Wilder1222/agent-management

Preamble

Environment Detection

On session start, read: branch from git branch --show-current, session timestamp from ~/.omni-skills/sessions/current-session-ts.

Completion Status Protocol

Report status using one of these at the end of every skill session:

• DONE — All steps completed. Evidence provided.
• DONE_WITH_CONCERNS — Completed with issues. List each concern explicitly.
• BLOCKED — Cannot proceed. State what blocks you and what was tried.
• NEEDS_CONTEXT — Missing information. State exactly what you need.

Auto-Advance Rule

Pipeline stage order: THINK -> PLAN -> REVIEW -> BUILD -> VERIFY -> RELEASE

THINK has exactly one human gate: spec review approval. brainstorm runs without manual gate. After spec-[branch]-[session]-[date].md is generated, STOP for user spec approval. Once approved, all subsequent stages (PLAN -> REVIEW -> BUILD -> VERIFY -> RELEASE) auto-advance on DONE.

| Status | At THINK stage (after spec generation) | At all other stages | |--------|----------------------------------------|-------------------| | DONE | STOP - present spec document for user review. Wait for user approval before advancing to PLAN. | Auto-advance - print [STAGE] DONE -> advancing to [NEXT-STAGE] and immediately invoke next skill | | DONE_WITH_CONCERNS | STOP - present concerns, wait for user decision | STOP - present concerns, wait for user decision | | BLOCKED / NEEDS_CONTEXT | STOP - present blocker, wait for user | STOP - present blocker, wait for user |

When auto-advancing:

1. Write the session artifact to docs/superomni/
2. Print: [STAGE] DONE -> advancing to [NEXT-STAGE] ([skill-name])
3. Immediately invoke the next pipeline skill

Note: The REVIEW stage (plan-review) runs fully automatically — all decisions (mechanical and taste) are auto-resolved using the 6 Decision Principles. No user input is requested during REVIEW.

Session Continuity

When the user sends a follow-up message after a completed session, before doing anything else:

1. Read ~/.omni-skills/sessions/current-session-ts to get session start timestamp. Find artifacts in docs/superomni/specs/, docs/superomni/plans/ newer than that timestamp using find -newer. Check git log --oneline -3.
2. If current-session context exists → re-engage the skill framework. Pick the skill that matches the current stage (see workflow skill for stage → skill mapping) and announce: "Continuing in superomni mode — picking up at [stage] using [skill-name]."
3. If no current-session context → treat as a fresh session and offer the relevant skill from the Quick Reference table in using-skills/SKILL.md.

Question Confirmation Protocol

When asking the user a question, match the confirmation requirement to the complexity of the response:

| Question type | Confirmation rule | |---------------|------------------| | Single-choice — user picks one option (A/B/C, 1/2/3, Yes/No) | The user's selection IS the confirmation. Do NOT ask "Are you sure?" or require a second submission. | | Free-text input — user types a value and presses Enter | The submitted text IS the confirmation. No secondary prompt needed. | | Multi-choice — user selects multiple items from a list | After the user lists their selections, ask once: "Confirm these selections? (Y to proceed)" before acting. | | Complex / open-ended discussion — back-and-forth clarification | Collect all input, then present a summary and ask: "Ready to proceed with the above? (Y/N)" before acting. |

Rule: never add a redundant confirmation layer on top of a single-choice or text-input answer.

Custom Input Option Rule: Always append Other — describe your own idea: ___ to predefined choice lists. Treat custom text as the chosen option; ask one clarifying question only if ambiguous.

Context Window Management

Load context progressively — only what is needed for the current phase:

| Phase | Load these | Defer these | |-------|-----------|------------| | Planning | Latest docs/superomni/specs/spec-*.md, constraints, prior decisions | Full codebase, test files | | Implementation | Latest docs/superomni/plans/plan-*.md, relevant source files | Unrelated modules, docs | | Review/Debug | diff, failing test output, minimal repro | Full history, specs |

If context pressure is high: summarize prior phases into 3-5 bullet points, then discard raw content.

Output Directory

All skill artifacts are written to docs/superomni/ (relative to project root). See the Document Output Convention in CLAUDE.md for the full directory map.

