hallouminate 0.1.0

A markdown corpus indexer for LLMs to build and query their own per-repo wikis.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
//! Tool registrations for the hallouminate MCP server. Every stateful tool
//! is a proxy to the local daemon: opening a `DaemonClient` and dispatching
//! one RPC per call. Keeps the daemon as the canonical owner of the LanceDB
//! ground directory and per-corpus mutation locks per the spec's Approach.
//!
//! When the daemon is unreachable (e.g. no `hallouminate daemon` running),
//! tool calls return JSON-RPC `-32603 internal_error` with the documented
//! "daemon unavailable" hint instead of silently opening a local LanceDB
//! handle — that fallback is exactly the multi-process race the daemon is
//! built to remove.

use rmcp::handler::server::router::tool::ToolRouter;
use rmcp::handler::server::wrapper::Parameters;
use rmcp::model::{CallToolResult, Content, ErrorData, ServerCapabilities, ServerInfo};
use rmcp::{ServerHandler, tool, tool_handler, tool_router};
use schemars::JsonSchema;
use serde::Deserialize;
use std::path::{Path, PathBuf};

use crate::app::daemon::{
    AddMarkdownRequest, AddMarkdownResult, DaemonClient, DaemonRequest, DaemonRequestPayload,
    DaemonRpcError, DeleteMarkdownRequest, DeleteMarkdownResult, ErrorKind,
    GlobalizeMarkdownRequest, GlobalizeMarkdownResult, GroundRequest, GroundResult, IndexRequest,
    ListCorporaResult, ListFilesRequest, ListFilesResult, ListTreeRequest, ListTreeResult,
    ReadMarkdownRequest, ReadMarkdownResult, client_for,
};

const SERVER_INSTRUCTIONS: &str = "\
Hallouminate stores per-repository markdown wikis on disk and exposes them \
for semantic search. Each `[[repository]]` entry derives a `repo:{name}:wiki` \
corpus rooted at `<repo>/.hallouminate/wiki/`. The wiki is the canonical \
place for cross-session knowledge: architecture, conventions, gotchas, \
\"why this design not that one\" notes.

Two audiences use this server:
- CONSUMERS (grounding agents) read the wiki via `ground` / `read_markdown` \
  / `list_files` / `list_tree`.
- AUTHORS (curator agents) write entries via `add_markdown` / overwrite via \
  `read_markdown` + `add_markdown { overwrite: true }`.

Default corpus: tool calls that omit `corpus` default to the wiki for the \
repository containing the daemon's current working directory. Pass `corpus` \
explicitly to target another wiki, the repo's source corpus \
(`repo:{name}:corpus`), or a user-declared `[[corpus]]` entry. \
`list_corpora` enumerates everything available.

Tools:
- `list_corpora` — every configured corpus name.
- `list_files` — flat list of relative paths in a corpus.
- `list_tree` — same files grouped into a directory tree with subdirs; use \
  this to navigate progressively-disclosed wikis instead of reading every \
  index.md in sequence.
- `read_markdown` — verbatim file contents (UTF-8). Call before overwriting.
- `add_markdown` — write a file (atomic, no-symlink-follow). Reindexes that \
  file AND walks ancestor `index.md`s to refresh the link tree.
- `delete_markdown` — unlink file + prune index rows + refresh ancestor \
  indexes.
- `ground` — semantic search; returns ranked chunks with snippet, \
  heading_path, line_range, score.
- `index` — bulk (re)index a corpus or all corpora.

Filesystem is the source of truth; LanceDB rows are derived and refreshed \
after `add_markdown` / `delete_markdown`. `index` is the only way to pick \
up edits made outside hallouminate.

# Authoring conventions (REQUIRED for `add_markdown`)

ONE TOPIC PER FILE. A wiki entry is a slice of knowledge with a clear \
scope. The chunker splits on headings — two unrelated topics in one file \
make `ground` rank both sections together, which is rarely what you want.

FIRST NON-BLANK LINE IS H1. Every file's first non-blank line must be \
`# Topic Name`. The chunker uses the H1 as the breadcrumb root; without \
it, search results lose navigability. The H1 is also what the auto-index \
quotes as each entry's gloss.

