perf(server): stop loading the whole database on hot orchestration paths

[leoisadev1]

Several server hot paths did work proportional to the entire environment instead of the thread at hand. Starting a Codex or Kimi session loaded every project, bot, thread, message, and activity through a full projection snapshot just to find one thread. Every streamed assistant token re-read secret files from disk to check one settings flag. Each persisted event wrote all 14 projector cursors and reloaded the thread row, and checkpoint capture loaded a thread's full message history about four times per turn. Since node:sqlite is synchronous, these all block the event loop and grow with total chat history.

Session starts now use targeted by-id lookups with new narrow queries for delegations, groups, and bots. Materialized settings are cached and invalidated on writes, so secrets are read once instead of per token. Projectors declare the event types they handle and unrelated ones are skipped, with all cursors advancing in one batched write. Checkpoint capture uses the lightweight checkpoint-context query, ingestion caches the per-thread runtime context across content deltas, and the shell snapshot keeps open plus the 20 most recent finished delegations per thread with result text capped at 2,000 characters. Two behavior boundaries: the session-start lookup excludes archived and deleted threads, which the old snapshot included, and a thread whose project row is missing no longer gets a checkpoint.

Verified with 401 passing tests across the touched orchestration and provider suites, including new tests for the settings cache, projector skipping, and checkpoint linkage, plus 241 server, decider, and routine tests and a clean workspace typecheck. Fixing a pre-existing race in the ProviderCommandReactor test drain was required: the faster pipeline exposed that drain could return before subscribed events reached the worker.

Created with Claude Fable 5.1 in Claude Code.

🤖 Generated with Claude Code

---

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
akeru-bot-landing	Ready	Preview	Sep 26, 2026 5:52pm UTC

[greptile-apps]

Confidence Score: 5/5

[High risk] Refactors database queries on hot orchestration paths.

No outstanding blocking findings; the PR is safe to merge based on the reviewed findings.

Summary

This PR replaces full-state reads on server orchestration paths with targeted queries, caches settings, reduces projection work, and limits delegation data sent to shell clients. The latest change makes reactor startup fail rather than lose provider work when it cannot replay events committed during startup.

_{Reviews (5) · Last reviewed commit: "fix(server): read the reactor startup ga..."}