start
Recover a HADES session: read .hades/session.md TL;DR, check git status, identify active plan, present concise session-resume summary, and await operator direction.
- License 许可
- Apache-2.0
- 最近更新
- 2026-05-29
- Stars
- ★ 3
- Forks
- 0
01 前言 Front Matter
- name
- start
- license
- Apache-2.0
- version
- 0.12.0
- keywords
- [ "hades", "session-resume", "handoff", "recovery", "context-restoration" ]
- description
- Recover a HADES session: read .hades/session.md TL;DR, check git status, identify active plan, present concise session-resume summary, and await operator direction.
- agentskills_version
- 1
03 SKILL.md 正文
HADES — Session start / recovery skill
Use this skill at the beginning of a Hermes session in a HADES project,
or after a /clear that wiped conversation context. Synthesizes the
session-resume documents into a concise restart point.
Operator entry point: invoke /hades:start (slash command registered
by the plugin). The slash command handler (in
plugin/hades/commands/start.py) executes the procedure below.
Note: Steps 1, 3, 4 are mechanized in the slash command handler (
commands/start.py); steps 2 (project memory) + 5 (5-line summary presentation) + 6 (await direction) apply to the LLM consuming this skill in conversation context.
Inputs
None — this skill reads project files autonomously.
Steps
-
Read .hades/session.md (root of repo). Extract:
## TL;DR— current state in 1-3 sentences## Repo state— last commit + branch + tag (if any)## Active plan status— which plan/stage is in flight## Pending dispatches— background subagents not recoverable## Pending operator actions— items awaiting operator## Suggested first-message— primer for operator response
-
Read project memory at:
local agent memory. Confirm methodology + doctrine memory entries present. -
Check git state:
git status— uncommitted changesgit log --oneline -10— recent commitsgit tag --list 'v*' --sort=-creatordate | head -5— recent tags
-
Check daemon state (if applicable):
pgrep -f hades-ctld— daemon running- If running:
curl --unix-socket /tmp/hades-system.sock http://localhost/v1/health
-
Present 5-line session summary to operator:
## HADES session resume - **State**: [from TL;DR] - **Repo**: branch <name>, last commit <hash> "<subject>" - **Active plan**: release item stage — <status> - **Pending operator actions**: <list or "none"> - **Suggested next**: <from .hades/session.md "Suggested first-message"> ¿procedo con <suggested next>, o cambiamos prioridad? -
Await operator direction.
Doctrine reminders (apply to ALL responses post-/start)
- Max-scope always — pick most complete solution
- Build the final product, not the stages — refactor toward final shape
- No stubs — production code complete day 1
- No defer, no tech debt — fix coverage gaps + missing tests before next task
- No Claude attribution in commits (invariant; gated by pre_tool_call callback)
- Tag safety gate: NEVER push tags without operator approval
Edge cases
- .hades/session.md missing: instruct operator to invoke
/hades:handoffin a prior session, OR proceed withgit status + logsummary alone - Methodology memory missing or stale (>14d): surface to operator,
recommend re-reading
docs/METHODOLOGY.md - Daemon not running: state explicitly; some operations (caronte code-graph,
/v1/events) unavailable until
hades-ctldstarted
See also
/hades:handoff— sister skill: snapshot state to .hades/session.md + commitlocal agent memory- Project project instructions (this skill replaces the CC-specific
local agent configunder Hermes substrate)
04 同仓库更多 Skills
handoff
Snapshot HADES session state to .hades/session.md, commit with conventional message (no AI attribution), prepare for /clear or session end. Sister skill to /hades:start.
hades
Canonical HADES system project skill: doctrine reminders (max-scope, no defer, no tech debt, no AI-attribution in commits), workflow patterns (brainstorm → spec → write-plan → execute-plan → release), hard rules, and pointers to project memory.
agente-ejecutor
暂无简介
research-cheap
暂无简介