Mediabot development journal · MB1001–MB1303 · 14 November 2026
When we last opened the spellbook, Mediabot had reached MB1000: a thousand numbered checkpoints, a competitive word plugin, French Trivia, and a growing collection of lessons about what can go wrong between receiving a command and delivering its answer.
Since then, the castle has gained new windows, a better control room, and considerably stronger wards. Mediabot 3.7 reached its stable release, development moved to 3.8dev, and the work gradually shifted from correcting individual spells to testing how the whole apparatus behaves under pressure.
There are more than three hundred checkpoints to catch up on. Fortunately, nobody is being assigned three hundred pages of History of Magic. Here are the changes that matter. 🏰
Shared links and searches have become richer without asking the IRC loop to wait for distant websites.
MixCloud gained opt-in previews for shows and public creator profiles, including available titles, authors, durations and profile details. TikTok now has a dedicated asynchronous preview path for supported video, photo, profile and other public links, instead of relying on a generic landing-page title.
Wikipedia searches follow the channel’s language. DuckDuckGo searches can deliver three distinct organic results, each with its own title, excerpt and destination. A missing result cannot quietly borrow its neighbour’s description, and a blocked search is not dressed up as a successful ranked answer.
The smaller corrections matter too: Spotify metadata survives quoted HTML attributes, namespaced RSS and Atom entries are recognised, YouTube preview authors lose an unnecessary terminal “- Topic” suffix, and shorter MetaAI source links remain readable without needless shortening.
The guiding charm is simple: send useful information, preserve its source, and say when the owl could not complete its journey.
Partyline gained an internal command shell, entered with .command. Operators can navigate virtual task directories, inspect workers and scheduled jobs, filter listings, and follow bounded live observations. A selected channel gives those observations an explicit context.
This is an administrative view of Mediabot’s own machinery, not an operating-system shell. Access belongs to authenticated, authorised sessions, and watches stop when their owner or context is no longer valid.
An importable Grafana dashboard adds another view of the castle: IRC and database status, activity over time, and recent channel and command rankings. Setup documentation covers the path from the metrics endpoint to the dashboard.
For users, the new public !ping reports Mediabot’s local response-preparation time, including time spent waiting in its output queue. It is useful evidence of bot responsiveness, rather than a claim to measure the user’s network round trip. ⏱️
A repeated “continue” should continue the right conversation. AI cache identities now account for the complete request: history, instructions, model and generation parameters, within the caller’s conversation.
Resetting or forgetting a conversation also retires its pending answer ownership. An old callback cannot publish into a replacement conversation or unlock the newer request. MetaAI checks whether its work is still current before launching further research, contribution or synthesis stages.
That does not undo a provider request already sent, but it prevents obsolete work from acquiring fresh authority inside the bot.
Hailo and Spark received related repairs around queue identity, timing and text handling. Negation, explicit numeric signs and Unicode characters deserve to survive an edit; malformed generation identifiers do not deserve to revive an abandoned event. 🕯️
A substantial part of this chapter concerns changes that should be invisible when everything works normally.
Role changes now invalidate the relevant cached authority. Account deletion revokes sessions after the transaction commits. Partyline sessions lose obsolete privileges and pending privileged work when their authority changes or can no longer be confirmed.
Channel identity is more consistent across letter-case variants and supported IRC channel prefixes. Polls, karma, achievements, membership observations and private replies received focused corrections so that similar spellings cannot accidentally become different rooms—or send an answer to the wrong person.
Connection ownership received the same attention. Old JOIN, WHO, NAMES and delayed-output callbacks must remain attached to their original connection. Replacing a transport retires its pending work instead of allowing yesterday’s ghost to speak through today’s socket.
PONG traffic can now overtake pending ordinary frames, while ordinary messages retain their order. Keepalive recovery distinguishes an unanswered exchange from a PING that never left the local output queue.
AntiFlood also became more precise: it charges physical IRC lines, avoids spending global allowance on locally refused attempts, refills channel credit progressively, and expires stale deferred responses after a bounded lifetime. Long replies still need orderly admission; retrying them should not manufacture another flood incident.
Timers, plugin jobs and child processes now have clearer ownership through replacement, cancellation and shutdown. Generic workers retain responsibility for collecting their operating-system children even after logical completion. Achievement workers, paced replies and nicklist timers received similar lifecycle repairs.
Rehash handling preserves confirmed channel information and active log destinations when replacement preparation fails. Incomplete cleanup is reported as incomplete, rather than hidden behind a cheerful success message.
The architecture work gave several responsibilities clearer homes: Partyline plugin administration, output admission, channel data access, version presentation and user selection. The aim was to make the next correction easier to place and test. The extraction chapter also acquired a stopping point; moving staircases are entertaining at Hogwarts, less so in a maintenance branch.
Everyday commands benefited along the way: Trivia deadlines and score persistence, poll expiry, reminder cancellation, Unicode-safe replies, birthday validation, retained unsent RSS articles, and more dependable database-error handling.
The final stretch turned those improvements into a five-part qualification campaign:
echo, spike and addbadword, including Unicode, caller identity, revoked rights and SQL recovery.These tests run without asking production channels to become a Defence Against the Dark Arts classroom. Where a boundary is simulated—such as an HTTP response or an identity lookup—the qualification record says so.
The endurance run on the development host completed 300 cycles and delivered 1,200 messages in about twenty minutes. Post-drain notifier and descriptor counts stayed at their baselines. Peak RSS growth after warmup was approximately 2.41 MiB, within the declared 16 MiB envelope.
| Evidence | Result |
|---|---|
| MB1302 host endurance | Passed: 300 cycles, 1,200 deliveries |
| MB1302 host full suite | 43,027 / 43,027 assertions |
| MB1303 targeted native validation | 244 / 244 assertions |
| MB1303 fresh qualification CI | 43,044 / 43,044 assertions across 1,139 files |
| MB1303 CI, Debian 13 gate and fresh qualification | All passed |
The test suite itself received housekeeping: redundant checks were consolidated, independent cold-import probes were batched within a small process limit, and targeted checks precede a single full pre-commit run. The objective is useful regression coverage, not a larger pile of green counters.
MB1303 closes the implementation of this planned campaign, with its three CI workflows now green. We have a repeatable way to examine the difficult moments: a revoked permission, a failed read, a replaced connection, a delayed reply or a worker that must be collected.
The limits remain explicit. Three representative commands are not every command. Twenty minutes of local traffic are not several days of production load. A fresh dependency inventory is not a fully locked dependency universe.
But the next bug now arrives in a castle with better maps, clearer ownership and a proper testing chamber.
Raise a butterbeer to 3.7, keep a wand ready for 3.8dev, and let the next change earn its place in the spellbook. 🍺🪄
You must be logged in to reply.