@saasontools/strauss-kb 0.1.1 → 0.1.3

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -0,0 +1,85 @@
1
+ # Architecture
2
+
3
+ The README says what the format is and how to use it. This says why it is
4
+ shaped that way, and which alternatives were tried and dropped — the decisions
5
+ a later reader would otherwise reopen.
6
+
7
+ ## One record per file
8
+
9
+ The filename is the identity, so two writers never merge; they only choose
10
+ distinct names. Publication uses `link`, which fails when the name is taken, and
11
+ a collision surfaces as a 409 the caller has to answer rather than a silent
12
+ last-write-wins.
13
+
14
+ Read-modify-write (`setStatus`, `answer`) checks a content digest immediately
15
+ before publishing. That narrows the lost-update window to two adjacent syscalls
16
+ rather than closing it. A lock would close it and add a stale-hold failure worse
17
+ than the residue: a crashed holder blocks every later writer, where a lost
18
+ update costs one retry.
19
+
20
+ ## Read for a question, not for a session
21
+
22
+ A base loaded at the start of a long conversation is summarised away by the end
23
+ of it, and nothing keeps it alive. So no consumer loads it that way:
24
+
25
+ | Consumer | How it reads |
26
+ | ------------------------------- | ---------------------------------------------------------------------------- |
27
+ | Diff annotation | `matchToDiff` — deterministic, no context involved |
28
+ | "Has this been decided?" | a fresh short-lived reader, given the base and the question, discarded after |
29
+ | An implementor writing a record | a point query at the moment of writing, not a load an hour earlier |
30
+
31
+ Each has clean context by construction. Where a base genuinely must stay
32
+ resident it will drift, and there is no defence — the mitigation is that
33
+ reloading costs about three thousand tokens, so read it again at the point of
34
+ use rather than trying to keep it.
35
+
36
+ ## What happens when a base outgrows a context
37
+
38
+ Loading stops working somewhere above a few hundred records. What replaces it is
39
+ not a different answer to the same question — it is the same reader, given
40
+ candidates instead of everything:
41
+
42
+ ```
43
+ vector recall → top ~20 candidates, with their scores
44
+
45
+ reader judges → picks what answers, or says nothing does
46
+ ```
47
+
48
+ The reader stays the judge in both regimes, which is what preserves the two
49
+ structural wins the README reports: it can say no record answers the question,
50
+ and it picks the record that answers rather than the one nearest the topic.
51
+ Neither survives if a ranker's top hit is taken as the answer.
52
+
53
+ **A score threshold is not the growth path.** The wrong `audit trail` hit scored
54
+ 0.318 against the correct `race condition` hit at 0.295; any cut that drops the
55
+ first drops the second. A threshold excludes the absurd — an unrelated query
56
+ scored 0.091 — and nothing else.
57
+
58
+ **Tags are not the growth path either.** The field exists, is written, and is
59
+ read by nothing but an index line, deliberately. Free-text tags drift the way
60
+ `auth` / `authentication` / `authn` drift, which is the failure this format was
61
+ rewritten to remove; enforcing a vocabulary would make them a closed enum, which
62
+ `type` already is. The labels that matter here are already verifiable —
63
+ `strauss_anchors` names a file and a symbol, which either match the repository
64
+ or do not, where a tag can be wrong forever. If narrowing ever matters, measure
65
+ tag narrowing against vector recall on a real base rather than adding both.
66
+
67
+ ## Rejected: a format that needs a parser
68
+
69
+ This was broken twice. A hand-rolled frontmatter reader could not express nested
70
+ maps, so it misread every OKF `generated`, `sources[]`, and `verified[]`. Its
71
+ replacement's first log format was `·`-delimited, with a splitter to read it
72
+ back. Both are gone: the log is JSONL and the schema is emitted from Zod, so
73
+ `strauss-kb schema` is the contract rather than a description of one.
74
+
75
+ ## Rejected for now: a base registry
76
+
77
+ Cross-base questions are unaskable by construction — supersession, traces, and
78
+ search stop at the directory boundary. That is the price of a base that can be
79
+ copied, deleted, or handed over whole, and it is what keeps the search index
80
+ disposable.
81
+
82
+ If cross-base ever becomes the common case, the cheap escape is a registry: a
83
+ list of paths a caller may name explicitly, queried one at a time and merged
84
+ only for display. It is deliberately unbuilt. Adding it early would drag back
85
+ the cross-scope machinery this model exists to avoid.
package/README.md CHANGED
@@ -146,12 +146,21 @@ the name is taken — two writers choosing one concept id is a 409 the caller mu
146
146
  answer, by picking a more specific slug or by saying it meant to replace.
147
147
  `rename` is used only when the caller passes `overwrite`.
148
148
 
