菜单
Resume session from CONTEXT-llm.md. Use when resuming work, loading saved context, continuing a previous session. Triggers include "load context", "resume session", "continue where I left off".
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。
基于 SOC 职业分类
Diagnose an agent FAILURE and prescribe the cheapest/earliest guardrail to add NEXT — then scaffold it ready to paste. Craft-on-the-fly, friction-driven, minimal: grow a harness to fit the real failure mode instead of designing one upfront. Sibling of /pick-model (picks the model) and /pick-workflow (picks the topology); this picks the GUARDRAIL. Domain-general — any agentic task, not just code. Use when 'agent keeps failing at X', 'it hallucinated/broke the contract/wasted tokens', 'what guardrail do I add', 'harness for X', 'how do I stop it doing Y', or when starting a new agentic task and you want a minimal starter harness. Recommends + scaffolds; does NOT run the task.
Decide HOW work should EXECUTE — linear vs parallel fan-out, sub-agents vs teams vs dynamic Workflow, and the per-step seam — instead of defaulting to linear+single-model. Produces an execution design, does NOT run the task. Delegates the per-step model+effort call to a model-picker skill when the environment has one. Use when authoring/challenging a skill or agent, AND mid-task when deciding how to run real work — 'should I fan this out', 'should this fan out', 'parallel or linear', 'sub-agents or workflow', 'per-step seam', 'execution topology', 'execution architecture'. For a trivial few-item call, early-exit to linear cheaply rather than a full analysis.
Save session to CONTEXT-llm.md with conversation summary. Use when saving work, checkpointing progress, preserving session state. Triggers include "save context", "save session", "checkpoint", "save my progress".
Challenge, push back, play devil's advocate on AI output or intentions. Use when: challenge this, are you sure, push back, prove it, what if you're wrong, devil's advocate, stress test, poke holes, second opinion, sanity check, too confident, really?, question this decision, avant de commencer, je veux faire X, aide-moi à décider. Modes: route (default — any invocation whose first word isn't a recognized subcommand, including a free-text description), forward (before starting), anchor (committed too fast), verify (facts wrong?), framing (wrong problem?), deep (full devil's advocate in fresh context).
Capture a session's verbatim "meat" — the exact phrasing of constraints, rejected paths, corrections, live reasoning, learnings, pivots, open questions — into CHECKPOINT-*-llm.md, plus a synthesized summary of state. Use before /clear on a long reasoning session, when exact wording matters, or when saving a session you'll resume cold. Triggers: "checkpoint", "save the meat", "capture verbatim", "checkpoint before clear".
Generate a copy-paste instruction block to steer Claude Code's /compact so the summary keeps what matters. Use when about to compact context intentionally, before a long session gets summarized, or when the user says "instruct compact", "compaction instructions", "steer compact", "custom compact", "prepare compaction".
| name | load-context |
| description | Resume session from CONTEXT-llm.md. Use when resuming work, loading saved context, continuing a previous session. Triggers include "load context", "resume session", "continue where I left off". |
| argument-hint | [stream-name] [--as <role>] [--full] |
| allowed-tools | ["Bash","Read","AskUserQuestion"] |
| model | haiku |
| context | main |
| user-invocable | true |
Load session state from CONTEXT-{stream}-llm.md and optionally expand full resources.
Speed: < 3 seconds (default), 5-8 seconds (--full)
CRITICAL: After EVERY AskUserQuestion call, check if answers are empty/blank. Known Claude Code bug: outside Plan Mode, AskUserQuestion silently returns empty answers without showing UI.
If answers are empty: DO NOT proceed with assumptions. Instead:
rtk for ALL shell commandsBash: rtk ls -t CONTEXT-*llm.md done/CONTEXT-*llm.md 2>/dev/null || true
Read: CONTEXT-{stream}-llm.md (if stream known from $ARGUMENTS)
If not found in project root, check done/ subfolder. If found there, note 📦 (from done/) in report.
If multiple streams and no selection → AskUserQuestion with options (mark done/ files with 📦).
Filename: "default" → CONTEXT-llm.md, "{name}" → CONTEXT-{name}-llm.md
Role: if --as <role> is passed, record it as this session's active role. A later /save-context call in the same session reuses it without needing --as again.
Parallel Read: OpenSpec project/proposal/tasks.md, hot files, manifest.yaml.
Hot file selection: --full reads all hot files (all roles); default (lean) reads [P1]-tagged only, across ALL roles, not just mine (untagged = P2 = lean-skip). Fallback if no tags present: top 3.
DO NOT restore tasks — informational only.
Thinking Artifacts (if ## Thinking Artifacts section exists in CONTEXT file):
--full mode: Read referenced $PRAXIS_DIR/thinking artifacts and include brief summaries in report$PRAXIS_DIR is unset: display paths as-is, note ⚠️ $PRAXIS_DIR not setPhase 1's Read already put the whole file in context — re-rendering it would pay those tokens twice and risk a paraphrase drifting from the source. The report orients, never re-explains.
Output exactly 4 lines — no section blocks, no re-synthesis:
🔄 {stream} · {status} · sauvé {relative time} par [{role}]
🎯 {goal, one line}
▶️ {Next[0]}
⚠️ {N} lignes nouvelles de [{other role}] depuis mon dernier save ← omit if none
Solo session (no role recorded) → drop par [{role}] from line 1; line 4 never applies.
## Standing lines: render as rules in force for THIS resumed session, not informational bullets.
[constraint] → active rule the model must honour now[assumption] → premise to re-verify before relying on it[definition] → term binding for this sessionSee reference.md for section mapping, delta-line computation, error messages, and role resolution.