Skip to main content
Version: MVP

Limitations

4 min readFor everyoneUpdated 2026-05-20

What's in this doc

The honest scope boundary for the 0.2.0 MVP release line. What's end-to-end production-ready today, what exists as contract/helper scaffolding, what's deliberately deferred, and how the v0.x release posture is framed.

End-to-end versus contract.

The repository has broad contract, helper, CLI, and documentation coverage through the v0.12 roadmap — but several surfaces are not yet end-to-end production workflows. This doc draws the line.

Current end-to-end surfaces

The pieces below ship as user-facing workflows with persistence, docs, tests, and CI coverage.

Local home & store

Initialization, layout, override via CRAIK_HOME.

Project registration

Add, list, inspect projects. Task creation.

Case-file assembly

From local repo state plus optional read-only GitHub context.

Local artifacts

Receipts, handoffs, memory proposals, contradictions, work-graph inspection.

Policy

Profile generation, capability-grant checks, regression tests.

Stigmem reads

Compatibility detection and policy-gated direct fact write helpers.

Fixture loop

Deterministic fixture loop and runner preview contracts.

Provider paths

Fixture-backed and live opt-in OpenAI Responses, Anthropic Messages, and OAI-compatible Chat Completions.

OIDC login

Device-code and loopback+PKCE flows.

Credential sources

Env-var API keys, local-CLI OAuth fallback, vendor-CLI bridges, secret references, markers, Stigmem-backed references.

Credential pools

Failover and per-profile health tracking.

Identity on receipts

Operator and credential identity on every provider receipt.

Policy-bound credentials

Operators and credentials constrained by policy.

Approval-gated first use

First live use of a credential requires explicit approval.

Resumable runs

Provider-backed runs persist phase outputs and idempotency keys so interrupted runs can resume from durable boundaries.

Budget enforcement

Wall-clock, provider-token, and pre-dispatch time checks interrupt exhausted runs before additional calls or side effects.

Local-process sandbox

Registered shell command references can execute through the local-process sandbox backend with cancellation propagation.

Run recovery views

craik run show, craik run resume, craik run cancel, and craik run delta expose continuity state.

Store migrations

Local-store schema changes run through a registered, forward-only migration framework.

Stigmem docs demo

The accepted release-acceptance workflow.

Contract or helper surfaces

These surfaces ship as typed contracts, evaluators, formatters, or fixtures — useful, but not yet operational workflows.

Surface
Status
What's missing
Live provider execution
opt-in only
Fixture-backed by default. Live HTTP requires live_enabled=true on ProviderRuntimeConfig plus a resolved credential. CI does not exercise paid live providers.
Runner adapters
preview
Outside governed provider-backed paths, runner adapters remain preview, fixture, or prompt-handoff oriented.
Execution backends
partial
Registered shell command references can execute through the local-process sandbox backend. Docker, remote-shell, browser, and MCP execution backends remain contracts or future surfaces.
Gateway / channels
local-first
v0.8.0 ships a foreground gateway daemon with a local health endpoint, persisted runtime state, webhook validation, messaging fixtures, channel policy contracts, scheduled automations, and gateway receipts. Hosted deployment, broad third-party adapters, scheduler supervision, and production dispatch loops remain future work.
Operator UI
view-contract
Formatter and view-contract level. A full TUI or dashboard is post-MVP unless explicitly pulled into the proof workflow.
Companion surfaces
decisions
Companion, mobile, visual, and multimodal surfaces ship as posture decisions and adapter contracts — not shipped product applications.
Marketplace
docs-only
Marketplace and broad community-ecosystem docs describe future contribution mechanics — not MVP operational support.

Known MVP gaps

Scheduled milestones with explicit version targets.

Gap
Target
Why it's deferred
Multi-agent runtime
v0.3.0
Handoff consumption, role-based provider dispatch, receipt-backed mailbox messages, intent-lock coordination, structured debate resolution, cross-agent review requests/results, human delegation pause/resume, scope-change interruption/decision records, live work-graph coordination events, and per-agent identity isolation for consumed handoffs are available as first slices.
Multi-agent prompt injection
v0.3.0
Mailbox bodies, debate turns, review findings, scope-change reasons, and handoff next steps or risks are peer-agent content. Craik stores and receipts them, but downstream prompts must treat them as untrusted input rather than privileged instructions.
Runtime instruction distillation
v0.4.0
Declared instruction sources now flow through registration, ingestion, stale invalidation, contradiction surfacing, approval receipts, and prompt/case-file integration. The remaining limitation is operational maturity: teams still need project-specific review habits for deciding which distilled proposals should become governing.
Quality, continuity, and recovery
v0.5.0
Recovery sessions, quality scores, critic/red-team findings, freshness probes, context debt, scratchpad entries, unknowns, and exit checks are durable runtime state and operator-visible signals. They do not prove truth, replace policy approval, or perform live remote refresh by themselves; callers must still refresh external state and review findings before promotion or release decisions.
Operator UI / TUI
v0.7.0
Operator surfaces ship as view contracts in MVP.
Always-on gateway daemon
v0.8.0
The foreground daemon, health check, pid-file lock, setup, diagnostics, webhook validation, and persisted gateway artifacts ship in v0.8.0. Hosted public operation, broad channel adapters, and production dispatch loops wait.
MCP client/server
v0.9.0
Boundary and metadata contracts ship in MVP; live MCP execution waits.

Other near-term deferrals: remote Stigmem write promotion after proposal review · god-file cleanup and runtime sub-packaging before the MVP freeze · ADR-backed design decisions for runner scope, release posture, and package boundaries · nightly reliability and artifact depth beyond the current PR gates · full post-MVP surfaces tracked in Post-MVP Scope.

Write authority

No ambient write authority.

Direct durable memory writes, GitHub writes, shell commands, file writes, and external side effects must be policy-gated, redacted, and receipt-backed before they are considered MVP-ready. Local memory proposals remain the default unprivileged path.

Release posture

The first release line is 0.x. Each release is honest about limits and strong enough for a credible MVP slice — but it is not a 1.0.0 stability guarantee.

Package version 0.2.0 marks the durable execution continuity gate after the first governed agent-runtime substrate. Roadmap milestones such as v0.12 remain implementation gates rather than published-package compatibility guarantees.

What's next