149
+ Read-modify-write (`setStatus`, `answer`) checks a content digest immediately
150
+ before publishing, which narrows the lost-update window rather than closing it.
151
+ [ARCHITECTURE.md](./ARCHITECTURE.md) says why a lock was rejected.
152
+
149
153
  `supersede` writes both directions, so a backlink cannot drift in normal use and
150
154
  `validate` drops to catching hand-edits.
151
155
 
152
156
  Records are never deleted. Superseding keeps the earlier reasoning inspectable,
153
157
  which is what a later `trace` reads.
154
158
 
159
+ `no-decision` records the explicit claim that a piece of work had nothing to
160
+ decide (an idempotent `decision.none` record). It exists for workflow gates:
161
+ "did you write a decision?" rewards writing a junk one, "did you answer?"
162
+ does not — so silence has to be expressible.
163
+
155
164
  ## CLI
156
165
 
157
166
  ```
@@ -172,6 +181,11 @@ strauss-kb [--bundle PATH] <command> [args]
172
181
  validate Cross-record checks. Exits 1 when it reports a problem.
173
182
  schema JSON Schema for the format.
174
183
  types The twelve types, their sections and initial status.
184
+ pin [bundle-path] [flags] Pin a base. --mode, --profiles, --frozen; --local/--user pick the layer.
185
+ unpin [bundle-path] Remove a base from every manifest layer that holds it.
186
+ pins Every pinned base, with whether it resolves to records.
187
+ context [--profile NAME] [--budget N] The pinned-base index block, for injection at context birth.
188
+ sync-instructions <file> Plant the context block between sentinels in an instruction file.
175
189
 
176
190
  --bundle PATH defaults to ./.strauss/kb
177
191
  STRAUSS_KB_ACTOR names the writer in the log
@@ -202,9 +216,12 @@ strauss-kb validate || echo "problems above"
202
216
  `strauss-kb-mcp` speaks stdio and takes no API key and no required environment.
203
217
  Every CLI verb is a tool: `kb_write`, `kb_write_decision`, `kb_no_decision`,
204
218
  `kb_status`, `kb_supersede`, `kb_answer`, `kb_load`, `kb_query`, `kb_trace`,
205
- `kb_list`, `kb_index`, `kb_log`, `kb_validate`, `kb_schema`, `kb_types`. Every
206
- tool but `kb_schema` and `kb_types` takes a `bundlePath`; those two describe the
207
- format rather than any one base.
219
+ `kb_list`, `kb_index`, `kb_log`, `kb_validate`, `kb_schema`, `kb_types`,
220
+ `kb_pin`, `kb_unpin`, `kb_pins`, `kb_context`. Most tools take a `bundlePath`;
221
+ `kb_schema` and `kb_types` describe the format rather than any one base, and
222
+ `kb_pins` and `kb_context` read the workspace pin manifests instead. The one
223
+ CLI verb with no tool is `sync-instructions` — file plumbing for hooks, not an
224
+ agent capability; the capability is `kb_context`.
208
225
 
209
226
  ```json
210
227
  {
@@ -265,12 +282,20 @@ base answered eight; embedding search over the same records answered four. Two
265
282
  of those differences are structural rather than matters of degree: a reader can
266
283
  say no record answers the question, where vector search returns its nearest
267
284
  neighbour whatever the distance; and a reader picks the record that answers the
268
- question rather than the one nearest the topic.
285
+ question rather than the one nearest the topic. The mechanism is simple:
286
+ retrieval makes similarity the gatekeeper, and a match the ranker misses never
287
+ reaches the model. A full read lets the model do the matching itself —
288
+ synonymy, implication across records, aggregation — which no ranker does.
289
+
290
+ Read for a question, not for a session: a base loaded at the start of a long
291
+ conversation is summarised away by the end of it, and reloading costs about
292
+ three thousand tokens. Read it again at the point of use.
269
293
 
270
294
  `load` refuses rather than truncating when a base exceeds its budget (25,000
271
295
  tokens by default). A truncated base is indistinguishable from a complete one,
272
296
  so a caller would answer "that was never decided" from a slice it did not know
273
- was a slice. Superseded records come back as name, replacement and date only
297
+ was a slice. `context` refuses the same way at its own, tighter budget (4,000
298
+ by default). Superseded records come back as name, replacement and date only —
274
299
  their bodies no longer hold, and a body read later in a long session outlives
275
300
  the qualifier that said so. `trace` still reaches them by id.
276
301
 
@@ -294,6 +319,80 @@ fact; a missing replacement is `broken-chain` with no head — the case that nee
294
319
  the most care, because returning the stale record unmarked looks exactly like
295
320
  success.
296
321
 
