chore(instructions): reduce bootstrap payload by extracting reusable …

[scappuccino-grid]

…content into skills

Extract reusable content from bootstrap rules into standalone skills loaded via ACQUIRE, reducing bootstrap payload from 556 to ~200 lines (64% reduction).

New skills: hitl, subagent-orchestration, execution-discipline, risk-assessment, prep-steps.
Updated: questioning skill, load-context skill, CLAUDE.md, plugin bootstrap mirrors.

Made-with: Cursor

[github-actions]

Rosetta Triage Review

Summary: This PR refactors the bootstrap instruction system by extracting large inline sections from bootstrap rules into five standalone skill files (hitl, subagent-orchestration, execution-discipline, risk-assessment, prep-steps), reducing the bootstrap payload by 64% (from ~556 to ~200 lines). Agents now load these skills on-demand via ACQUIRE rather than receiving the full content at bootstrap.

Findings:

• Architecture alignment: Change follows Rosetta's progressive disclosure principle correctly — skills are loaded only when needed, keeping context lean.
• Plugin sync: All plugin mirrors (plugins/core-codex/, plugins/core-copilot/) are updated in sync with source instructions, as required by the architecture.
• Skill structure: New SKILL.md files have proper frontmatter, when_to_use_skill sections, and correct tag definitions.
• DRY concern: prep-steps/SKILL.md contains both a <plan_manager_sequence> summary block AND full <prep_step_2>/<prep_step_3> detail blocks, duplicating content within the same file.
• plan_manager dependency: The prep-steps skill requires plan_manager, which is an opt-in feature in Rosetta MCP. Agents deployed without plan_manager enabled will fail to execute the skill as written.
• Increased MCP call frequency: The new skill_check_loop in CLAUDE.md marks hitl and execution-discipline as ALWAYS-triggered skills, meaning every agent response or action will now trigger additional ACQUIRE calls — consider the latency and token cost implications.
• No simulation evidence: The PR description does not include results from a bootstrap flow simulation with the updated instructions, which the triage prompt for this repo identifies as required when instruction files are modified.
• Behavioral change summary: Guardrails and orchestration rules previously inlined in bootstrap are now only accessible if the agent explicitly loads the corresponding skills. If a skill is unavailable (unpublished, network issue), those rules silently disappear from the agent's context.

Suggestions:

• Consider resolving the DRY issue in prep-steps/SKILL.md by replacing the inline <prep_step_2>/<prep_step_3> sections with a reference to the <plan_manager_sequence> only, or vice-versa.
• Add a note or guard in prep-steps/SKILL.md for the case where plan_manager is unavailable (fallback to direct execution of the listed steps).
• Include a brief simulation log or validation result in the PR description showing the bootstrap flow working correctly with the new skill-based approach.

Automated triage by Rosetta agent

[isolomatov-gd]

The task was to extract if-then style skills from bootstrap and reduce duplication. "Execution-discipline" is not a skill, it is not going to be invoked by AI reliably, mostly because it does not provide a need.

Good examples of skills:

1. Deviation (if you deviate or cannot achieve original goal...)
2. Risk assessment (step before execution, we can directly invoke)
3. Self-organization (if file is large, ...)
4. Self-learning and memory (if user is unhappy or you made a mistake, ...)
5. Sensitive data handling (if you see sensitive or even possibly sensitive data, ...)
6. HITL (if not disabled explicitly by user with exactly "No HITL", ....)
7. Subagent/Orchestrator rules (if you are subagent ..., if you are orchestrator ... + some rules are duplicate)
8. Dangerous Actions (if action or consequence ...)

These are good skills => AI knows what to look for AND will engage whenever it is called.