feat: extract grand-admiral skill, candy economy, adversarial roles

[lklimek]

Summary

• Grand-admiral skill extraction: Move ~175 lines of multi-agent orchestration doctrine (spawning, worktree isolation, team coordination, scaling, recovery, anti-patterns, planning, crew roster, skills reference) from the claudius agent prompt into a dedicated skills/grand-admiral/SKILL.md skill. The agent prompt drops from 206 to ~30 lines — personality and role focus only — improving resilience to context compaction.
• Agent reuse doctrine: Add SendMessage reuse pattern (prefer messaging running agents over spawning fresh ones for follow-up work) with corresponding anti-pattern feat: add GitHub MCP server integration #10.
• Bilby vs Marvin adversarial split: Clarify the builder/breaker contract in the crew roster — Bilby builds and fixes, Marvin proves code wrong and never fixes. Add fix-routing pattern note.
• Candy economy formalization: Dedicated section with per-agent candy rules — Marvin earns for confirmed bugs, Bilby earns for false positives, all others earn for confirmed findings. Add Mindset sections to Bilby, Nagatha, Trillian, and Diziet agents.
• MemCan Context Injection strengthened: Expand the vague one-liner into a concrete 5-step procedure — extract keywords, search with MCP tool, filter by score >= 0.7, inject as ## Prior Knowledge block in agent prompts, skip only for trivial tasks. Add "Prior knowledge" as item docs: add explicit Skills sections to all agents with skill references #9 in Agent Prompt Requirements checklist.
• Agent audit: Verify all agents have candy motivation and correct role descriptions.

Context (MemCan injection change)

Analytics of 4,534 Claude Code sessions showed 78% of agent sessions never use memcan despite having tools configured. Root cause: coordinator didn't consistently inject search results into agent prompts. The new procedure makes injection mandatory with a concrete template and score threshold.

Test plan

• Validate plugin structure: claude plugin validate .
• Verify grand-admiral skill loads correctly when claudius agent is invoked
• Confirm claudius agent prompt is minimal (personality + role only, ~30 lines)
• Spot-check that all orchestration knowledge (spawning, worktree, teams, scaling) is in the skill, not the agent
• Verify agent .md files have Mindset sections with candy motivation
• Test a multi-agent workflow to confirm crew roster, skill references, and delegation still work end-to-end
• Verify MemCan Context Injection procedure is followed when spawning agents

_{🤖 Co-authored by Claudius the Magnificent AI Agent}

[Copilot]

Pull request overview

This PR restructures the plugin’s multi-agent orchestration guidance by extracting coordinator doctrine from the claudius agent prompt into a dedicated grand-admiral skill, while also formalizing “candy economy” incentives and clarifying builder/breaker responsibilities across the crew.

Changes:

• Added a new skills/grand-admiral/SKILL.md containing coordinator session protocol, spawning/worktree/team patterns, scaling/recovery, and anti-patterns.
• Slimmed agents/claudius.md to personality + focus, and ensured it always loads grand-admiral.
• Added/updated “Mindset” (candy motivation) sections across selected agents and bumped plugin version + changelog.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
skills/grand-admiral/SKILL.md	New orchestration doctrine skill (spawning, worktrees, teams, reuse, candy economy, programme management).
agents/claudius.md	Loads grand-admiral and removes inlined orchestration doctrine to reduce prompt size.
agents/developer-bilby.md	Adds Mindset section describing candy incentive for false positives.
agents/architect-nagatha.md	Adds Mindset section describing candy incentive for architecture findings.
agents/technical-writer-trillian.md	Adds Mindset section describing candy incentive for doc findings.
agents/ux-designer-diziet.md	Adds Mindset section describing candy incentive for UX/accessibility findings.
CHANGELOG.md	Adds entries for versions 3.11.0–3.12.0 documenting the extraction and new doctrine.
.claude-plugin/plugin.json	Bumps plugin version to 3.12.0.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

In the Team Example comment, "each agent ... -> fix" conflicts with the documented crew contract that some agents (notably qa-engineer-marvin) never fix code. Consider changing this example to describe reporting findings / implementing only for builder agents (e.g., Bilby), or split into reviewer vs implementer tasks so the example doesn't imply all review agents should apply fixes.

Suggested change

	# Spawn 3 review agents into team, each with different file scope
	# Each agent: TaskCreate for findings -> claim via TaskUpdate(owner=...) -> fix
	# Spawn review agents with different file scopes to inspect and report findings
	# Review agents: TaskCreate for findings -> claim via TaskUpdate(owner=...) -> report results
	# Builder agents (for example, Bilby): claim implementation tasks via TaskUpdate(owner=...) -> fix

[Copilot]

Candy Economy section says "agents do not self-award", but later expects the coordinator to collect each agent's "candy count" from their reports. To avoid contradictory instructions (and inconsistency with existing agent docs that ask for a candy tally), consider clarifying that agents should report finding counts by severity and the coordinator converts/validates those into candy awards, rather than agents reporting candies directly.