chore: agentic OS polish — coding rules and reliability fixes

[rgxdev]

Summary

• CLAUDE.md: persistent project-level AI coding rules (Conventional Commits, minimal/typed/secure code, no UI churn, targeted-only validation). A matching global ~/.claude/CLAUDE.md was also installed in this environment (not part of this repo/diff).
• fix: truncate get_terminal_output before it enters model context — terminal_output returned a live terminal session's full buffer uncapped, unlike terminal_run/task_run/test_run, which already tail-slice to the last 4000/8000 chars. A long-running session's output could blow the model's context budget just from re-reading it. Now capped to the last 8000 chars, matching the existing pattern.
• fix: prevent duplicate concurrent runs of the same automation — listDueAutomations only advances nextRunAt once the run fully completes, so an automation run that outlives one 30s scheduler poll tick was still "due" on the next tick and got dispatched a second time in parallel (same race let a manual runNow overlap a poll-triggered run). Added a process-local in-flight guard, mirroring the existing inFlightSyncs pattern in knowledge-base.ts, so a second concurrent dispatch is skipped instead of double-running the automation's agent/workflow.

Investigated but found already well-handled (no change needed): knowledge-base prompt-context injection already has a char budget + env-configurable caps; text_search/codebase_search/usages already cap aggregate output bytes; dead-letter/transient-retry handling and stale-run lease recovery already exist in agent-runtime.ts/scheduler.ts; approval resolution already uses an atomic CAS to prevent double-approval.

Files changed

• CLAUDE.md (new)
• apps/server/src/tools-builtin/read.ts
• apps/server/src/tools-builtin/read.test.ts (new)
• apps/server/src/scheduler.ts
• apps/server/src/scheduler.test.ts

Test plan

• bun test apps/server/src/tools-builtin/read.test.ts apps/server/src/tools-builtin/search.test.ts — pass
• bun test apps/server/src/scheduler.test.ts — pass (5/5, including the new in-flight-guard test)
• bunx turbo run typecheck --filter=./apps/server — pass
• bunx biome check on all touched files — clean
• Full suite / broad UI review — intentionally not run, per scope (targeted validation only)

Remaining risks

• The automation in-flight guard is process-local only; it doesn't protect against duplicate dispatch across multiple server replicas in a multi-instance deployment (would need a DB-level atomic claim, out of scope for this pass).
• get_terminal_output's new cap is a fixed 8000 chars (matching the existing task_run/test_run tail-slice constant); no configurability was added, consistent with "prefer simple truncation over complex summarization."

---

Generated by Claude Code