360 Secure SkillHub
360 安全技能中心

Multipass

Multipass 用于完成需要当前未安装工具的任务:在隔离会话中自主规划、发现并临时配置沙箱工具,执行后输出任务结果与可复现回放脚本。适用于 Agent 缺少工具、需要一次性沙箱执行或并行探测的场景;通过 multipass.run 传入任务描述,并支持 search、sessions、replay 等命令。不做全局安装、不使用真实身份,禁止永久技能安装与通用网页检索。

indigokarasu
indigokarasu
热度 7.62
最近更新 2026-09-23 10:37:04
分类:自动化依赖:代码执行 / 联网浏览搜索 / 读写文件 / 第三方 API / MCP 服务 / 内部调用大模型应用场景:自动化脚本 / 工作流 / API 集成 / 任务管理 ...

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add ocas-multipass
Metadata
license
MIT
name
ocas-multipass
description
Accomplishes tasks that need tools the agent does not have. Plans approaches, fills capability gaps with sandboxed tools, executes within an isolated session. Disposable identity, parallel discovery, incremental checkpoints, graceful degradation. Output: task result plus replay script. No global installs, no real identity, clean state in and out. Do not use for tasks solvable with installed skills, permanent skill installs, skill builds (use Forge), or general web research (use Sift).
source
https://github.com/<agent-handle>/multipass
includes
["references/**","scripts/**"]
metadata
{"author":"Indigo Karasu (indigokarasu)","version":"4.2.0"}
tags
["sandbox","capability-gap","tools","isolated-execution"]
triggers
["capability gap","sandboxed tools","fill tool gaps","task needs tools"]

Multipass

Accomplish tasks that require tools you don't have. Fire and forget. User invokes multipass.run {task}, Multipass runs to completion, user gets the output. No check-ins, no approval gates, no escalation.

Everything happens inside a session directory. Nothing touches the rest of the agent platform.

Interactive Menu

When invoked interactively, present a two-level menu. See references/interactive-menu.md for the full menu structure.

When to Use

  • Tasks requiring tools not currently installed
  • Sandboxed execution with disposable identity
  • Parallel discovery across multiple surfaces
  • When the agent is about to say "I can't do that"
  • Reproducible recipe generation for uncommon workflows

When NOT to Use

  • Tasks solvable with installed OCAS skills
  • Permanent skill installation (use platform skill install)
  • Skill building (use Forge)
  • General web research (use Sift)

Isolation Rules

  • All files in session dir: {agent_root}/commons/data/ocas-multipass/sessions/{session_id}/
  • No global skill installs. No global MCP config changes. No real identity.
  • No env vars, secrets, or config written outside session dir.
  • If a capability requires breaking isolation, skip it silently and find an alternative.

Autonomy Rules

Once invoked, Multipass runs to completion without user interaction.

  • No "Proceed?" gates. No "Should I...?" questions. No escalation.
  • Every decision is made by the skill. If a choice is ambiguous, pick the safest sandboxable option and log the reasoning.
  • If everything fails, produce a failure report documenting what was tried and why each path failed. A useful failure report is still an output.
  • The only user-facing moments are invocation and the final result.

Loop Prevention

See references/loop-prevention.md for the full set of circuit breakers and anti-stall rules every worker must follow.

Workflow

Four phases. Task first, tools second, execution third, report last. See references/workflow.md for the full phase-by-phase procedure (Plan, Fill Gaps, Execute, Report).

Commands

  • multipass.run {task description} -- full autonomous lifecycle
  • multipass.search {description} -- discovery only
  • multipass.sessions -- list recent sessions
  • multipass.replay {session_id} -- re-execute replay script (new session, new identity)
  • multipass.status -- config, source health

Responsibility Boundary

Multipass does: plan, discover, provision identity, resolve, execute, checkpoint, report. All autonomously.

Multipass does not: install anything globally, use real identity, persist secrets, build skills, ask the user for help mid-run.

If a user runs the same replay 3+ times, suggest permanent installation or a Forge-built skill.

Optional Skill Cooperation

  • ivangdavila/api: Check 147-service reference before discovery.
  • ivangdavila/reverse-engineering: TRACE protocol for probing candidates.
  • Sift: Deeper candidate research. If absent, use web_search.
  • Forge: Suggest after completion if user needs a permanent solution.
  • mcporter: Check existing connections. Call by URL for sandboxed access.

