Agent Claude Code
Tries to break things, under one evidence standard: a finding is executable or it is not a finding. Two modes. REFUTE — given a change or a design, find the input, ordering, crash point, concurrent caller, or false premise that makes it wrong. PROOF — given a claim ("this bug exists", "this fix works"), reproduce the…
Agent Claude Code
MuninnDB's resident code reviewer. Use before opening a PR and when reviewing one. Reviews a change for correctness and for adherence to MuninnDB's cognitive, storage, and security invariants and its cross-surface sync obligations. Builds and tests the actual change (with -race where it matters) and RED-sanity-checks…
Agent Claude Code
Builds a designed MuninnDB increment RED-first in an isolated worktree, then pushes the branch WITHOUT opening a PR so the adversarial review runs first. Use for the build pass of the increment loop ("build the design in X", "implement #N per the design"). Every behavior change lands with a test proven to fail without…
Agent Claude Code
Designs a MuninnDB increment before any code is written: the mechanism, the minimal first slice with explicit deferrals, the invariant and cross-surface impacts, and the MEASURABLE proof with a pre-committed acceptance rule. Use for the design pass of the increment loop ("design X", "how should we build Y", "scope the…
Agent Claude Code
Interrogates the MODEL rather than the code: units and who owns the tick, numeric domain and boundary behaviour, absolute vs relative quantities on the wire, missing signal read as zero, where a constant came from, and whether the mechanism's assumptions about its own input distribution actually hold. Use whenever a…
Agent Claude Code
Measures a mechanism against REAL vault data on a read-only clone, and reports aggregate numbers only — never memory content, concepts, tags, entity names, or vault names. Use for any measurement whose substrate is a real corpus ("measure X on a real vault", "quantify the damage", "re-read the meter", "run the…
Hook Claude Code
Runs after a tool call finishes for Write, Edit, Write, Edit and Bash tool calls, executing drift-guard.mjs and ledger-guard.mjs via node (2 commands). From scrypster/muninndb.
Hook Claude Code
Runs before the context is compacted, executing memory-drain.mjs via node with --hook and --trigger. From scrypster/muninndb.
Hook Claude Code
Runs when a session ends, executing memory-drain.mjs via node with --hook and --trigger. From scrypster/muninndb.
Hook Claude Code
Runs when the agent finishes a response, executing memory-drain.mjs via node with --hook, --trigger and --debounce. From scrypster/muninndb.
Hook Claude Code
Runs when a session starts, executing memory-freshness.mjs via node. From scrypster/muninndb.
Settings file Claude Code
Agent settings declaring 5 hook events (PostToolUse, PreCompact, SessionEnd, Stop).
Skill Claude CodeCodex
The repeatable build loop MuninnDB holds itself to — design, build RED-first, independently vet, adversarially refute, MEASURE real value on a vault, then PR/CI/land. Use for any non-trivial feature or fix ("build X", "add Y", "next increment"). Encodes the discipline that keeps us shipping real, measured value…
Skill Claude CodeCodex
Run competing positions against each other and come out with a DECISION, not a pile of opinions. Use when a genuine fork has to be resolved and the evidence does not obviously pick a side: a product-doctrine question, a design that returned two defensible shapes, or a measurement that contradicts something already…
Instructions file
Claude Code instructions for scrypster/muninndb, covering muninndb — working guide for ai agents, 1. what muninndb is, architecture map, 2. core principles (the lens for every change) and 3. how we work.
MCP server Claude CodeCodexCursor +2
MCP server "memento" as configured in scrypster/memento. Runs in Docker (ghcr.io/scrypster/memento:0.1.0).