Skip to content

CLAUDE.md

Latest commit

 

History

History
68 lines (52 loc) · 4.97 KB

CLAUDE.md

File metadata and controls

68 lines (52 loc) · 4.97 KB
Raw

AutoBot Development Instructions

Rules: docs/developer/CLAUDE_RULES.md — read when starting a new task Workflow: docs/developer/CLAUDE_WORKFLOW.md — labels, gh workarounds, CI diagnosis, pre-merge gates Git & Worktrees: docs/developer/CLAUDE_GIT.md — branching, safety, push recovery Batch & Agents: docs/developer/CLAUDE_BATCH.md — parallel work, sub-agent rules Code Review: docs/developer/CLAUDE_REVIEW.md — review methodology, merge checklist Issue Closure: docs/developer/CLAUDE_CLOSURE.md — closure gate, discovery issues Reference: docs/developer/AUTOBOT_REFERENCE.md Architecture exceptions: docs/developer/ARCHITECTURE_EXCEPTIONS.md

Read a pointed doc only when its topic triggers. Universal rules (worktree mandate, issue decomposition, verification/evidence, CI diagnosis, no temp fixes) live in the global instructions and the docs above — not repeated here.

Engineering Standard

Correctness → Speed → Maintainability. No wasted motion. No speculative work.

  • Parallelize independent tool calls, file reads, agent tasks
  • Minimal surface area — least code that fully solves the problem
  • Async-first — never add sync calls to async paths
  • Decision speed: ≤3 exploration commands then act; if stuck, escalate with findings

Core Rules

Every task must: link to GitHub issue · search memory first · break into subtasks · use code-reviewer · update issue throughout · verify before closing.

  1. Check Before Writing — search existing code/docs/PRs before creating anything
  2. Reuse — import from autobot_shared/; never duplicate or hardcode
  3. Standardize — ≤30-line functions; no _v2/_fix suffixes
  4. Clarify — confirm architecture before coding
  5. Verify — show evidence (test output, curl, build) before claiming done
  6. Report & Fix — file GitHub issues for every discovered bug, even off-task, and fix pre-existing issues discovered along the way (same PR when in-scope; fast-follow when larger). ASK only for large refactors / product decisions / risky blast radius. See CLAUDE_RULES.md Rule 6.

Essential Patterns

What How
Redis from autobot_shared.redis_client import get_redis_client
Config from autobot_shared.ssot_config import config / import { getBackendUrl } from '@/config/ssot-config'
Logging logging.getLogger(__name__) / createLogger('Name') — no print() or console.*
Encoding Always encoding='utf-8' explicitly
Cache TTL Never hard-code — use module-level constant from env var (see chat_history/cache.py)
LEDGER/EXECUTOR Coordination tools complete instantly — do NOT wait; continue immediately with execution tools
Copyright mrveiss is sole owner/author

Workflow Quick Rules

  • Branch target: Dev_new_gui for all PRs
  • Commit format: <type>(scope): <description> (#issue-number)
  • Security reviews are findings-first: read the diff only, emit a severity/file:line/issue/fix table within 3 tool calls, verify after — never explore before the table lands (skill: secreview)
  • Long analyses go to a file, not the response: research/audit/comparison output is written to docs/research/<topic>.md (or docs/audit/) incrementally as it is produced; the reply is the path plus a short summary — an interrupted or token-capped response must never lose the work
  • Never --no-verify — PostToolUse hook auto-formats .py
  • Protected branches: main/master blocked by pre-commit hook → use issue-* or hotfix-*
  • PR template headings: Thinking Path · What Changed · Verification · Model Used
  • No internal info in outward artifacts: never expose IPs, hostnames, secrets/tokens, or internal filesystem paths in GitHub issues/PRs/comments/logs — redact to generic role/node refs
  • PR queue limit: ≥5 open PRs → defer implementation and notify
  • Codebase is source of truth: never edit /opt/autobot/ or /var/log/autobot/
  • System updates (test AND prod): ONLY via the builtin updater a user reaches at /slm/maintenance/updates/code-sync (code-sync API / self-update); if the builtin can't do it, fix that gap (issue + PR) — never side-channel via ad-hoc ansible/shell
  • Batch similar-scope issues per PR: one PR may close several similar-scope issues (Closes #A, #B, …) — each MUST be fully delivered (closure Gate 3: partial delivery never closes); genuinely independent or different-risk changes still get separate PRs
  • Model tiers: Haiku (claude-haiku-4-5-20251001) for status/labels/simple transitions; Sonnet for implementation and reviews