FILE STEM MATCHES THE SLUG. \"Corpus walker\" → `corpus-walker.md`. \
Lowercase, kebab case. No spaces, no capitals, no extensions other than \
`.md`.

LEAD WITH THE CONCLUSION. Don't bury what the file is about under \
preamble. Cite files and line ranges by path: \
`src/domain/corpus/walker.rs:42`. Prefer concrete examples to abstract \
description. ~50-150 lines per entry is the right band.

# Tree layout & linking

Subdirectories work — write `architecture/dataflow.md` and the daemon \
creates `architecture/` for you. Use them to give a wiki shape:

- Top-level files for foundational topics (`architecture.md`, \
  `mcp-surface.md`).
- Subdirectories for related entries (`adapters/lance.md`, \
  `adapters/mcp.md`, `adapters/index.md`).

Each directory carries an `index.md`. The daemon scaffolds and maintains \
the LINK LIST inside it between markers, but you OWN the prose outside.

After `add_markdown` / `delete_markdown`, the daemon walks from the corpus \
root down to the changed file's parent and refreshes the link list inside \
`<!-- HALLOUMINATE:INDEX-START -->` / `<!-- HALLOUMINATE:INDEX-END -->` \
in each ancestor `index.md`. A missing `index.md` is scaffolded; prose \
outside the markers is preserved verbatim. To opt OUT, remove the markers \
— the daemon will then leave the file alone.

Link convention: `[stem](./stem.md)` for files, \
`[subdir/](./subdir/index.md)` for directories. Use relative paths so the \
links survive moves of the whole wiki.

# Authoring loop

```
1. list_tree                          (see what's already there)
2. ground \"<topic-adjacent search>\"   (find related entries to cross-link)
3. read_markdown index.md             (confirm naming + style)
4. draft (H1 first line, kebab slug, link adjacent entries)
5. add_markdown { corpus, path, content, overwrite: false }
6. (the daemon updates ancestor index.md link lists for you)
```

# Examples

Add a top-level entry:
```
add_markdown {
  corpus: \"repo:myrepo:wiki\",
  path: \"corpus-walker.md\",
  content: \"# Corpus walker\\n\\nGitignore-aware...\\n\",
  overwrite: false,
}
```

Add a nested entry — daemon creates `adapters/index.md` if missing:
```
add_markdown {
  corpus: \"repo:myrepo:wiki\",
  path: \"adapters/lance.md\",
  content: \"# LanceDB adapter\\n\\n...\\n\",
  overwrite: false,
}
```

Update with rollback safety:
```
read_markdown { corpus: \"repo:myrepo:wiki\", path: \"corpus-walker.md\" }
// edit content
add_markdown { ..., overwrite: true }
```

Multi-root corpora write all `add_markdown` / `delete_markdown` to the \
FIRST configured root only — keep one root if you can.
";

/// Build a `CallToolResult` with both a human-readable text content block
/// and a `structured_content` JSON payload. `CallToolResult` is
/// `#[non_exhaustive]` in `rmcp`, so we must construct via the provided
/// `success(...)` constructor then mutate public fields.
fn tool_ok(text: String, structured: serde_json::Value) -> CallToolResult {
    let mut result = CallToolResult::success(vec![Content::text(text)]);
    result.structured_content = Some(structured);
    result
}

/// Render a `TreeNode` as an indented ASCII outline — subdirs first, then
/// files, with one entry per line and two-space indents per depth level.
/// Mirrors the structured tree for clients that only want the text block.
fn render_tree_outline(
    node: &crate::domain::corpus::sandbox::TreeNode,
    depth: usize,
    out: &mut String,
) {
    let indent = "  ".repeat(depth);
    let label = if node.path.is_empty() {
        ".".to_string()
    } else {
        Path::new(&node.path)
            .file_name()
            .and_then(|s| s.to_str())
            .unwrap_or(&node.path)
            .to_string()
    };
    out.push_str(&format!("{indent}{label}/\n"));
    for sub in &node.subdirs {
        render_tree_outline(sub, depth + 1, out);
    }
    let file_indent = "  ".repeat(depth + 1);
    for f in &node.files {
        let name = Path::new(&f.path)
            .file_name()
            .and_then(|s| s.to_str())
            .unwrap_or(&f.path);
        out.push_str(&format!("{file_indent}{name}\n"));
    }
}

fn internal_error(msg: impl Into<String>) -> ErrorData {
    ErrorData::internal_error(msg.into(), None)
}