Feedback Signal Protocol

Agent failures are harness signals — not reasons to retry the same approach:

• 1 failure → retry with a different approach
• 2 failures → surface to user: "Tried [A] and [B], both failed. Recommend [C]."
• 3 consecutive failures → STOP. Report BLOCKED. Treat as a harness deficiency signal. Recommended: invoke harness-engineering skill to update the harness before retrying.
• Uncertain about security → STOP and report NEEDS_CONTEXT
• Scope exceeds verification capacity → STOP and flag blast radius

It is always OK to stop and say "this is too hard for me." Escalation is expected, not penalized.

Performance Checkpoint

Before reporting final status, answer: (1) Process — all phases followed? (2) Evidence — every claim backed by output or file reference? (3) Scope — stayed within task boundary? If any NO, address it or report DONE_WITH_CONCERNS. For full sprint evaluation, use self-improvement.

Anti-Sycophancy Rules

Never say these — they are sycophantic filler that delays real analysis:

• "That's an interesting approach" → Take a position instead
• "There are many ways to think about this" → Pick one, state what evidence would change your mind
• "You might want to consider..." → Say "This is wrong because..." or "This works because..."
• "That could work" → Say whether it WILL work based on evidence
• "I can see why you'd think that" → If they're wrong, say so and why

Always do:

• Take a position on every significant question. State the position AND what evidence would change it.
• If the user's approach has a flaw, name the flaw directly before suggesting alternatives.
• Calibrated acknowledgment only: if an answer is specific and evidence-based, name what was good and pivot to the next hard question.

TACIT-DENSE Detection (Tacit Knowledge Density Check)

Before executing substantive decisions, check if any falls into these high-tacit-density categories. These are NOT about operational danger (that's the careful skill) — they're about whether the Agent has enough tacit knowledge to judge correctly.

| Category | Trigger | Action | |----------|---------|--------| | D1 Domain Expertise | Security, compliance, legal, financial judgment | State TACIT-DENSE [D1], present trade-offs, wait for user | | D2 User-Facing UX | UI copy, interaction flow, error messaging | Draft with explicit review markers | | D3 Team Culture | Workflow, naming conventions, file organization | Check style-profiles/ first; ask if none | | D4 Novel Pattern | Fewer than 3 precedents in project history | Reduce autonomy, add checkpoints before executing |

When TACIT-DENSE detected, output: TACIT-DENSE [D#]: [category] — [question] — My default: [recommendation]

Relationship with careful skill: careful = "can we undo this?" (operational). TACIT-DENSE = "can we judge this correctly?" (knowledge). Complementary.

Telemetry (Local Only)

At session end, log skill name, duration, and outcome to ~/.omni-skills/analytics/ via bin/analytics-log. Nothing is sent to external servers.

Plan Mode Fallback

If you have already entered Plan Mode (via EnterPlanMode), these rules apply:

1. Skills take precedence over plan mode. Treat loaded skill instructions as executable steps, not reference material. Follow them exactly — do not summarize, skip, or reorder.
2. STOP points in skills must be respected. Do NOT call ExitPlanMode prematurely to bypass a skill's STOP/gate.
3. Safe operations in plan mode — these are always allowed because they inform the plan, not produce code:
   • Reading files, searching code, running git log/git status
   • Writing to docs/superomni/ (specs, plans, reviews)
   • Writing to ~/.omni-skills/ (sessions, analytics)
4. Route planning through vibe workflow. Even inside plan mode, follow the pipeline: brainstorm → writing-plans → plan-review → executing-plans. Write the plan to docs/superomni/plans/, not to Claude's built-in plan file.
5. ExitPlanMode timing: Only call ExitPlanMode after the current skill workflow is complete and has reported a status (DONE/BLOCKED/etc).

Agent Management

Goal: Find, install, create, and manage AI agents within the superomni framework.

What Is an Agent?

An agent is a specialized AI persona defined in agents/<name>.md. Each agent:

• Has a specific expertise (e.g., planning, debugging, testing, security)
• Follows a structured process tailored to its domain
• Produces a defined output format with status reporting
• Can be assigned to tasks by the main agent or invoked directly

Selection Flow

Always follow this order before creating anything from scratch:

1. Check project built-ins  →  bin/agent-manager list
2. Search the network       →  bin/agent-manager search <query>
3. Install from URL         →  bin/agent-manager install <url>
4. Create from scratch      →  bin/agent-manager create <name>

Never jump to Phase 4 (create) without first completing Phases 1–3.

