v0.87.0
Scramjet 0.87.0 adds focused command-authoring support built around commands as lightweight generalized plans executed by capable agents.
Highlights
- Six read-only command specialists: Add focused agents for command-set exploration, architecture, instruction semantics, concrete failure analysis, holistic review, and independent finding assessment.
- Shared command-writing guidance: Package the progressively loaded
writing-scramjet-commandsskill covering outcomes, authority, handoffs, informed user alignment, approval-only coaching exceptions, context economy, and evidence from actual use. - Deterministic command linting: Add
scramjet-command-lintfor runtime-derived recognition and collision diagnostics plus structural Goals and context-framing checks. It does not establish semantic quality or operational effectiveness. - Mach 12 command-work routing: Use Scramjet specialists for command work, retain code specialists for runtime work, partition mixed briefs, and keep mutation, project-tool execution, interaction, and publication with the parent agent.
- Project-native checks and contextual continuation: Teach Mach 12 to discover repository-authorized development tools and recommend next steps from completed-work context rather than fixed array positions.
- Informed alignment: Require consequential user asks to include relevant evidence, alternatives, consequences, uncertainty, and a recommendation without ceremonial confirmation.
Known limitations
- Real-world effectiveness of the final six-agent architecture remains unproven; structural and synthetic checks do not establish product value.
- Project-native tool discovery and transport were not exercised through a final production Mach 12 command path.
- A live command-only planning run showed that the one-architect policy can collapse materially different alternatives too early; that limitation remained unresolved at release.
Closes #474 through the six-agent redesign.
Reviewed by GPT-5.6 Sol