/// JSON-RPC -32602: surface caller-supplied input failures (bad corpus name,
/// unsafe path, missing required argument) distinctly from server faults so
/// MCP clients can route them as user errors instead of retries.
fn invalid_params(msg: impl Into<String>) -> ErrorData {
    ErrorData::invalid_params(msg.into(), None)
}

/// Open a `DaemonClient` for one tool call, surfacing a clear
/// "daemon unavailable" error if the socket is missing. The MCP server is
/// long-lived but the daemon may not be, so we dial per-call instead of
/// caching a client across requests. Production callers go through
/// `daemon_socket_path()` (which respects `HALLOUMINATE_SOCKET`); the env
/// is the only override hook the MCP transport needs.
async fn daemon_for_tool() -> Result<DaemonClient, ErrorData> {
    // Per the spec's Non-goals: do NOT auto-start the daemon from MCP.
    client_for(None)
        .await
        .map_err(|e| internal_error(format!("{e:#}")))
}

/// Translate a daemon RPC error into the MCP transport's `ErrorData` shape.
/// Daemon `InvalidParams` becomes `-32602`, `Internal` becomes `-32603`, and
/// transport / decode failures (already `anyhow::Error` by the time we get
/// here) collapse to `-32603` so MCP clients don't misinterpret a network
/// flake as a user error.
fn map_daemon_err(err: anyhow::Error) -> ErrorData {
    if let Some(rpc) = err.downcast_ref::<DaemonRpcError>() {
        return match rpc.kind {
            ErrorKind::InvalidParams => invalid_params(rpc.message.clone()),
            ErrorKind::Internal => internal_error(rpc.message.clone()),
        };
    }
    internal_error(format!("{err:#}"))
}

#[derive(Debug, Deserialize, JsonSchema)]
pub struct GroundParams {
    /// Free-text query to embed and search against the index.
    pub query: String,
    /// Optional corpus name; required when more than one is configured.
    #[serde(default)]
    pub corpus: Option<String>,
    /// Max number of files in the rolled-up result.
    #[serde(default)]
    pub top_files: Option<usize>,
    /// Max chunks returned per file.
    #[serde(default)]
    pub chunks_per_file: Option<usize>,
    /// Internal raw-hit cap before per-file rollup.
    #[serde(default)]
    pub limit: Option<usize>,
    /// Trim each chunk's snippet to N chars in both the outline and the
    /// structured response. Orthogonal to format selection.
    #[serde(default)]
    pub snippet_chars: Option<usize>,
}

#[derive(Debug, Deserialize, JsonSchema)]
pub struct IndexParams {
    /// Optional corpus name; omit to index every configured corpus.
    #[serde(default)]
    pub corpus: Option<String>,
}

#[derive(Debug, Deserialize, JsonSchema)]
pub struct ListCorporaParams {}

#[derive(Debug, Deserialize, JsonSchema)]
pub struct ListFilesParams {
    /// Corpus name; defaults to the wiki for the repo containing the
    /// daemon's cwd. Required only when no default applies and multiple
    /// corpora are configured.
    #[serde(default)]
    pub corpus: Option<String>,
}

#[derive(Debug, Deserialize, JsonSchema)]
pub struct ListTreeParams {
    /// Corpus name; defaults to the wiki for the repo containing the
    /// daemon's cwd. Required only when no default applies and multiple
    /// corpora are configured.
    #[serde(default)]
    pub corpus: Option<String>,
}

