LGTM

"Review as if you have to maintain it forever."

LGTM (opencode-lgtm) is an OpenCode plugin designed for experienced engineers. It focuses on high-level architectural design, code quality, and rigorous code reviews.

Vision

Unlike generic coding assistants that focus on writing code quickly, LGTM focuses on writing code correctly. It acts as a digital Staff Engineer or Maintainer.

Key Features

• Architectural Guardrails: Analyzes changes against project-specific design patterns.
• Deep Code Review: Focuses on edge cases, security, and maintainability rather than just syntax.
• Maintainer Persona: Provides rationales for suggestions, fostering a culture of mentorship.
• Review-Only Mode: By default, it suggests and explains rather than blindly editing files.

Philosophy

• Quality over Speed: Better to have no code than bad code.
• Explicit over Implicit: Use clear types and patterns.
• OSS Spirit: Open, collaborative, and opinionated.

Development

• Run bun run check before opening a PR.
• specs/ is a git submodule; keep it pinned to the intended lgtm-specs commit.

Agents

• Primary entrypoint: @review (LGTM Review Lead).
• roles/* agents (e.g. roles/planner, roles/reviewer/logic) are internal subagents used by the lead via task.
• If you manually select a roles/* agent (e.g. @roles/planner), LGTM will warn and you should switch back to @review.

Tools

• lgtm status: shows detected language and whether config loaded.
• lgtm clean_cache: clears LGTM cache.
• lgtm gen_policy: creates/updates .lgtm/policy.md for the current repo.
  • On update, writes a backup to .lgtm/policy.md.bak.
  • Add repo-specific guidance under "Local Notes (Preserved)" (kept on reruns).

Debugging

• Inspect rendered prompts/models:
  • opencode debug agent review
  • opencode debug agent roles/explorer
• Inspect what was sent in a run:
  • JSONL logs: ~/.local/state/opencode/lgtm/logs/*.jsonl (or XDG_STATE_HOME/opencode/lgtm/logs)
  • Export a session: opencode export <sessionID>

Default Config

• The plugin ships a default config template at config/default.json.
• Full config reference: config/README.md.
• On startup, LGTM loads this bundled default first, then overlays user config from ~/.config/opencode/lgtm/config.json (or XDG_CONFIG_HOME/opencode/lgtm/config.json).
• To customize locally, copy and edit:
  • cp config/default.json ~/.config/opencode/lgtm/config.json
• User config is treated as the source of truth for explicitly configured capabilities. By default (skip_model_validation: false), startup fails fast when a configured capability cannot resolve to any usable model.

Interactive Model Setup

• Run bun run setup:models to launch a terminal wizard.
• The wizard reads available provider/model IDs from OpenCode config (~/.config/opencode/config.json or XDG_CONFIG_HOME/opencode/config.json) and writes LGTM config to ~/.config/opencode/lgtm/config.json.
• If strict validation fails, rerun bun run setup:models or manually update models.capabilities so each configured capability has at least one usable model.

Config Parameters

• Configure runtime policy in defaults inside ~/.config/opencode/lgtm/config.json.
• defaults.review_bash_policy
  • Default: relaxed
  • strict: enforce single-command, read-only git/gh usage in @Review workflows.
• defaults.allow_general_fallback
  • Default: false
  • true allows capability resolution fallback to general when direct capability candidates are unavailable.
  • Security posture: keep disabled unless you explicitly want resilience over strict capability isolation.
• defaults.debug_logging
  • Default: false
  • true enables LGTM debug JSONL logging.

Debug Payload Guards

• LGTM debug logs record full prompts/outputs for traceability.
• Some upstream tool outputs may still be truncated before LGTM sees them (e.g. large command output payloads).