conventions

Scan diffs and the working tree for personal information (emails, personal paths, API keys, phone numbers) before commit or push.

The phase-gated developer loop — /setup /plan /work /review /release /bugfix — extracted from agentm as a standalone native plugin. The base other dev plugins enhance.

Design-doc and ADR authoring — the /design command (author → translate → sequence) and the ADR skill for capturing architectural decisions with re-audit triggers.

The Obsidian/Google-Drive vault storage backend for the agentm memory engine — re-homed out of the kernel behind the V5-1 storage seam. Ships the `vault` named backend (and, across V5-2, the vault-specific machinery) as an agentm-discovered scripts/ payload.

CI workflows + dependency-update tooling — the dependabot-fixer skill that repairs breakage on Dependabot PRs.

Deterministic transcript-JSONL cost analyzer — per-message cost curve, 5h rolling window, floor breakdown, and cache-read vs cache-write vs fresh-input split. Prerequisite for Part C (status-line meter) and Part G (heat-based floor).

Opinionated, template-driven wiki maintenance — the wiki-author authoring/learning skill, the documenter agent, the diataxis-evaluator, recent-wiki-changes, and the check-wiki.py linter. Maintains a wiki in the operator's voice; enhances developer-workflows' phase-boundary documentation when both are installed.

Standalone adversarial code review of any diff or PR — the adversarial-reviewer (+ cross-model Gemini) agents, the evidence-tracker hook, a /code-review command, a /simplify command, and security-review + testing-strategy skills. Enhances developer-workflows' review capability when both are installed.

Discovery-to-watchlist research primitives. idea-search is a read-only scan over the agentm recall engine surfacing existing vault/codebase entries relevant to a question. learn-forward leans by name on agentm's approved-source forward-learning pipeline to mine operator-configured sources onto the watchlist. codebase-improvement applies a research insight's stale pattern against the operator's own repo, surfacing exactly one watchlist finding per match -- never auto-editing. All three are strictly discovery-surfacing: findings land in the watchlist for operator review, nothing is ever auto-adopted.

Universal control + safety layer — kill-switch, steer, commit-on-stop hooks plus commit-no-coauthor / worktrees-operator-initiated conventions. Usable in any session; enhances developer-workflows (its hooks engage across the phase loop) when both are installed.