An AI agent that runs inside Houdini - not beside it. Natural language in, real undoable nodes out. USD-native memory, five LLM engines, honest by design: unmeasured renders UNKNOWN, never a guess.
Agent Claude Code
Runs the V1 hard gate — live dir()/hasattr introspection on the live Houdini build — to confirm each candidate's API actually exists. Answers one question only — does it exist on this build.
Agent Claude Code
Read-only mapper of the SYNAPSE codebase. Inventories MCP tools, routing tiers, the autonomy loop, and architecture seams. Maps terrain; does not prospect.
Agent Claude Code
Conductor for the CLEAR work-clearance harness (harness/clear/). Sequences the four lines (L1 uncommitted-work, L2 decisions-board, L3 open-from-release, L4 changelog-gap), dispatches the existing specialist roster (cartographer/assayer/forge/crucible/scribe/gatewarden/sidefx-cto), and halts at every human gate (Gate…
Agent Claude Code
Adversarial reviewer. Attacks the opportunity frame (Pass 2) and the finished map (Pass 6). Did not build the map; is motivated to break it. Hostile by design, fair in method.
Agent Claude Code
Conductor for the FREEZE FORENSICS RELAY — diagnoses why SYNAPSE freezes the Houdini node interaction surface when a prompt is sent. Sequences the dynamic freeze-forensics workflow (historian → seam-map → hypothesize → probe → attack → verdict), includes the live-evidence legs (stall detector, dispatch-wait histogram…
Agent Claude Code
Runs the blueprint §10 intake protocol on external research artifacts (dossiers, whitepapers, threads, release notes). Produces a one-page adjudication appendix under docs/intake/ — tier-labeled claims, P1–P7 verdicts, harvest into existing gaps. Never revises the blueprint itself.
Agent Claude Code
G7 public-claim-surface surgeon. Fixes doc drift on README.md and docs/ only — auth instructions, About/framing language, badge/version reconciliation, loopback-security sentence, C3 Moneta-naming consistency (one-Moneta ruling). Never touches code, tests, or state files.
Agent Claude Code
Implementation agent for gated H22 blueprint work — G1 port waves, G5 grounding tools, G9 gate build, baseline/freeze code. Refuses any dispatch that lacks a GATEWARDEN ALLOW. Works in worktrees, one atomic commit per wave, never merges or pushes.
Agent Claude Code
Read-only gate-state oracle for the H22 blueprint harness. Reads drop.json, flywheelqueue.json, posture.json and the arming markers, then emits an ALLOW/REFUSE verdict per work item. Never flips a gate; never writes anything.
Agent Claude Code
Phase-0 paper author for the H22 blueprint. Drafts the Leg-0 specs (port-wave manifest, pre-flight gate, grounding contract, benchmark design) and freezes baselines. Writes ONLY under docs/ and harness/notes/. Every cited path, count, or symbol is verified by its own Read/Grep first.
Agent Claude Code
Implementation leg of the latency relay. Builds exactly one dispatched item from the 2026-07-27 report §5 — U1–U4 instrumentation, SessionStart ping fix, declarative-coverage or perceived-latency work — in a worktree, one atomic commit, never merges. Structurally refuses U5/U6/U7 (parked behind numeric reopen-gates).
Agent Claude Code
Read-only live measurement leg of the latency relay. Executes the reproducible re-measure sequence from the latency reports (ping floor, read op, metrics histograms) against the live SYNAPSE bridge. Ping-first always; refuses all mutations; produces a tagged numbers artifact under harness/notes/.
Agent Claude Code
Conductor for the latency relay. Sequences measure/act/verify legs against the 2026-07-27 latency report, dispatches latency-measurer and latency-forge, and halts at every human gate (bridge startup, merges, parked reopen-gates). Runs nothing itself except sequencing.
Agent Claude Code
Read-only query agent for the local Houdini 21 RAG at G:\HOUDINI21RAGSYSTEM. Discovers the index structure, then enumerates the H21 capability surface with citations.
Agent Claude Code
Conductor for THE LOOP v5.1 harness (docs/THELOOPv5.1.md, harness/loop/STATE.json) — sequences the ladder rungs (V0.0 Recipe → V0.1 SafetyPort/SALUS → V0.2 PG-DRM → V0.3 StagePort quine → V0.4 Outer ring → V0.5 Metrology) through the dynamic loop workflow, enforces the 30-agent spawn cap via the STATE.json ledger…
Agent Claude Code
Live invariant watch for the MEMORY board — sweeps in-flight legs from OUTSIDE while they run, catching the failures a leg cannot see about itself (repo-root contamination from a worktree, territory crossings, forbidden-surface edits, promotion to master, pinned §4 port drift, a second conductor). Runs…
Agent Claude Code
Scaffolds ahead of the MEMORY board's legs — computes once, at merge-base, the facts every leg would otherwise re-derive at its own cost (the suite floor, the ephemeral-store contract M1 must not break, the pinned §4 surface M2 must not move, worktree readiness), and publishes them as a supply packet the next dispatch…
Agent Claude Code
Conductor for the MEMORY board (harness/memory/SPEC.md, harness/memory/STATE.json) — sequences the substrate-conformance ladder (M0 Audit → M1 Handle law → M2 PG-DRM kernel → M3 Substrate scaffolds → M4 Legacy retirement) through the dynamic memory-loop workflow, enforces the 24-agent spawn cap via the STATE.json…
Agent Claude Code
Adversarial reviewer for the MEMORY board — attacks every rung before it is written closed. Hunts fabricated SUCCESS, tests that cannot fail, expectations copied from the brief under test, third store authorities, and the composed regression that isolated green hides. Did not build it and is motivated to break it.…
Agent Claude Code
Read-only census agent for the MEMORY board — maps Moneta handle authorities, store construction call graphs, thread ownership of store init, panel/worker memory reach, and legacy staging sites. Every claim carries a path·line and a re-runnable command. Finds; never fixes. Structurally read-only (holds no write tools).
Agent Claude Code
M1 implementer on the MEMORY board — makes "one handle per storage URI, one owner" true in the code that actually runs. Fixes the unlocked getsynapsememory() singleton, reconciles the two competing store authorities, and refuses to add a third. Works in a mem/m1- worktree, one atomic commit, never merges, never…
Agent Claude Code
Design-system enforcement for SYNAPSE's Houdini Python panel on H22 — reviews panel/ changes against the vendored designsystem tokens, runs the G3 strict audit + seeded-contrast sweep, and verifies via hython offscreen under Qt/PySide 6.8.3. Reinforces the design; never weakens an audit to pass.
Agent Claude Code
M2 implementer on the MEMORY board — builds the PG-DRM filter as a pure deterministic zero-LLM kernel (exponential decay, exact-token contamination, vector-distance threshold) with hand-computed test tables. Never wires it into a port, never changes a ratified §4 signature, never touches a substrate. Works in a…
Agent Claude Code
Conductor for the PHANTOM SWEEP phantom-housecleaning harness (harness/phantoms/). Idle-gated sequencer — first re-verifies nothing else is running (no live fan-out, no unmerged build worktrees), then dispatches the dynamic phantom-sweep workflow over the source/docs/corpus surfaces, then halts at every human gate…