.
Model: opus-5-5
Co-authored-by: clawbot <sneak+clawbot@sneak.cloud>
The default database URL asked for WAL mode with `_journal_mode` and
`_busy_timeout`, which the SQLite driver does not read, so the database
ran in rollback-journal mode. The default now uses the driver's
`_pragma=journal_mode(WAL)` and `_pragma=busy_timeout(5000)` parameters,
in one constant that both the config default and the database fallback
use. The database no longer sets its own busy timeout after opening, so
the URL's value applies. A new test opens a file database with those
parameters through db.New and checks the journal mode and busy timeout.
README.md's database text now matches, and an in-memory test database
drops the same ignored parameter.
Model: opus-5-5
Co-authored-by: clawbot <sneak+clawbot@sneak.cloud>
Each test environment in internal/ircserver named its shared in-memory
database after the test. A repeat run of a test in the same process
could open the database that an earlier run's server was still
holding, and start on that run's users. The database now gets a random
name, and a new test checks that two environments open at once do not
share one.
The database close and the server stop are now registered as soon as
each is opened, so a setup that fails part-way still closes its
database.
Model: opus-5-5
Every command a registered IRC listener client sends, PING and PONG
included, now sets last_seen on its session and its client, so the
idle cleanup no longer removes an IRC user who is still active; one
who sends nothing for SESSION_IDLE_TIMEOUT is still removed. The two
UPDATEs that GetSessionByToken ran for every authenticated HTTP API
request move into UpdateLastSeen, which both transports now call. The
IRC listener refreshes on every command, with no interval between
writes: that is what every HTTP API request already costs. README.md
describes last_seen to match.
Model: opus-5-5
Co-authored-by: clawbot <sneak+clawbot@sneak.cloud>
USERHOST, VERSION, ADMIN, INFO, TIME, KILL and WALLOPS work on the
HTTP API and on the IRC listener. The USERHOST reply, the VERSION,
ADMIN and INFO text, KILL and WALLOPS are each one service function
that both transports call. User mode +w selects who receives WALLOPS.
MODE on your own nick queries or changes your user modes through one
parser on both transports: the whole mode string is checked before
anything is stored, and both flags are stored in one transaction. KILL
closes the victim's IRC connection after sending it KILL and ERROR,
without the operator waiting on it. The sessions table gains
is_wallops in 001_initial.sql, as the pre-1.0 rule in REPO_POLICIES.md
asks.
Model: opus-5-5
Co-authored-by: clawbot <sneak+clawbot@sneak.cloud>
The shared files are fetched from sneak/prompts dd4027b, with this
repository's own entries after the shared content in .gitignore,
.editorconfig and .dockerignore.
Lint and tests are Dockerfile phases (golangci-lint v2.14.0, Debian Go
1.24.13) that the build stage depends on, and the Makefile targets call
the script/ entrypoints. make fmt also formats Markdown with prettier.
Fixes for the new lint findings: the auth cookie is always Secure, an
IRC connection's relay goroutine stops when the connection closes, and
repeated strings are constants. neoirc-cli treats a plain-HTTP server on
localhost or a loopback address as secure, so local use keeps its session.
Whether the 60-second test cap covers building the test phase is open
on sneak/prompts issue 113.
Model: opus-5-5
The HTTP server built its router inside the goroutine that starts
serving, so the start hook returned before the router existed, and
the internal/handlers tests raced with it or hit a nil router. The
start hook now runs configure, enableSentry and SetupRoutes, in that
order, then serves in the background.
The goroutine that sends queued messages to an IRC client read c.nick
without c.mu while NICK changed it. Every such read now takes the lock.
Under -race in the Docker build, internal/handlers takes over 30s on
database work, not clock waits, so both go test runs in make test use
-timeout 120s. The || retry stays
(#101).
Model: opus-5-5
Changes the Go module path from `git.eeqj.de/sneak/neoirc` to `sneak.berlin/go/neoirc`.
All occurrences updated:
- `go.mod` module directive
- All Go import paths across 35 `.go` files (107 import statements)
- All JSON schema `$id` URIs across 30 `.json` files in `schema/`
No functional changes — this is a pure rename of the module path.
`docker build .` passes clean (formatting, linting, all tests, binary build).
closes#98
Co-authored-by: clawbot <clawbot@users.noreply.git.eeqj.de>
Reviewed-on: #99
Co-authored-by: clawbot <clawbot@noreply.example.org>
Co-committed-by: clawbot <clawbot@noreply.example.org>
clawbot
