ai-bot: fulfill readRealmFile via Matrix attachments on the host-command path

[lukemelia]

Stacked on #5344. Base is that PR's branch, so the diff here is only the delta. Review/merge #5344 first.

What this explores

An alternative architecture for the readRealmFile feature: instead of ai-bot running reads in an inline, same-turn loop and feeding content to the model in-process, the read content lives in Matrix as an attachment, and the continuation rides the existing host-command result path. This is the "content in Matrix, attachment variant" we wanted to evaluate against the same-turn approach in #5344.

How it works now

1. The model calls readRealmFile. It surfaces as a normal command request on the bot's message, tagged executedBy: 'ai-bot' (the host records it, never runs it — those guards from ai-bot: the assistant can read a skill directly from the realm server #5344 are unchanged).
2. After the answer streams, ai-bot fetches each file, uploads it to the Matrix media repo, and posts a command-result event carrying the file as data.attachedFiles (success) or an invalid result with a reason (failure).
3. That result event re-enters the handler (a narrow exception lets the bot's own command-result events through; everything else it posts is still ignored). getShouldRespond then waits until every request — reads and any host commands — has a result, exactly as it already does for host commands.
4. On the next turn, the existing reconstruction (toResultMessages → buildAttachmentsMessagePart) downloads the attachment and feeds its content to the model — the same path the host readFile uses.

What this deletes

Because reads now resolve next-turn like host commands, the two timing models collapse into one. Gone:

• the inline for(;;) generation loop in main.ts and messagesOverride;
• the mixed-round rejection (readRealmFile can now coexist with host commands in one answer);
• the "thinking"-event indicator rotation (beginCommandResultIndicator / sendCommandResultIndicator / resetForNextEvent);
• buildReadRealmFileFollowup / in-memory tool-result splicing.

Net −90 lines, with main.ts substantially simpler.

Storage / dedup

Uploads dedupe by content hash, so a skill read repeatedly is stored in Matrix once, and a changed file misses the cache and re-uploads (dedup without staleness). Within a room, a file already attached isn't re-read at all (the existing isFileAttachedInRoom path). Matrix media itself is not content-addressable, so this app-level cache is what avoids re-storing identical bytes.

Trade-offs to weigh vs #5344

• (+) Far less bespoke machinery; reads and host commands share one path and mix freely.
• (+) Reads are transparent in the timeline as real command requests + results.
• (−) File content now persists in Matrix (media repo), where ai-bot: the assistant can read a skill directly from the realm server #5344 deliberately kept it out (privacy/size).
• (−) "Always live" softens to "live as of when read" — a later turn reconstructs from the attachment snapshot rather than re-fetching.
• (−) Each read batch costs an extra Matrix round-trip + a per-turn credit check (the continuation is a normal billed turn).

Notes on correctness

• Billing / on-behalf-of: a self-triggered continuation arrives sender = bot; it's re-attributed to the single human in the room (reads only happen in single-human rooms, so this is unambiguous), so billing and the delegated read act on the user's behalf.
• No double-response: getRoomEvents slices history at the trigger event, so for a batch of reads only the handler whose result completes the set passes getShouldRespond; the bot's answer is posted after all results and is never in a result-handler's sliced history. Same mechanism that protects the host multi-command flow.

Tests

Unit tests for the fulfillment (applied result + attachment on success, invalid + reason on failure, malformed args, and content-hash dedup → one upload), plus classifyToolCalls (reads and host commands coexist) and fileLabelFromUrl. All green; ai-bot type-check and lint clean.

🤖 Generated with Claude Code

[jurgenwerk]

Getting this problem, the bot won't continue after reading. Will try to fix

[jurgenwerk]

Added a fix, tested it and it looks good!