AI Companion – LLM Player Bot (Ollama, ChatGPT)

AI Companion 0.1.6

⚠️ Existing users: your skill files in config/aicompanion/skills/ are not updated automatically — the mod never overwrites files you might have edited. Run /companion skills reset after updating to pick up the rewritten farming and harvest skills and the new fishing one. Your old copies are saved as .md.bak.

On the work skills: farming, harvesting and fishing are a work in progress. This release makes them a lot more reliable than they were, but the companion still needs supervision — it will occasionally narrate a job instead of doing it, build a field that isn't quite right, stall partway through, or need a nudge to empty its pockets. Treat them as something to play with rather than something to leave running unattended. Feedback and logs are genuinely useful.

New

• Token usage HUD — a panel in the top-left shows your session's LLM spend at a glance: total tokens, the input/output split, request count, and a tokens-per-minute graph for the last 30 minutes. It appears only while a companion is out, and F1 hides it with the rest of the interface. The graph is the useful part on a paid endpoint — a companion answering the odd question looks nothing like one stuck in a loop, and you'll spot the difference within a minute instead of at the next usage report. This one's solid.
• Fishing skill — /companion skill fishing. The companion finds water, casts, and keeps fishing. It'll fetch a rod first if it hasn't got one. Early days: it decides for itself when its pockets are full, so it sometimes keeps going longer than you'd like, and it won't stash the catch unless you ask.
• /companion skills reset [name] — restores the built-in skills to their latest versions after a mod update, backing up your edits to .bak. Takes effect immediately, no restart.

Fixed

• The companion no longer goes quiet or stands around mid-job. Three separate causes: telling it to stop left it with nothing queued up so it never spoke again; and a simple wave or an "I'm already on it" reply to casual chat would silently cancel whatever it was doing — after which it happily kept reporting progress on a job that had already ended.
• The fishing float is reeled in when fishing stops. It used to be left bobbing in the water forever.
• farm no longer fails when the companion forgets the range — it defaults to 16 blocks instead of erroring and wasting a turn.

Improved

• Farms are much more likely to sit level with the ground. They were reliably ending up on a raised dirt platform with the water channel pouring off the sides and flooding everything nearby. The companion is now told how to read ground level and to keep water contained, and most fields come out flush — but this is guidance to a language model rather than a hard rule, so you'll still get the occasional raised or flooded field. Ask it to clear the field and try again if you get a bad one.
• Farming and harvest skills rewritten. They previously described tilling soil with a hoe, digging channels and planting seeds by hand — none of which the companion can actually do, so it would describe the work convincingly and then stand still. Both now use only real actions and check whether they're genuinely working before claiming to be. Much better, not yet bulletproof.
• Better builds for non-buildings. Fields, paths, walls and bridges are no longer treated like houses, so you don't get furniture and room dividers in a wheat field.

Known rough edges

• Field layouts vary between builds. The same request can produce a flush field one time and a raised or badly watered one the next. Building is model-driven, so results depend on the day and the model.
• Harvesting can stall on a large field — the companion stands among ripe crops doing nothing. Diagnostics were added this release to pin down the cause; if you hit it, the log now records why it stopped.
• Harvesting is slow by design. It walks to each crop one at a time, so a big field takes many minutes. Progress that looks like nothing is often just progress.
• Nothing is stashed automatically. Ask it to put the wheat or fish in a chest and it will, but that interrupts the job and you'll need to tell it to resume.
• Results depend on your model. A weaker or heavily quantised local model is more likely to talk about a job than do it. If a skill misbehaves, try a more capable model before assuming the skill is broken.