322
+ ## Living in an agent session
323
+
324
+ Long sessions lose a knowledge base twice over: attention decays, and
325
+ compaction summarises away both the records loaded early and the instruction
326
+ that said to consult them. The fix is two-tier: a small index is re-injected
327
+ at every context birth, and record bodies are fetched by tool when a question
328
+ actually needs them.
329
+
330
+ ```bash
331
+ strauss-kb pin docs/kb # mark a base every session should see
332
+ strauss-kb context # emit the pinned index block
333
+ strauss-kb sync-instructions AGENTS.md # or keep it in an instruction file
334
+ ```
335
+
336
+ Pins live in `.strauss/kb-pins.json`, committed with the repo. Two more
337
+ layers exist: `.strauss/kb-pins.local.json` (personal, gitignore it) and
338
+ `~/.strauss/kb-pins.json` (every workspace). Nearest layer wins per base,
339
+ `--local`/`--user` write the other layers, and `unpin` removes from all
340
+ three. A malformed layer is skipped on read and refused on write.
341
+
342
+ Per pin:
343
+
344
+ - `--mode full` — inject the records themselves, not just the index. For
345
+ small or critical bases (ADRs). Falls back to a labelled index when it
346
+ cannot fit the block budget.
347
+ - `--mode index` — never inject bodies.
348
+ - `--profiles a,b` — only inject in the named profiles.
349
+ - `--frozen` — the base is concluded; write commands refuse until `--unfreeze`.
350
+
351
+ Budgets are named profiles — `session-start`, `compact`, `turn` — with
352
+ per-repo overrides in the manifest, so hook commands never carry numbers:
353
+
354
+ ```json
355
+ {
356
+ "pins": [{ "path": "docs/adr", "mode": "full" }],
357
+ "context": { "compact": { "budgetTokens": 1500 } }
358
+ }
359
+ ```
360
+
361
+ Flags beat the manifest, the manifest beats the built-ins, and invalid values
362
+ fall back to defaults instead of silencing the index. Past its budget,
363
+ `context` refuses like `load` does — never truncates — and its refusal says
364
+ what to load directly and how to shrink the block.
365
+
366
+ `sync-instructions <file>` keeps the same block between
367
+ `<!-- strauss-kb:begin/end -->` sentinels in AGENTS.md or CLAUDE.md, touching
368
+ nothing outside them. Re-run it when pins change; it is idempotent. This is
369
+ the mechanism for runtimes without a reliable post-compaction hook, since
370
+ instruction files are re-read where conversation history is not.
371
+
372
+ What each runtime gets (configs in the
373
+ [plugin's adapters](../../plugins/strauss-kb/adapters/)):
374
+
375
+ | Layer | Claude Code | Codex CLI | Antigravity CLI |
376
+ | ------------------------- | ------------------ | ------------------------------------------- | -------------------------- |
377
+ | MCP tool descriptions | ✓ | ✓ | ✓ |
378
+ | Session-start injection | SessionStart hook | SessionStart hook | PreInvocation, per turn |
379
+ | Post-compact re-injection | ✓ `compact` source | ✓ client-side; instruction-only when hosted | moot — injected every turn |
380
+ | File-read blocking | opt-in PreToolUse | ✗ (shell is the side door) | opt-in PreToolUse, JSON |
381
+ | Instruction file | CLAUDE.md | AGENTS.md | AGENTS.md + rules/ |
382
+
383
+ One more thing agents add: file tools. A raw read of a record file bypasses
384
+ standing entirely — a superseded record reads exactly like a current one — so
385
+ bases are read through the tools, and a workspace can enforce that with deny
386
+ rules or the plugin's opt-in PreToolUse script:
387
+
388
+ ```json
389
+ {
390
+ "permissions": {
391
+ "deny": ["Read(.strauss/kb/**)", "Read(**/.strauss/kb/**)"]
392
+ }
393
+ }
394
+ ```
395
+
297
396
  ## Optional search tier
298
397
 
299
398
  `@tobilu/qmd` is an **optional peer dependency** providing BM25 (`searchLex`,
@@ -333,7 +432,8 @@ hold is the exception — no invariant, deterministic path.
333
432
  the directory boundary. "Was this settled somewhere else?" is answered by a
334
433
  person choosing which base to open. That is the price of a base that can be
335
434
  copied, deleted, or handed over whole, and it is what keeps the search index
336
- disposable.
435
+ disposable. [ARCHITECTURE.md](./ARCHITECTURE.md) covers the registry that would
436
+ lift it, and why it is unbuilt.
337
437
 
