One memory for every coding agent — plain markdown in your repo. Injected on Claude Code & Cursor, skills on Codex & Copilot, plugin for OpenCode. Dated, capped, audited: agent memory that doesn't rot. Session handoffs, human-approved promotion into knowledge and rules, plus orchestration and QA layers.
Agent
Builder to an ALREADY-DECIDED spec (the model split: the main session designs; the executor builds). Use for: building a designed slice/chunk, test writing, mechanical refactors with an objective gate (typecheck, test suites, byte-replay). The spec — normally a file at projects/ /plans/YYYY-MM-DD- .md, named in the…
Agent
Isolated Claude instance that critiques ideas, designs, patterns, hypotheses, or plans without context from the parent conversation. Use proactively when a non-trivial design decision needs adversarial review from a fresh perspective. Pair with an external-family model (parallel call) when you have one — different…
Agent
QA lens agent: probes the RUNNING product — web UI via Playwright MCP, API via curl/HTTP — through ONE assigned lens (user-flow · edge-state · honesty · contract · ux-critique) and returns STRUCTURED FINDINGS with repro steps + evidence. Observation-only: it never decides, never fixes, never edits files, never clicks…
Agent
Read-only fact-gatherer (the model split: the main session designs; recon gathers — breadth over depth). Use for: mapping a codebase surface before a design, input-availability tables, verifying a doc claim against code, sweeping many files for a fact, external fresh-checks against official sources. Returns RAW FACTS…