@jcode.labs/ragmir 2.10.0 → 2.11.0

package/skills/ragmir/SKILL.md

@@ -43,11 +43,15 @@ acceptable. Normal confidential indexing keeps remote model loading disabled.
 
 | Need | Command or MCP tool |
 | --- | --- |
+| Active base, readiness, and capabilities | Read `ragmir://context` or run `rgr status --json` |
+| Bounded source coverage and index drift | Read `ragmir://sources` or run `rgr audit` |
 | Readiness | `rgr doctor` or `ragmir_status` |
 | Repair setup or stale data | `rgr doctor --fix` |
 | Check unindexed or unsupported files | `rgr audit --unsupported` |
+| Preview redacted chunks before indexing | `rgr preview --path <prefix> --json` |
 | Check privacy posture | `rgr security-audit` |
 | Retrieve exact passages | `rgr search "query" --compact` or `ragmir_search` |
+| Explain hybrid ranking | `rgr search "query" --explain` or `ragmir_search` with `explain: true` |
 | Expand one returned citation | `ragmir_expand` |
 | Gather broad cited evidence | `rgr research "topic" --compact` or `ragmir_research` |
 | Return deterministic context | `rgr ask "question"` or `ragmir_ask` |
@@ -57,6 +61,21 @@ Use `rgr route-prompt "..." --json` or `ragmir_route_prompt` only when it is unc
 current request needs the local corpus. The router is deterministic and does not retrieve or store
 the prompt.
 
+## Monorepo routing
+
+When a monorepo has a root Ragmir base and nested app bases, select the base before retrieval:
+
+1. Run `rgr bases --json` from the directory in scope. The nearest configured ancestor is
+   `activeId`.
+2. Use the root base for shared architecture or cross-app questions. Use the nested base for an
+   app-specific question.
+3. If the shell working directory is not inside the intended base, pass
+   `--project-root /absolute/path/to/base` before the command.
+4. For MCP, call `ragmir_status` when the active base is uncertain and verify `knowledgeBaseId`.
+   Generated helpers pin `RAGMIR_PROJECT_ROOT`; nested bases receive distinct server names.
+5. Never silently combine citations from different bases. Label each base when a task genuinely
+   requires evidence from more than one.
+
 ## Indexing
 
 Use `.ragmir/config.json` for sources, with paths, globs, and `!` exclusions:
@@ -80,6 +99,11 @@ OCR stays opt-in and never calls a remote service by default.
 `local-hash` is the default retrieval provider. It supports cited local retrieval but is not
 semantic embeddings. Use the Transformers provider only after an explicit model preload.
 
+Before a costly ingest or after changing chunk settings, run `rgr preview --json`. Review redaction
+counts, citations, `contextPath`, omitted chunks, and p50/p95 sizes. Search, ask, research, MCP
+retrieval, and golden queries can restrict retrieval with source paths and structural context paths.
+Use explanations to inspect RRF ranks and contributions, not as an independent relevance score.
+
 ## MCP
 
 `rgr setup` writes local MCP helpers under `.ragmir/` for Claude Code, Codex, Kimi, OpenCode, and
@@ -105,6 +129,10 @@ only when the exact chunk or neighboring context is needed. Retrieval tools acce
 the configured `mcpMaxOutputBytes` remains the hard ceiling. Inspect `_meta["ragmir/output"]` to see
 whether the response was compacted or truncated.
 
+When MCP resources are supported, read `ragmir://context` first for a bounded identity, readiness,
+freshness, coverage, and capability overview. Read `ragmir://sources` only when source coverage or
+index drift matters; its per-file lists are capped while totals remain complete.
+
 MCP is read-focused. Only remove an index with the explicit shell command:
 
 ```sh