#[derive(Debug, Deserialize, JsonSchema)]
pub struct AddMarkdownParams {
    /// Corpus that owns the markdown file.
    pub corpus: String,
    /// Relative path under the corpus' FIRST configured root (multi-root
    /// corpora ignore paths[1..] for writes). The caller owns the directory
    /// structure and markdown shape — convention: `<slug>.md` or
    /// `<category>/<slug>.md`, first line `# Title`.
    pub path: String,
    /// Markdown bytes to write as UTF-8 text. Stored verbatim — hallouminate
    /// does not template or validate the markdown format.
    pub content: String,
    /// Replace an existing file. Defaults to false to avoid accidental
    /// clobber; use `read_markdown` first to inspect, then re-call with
    /// overwrite=true.
    #[serde(default)]
    pub overwrite: bool,
}

#[derive(Debug, Deserialize, JsonSchema)]
pub struct ReadMarkdownParams {
    /// Corpus that owns the markdown file.
    pub corpus: String,
    /// Relative path under the corpus' first configured root, same shape as
    /// `add_markdown`. Symlinks are rejected.
    pub path: String,
}

#[derive(Debug, Deserialize, JsonSchema)]
pub struct DeleteMarkdownParams {
    /// Corpus that owns the markdown file.
    pub corpus: String,
    /// Relative path under the corpus' first configured root, same shape as
    /// `add_markdown`. Symlinks are rejected. Irreversible.
    pub path: String,
}

#[derive(Debug, Deserialize, JsonSchema)]
pub struct GlobalizeMarkdownParams {
    /// Corpus the entry currently lives in.
    pub source_corpus: String,
    /// Relative path of the entry within `source_corpus`.
    pub path: String,
    /// Destination path within the global corpus. Defaults to `path`.
    #[serde(default)]
    pub dest_path: Option<String>,
    /// Replace an existing destination entry. Defaults to false.
    #[serde(default)]
    pub overwrite: bool,
}

/// Long-lived MCP server handle. Every tool method dials the daemon over a
/// fresh `UnixStream`, so the server is stateless beyond `tool_router`
/// and the captured client cwd.
#[derive(Debug, Clone)]
pub struct HallouminateTools {
    // The `tool_router` field is read by `#[tool_handler]`-generated code
    // when dispatching `tools/call`; rustc's dead-code pass doesn't see the
    // macro expansion, so silence the warning here.
    #[allow(dead_code)]
    tool_router: ToolRouter<HallouminateTools>,
    /// CWD captured once at MCP server startup, forwarded on every daemon
    /// hop so repo discovery resolves against the client's workspace.
    cwd: PathBuf,
}

#[tool_router]
impl HallouminateTools {
    pub fn new(cwd: PathBuf) -> Self {
        Self {
            tool_router: Self::tool_router(),
            cwd,
        }
    }

    #[tool(
        description = "Semantic search over a markdown corpus. `content` is a ripgrep-style outline (path, summary, line_range, score, snippet). `structuredContent.docs` maps absolute_path → { corpus, score, summary, keywords, mtime, chunks: [{chunk_id, heading_path, line_range, score, snippet}] }. Defaults from config: top_files=10, chunks_per_file=3, limit=50. Snippets are full chunk text unless `snippet_chars` is set."
    )]
    pub async fn ground(
        &self,
        Parameters(params): Parameters<GroundParams>,
    ) -> Result<CallToolResult, ErrorData> {
        let client = daemon_for_tool().await?;
        let req = DaemonRequest {
            cwd: self.cwd.clone(),
            payload: DaemonRequestPayload::Ground(GroundRequest {
                query: params.query,
                corpus: params.corpus,
                top_files: params.top_files,
                chunks_per_file: params.chunks_per_file,
                limit: params.limit,
                snippet_chars: params.snippet_chars,
            }),
        };
        let result: GroundResult = client.call(req).await.map_err(map_daemon_err)?;
        let structured =
            serde_json::to_value(&result.response).map_err(|e| internal_error(e.to_string()))?;
        Ok(tool_ok(result.outline, structured))
    }

