菜单
Standardizes agent output format with key decisions, artifacts, status, and deferred items. Used when any agent completes a task and needs to report results to the orchestrator.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Evidence-based verification with 2 quality gates (pre-check and post-check), anti-rationalization enforcement, and 3-step retry escalation. Used when completing tasks, verifying implementations, or before claiming work is done.
Routes work through MaxsimCLI commands based on project state and user intent. Provides command reference and decision routing table. Used when determining which MaxsimCLI command to use or when starting a new session.
Coordinates Tier 2 Agent Teams with TeamCreate, SendMessage, competitive implementation, multi-reviewer code review, and collaborative debugging patterns. Used when workflows require inter-agent communication, adversarial debate, or shared task lists.
Autonomous optimization loop with reference workflows. Powers /maxsim:improve, /maxsim:fix-loop, /maxsim:debug-loop, /maxsim:security. Used when running autonomous optimization, error repair, bug hunting, or security audit loops.
GitHub-native persistence for project learnings, decisions, and patterns using issue labels and structured comments. Used when recording what worked, what failed, or architectural decisions that should persist across sessions.
Orchestrates parallel agent execution using worktree isolation following Anthropic's batch pattern. Used when multiple independent tasks can be executed simultaneously.
基于 SOC 职业分类
| name | handoff-contract |
| description | Standardizes agent output format with key decisions, artifacts, status, and deferred items. Used when any agent completes a task and needs to report results to the orchestrator. |
Every agent output includes exactly four sections. The orchestrator reads these sections for state tracking, artifact management, and pipeline decisions. Missing sections break the pipeline.
What was decided during execution and why. Downstream agents depend on this context.
### Key Decisions
- Chose X over Y because [reason with evidence]
- Deferred Z to [phase/plan] because [scope constraint]
- Interpreted ambiguous requirement as [interpretation] because [reasoning]
Include: technology choices, scope adjustments, interpretations of ambiguous requirements, tradeoffs made.
Omit: routine implementation steps, decisions with no downstream impact.
All files created or modified, grouped by action.
### Artifacts
- Created: /absolute/path/to/new-file.ts
- Created: /absolute/path/to/another-file.md
- Modified: /absolute/path/to/existing-file.ts
- Deleted: /absolute/path/to/removed-file.ts
Use absolute paths. Include every file touched — not just primary deliverables. Config changes, test files, and generated files all count.
One of three values with supporting evidence:
| Status | Meaning | Required Evidence | Orchestrator Action |
|---|---|---|---|
PASS | All tasks done, verification passed | Test output, build output, or explicit verification | Advance to next stage |
FAIL | Could not complete; stopped | What failed, what was attempted | Escalate to user |
PARTIAL | Some tasks done; stopped at checkpoint | Which tasks passed, which remain | Resume from checkpoint |
### Status
PASS
Evidence:
- Tests: 47 passed, 0 failed (npm test output)
- Build: exit code 0
- All 4 plan tasks completed
Evidence is not optional. A status claim without evidence is treated as FAIL.
Work discovered but not implemented. Captures scope that would otherwise be lost.
### Deferred Items
- [feature] Add caching layer -- not in current plan scope
- [bug] Race condition in parallel writes -- needs investigation in Phase 5
- [refactor] Extract shared validation logic -- deferred to cleanup phase
- [investigation] Memory growth after 100+ concurrent connections -- repro needed
Categories: feature, bug, refactor, investigation
If none: write ### Deferred Items\nNone
### Key Decisions
- [decision]: [reason]
### Artifacts
- Created: /path/to/file
- Modified: /path/to/file
### Status
PASS | FAIL | PARTIAL
Evidence:
- [verification output or description]
### Deferred Items
- [category] [description] -- [reason deferred]
An orchestrator receiving output without all four sections treats the result as PARTIAL and requests re-submission. Agents cannot mark a task complete without producing a conforming handoff.