Agent Integration — Full Pipeline

Run all three phases of agent integration in a single session. Parameters are collected once and reused across all phases.

Parameters

Collect these before starting (ask the user if not provided):

Parameter	Description	How to derive
AGENT_NAME	Human-readable name (e.g., "Gemini CLI")	User provides
AGENT_PACKAGE	Go package dir name — no hyphens	Lowercase, remove hyphens/spaces
AGENT_KEY	Registry key for agent.Register() and entire enable	Check existing patterns in cmd/entire/cli/agent/registry.go
AGENT_SLUG	Filesystem/URL-safe slug (kebab-case) used in E2E runner filenames and script names	Kebab-case of agent name; align with existing entries in e2e/agents/
AGENT_BIN	CLI binary name	command -v <binary>
LIVE_COMMAND	Full command to launch agent	User provides
EVENTS_OR_UNKNOWN	Known hook event names, or "unknown"	From agent docs or "unknown"

Note: These identifiers can differ. Run grep -r 'AgentName\|func.*Name()' cmd/entire/cli/agent/*/ and e2e/agents/ to see how existing agents handle the split.

Architecture References

These documents define the agent integration contract:

• Implementation guide: docs/architecture/agent-guide.md — Step-by-step code templates, event mapping, testing patterns
• Integration checklist: docs/architecture/agent-integration-checklist.md — Design principles and validation criteria

Scope

This skill targets hook-capable agents — those that support lifecycle hooks (implementing HookSupport from agent.go). Agents that use file-based detection (implementing FileWatcher) require a different integration approach not covered here. Check agent.go for the current interface definitions.

Core Rule: E2E-First TDD

This skill enforces strict E2E-first test-driven development. The rules:

1. E2E tests are the spec. The existing ForEachAgent test scenarios define what "working" means. The agent runner makes those tests runnable for the new agent.
2. Run E2E tests at every step. Each implementation tier starts by running the E2E test and watching it fail. You implement until it passes. No exceptions.
3. Unit tests are written last. After all E2E tiers pass (Step 14), you write unit tests using real data collected from E2E runs as golden fixtures.
4. If you didn't watch it fail, you don't know if it tests the right thing. Never write a test you haven't seen fail first.
5. Minimum viable fix. At each E2E failure, implement only the code needed to fix that failure. Don't anticipate future tiers.
6. /e2e:debug is your debugger. When an E2E test fails, use the artifact directory with /e2e:debug before guessing at fixes.

Pipeline

Run these three phases in order. Each phase builds on the previous phase's output.

Phase 1: Research

Discover the agent's hook mechanism, transcript format, and configuration through binary probing and documentation research. Produces an implementation one-pager at cmd/entire/cli/agent/$AGENT_PACKAGE/AGENT.md that the other phases use as their single source of agent-specific information.

Read and follow the research procedure from .claude/skills/agent-integration/researcher.md.

Expected output: Implementation one-pager at cmd/entire/cli/agent/$AGENT_PACKAGE/AGENT.md and a test script at scripts/test-$AGENT_SLUG-agent-integration.sh.

Commit: After the research phase completes, use /commit to commit all files.

Gate: If the verdict is INCOMPATIBLE, stop and discuss with the user before proceeding.

Phase 2: Write E2E Runner

Create the E2E agent runner so existing test scenarios can exercise the new agent. No unit tests are written in this phase — no new test scenarios either (existing ForEachAgent tests are the spec).

Read and follow the procedure from .claude/skills/agent-integration/test-writer.md.

Expected output: E2E agent runner at e2e/agents/$AGENT_SLUG.go that compiles and registers with the test framework.

Commit: After the E2E runner compiles and registers, use /commit to commit all files.

Phase 3: Implement (E2E-First, Unit Tests Last)

Build the Go agent package using strict E2E-first TDD. E2E tests drive development at every step — run each tier, watch it fail, implement the minimum fix, repeat. Unit tests are written only after all E2E tiers pass, using real data from E2E runs as golden fixtures.

Read and follow the implement procedure from .claude/skills/agent-integration/implementer.md.

Expected output: Complete agent package at cmd/entire/cli/agent/$AGENT_PACKAGE/ with all E2E tiers passing and unit tests locking in behavior.

Note: AGENT.md is a living document — Phases 2 and 3 update it when they discover new information during testing or implementation.

Final Validation

After all three phases, run the complete validation:

mise run fmt      # Format
mise run lint     # Lint
mise run test:ci  # All tests (unit + integration)

Summarize:

• Compatibility verdict from Phase 1
• Files created in Phases 2 and 3
• Test coverage
• Any remaining TODOs or gaps