    #[tool(
        description = "Build or refresh the LanceDB index for one or all configured corpora. Returns a one-line summary in `content` and the per-corpus IndexReport in `structuredContent`."
    )]
    pub async fn index(
        &self,
        Parameters(params): Parameters<IndexParams>,
    ) -> Result<CallToolResult, ErrorData> {
        let client = daemon_for_tool().await?;
        let req = DaemonRequest {
            cwd: self.cwd.clone(),
            payload: DaemonRequestPayload::Index(IndexRequest {
                corpus: params.corpus,
                paths_from: None,
            }),
        };
        let report: crate::app::cli::IndexReport =
            client.call(req).await.map_err(map_daemon_err)?;
        let summary = report
            .corpora
            .iter()
            .map(|c| {
                format!(
                    "{}: upserted={} touched={} deleted={} chunks+={}",
                    c.name, c.files_upserted, c.files_touched, c.files_deleted, c.chunks_inserted
                )
            })
            .collect::<Vec<_>>()
            .join("\n");
        let structured =
            serde_json::to_value(&report).map_err(|e| internal_error(e.to_string()))?;
        Ok(tool_ok(summary, structured))
    }

    #[tool(
        description = "List the corpus' files as a directory tree. `content` is an indented ASCII outline (subdirs first). `structuredContent` is { corpus, root: {path, absolute_path, files: [...], subdirs: [...]} } — recursive so an LLM can navigate progressively-disclosed wikis without reading every index.md. Defaults to the wiki for the repo containing the daemon's cwd when `corpus` is omitted."
    )]
    pub async fn list_tree(
        &self,
        Parameters(params): Parameters<ListTreeParams>,
    ) -> Result<CallToolResult, ErrorData> {
        let client = daemon_for_tool().await?;
        let req = DaemonRequest {
            cwd: self.cwd.clone(),
            payload: DaemonRequestPayload::ListTree(ListTreeRequest {
                corpus: params.corpus,
            }),
        };
        let result: ListTreeResult = client.call(req).await.map_err(map_daemon_err)?;
        let mut outline = String::new();
        render_tree_outline(&result.root, 0, &mut outline);
        let structured =
            serde_json::to_value(&result).map_err(|e| internal_error(e.to_string()))?;
        Ok(tool_ok(outline, structured))
    }

    #[tool(
        description = "List files currently visible in a corpus, honoring paths/globs/exclude rules. `content` is newline-separated relative paths. `structuredContent` is { files: [{path, absolute_path}, …] }. Paths are relative when the file lives under a configured corpus root, absolute otherwise. Defaults to the wiki for the repo containing the daemon's cwd when `corpus` is omitted."
    )]
    pub async fn list_files(
        &self,
        Parameters(params): Parameters<ListFilesParams>,
    ) -> Result<CallToolResult, ErrorData> {
        let client = daemon_for_tool().await?;
        let req = DaemonRequest {
            cwd: self.cwd.clone(),
            payload: DaemonRequestPayload::ListFiles(ListFilesRequest {
                corpus: params.corpus,
            }),
        };
        let entries: ListFilesResult = client.call(req).await.map_err(map_daemon_err)?;
        let text = entries
            .iter()
            .map(|e| e.path.clone())
            .collect::<Vec<_>>()
            .join("\n");
        let structured = serde_json::json!({ "files": &entries });
        Ok(tool_ok(text, structured))
    }

    #[tool(
        description = "Write a markdown file under the corpus' FIRST configured root, creating parent directories as needed, then refresh just that file's LanceDB rows. Atomic write, no-symlink-follow. Stores content verbatim — no markdown schema imposed. For updates, call `read_markdown` first, then re-call with `overwrite=true`."
    )]
    pub async fn add_markdown(
        &self,
        Parameters(params): Parameters<AddMarkdownParams>,
    ) -> Result<CallToolResult, ErrorData> {
        let client = daemon_for_tool().await?;
        let req = DaemonRequest {
            cwd: self.cwd.clone(),
            payload: DaemonRequestPayload::AddMarkdown(AddMarkdownRequest {
                corpus: params.corpus,
                path: params.path,
                content: params.content,
                overwrite: params.overwrite,
            }),
        };
        let response: AddMarkdownResult = client.call(req).await.map_err(map_daemon_err)?;
        let text = format!(
            "wrote {} and refreshed corpus {}",
            response.path, response.corpus
        );
        let structured =
            serde_json::to_value(&response).map_err(|e| internal_error(e.to_string()))?;
        Ok(tool_ok(text, structured))
    }

