The build-out runs in three gates. One gate = one named milestone: it ships when its scenarios pass and their evidence posts (ADR-0002). The live board carries the same set of cards: agentpcap — Roadmap & Build Board.
Current phase
AG-0.1 build-out. The specification is in force and renders in CI; the implementation starts with the workspace skeleton and the stdio class.
Release gates
| AG-0.1 |
one lab host, one stdio MCP client and server |
ledger chain verifies after rotation; no path writes; zero arguments stored |
| AG-0.2 |
the fleet plus remote Streamable HTTP and one attributed-mode host |
M3 coverage matrix; M4 capability drift |
| AG-1.0 |
full fleet with seeded shadow deployments |
M1 recall, M2 overhead delta, M3 full matrix, M4 fidelity, M5 triage precision, M6 handoff latency — gate rule: M1 passing with M3 at M2 cost |
AG-0.1 — MVP: the stdio class, end to end
| #10 |
feat: cargo workspace and dependency skeleton |
build,mvp |
| #11 |
feat(tap): pipe collector frames stdio JSON-RPC in both directions |
tap,mvp |
| #12 |
feat(ledger): append-only JSONL writer with prev-hash chain and rotation |
ledger,mvp |
| #13 |
feat(cli): record and report subcommands |
cli,mvp |
| #14 |
test(ledger): conformance fixtures for chain, rotation, tamper, access log |
ledger,mvp,test |
| #31 |
release: AG-0.1 gate |
mvp,release-gate |
AG-0.2 — The wire class and both modes
| #15 |
feat(tap): eBPF socket collector with cgroup and lineage filters |
tap |
| #16 |
feat(tap): TLS-boundary uprobe collector for attributed mode |
tap |
| #17 |
feat(reassemble): MCP session reassembly — init, tools-list, tools-call, notify |
reassemble |
| #18 |
feat(classify): unidentified-flow events with evidence features |
classify |
| #19 |
feat(cli): diff subcommand for capability drift |
cli |
| #20 |
feat(cli): export subcommand for the gate stream |
cli |
| #21 |
feat(deploy): systemd unit, launchd template, bounded capability set |
deploy |
| #22 |
feat(deploy): per-node DaemonSet manifest with node-scope ledger |
deploy |
| #23 |
feat: coexistence export consumption with the dedupe key |
ledger |
AG-1.0 — The evaluation gate
| #24 |
feat(eval): Python harness scaffold over the CLI and export surfaces |
eval |
| #25 |
perf(eval): M2 tap overhead — tap-on versus tap-off delta against the sub-3% figure |
eval,performance |
| #26 |
perf(eval): M1 shadow-discovery recall on the fixture fleet |
eval,performance,release-gate |
| #27 |
perf(eval): M3 transport-coverage matrix run |
eval,release-gate |
| #28 |
perf(eval): M4 ledger fidelity on capability drift |
eval,performance |
| #29 |
perf(eval): M5 unidentified-flow triage precision |
eval,performance |
| #30 |
perf(eval): M6 gate handoff latency against the 2-second budget |
eval,performance |
| #32 |
release: AG-1.0 gate |
release-gate |
The measurement set
The six measurements map to the evaluation issues above: M1 → #26, M2 → #25, M3 → #27, M4 → #28, M5 → #29, M6 → #30. Baselines come from the published figures of the research note.