How messages, models, tools and memory fit together

1. Receive and enqueue

   The platform adapter turns text, mentions, quotes, images, files and sender details into a single MessageEvent, which is written to a SQLite inbound queue before processing, so a crash or restart does not lose it; failures are retried with backoff.

2. Backfill after reconnect

   After a reconnect Diana pulls the history it missed and checks group seq numbers for gaps. Only the newest few messages per session that would have triggered a reply directly enter the reply path; the rest only fill in history.

3. Deterministic gates

   Checks group availability, whether the bot and group are enabled, mutes, model quota, member level, reply suppression and bot-to-bot suppression, recording the reason at every step.

4. Trigger detection

   Direct messages are handled directly; groups look for explicit mentions, quotes, aliases, links and plugin commands. A follow-up or correction the same person sends while a reply is generating is merged into it.

5. Unprompted reply rating

   Group messages that were not directly triggered are rated by the router model — "is this addressed to the bot?" plus a 0–1 chat score — then passed by chat level, cooldown and the bot's share of recent messages. Errors, timeouts (60 seconds) and unparsable output all mean silence.

6. Assemble context

   The stable system prompt and history come first; memory, tool evidence and other per-turn content sit after the cache breakpoint, all trimmed to the token budget.

7. Model and agent loop

   The agent runs one tool per step within step, per-tool timeout and finalization-reserve limits; when a limit is hit it is forced to finalize, and the reply is checked for line breaks, evidence and links before it returns.

8. Pre-send review

   Model silence, semantic dedup, one audit model call (private closing, group stop, loops, account safety, unprompted reply quality), and send-time checks for interrupts, idempotency and pauses.

9. Send and record

   Replies go back to the source channel in chunks, with only the first quoting and mentioning; group sends back off on separate text and media lanes. Outcome codes, reasons, delivery stages, latency and tokens are all persisted.

Flowcharts, conditions and outcome codes for every step are on the Pipeline page.

Retrieval narrows with structured filters first, then ranks by textual relevance and time decay, and finally passes the model only the matching evidence. Summaries are not recomputed when configuration reloads — they expire only when the input history changes or the summary version is bumped.

• Ordinary images are converted into whatever the provider accepts; the media layer does not shrink originals unconditionally.
• Text-heavy screenshots use both a high-detail vision input and OCR evidence.
• History image descriptions are cached by content SHA-256 and algorithm version.
• Image count and vision tokens are budgeted so base64 history cannot grow without bound.
• Video contributes a limited number of context frames, and temporary files are cleaned up afterwards.