Known Limitations¶
Aggregated directly from the "MVP Limitations" sections of each architecture document — nothing below is invented for this page. Follow the links for full context.
Execution¶
- Execution is synchronous and single-process; the model field for "parallel" execution exists conceptually in the stage list but has no parallel implementation.
- No retry/backoff policy is implemented beyond the
retry_countfield existing onAgentRunResult; retries default to zero and are not yet exposed via CLI. - There is no persisted "project-level default risk policy" — risk level is set per mission at creation time.
(Mission Orchestrator, Adapter Architecture)
Roles¶
- The 27 built-in roles are fixed at this MVP stage; there is no CLI
command yet to add or edit a custom role (
assync role list/show/validateare read-only). default_adapteris currently informational only — the orchestrator requires an explicit run-config entry per selected role.
Authorization¶
- There is no authorization renewal flow yet — an expired authorization must be re-issued from scratch (via a new human decision).
- Scoping authorization to a specific
target_run_idis modeled but not yet enforced beyondtarget_mission_idmatching.
Decisions¶
- There is no mission-side automatic re-evaluation after a decision
resolves — a rejected risk-acceptance decision requires a separate,
explicit
assync mission abortcall, rather than an automatic mission-state transition. Approving a decision does not, by itself, move the mission out ofWAITING_FOR_AUTHORIZATIONeither — see the Quick Start. required_authorityis currently a free-text field (default"governance_authority"), not yet cross-checked against the role registry's non-executable Human Governance Authority entry.
Synthesis¶
- Evidence-classification precedence is currently applied per-finding,
not used to resolve two findings that directly contradict each other
with different evidence strength — synthesis currently surfaces this
via
conflictsrather than auto-resolving it (see thecontradictionmock scenario).
Storage¶
- No schema-migration tooling exists yet; a
schema_versionbump would currently require a manual, out-of-band migration. - The cross-project registry store and a project's own runtime store are
both
FilesystemStoreinstances but are never the same instance — there is no automated check today that a registry root and a project's own runtime root can never accidentally coincide, beyond the nested-path rejection inprojects/service.py::register_project.
Security¶
- No production credentials, autonomous production actions, or release execution exist anywhere in this codebase — by omission, not by a runtime guard that could be bypassed.
- The credential-redaction regex in the local-command adapter is a defense-in-depth measure, not a guarantee — callers must still avoid passing secrets into adapter input in the first place.
- No sandboxing (containers, seccomp, etc.) is applied to the local command adapter beyond the executable allowlist, timeout, and output-cap controls.
- The local command adapter's subcommand denylist currently only covers
git; a real provider-CLI integration would need its own equivalent review before being added.
(Security Model, Adapter Architecture)
Not implemented at all¶
These are documented extension points, not partially-built features — implementing any of them was explicitly out of scope for this MVP:
- Remote provider adapters (Claude Code, ChatGPT/OpenAI, Antigravity, local model) — only the deterministic mock adapter and the restricted local-command adapter exist today.
- A
ReviewPublisher(GitHub Checks / draft PR publisher) — no GitHub API is called anywhere in this codebase. - A webhook receiver, web application API, worker queue, or containerized execution.
- A published PyPI package.