338
438
  ## License
339
439
 
@@ -1,7 +1,7 @@
1
1
  import {
2
2
  KB_COMMANDS,
3
3
  KbStore
4
- } from "./chunk-ZSYSHJVZ.js";
4
+ } from "./chunk-HYNAEAPM.js";
5
5
 
6
6
  // src/mcp.ts
7
7
  import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
@@ -18,6 +18,7 @@ function createKbMcpServer() {
18
18
  now: () => (/* @__PURE__ */ new Date()).toISOString()
19
19
  };
20
20
  for (const command of KB_COMMANDS) {
21
+ if (!command.tool) continue;
21
22
  server.registerTool(
22
23
  command.tool,
23
24
  { description: command.description, inputSchema: command.input.shape },
@@ -44,4 +45,4 @@ export {
44
45
  createKbMcpServer,
45
46
  runKbMcpServer
46
47
  };
47
- //# sourceMappingURL=chunk-WFHYWZX5.js.map
48
+ //# sourceMappingURL=chunk-FSI4Q2FD.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"sources":["../src/mcp.ts"],"sourcesContent":["import { McpServer } from \"@modelcontextprotocol/sdk/server/mcp.js\";\nimport { StdioServerTransport } from \"@modelcontextprotocol/sdk/server/stdio.js\";\nimport { KB_COMMANDS } from \"./commands/index.js\";\nimport { KbStore } from \"./kb-store.js\";\n\n/**\n * A knowledge base's own MCP server, over stdio.\n *\n * Standalone because a base is self-contained: a directory of markdown that\n * needs no database, no HTTP surface, and no running service to read. Folding\n * these tools into a larger server would make every consumer start that server\n * to open files it could open itself.\n *\n * Every tool is a projection of `KB_COMMANDS`, which the CLI also projects, so\n * the two cannot drift.\n */\nexport function createKbMcpServer(): McpServer {\n const server = new McpServer({ name: \"strauss-kb\", version: \"0.1.0\" });\n const store = new KbStore({\n warn: (entry) => process.stderr.write(`${JSON.stringify(entry)}\\n`),\n });\n const ctx = {\n store,\n actor: process.env.STRAUSS_KB_ACTOR ?? \"mcp\",\n now: () => new Date().toISOString(),\n };\n\n for (const command of KB_COMMANDS) {\n // CLI-only plumbing (sync-instructions) edits files for hooks and\n // instruction blocks; the agent capability it serves is kb_context.\n if (!command.tool) continue;\n server.registerTool(\n command.tool,\n { description: command.description, inputSchema: command.input.shape },\n async (args: unknown) => {\n const result = await command.run(ctx, command.input.parse(args));\n return {\n content: [\n {\n type: \"text\" as const,\n text:\n typeof result === \"string\"\n ? result\n : JSON.stringify(result, null, 2),\n },\n ],\n };\n },\n );\n }\n\n return server;\n}\n\nexport async function runKbMcpServer(): Promise<void> {\n await createKbMcpServer().connect(new StdioServerTransport());\n}\n"],"mappings":";;;;;;AAAA,SAAS,iBAAiB;AAC1B,SAAS,4BAA4B;AAe9B,SAAS,oBAA+B;AAC7C,QAAM,SAAS,IAAI,UAAU,EAAE,MAAM,cAAc,SAAS,QAAQ,CAAC;AACrE,QAAM,QAAQ,IAAI,QAAQ;AAAA,IACxB,MAAM,CAAC,UAAU,QAAQ,OAAO,MAAM,GAAG,KAAK,UAAU,KAAK,CAAC;AAAA,CAAI;AAAA,EACpE,CAAC;AACD,QAAM,MAAM;AAAA,IACV;AAAA,IACA,OAAO,QAAQ,IAAI,oBAAoB;AAAA,IACvC,KAAK,OAAM,oBAAI,KAAK,GAAE,YAAY;AAAA,EACpC;AAEA,aAAW,WAAW,aAAa;AAGjC,QAAI,CAAC,QAAQ,KAAM;AACnB,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,EAAE,aAAa,QAAQ,aAAa,aAAa,QAAQ,MAAM,MAAM;AAAA,MACrE,OAAO,SAAkB;AACvB,cAAM,SAAS,MAAM,QAAQ,IAAI,KAAK,QAAQ,MAAM,MAAM,IAAI,CAAC;AAC/D,eAAO;AAAA,UACL,SAAS;AAAA,YACP;AAAA,cACE,MAAM;AAAA,cACN,MACE,OAAO,WAAW,WACd,SACA,KAAK,UAAU,QAAQ,MAAM,CAAC;AAAA,YACtC;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;AAEA,eAAsB,iBAAgC;AACpD,QAAM,kBAAkB,EAAE,QAAQ,IAAI,qBAAqB,CAAC;AAC9D;","names":[]}