Implementation status¶
Snapshot of the 0.8.0 research beta (tag v0.8.0), an earlier, obsolete iteration of the thesis;
the thesis evaluates 2.0.1. It is kept as that cut's record and is not updated with later releases.
Note
The current version is 2.0.1. Several rows below have changed since 0.8.0: built-in adapters
for the Claude Code, Codex, Gemini CLI and aider CLIs (agentProviders.<id>.kind), an MCP
server for embedded mode (harness mcp serve), token authentication and roles in the local API
(#18), retention applied by harness gc, the closed thesis-impact defects (#3–#8) and 80 CLI
commands. See the README, the
CLI reference, the
changelog and the
trust model.
Complete in this repository¶
| Area | Status | Notes |
|---|---|---|
| Domain model | Complete for current MVP | Typed entities and normalized states for task, run, phases, changes, evidence, findings, decisions, invocations, memory and retrospective |
| Normative workflow | Complete | Nine required phases plus separate retrospective |
| State machine | Complete | Strict order, correction path, invalidation, cancellation and resumption |
| Configuration snapshots | Complete | Project/profile/workflow/policy digests stored per execution |
| Event store | Complete local implementation | SQLite WAL, append-only API, hash chain, JSONL export |
| Artifact store | Complete local implementation | Content addressed, atomic, redacted and digest verified |
| ChangeSet | Complete for structured tasks | Baseline, task-owned paths, deterministic digest, unrelated-change exclusion |
| Capabilities | Complete application-level model | Actor, scope, expiration, deny-by-default and command authorization |
| Process runner | Complete local adapter | No shell, timeout, cancellation, output limits; not a strong sandbox |
| Agent provider 1 | Complete | Deterministic simulated provider |
| Agent provider 2 | Complete generic adapter | Structured external command/JSON provider |
| Validators | Complete foundational set | Generic command validator plus independent deterministic review |
| Technology profiles | Complete initial proof | Python and Node.js |
| Gates | Complete | Fail closed, typed severities, mandatory-check semantics |
| Human decisions | Complete | Approve, reject, changes requested and explicit exception; exact digest binding |
| Revalidation | Complete | Manual correction + continue; stale approvals rejected |
| Memory | Complete deterministic MVP | Five levels, provenance, approval, expiry and scoped selection |
| Telemetry | Complete basic projection | Timings, invocations, retries, validation cycles, findings, changes; no invented tokens/cost |
| Retrospective | Complete deterministic MVP | Evidence-based observations and recommendations; never auto-applied |
| CLI | Complete MVP | Initialization, tasks, runs, decisions, status, evidence, findings, trace, doctor, benchmarks and API |
| API/web | Complete local MVP | Read status/trace/findings/evidence/retrospective and issue digest-bound decisions |
| Reporting | Complete MVP | Markdown, JSON, JSONL and SARIF |
| Plugin protocol | Complete v1 prototype | One-shot external JSON request/response, client validation and SDK echo server |
| Packaging | Complete | Wheel, source distribution, Dockerfile and CI definition |
| Tests | Complete for declared beta scope | Unit, contract, integration, security, E2E and performance smoke tests |
| Benchmarks | Complete reproducible baseline | Microbenchmarks and direct-vs-governed synthetic Python/Node scenarios |
Partial or intentionally constrained¶
| Area | Current implementation | Remaining work for production |
|---|---|---|
| Sandboxing | Application-level capabilities, path checks, timeout/cancellation; write confinement of agent providers (runtime.agentSandbox: sandbox-exec on macOS, bwrap on Linux) |
Container/VM/OS sandbox for validators and plugins; read, network, CPU/memory/PID controls |
| Git | Read-only baseline metadata and workspace diff | Stage/commit/push adapters, ownership by hunk and signed attestations |
| Coverage | Can run profile commands; whole-project tools can be configured | Normalized diff-coverage adapter per stack |
| SAST/SCA | External command/plugin contract supports them | Built-in parsers and curated default policies |
| DAST | Representable as validator/plugin | Lifecycle, target isolation, probes and generic normalization |
| Real model providers | Generic command adapter | Provider-specific session, resume, event and token APIs |
| Usage/cost | Records provider-reported values when available | Provider-specific collection and pricing snapshots |
| Multi-unit projects | Domain/configuration structures exist | Impact graph and parallel scheduler across units |
| Plugins | Protocol and capability-aware launcher | Persistent lifecycle, registry, signature verification and isolation |
| Web | Functional local dashboard | Authentication, multiuser authorization, accessibility hardening and richer evidence views |
| Retention | Configuration fields and local stores | Automated retention, encryption and legal hold |
| Replay | Trace and snapshots are available | Environment/plugin lock reconstruction and comparative replay command |
| Attestation | Digests and hash chain | External signature/OIDC/in-toto compatible attestation |
Explicitly deferred¶
- GraphRAG and semantic graph memory.
- Automatic workflow generation.
- Automatic changes to phases, gates, permissions, prompts or normative memory.
- Distributed execution and remote workers.
- Multi-database or multi-tenant platform.
- Plugin marketplace.
- Generic production DAST.
- Broad language catalog beyond Python and Node.js.
- Production MCP server lifecycle and transport.
These items are deferred because they are not necessary to demonstrate the core research hypothesis and would materially increase security and evaluation complexity.