hydra

分类: 工具与效率 | 上传者: blueberrycongee | 下载: 0 | 版本: v1.0（最新）
当任务应该在隔离的工作树中运行 Hydra 的文件合同工作流时，或者必须检查、重试或清理现有的 Hydra 工作流时使用。
更新日志: Source: GitHub https://github.com/blueberrycongee/termcanvas
目录结构

当前层级: tree/main/skills/skills/hydra/
📄 SKILL.md 5.2 KB
SKILL.md

---
name: hydra
description: Use when a task should run through Hydra's file-contract workflow in an isolated worktree, or when an existing Hydra workflow must be inspected, retried, or cleaned up.
---

# Hydra Sub-Agent Tool

Use this skill after routing has already determined that Hydra is the right
execution path. Hydra is a strict file-contract workflow engine:
`handoff.json`, `task.md`, `result.json`, and `done` are authoritative.
Terminal conversation is not a source of truth.

## Choose the path

- `hydra run --task "..." --repo . --template single-step`
  - one implementer handoff
  - use for clear implementation work that still needs worktree isolation and
    `result.json` / `done` evidence
- `hydra run --task "..." --repo .`
  - default planner -> implementer -> evaluator workflow
  - use for ambiguous, risky, PRD-driven, or long-running tasks
  - if the user wants one provider for all roles, pass `--all-type <provider>`
  - if the user wants a mix, pass `--planner-type`, `--implementer-type`, and
    `--evaluator-type`
  - if the user does not specify providers, inherit the current terminal type
    when available rather than hard-coding Claude or Codex

## Agent characteristics (soft guidance, not hard rules)

When choosing providers for each role, consider these observed tendencies:

**Claude** — stronger at reasoning, planning, and architectural judgment.
- Good fit for: planner, evaluator
- Watch out for: stub implementations (code that looks complete but doesn't work),
  context-window anxiety (rushing to finish when context fills up)

**Codex** — stronger at code generation, tends to complete the full task.
- Good fit for: implementer
- Watch out for: over-engineering (excessive try-catch, unnecessary boundary checks),
  test hacking (over-mocking, tests that pass without exercising real code)

These are defaults from experience, not constraints. The user can override freely
with `--all-type` or per-role flags. When the user does not specify, a reasonable
default is `--planner-type claude --implementer-type codex --evaluator-type claude`.

## Direct worker primitive

- `hydra spawn --task "..." --repo .`
  - one direct isolated worker terminal
  - use when the split is already known and only a separate worker is needed
  - this is not a full workflow run
  - use `--worker-type <provider>` when the user explicitly names the worker
    provider

## Quality bar

- Root cause first. Fix the real implementation problem before changing tests.
- Do not hack tests, fixtures, snapshots, or mocks to force a green result.
- Do not add silent fallbacks, swallowed errors, or default-success paths.
- Tests prove correctness; they do not replace correctness.
- A handoff only passes when `result.json` and `done` both exist and the schema
  validates.

## Workflow control

1. Investigate first and write a concrete task description.
2. Start the chosen workflow or worker path:
   - Existing worktree / read-only workflow: `hydra run --task "..." --repo . --worktree .`
   - Existing worktree / read-only worker: `hydra spawn --task "..." --repo . --worktree .`
3. After `hydra run` or `hydra spawn`, immediately start polling with `hydra watch`. Do not ask whether to watch — always watch.
4. For workflow runs created by `hydra run`, advance or inspect with:
   - `hydra tick --repo . --workflow <workflowId>`
   - `hydra watch --repo . --workflow <workflowId>`
   - `hydra status --repo . --workflow <workflowId>`
   - `hydra retry --repo . --workflow <workflowId>`
4. For direct workers created by `hydra spawn`, use `hydra list` to inspect and
   `hydra cleanup <agentId>` to clean up.
5. Read failures from structured Hydra state; do not parse terminal prose.
6. Before deciding to keep waiting, retry, or take over a live workflow, query telemetry first:
   - `termcanvas telemetry get --workflow <workflowId> --repo .`
   - `termcanvas telemetry get --terminal <terminalId>`
   - check `last_meaningful_progress_at`, `turn_state`, `foreground_tool`, and contract presence
7. Treat telemetry as advisory truth before completion:
   - `awaiting_contract` means the agent turn ended but `result.json` / `done` is still missing
   - `stall_candidate` means "needs attention", not automatic failure
8. Treat `hydra watch` as the polling loop for the main brain:
   - each poll should prefer telemetry over PTY prose
   - if telemetry shows `thinking`, `tool_running`, `tool_pending`, recent meaningful progress, or a foreground tool, keep waiting
   - if telemetry shows `awaiting_contract`, the model turn is done but the file contract is still pending
   - if telemetry shows `stall_candidate`, inspect recent telemetry events before retry/takeover
9. Clean up after completion:
   - workflow: `hydra cleanup --workflow <workflowId> --repo .`
   - worker: `hydra cleanup <agentId>`

## Result contract

`result.json` must contain:
- `success`
- `summary`
- `outputs[]`
- `evidence[]`
- `next_action`

The `done` marker must point at the exact `result.json` path for the handoff.
If either file is missing or invalid, Hydra must fail the handoff rather than
assuming success.

## Auto-approve inheritance

When you are already running in auto-approve/full-access mode, pass
`--auto-approve` so Claude/Codex sub-agents inherit the same autonomy.
Do not pass it in restricted approval modes.
登录后下载/点赞/收藏 ❤ 71 | ★ 0
hydra

目录结构

SKILL.md

举报内容

提示
