菜单
Use when the user says: 'compute', 'crunch this', 'analyze logs', 'aggregate the data', 'run a script', 'dedupe', 'count by', 'top N', or any data-shaping ask. Replaces dumping raw data into context with sandboxed script execution.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
基于 SOC 职业分类
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/FerroxLabs/ijfw --skill ijfw-compute命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
Maintain canonical AGENTS.md (open spec). Trigger: 'agents.md', 'update AGENTS.md', or auto-fired by ijfw-team after agent generation.
First-class design intelligence. Dispatches to the best available design skill (ui-ux-pro-max, frontend-design, superpowers), then layers IJFW constraints on top. Triggers: 'design', 'redesign', 'UI', 'UX', 'dashboard', 'page', 'component', 'make it look better', 'polish', 'pretty', 'professional', 'user experience', 'layout', 'visual', 'mobile-first', 'dark mode', 'accessibility', 'colors', 'typography', 'brand'.
Check for and apply IJFW updates safely. Trigger: 'update ijfw', 'upgrade', 'latest version', 'is there a new version', /update
| name | ijfw-compute |
| description | Use when the user says: 'compute', 'crunch this', 'analyze logs', 'aggregate the data', 'run a script', 'dedupe', 'count by', 'top N', or any data-shaping ask. Replaces dumping raw data into context with sandboxed script execution. |
| context | fork |
| model | sonnet |
Compute over read. When data is large or repetitive, run a sandboxed script and surface only the result. The four trees below decide which lever to pull.
Use when the input is bigger than the answer: log files, CSV/JSON dumps,
file-tree walks, repeated string transforms, aggregate stats, deduping.
Call ijfw_run compute:python "<script>" for pandas / numpy / stdlib parsing.
Call ijfw_run compute:js "<script>" for JSON shape-checks, regex sweeps,
quick numeric work. Sandbox is allowlist filesystem (cwd + project root) +
best-effort OS-level network deny; opt-in with IJFW_COMPUTE_NET=1 if the
script needs egress. Default timeout 30s, hard cap 300s via
IJFW_COMPUTE_TIMEOUT_MS. Output cap 100MB; overflow lands in the on-disk log.
Example: a 40MB nginx log. Instead of reading 200k lines into context, run
ijfw_run compute:python "import collections,sys;c=collections.Counter();[c.update([l.split()[8]]) for l in open('access.log')];print(c.most_common(10))"
and surface the top-10 status-code summary.
Use when the file is small (<2k lines), the task is a code edit or config tweak, or the agent needs to reason about structure rather than aggregate content. Direct Read is cheaper than spinning a subprocess; compute has fixed startup overhead.
Example: editing a single function in mcp-server/src/server.js. Read the
file, edit it, move on. No compute call needed.
Use after a compute or research step produces a finding worth recalling
across sessions. Call ijfw_run index:source <kind> <body> to write into
the per-project FTS5 db at <project>/.ijfw/index/compute.db. Schema is
raw table (source_kind, source, session_id, project_root, body, ts).
Per-write PRAGMA quick_check guards integrity.
Citation provenance (C9.6): pass --source=<pointer> before the body to
attach an origin (file path / observation kind / skill name). Search hits
surface this pointer + the session_id so users can trace where each row
came from. Omitted -> source stays NULL.
Example: after analyzing the nginx log, index the verdict:
ijfw_run index:source compute_output --source=logs/access.log "Top error 502 from upstream X 2026-05-08; 4.1% of requests".
Next session can search for it via the search tree below.
Use before computing again. If a previous session already answered a similar
question, recall it instead of recomputing. Call
ijfw_memory_search compute:query "<query>" for top-k FTS5 hits scoped to
the current project. Each hit returns body + source_kind + source +
session_id + ts; the agent decides whether the cached finding is still
fresh and can cite the source pointer.
Stemmed BM25 (C9.4): the FTS5 tokenizer is porter unicode61. Morphological
variants collapse: "authenticate" / "authenticating" / "authentication"
share a stem; "configure" / "configured" / "configuring" share a stem.
Synonym expansion (C9.5): default-on. Bare tokens expand against ~80 coding-
domain pairs (db <-> database, auth <-> authentication, perf <-> performance,
etc.). The result envelope reports synonym_matches: { token: [expansions] }
so callers see what fired. Disable per-process via IJFW_SYNONYM_EXPAND=0.
Session filter (C9.6): append --session=<id> to a query to scope hits to
a single session. The envelope echoes the filter as session_filter.
Example: user asks "what was the top nginx error last week?" Run
ijfw_memory_search compute:query "nginx error rate" first; if a recent
indexed finding lands, surface it directly. If empty or stale, fall back to
the compute tree on fresh log data.