    #[tool(
        description = "Read verbatim UTF-8 contents of a markdown file in a corpus. `content` is the full file text; `structuredContent` is { corpus, path, absolute_path, content, bytes }. Symlinks are rejected. Returns the on-disk text, not the indexed/chunked view — call `ground` for semantic search."
    )]
    pub async fn read_markdown(
        &self,
        Parameters(params): Parameters<ReadMarkdownParams>,
    ) -> Result<CallToolResult, ErrorData> {
        let client = daemon_for_tool().await?;
        let req = DaemonRequest {
            cwd: self.cwd.clone(),
            payload: DaemonRequestPayload::ReadMarkdown(ReadMarkdownRequest {
                corpus: params.corpus,
                path: params.path,
            }),
        };
        let response: ReadMarkdownResult = client.call(req).await.map_err(map_daemon_err)?;
        let structured =
            serde_json::to_value(&response).map_err(|e| internal_error(e.to_string()))?;
        Ok(tool_ok(response.content, structured))
    }

    #[tool(
        description = "Unlink a markdown file from the corpus' first configured root and prune its rows from the LanceDB index. Irreversible. Symlinks are rejected. `content` is a one-line summary; `structuredContent` is { corpus, path, absolute_path, file_ref }."
    )]
    pub async fn delete_markdown(
        &self,
        Parameters(params): Parameters<DeleteMarkdownParams>,
    ) -> Result<CallToolResult, ErrorData> {
        let client = daemon_for_tool().await?;
        let req = DaemonRequest {
            cwd: self.cwd.clone(),
            payload: DaemonRequestPayload::DeleteMarkdown(DeleteMarkdownRequest {
                corpus: params.corpus,
                path: params.path,
            }),
        };
        let response: DeleteMarkdownResult = client.call(req).await.map_err(map_daemon_err)?;
        let text = format!("deleted {} from corpus {}", response.path, response.corpus);
        let structured =
            serde_json::to_value(&response).map_err(|e| internal_error(e.to_string()))?;
        Ok(tool_ok(text, structured))
    }

    #[tool(
        description = "Copy a single markdown entry from `source_corpus` into the corpus marked `global = true` in config, making it searchable everywhere. The source stays in place (copy, not move). `dest_path` defaults to the source path; `overwrite=false` (default) errors if the destination already exists. The global corpus row is reindexed synchronously. Errors if no global corpus is configured or if `source_corpus` is itself the global corpus. `structuredContent` is { source_corpus, source_path, global_corpus, dest_path, absolute_path, indexed }."
    )]
    pub async fn globalize_markdown(
        &self,
        Parameters(params): Parameters<GlobalizeMarkdownParams>,
    ) -> Result<CallToolResult, ErrorData> {
        let client = daemon_for_tool().await?;
        let req = DaemonRequest {
            cwd: self.cwd.clone(),
            payload: DaemonRequestPayload::GlobalizeMarkdown(GlobalizeMarkdownRequest {
                source_corpus: params.source_corpus,
                path: params.path,
                dest_path: params.dest_path,
                overwrite: params.overwrite,
            }),
        };
        let response: GlobalizeMarkdownResult = client.call(req).await.map_err(map_daemon_err)?;
        let text = format!(
            "globalized {} from {} into {} as {}",
            response.source_path,
            response.source_corpus,
            response.global_corpus,
            response.dest_path,
        );
        let structured =
            serde_json::to_value(&response).map_err(|e| internal_error(e.to_string()))?;
        Ok(tool_ok(text, structured))
    }

