AgentCrew v0.2.0

AgentCrew v0.2.0 expands the conversation-first product team workflow. Describe an idea or outcome; AgentCrew selects and coordinates the roles, while you retain consequential product, risk, cost, release, and merge decisions.

Highlights

• CEO-led discovery and planning, executive briefs, private preferences, and explicit execution approval.
• Software Architect Agent, architecture decisions, and conformance review routing.
• Evidence-backed development, testing, review, and bounded specialist rework.
• Continuations that preserve required specialists and the actual pending approval phase.
• Provider-neutral adapters, a generic host loader, and project-local state with containment checks.
• Hardened command execution and Git inspection, protocol validation, and budget controls for paid execution and advice.
• A repeatable offline agentcrew demo, plus a terminal GIF and transcript in the README.

Setup And Compatibility

• Direct engine runs require --backend or a backend in .agentcrew/config.yaml.
• Engine command execution requires Linux, Bubblewrap, and working user namespaces. The host-agent Markdown workflow remains portable.
• Product approval does not imply approval of a new cost preview. Hosts must obtain the relevant approval before supplying --approve-decision or --approve-cost.
• Cached, unsupported, or incompletely reported usage remains unresolved rather than appearing free. A configured daily cap cannot be verified until that uncertainty is resolved or explicitly accepted.
• Free advice creates no workflow state; paid engine advice writes a minimal billing ledger.

Validation

• 580 non-live tests passed locally; 2 live-provider tests were excluded.
• The offline demo reaches ready_for_human_approval with zero model API calls.
• Python compilation, shell syntax, dependency consistency, and diff checks passed.

See the changelog and engine setup for details.