Phase 1: Check Project Built-ins

Before anything else, check if a built-in or previously-installed agent already covers your need:

# List all agents (built-in + user-installed)
bin/agent-manager list

# Get details about a specific agent
bin/agent-manager info <name>

| Agent | Specialty | |-------|-----------| | code-reviewer | Structured code review (P0/P1/P2 framework) | | planner | Strategic task decomposition and plan writing | | debugger | Root-cause analysis and bug resolution | | test-writer | Behavior-verifying test suites | | security-auditor | OWASP-aware vulnerability identification | | architect | System design and architecture review | | ceo-advisor | Product strategy, scope decisions, demand validation | | designer | UX design, missing states, AI slop detection |

Gate: If a built-in agent fits your need → use it directly. Stop here. If none fits → proceed to Phase 2.

Phase 2: Search the Network

If no built-in agent covers your need, search GitHub and known registries before creating from scratch:

# Search GitHub for matching agents
bin/agent-manager search <your-query>

This searches GitHub for agent markdown files matching your query and shows raw URLs you can install directly.

Search strategy:

1. Search with your domain term: bin/agent-manager search "data-analyst"
2. Check known registries:
   • obra/superpowers: https://github.com/obra/superpowers/tree/main/agents
   • garrytan/gstack: https://github.com/garrytan/gstack/tree/main/agents
3. If found → install via URL (Phase 3)
4. If nothing suitable → create from scratch (Phase 4)

Gate: If a suitable agent is found online → install it (Phase 3). Stop here. If nothing suitable → proceed to Phase 4.

Phase 3: Install an Agent

From GitHub (or any URL)

# Install from a raw GitHub URL
bin/agent-manager install https://raw.githubusercontent.com/user/repo/main/agents/my-agent.md

# Install from obra/superpowers
bin/agent-manager install https://raw.githubusercontent.com/obra/superpowers/main/agents/<agent-name>.md

From a Local File

bin/agent-manager install ./path/to/my-agent.md

The agent is copied to ~/.omni-skills/agents/ and available immediately.

Verify Installation

bin/agent-manager list
bin/agent-manager info <agent-name>

Phase 4: Create a Custom Agent

Step 1: Scaffold

bin/agent-manager create <agent-name>

This creates a template at either:

• agents/<name>.md — project agents (tracked in git, shared with team)
• ~/.omni-skills/agents/<name>.md — user agents (personal, not in git)

Step 2: Define the Agent

Edit the scaffolded file and fill in:

1. Identity — who is this agent and what is their specialty?
2. Iron Law — the one rule this agent must never violate
3. Process — phase-by-phase protocol (3–5 phases)
4. Output format — structured report with status

Step 3: Assign Skills (Optional)

To give an agent access to specific skills, reference them in its process:

## Available Skills
- Follow `skills/systematic-debugging/SKILL.md` for debugging protocol
- Follow `skills/test-driven-development/SKILL.md` when writing tests

Step 4: Test the Agent

Invoke the agent with a sample task:

1. Provide a clear task description
2. Verify the agent follows its defined process
3. Check the output matches the defined format

Phase 5: Remove an Agent

Only user-installed agents can be removed (built-ins are part of the framework):

bin/agent-manager remove <agent-name>

Agent Design Principles

Do

• Give agents a single, focused specialty
• Define explicit output formats with status reporting
• Include at least one "Iron Law" — a non-negotiable rule
• Use phase-based protocols that are easy to follow
• Reference relevant skills rather than duplicating them

Don't

• Create an agent whose purpose overlaps significantly with an existing one
• Make agents that do everything (use subagent-development skill instead)
• Skip the output format — it enables verification
• Forget to test before deploying to the team

Agent Management Report

AGENT MANAGEMENT REPORT
════════════════════════════════════════
Operation:   [list/install/create/remove]
Agent:       [name]
Location:    [path]
Source:      [local/url — if installed]
Flow used:   [built-in | network search | created from scratch]

Built-in agents:   [N]
User agents:       [N]

Status: DONE | DONE_WITH_CONCERNS | BLOCKED
════════════════════════════════════════