    #[tool(
        description = "List corpora configured in the hallouminate config file, including derived `repo:{name}:wiki` / `repo:{name}:corpus` entries from `[[repository]]` declarations. `content` is newline-separated corpus names; `structuredContent` is { corpora: [{name, paths}, …] }. Run `hallouminate config validate` for a richer summary."
    )]
    pub async fn list_corpora(
        &self,
        _params: Parameters<ListCorporaParams>,
    ) -> Result<CallToolResult, ErrorData> {
        let client = daemon_for_tool().await?;
        let entries: ListCorporaResult = client
            .call(DaemonRequest {
                cwd: self.cwd.clone(),
                payload: DaemonRequestPayload::ListCorpora,
            })
            .await
            .map_err(map_daemon_err)?;
        let names = entries
            .iter()
            .map(|e| e.name.clone())
            .collect::<Vec<_>>()
            .join("\n");
        let structured = serde_json::json!({ "corpora": &entries });
        Ok(tool_ok(names, structured))
    }
}

impl Default for HallouminateTools {
    /// `#[derive(Default)]` would construct `ToolRouter::default()` (an empty
    /// router) and skip the `#[tool_router]`-generated registration. Manual
    /// impl routes through `new()` so `HallouminateTools::default()` exposes
    /// the same tool set as `new()`. The default cwd is empty — production
    /// callers go through `serve_stdio` which captures the real cwd.
    fn default() -> Self {
        Self::new(PathBuf::new())
    }
}

#[tool_handler]
impl ServerHandler for HallouminateTools {
    fn get_info(&self) -> ServerInfo {
        // `ServerInfo` and `Implementation` are both `#[non_exhaustive]` in
        // `rmcp`; construct via `Default::default()` and mutate the fields
        // we care about so we don't fight the crate's evolution constraints.
        let mut info = ServerInfo::default();
        info.server_info.name = env!("CARGO_PKG_NAME").into();
        info.server_info.version = env!("CARGO_PKG_VERSION").into();
        info.instructions = Some(SERVER_INSTRUCTIONS.to_string());
        info.capabilities = ServerCapabilities::builder().enable_tools().build();
        info
    }
}

#[cfg(test)]
mod tests {
    //! MCP-specific tests. The corpus-boundary helpers
    //! (`safe_relative_path`, `pick_corpus`, `ensure_corpus_allows_file`,
    //! `first_corpus_root`, `atomic_write_no_follow`, `list_corpus_files`)
    //! live in `crate::domain::corpus::sandbox` and are tested there once.
    //! Daemon RPC error mapping has its own contract — pin the JSON-RPC
    //! code each error variant maps to so a future refactor of the
    //! daemon's `ErrorKind` would have to deliberately update the test.

    use super::*;

    #[test]
    fn map_daemon_err_routes_invalid_params_to_minus_32602() {
        let err: anyhow::Error = DaemonRpcError::invalid_params("bad corpus name").into();
        let mapped = map_daemon_err(err);
        assert_eq!(mapped.code.0, -32602);
        assert!(mapped.message.contains("bad corpus name"));
    }

    #[test]
    fn map_daemon_err_routes_internal_to_minus_32603() {
        let err: anyhow::Error = DaemonRpcError::internal("disk on fire").into();
        let mapped = map_daemon_err(err);
        assert_eq!(mapped.code.0, -32603);
        assert!(mapped.message.contains("disk on fire"));
    }

    #[test]
    fn new_stores_cwd_for_daemon_hops() {
        // Pin the field plumbing: the cwd handed to `HallouminateTools::new`
        // at MCP startup must be the same value every tool handler clones
        // into its `DaemonRequest`. Testing that cwd actually flows over
        // the socket needs a daemon fixture and lives in the integration
        // suite — this guards the boring-but-easy-to-break wiring.
        let cwd = PathBuf::from("/test/cwd");
        let tools = HallouminateTools::new(cwd.clone());
        assert_eq!(tools.cwd, cwd);
    }

    #[test]
    fn map_daemon_err_routes_transport_failure_to_minus_32603() {
        // Anything that is not a DaemonRpcError (transport / decode failure
        // before we got a typed error envelope from the daemon) must NOT be
        // misinterpreted as a caller-input failure. JSON-RPC -32603 is
        // explicitly "internal error" for cases like this.
        let err: anyhow::Error = anyhow::anyhow!("daemon unavailable: socket missing");
        let mapped = map_daemon_err(err);
        assert_eq!(mapped.code.0, -32603);
        assert!(mapped.message.contains("daemon unavailable"));
    }
}