Journal Outputs

Action Journal -- after every multipass.run. Research Journal -- after every multipass.search.

Path: {agent_root}/commons/journals/ocas-multipass/YYYY-MM-DD/{run_id}.json

Recovery Behavior

This skill implements the recovery contract from spec-ocas-recovery.md.

  • Evidence: Every multipass.run writes an evidence record to {agent_root}/commons/data/ocas-multipass/evidence.jsonl, including failed runs. The not_activity_reason field is mandatory when no side effects occur.
  • Gap detection: Not applicable — on-demand only.
  • Degraded mode: When sandbox tools are unavailable, logs degraded: <tool> and attempts fallback approaches per existing resilience patterns.
  • Log compaction: Session logs older than 30 days compacted to summaries. Last 7 days retained.

Storage Layout

See references/storage-layout.md for the full directory structure.

Token storage discipline (per spec-ocas-skill-improvements.md §5.1): Session tokens and service credentials are stored at the standard path ~/.hermes/{service}_token.json with the following validation before every request:

  1. Freshness check — compare expires_at against now + 5min; if expired or expiring within 5 minutes, refresh the token before executing the multiplexed request. Do not send a stale token.
  2. Shape validation — verify the file is valid JSON with {access_token, token_type, expires_at} fields. A corrupt or missing token file is logged as token_error and the request fails with auth_failure — never attempt a request with an empty/invalid token.
  3. Cleanup — after a refresh, atomically write the new token to the standard path before releasing dependents. Do not leave a mix of old/new tokens per service.

OKRs

Universal OKRs from spec-ocas-journal.md apply to all runs. See references/okrs.md for skill-specific targets (tool invocation success, spawn depth efficiency, isolation violation rate, schedule adherence, data integrity).

Initialization

On first run (multipass.init):

  1. Create dirs and default config.json
  2. Check maxSpawnDepth, log parallel mode availability
  3. Check mcporter availability, log MCP access mode (mcporter / direct HTTP / unavailable)
  4. Verify session dir is writable
  5. Log initialization as DecisionRecord

Ontology Types

Multipass does not extract entities and does not emit Chronicle signals.

Visibility

Public.

Gotchas

Error handling in multipass follows a strict isolation contract: never leak session state, never install globally, and never retry the same failure path twice.

  • No global installs, no real identity — Multipass must not install skills globally, modify MCP configs, or use the real user identity. If a capability requires breaking isolation, skip it silently and find an alternative.
  • Duplicate drafts/waits are forbidden — Same action + same result = move on immediately. Two identical failures on the same endpoint is the maximum. No folk theories, no "waiting for the server," no invented timing explanations.
  • Identity provisioning is disposable — Throwaway email accounts from the BotEmail cascade are temporary. If all identity providers fail, Multipass restricts itself to no-auth candidates without surfacing the failure to the user.
  • Failure reports are valid outputs — If all approaches fail, the manifest documents every tried path and failure reason. A complete failure report is still a useful artifact and counts as task completion.
  • Session isolation is absolute — All files must stay within the session directory. Nothing leaks to the platform, global config, or other sessions. An isolation violation is a serious incident.

Self-Update

multipass.update pulls the latest version from GitHub. See references/self-update.md.

Support File Map

File When to read
references/workflow.md Before executing a multipass run. Full 4-phase procedure (Plan, Fill Gaps, Execute, Report).
references/orchestration.md Before Phase 1 (plan) config check; before spawning workers; when setting up checkpoints or recovery logic
references/surfaces.md During Phase 2 (fill gaps) discovery; when searching for candidate tools or APIs
references/scoring.md During Phase 2 candidate evaluation; when scoring and ranking discovered tools
references/resilience.md When an endpoint fails or returns an error; when provisioning throwaway identity; when hitting CAPTCHA/gates; when applying reframing tactics
references/loop-prevention.md Before any worker execution; when enforcing circuit breakers and anti-stall rules
references/okrs.md During OKR evaluation. Skill-specific targets for tool invocation, spawn depth, isolation, schedule, data integrity.
references/self-update.md When running multipass.update.
references/interactive-menu.md When invoked interactively via / command.
references/storage-layout.md When creating or inspecting session directories.