菜单
Run the full Recursive Language Model pipeline — prep, fan out across chunks, and recursively synthesize until results fit one context window
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。
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/jmagly/aiwg --skill rlm-search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
基于 SOC 职业分类
| namespace | aiwg |
| name | rlm-search |
| platforms | ["all"] |
| description | Run the full Recursive Language Model pipeline — prep, fan out across chunks, and recursively synthesize until results fit one context window |
| version | 1.1.0 |
The full Recursive Language Model pipeline in one command. Prepares content if needed, fans the query out across all chunks, and recursively synthesizes results until they fit in a single context window. Use this when you need to answer a question against content too large to read at once.
Alternate expressions and non-obvious activations:
..| Pattern | Example | Action |
|---|---|---|
| Whole-repo search | "search the entire codebase for all usages of deprecated API" | rlm-search "..." --source . |
| Directory search | "recursively search src/ for logging calls" | --source src/ |
| File search | "use RLM to analyze this 5000-line file" | --source path/to/file.ts |
| Budget limit | "search but cap at 200k tokens" | --budget 200000 |
| Depth limit | "search up to 2 levels deep" | --depth 2 |
| Skip re-prep | "search using the existing prep" | No re-prep if manifest exists |
When triggered:
Extract query and source — identify the natural language query and the source path (file or directory). Default source is . (current directory).
Check for existing prep — look for a valid manifest in .aiwg/rlm-prep/ matching the source. Reuse only when the prep index covers every source file, manifest, and chunk. If missing, stale, incomplete, or from an older single-chunk-dropping prep run, rebuild with rlm-prep.
Initial fanout (level 1) — dispatch the query across all chunks, up to --max-parallel subagents at a time. Collect results with provenance.
Check synthesis fit — measure the total size of all level-1 results. If they fit in a single context window, synthesize directly (base case). If not, recurse.
Recursive reduction — chunk the level-1 results into a new set of chunks and fan out again. Each level-N subagent synthesizes the results from one batch of level-(N-1) answers. Repeat until the output fits in one window.
Final synthesis — produce a single coherent answer from the last reduction level. Include provenance: trace each claim back to a source file and line range.
Cost summary — report total tokens consumed, number of subagents launched, recursion depth reached, and USD cost estimate.
Level 0 (root query)
└── Level 1 fanout: N subagents (one per chunk)
├── chunk-0001 → answer fragment A
├── chunk-0002 → answer fragment B
├── chunk-0003 → (no match)
└── chunk-0004 → answer fragment C
If A + B + C fit in one window:
└── Synthesize → Final Answer ✓
If A + B + C do NOT fit:
Level 2 fanout: chunk the level-1 results
├── [A + B] → synthesis fragment 1
└── [C] → synthesis fragment 2
└── Synthesize fragments 1 + 2 → Final Answer ✓
The default --depth 3 means the pipeline will recurse at most 3 times before forcing synthesis even if results are large.
RLM Search Complete
Query: "Where is rate limiting implemented?"
Source: src/ | Chunks: 47 | Depth reached: 1 | Subagents: 14
Answer:
Rate limiting is implemented in three places:
1. **API gateway level** — `src/gateway/rate-limit.ts` (lines 12-45) applies
a sliding window limiter using Redis. Limits are configured per route in
`config/rate-limits.yaml`.
2. **Auth service** — `src/auth/middleware.ts` (lines 88-102) imposes a
per-IP limit of 10 login attempts per minute using an in-memory store.
3. **WebSocket connections** — `src/realtime/server.ts` (lines 231-248)
limits new connections per second to prevent connection floods.
Cost summary: 47 subagents, 184,320 tokens (~$0.18), 1 synthesis pass
<query> — Natural language question or task (required)--source <file|dir> — Source content to search (default: .)--depth N — Maximum recursion depth before forcing synthesis (default: 3)--max-parallel N — Max parallel subagents per level (default: 4, bounded by context budget). Alias --parallel is also accepted by the CLI.--budget N — Token budget for the entire operation (default: 500000)Prep coverage note: files that fit within one chunk are still written to a manifest and included in the search plan. Existing prep indexes are validated before reuse so older partial indexes are rebuilt automatically.
User: "search the entire codebase for where authentication tokens are validated"
Action: Check for existing prep of ., fanout across all chunks, synthesize.
Response:
RLM Search Complete
Query: "where are authentication tokens validated?"
Source: . | Chunks: 84 | Depth: 1 | Subagents: 84
Answer:
Token validation occurs at two layers:
1. **HTTP middleware** — `src/auth/middleware.ts` lines 34-67: the
`validateToken` function decodes and verifies JWTs using the
`jsonwebtoken` library, checking signature and expiry.
2. **GraphQL context** — `src/graphql/context.ts` lines 18-31: calls
`validateToken` on every request and attaches the decoded payload
to the GraphQL execution context.
Cost: 84 subagents, 241,800 tokens (~$0.24)
User: "use RLM to find all compliance-relevant data handling in the entire codebase"
Action:
aiwg rlm-search "find all places where PII or sensitive data is stored, transmitted, or logged" --source .
Level-1 produces 28 matching fragments totaling 40,000 tokens (too large for one pass). Level-2 reduces to 4 synthesis fragments, then final synthesis produces the answer.
Response: "Depth reached: 2. Found 14 locations across 9 files. [Full provenance-tagged answer]"
User: "deep search src/payments/ for Stripe webhook handling, cap at 100k tokens"
Action:
aiwg rlm-search "how are Stripe webhooks handled?" \
--source src/payments/ \
--budget 100000 \
--max-parallel 4
Response: If budget would be exceeded, the pipeline pauses and reports: "Budget checkpoint: 82,400 tokens used. Continue (remaining budget: 17,600)? [y/n]"
User: "use RLM to analyze this 8,000-line migration file for rollback risk"
Action:
aiwg rlm-search "identify any irreversible operations with no rollback path" \
--source db/migrations/0099_big_schema.sql \
--depth 2
Response: Preps the single file into ~40 chunks, fans out, synthesizes. Reports all DROP, TRUNCATE, and ALTER TABLE ... DROP COLUMN statements with line numbers.
User: "quick RLM search: where is the database connection string set?"
Action:
aiwg rlm-search "where is the database connection string configured?" \
--source . \
--depth 1 \
--max-parallel 8
Response: Forces synthesis at depth 1 — faster but may miss cross-chunk context. Reports results within a single fanout pass.
If the user's intent is ambiguous: