pi-braid 0.40.3 → 0.40.6

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.
Files changed (85) hide show
  1. package/AGENTS.md +17 -18
  2. package/README.md +116 -130
  3. package/TOOLS.md +35 -35
  4. package/build-info.json +4 -4
  5. package/dist/extensions/core/agent-environment/index.js +4 -3
  6. package/dist/extensions/core/prompt-builder/index.js +12 -7
  7. package/dist/extensions/core/prompt-builder/instance-composition.d.ts +51 -10
  8. package/dist/extensions/core/prompt-builder/instance-composition.js +87 -14
  9. package/dist/extensions/core/prompt-builder/system-info-tool.d.ts +3 -3
  10. package/dist/extensions/core/prompt-builder/system-info-tool.js +3 -3
  11. package/dist/extensions/core/prompt-builder/tool-guidelines.js +1 -1
  12. package/dist/extensions/core/skills-tool/auto-inject.d.ts +178 -17
  13. package/dist/extensions/core/skills-tool/auto-inject.js +264 -23
  14. package/dist/extensions/core/skills-tool/index.d.ts +70 -2
  15. package/dist/extensions/core/skills-tool/index.js +149 -15
  16. package/dist/extensions/core/skills-tool/reach.d.ts +119 -0
  17. package/dist/extensions/core/skills-tool/reach.js +213 -0
  18. package/dist/extensions/core/skills-tool/skill-triggers.d.ts +74 -13
  19. package/dist/extensions/core/skills-tool/skill-triggers.js +105 -24
  20. package/dist/extensions/enforcement/handlers/register-session-compact.js +12 -1
  21. package/dist/extensions/enforcement/handlers/register-session-shutdown-resumption.js +15 -34
  22. package/dist/extensions/enforcement/handlers/register-session-start-discoverability.d.ts +2 -1
  23. package/dist/extensions/enforcement/handlers/register-session-start-discoverability.js +55 -28
  24. package/dist/extensions/enforcement/handlers/register-turn-end-hygiene.d.ts +14 -10
  25. package/dist/extensions/enforcement/handlers/register-turn-end-hygiene.js +88 -82
  26. package/dist/extensions/enforcement/resume-state.d.ts +76 -0
  27. package/dist/extensions/enforcement/resume-state.js +90 -0
  28. package/dist/extensions/enforcement/substate/context-hygiene-state.d.ts +9 -0
  29. package/dist/extensions/enforcement/substate/context-hygiene-state.js +8 -0
  30. package/dist/extensions/enforcement/whisper.d.ts +12 -0
  31. package/dist/extensions/enforcement/whisper.js +16 -1
  32. package/dist/extensions/heartbeat/heartbeat-scheduler/index.js +14 -6
  33. package/dist/extensions/heartbeat/heartbeat-scheduler/policies.d.ts +7 -0
  34. package/dist/extensions/heartbeat/heartbeat-scheduler/policies.js +372 -0
  35. package/dist/extensions/memory/context-manager/handlers/tools/context-compress.d.ts +8 -2
  36. package/dist/extensions/memory/context-manager/handlers/tools/context-compress.js +54 -21
  37. package/dist/lib/config/braid-paths.d.ts +14 -0
  38. package/dist/lib/config/braid-paths.js +18 -0
  39. package/dist/lib/config/braid-settings.d.ts +21 -0
  40. package/dist/lib/config/braid-settings.js +16 -0
  41. package/dist/lib/detection/context-tiers.d.ts +47 -3
  42. package/dist/lib/detection/context-tiers.js +122 -13
  43. package/dist/lib/events/turn-boundary.d.ts +78 -0
  44. package/dist/lib/events/turn-boundary.js +91 -0
  45. package/dist/lib/orchestrator/workers.d.ts +3 -2
  46. package/dist/lib/orchestrator/workers.js +22 -4
  47. package/dist/lib/state/global-state.d.ts +1 -1
  48. package/docs/CONTEXT-POLICY.md +8 -0
  49. package/docs/LOCAL_DEV.md +0 -1
  50. package/docs/ROADMAP.md +5 -5
  51. package/docs/SETUP.md +19 -18
  52. package/package.json +2 -1
  53. package/skills/herdr-topology/SKILL.md +6 -1
  54. package/skills/research/SKILL.md +32 -0
  55. package/skills/subagent-orchestration/SKILL.md +2 -2
  56. package/src/extensions/core/agent-environment/index.ts +4 -3
  57. package/src/extensions/core/prompt-builder/AGENTS.md +8 -4
  58. package/src/extensions/core/prompt-builder/index.ts +12 -7
  59. package/src/extensions/core/prompt-builder/instance-composition.ts +98 -14
  60. package/src/extensions/core/prompt-builder/system-info-tool.ts +3 -3
  61. package/src/extensions/core/prompt-builder/tool-guidelines.ts +1 -1
  62. package/src/extensions/core/skills-tool/auto-inject.ts +350 -27
  63. package/src/extensions/core/skills-tool/index.ts +217 -15
  64. package/src/extensions/core/skills-tool/reach.ts +256 -0
  65. package/src/extensions/core/skills-tool/skill-triggers.ts +130 -23
  66. package/src/extensions/enforcement/handlers/register-session-compact.ts +12 -1
  67. package/src/extensions/enforcement/handlers/register-session-shutdown-resumption.ts +23 -36
  68. package/src/extensions/enforcement/handlers/register-session-start-discoverability.ts +59 -29
  69. package/src/extensions/enforcement/handlers/register-turn-end-hygiene.ts +101 -81
  70. package/src/extensions/enforcement/resume-state.ts +127 -0
  71. package/src/extensions/enforcement/substate/context-hygiene-state.ts +17 -0
  72. package/src/extensions/enforcement/whisper.ts +21 -1
  73. package/src/extensions/heartbeat/heartbeat-scheduler/index.ts +20 -5
  74. package/src/extensions/heartbeat/heartbeat-scheduler/policies.ts +370 -0
  75. package/src/extensions/memory/AGENTS.md +25 -1
  76. package/src/extensions/memory/context-manager/handlers/tools/context-compress.ts +80 -27
  77. package/src/lib/config/braid-paths.ts +20 -0
  78. package/src/lib/config/braid-settings.ts +46 -0
  79. package/src/lib/detection/context-tiers.ts +186 -13
  80. package/src/lib/events/AGENTS.md +1 -0
  81. package/src/lib/events/turn-boundary.ts +115 -0
  82. package/src/lib/orchestrator/workers.ts +33 -12
  83. package/src/lib/state/global-state.ts +1 -0
  84. package/src/lib/utils/AGENTS.md +1 -1
  85. package/src/types/extension-context.d.ts +22 -79
@@ -10,7 +10,19 @@
10
10
  * The cache lives in this module and is reset each session.
11
11
  *
12
12
  * One action:
13
- * list — paginated catalog (name, filePath, description, source)
13
+ * list — paginated catalog (name, filePath, description, source, reach),
14
+ * ordered by relevance to the current task unless `rank: "catalog"`.
15
+ *
16
+ * Each row carries the price of the read it names (see reach.ts): reachable,
17
+ * lines, bytes, chars, a token estimate, the file's heading count, and as many
18
+ * of its headings as the caller asked for. Picking a skill is picking a read,
19
+ * and a price is what makes that pick rational.
20
+ *
21
+ * The ordering is the deterministic scorer's answer (see `resolveRanking`):
22
+ * auto-inject.ts owns the ranking, and this handler owns only what needs the
23
+ * extension context — the task text (from the session branch) and the
24
+ * `braid.skills.ranking` policy, which is read and echoed rather than acted on
25
+ * while no model tier exists.
14
26
  *
15
27
  * Skills discovery (list_skills tool) plus the scoring-based auto-inject
16
28
  * whisper supersede an LLM top-N recommender: a second relevance engine
@@ -20,17 +32,28 @@
20
32
  * Also registers `skill_stats` — the adoption counter surface.
21
33
  */
22
34
 
23
- import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
35
+ import type { ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-agent";
24
36
  import { Type } from "typebox";
25
37
  import { createLogger } from "../../../lib/core/logger.js";
38
+ import { getSkillsRankingMode } from "../../../lib/config/braid-settings.js";
39
+ import type { SkillsRankingMode } from "../../../lib/config/braid-settings.js";
26
40
  import {
27
41
  registerAutoInject,
28
42
  updateSkillCache,
29
43
  getSkillStats,
30
44
  getDurableSkillStats,
31
45
  resolveDiscoveryMode,
46
+ rankSkillsDeterministic,
32
47
  } from "./auto-inject.js";
33
- import { readSkillTriggers } from "./skill-triggers.js";
48
+ import type { DeterministicRanking, RankedSkill, RelevanceReason } from "./auto-inject.js";
49
+ import { readSkillTriggerEntries } from "./skill-triggers.js";
50
+ import type { TriggerSource } from "./skill-triggers.js";
51
+ import {
52
+ clampSectionsMax,
53
+ DEFAULT_SECTIONS_MAX,
54
+ SECTIONS_MAX_LIMIT,
55
+ readSkillReach,
56
+ } from "./reach.js";
34
57
  import { getMonthlyAdoptionRollup } from "./session-records.js";
35
58
 
36
59
  const log = createLogger("skills-tool");
@@ -51,6 +74,8 @@ interface CachedSkill {
51
74
  disableModelInvocation?: boolean;
52
75
  /** Frontmatter trigger phrases, resolved from SKILL.md at cache time. */
53
76
  triggers?: string[];
77
+ /** Which source authored each phrase (a ranking row reports it). */
78
+ triggerSources?: Record<string, TriggerSource>;
54
79
  }
55
80
 
56
81
  let cachedSkills: CachedSkill[] = [];
@@ -62,10 +87,12 @@ let cachedSkills: CachedSkill[] = [];
62
87
  * shape), and the memoized read makes repeat turns stat-only.
63
88
  */
64
89
  export function setCachedSkills(skills: CachedSkill[]): void {
65
- cachedSkills = skills.map((s) => ({
66
- ...s,
67
- triggers: readSkillTriggers(s.filePath, s.description),
68
- }));
90
+ cachedSkills = skills.map((s) => {
91
+ const entries = readSkillTriggerEntries(s.filePath, s.description);
92
+ const triggerSources: Record<string, TriggerSource> = {};
93
+ for (const entry of entries) triggerSources[entry.phrase] = entry.source;
94
+ return { ...s, triggers: entries.map((entry) => entry.phrase), triggerSources };
95
+ });
69
96
  // Sync with auto-inject subsystem (enriched cache — triggers ride along)
70
97
  updateSkillCache(cachedSkills);
71
98
  log.debug(`Cached ${skills.length} skills for list_skills tool`);
@@ -92,6 +119,137 @@ function formatSource(skill: CachedSkill): string {
92
119
  return src;
93
120
  }
94
121
 
122
+ // Relevance ranking: the deterministic order
123
+ //
124
+ // A model tier was built here and measured against this scorer: it never lost
125
+ // a case and won the two prose prompts the scorer misses, but it did not clear
126
+ // its pre-registered bar, and a tier that buys two wins with a call on every
127
+ // ambiguous read is a finding about the scorer, not a feature. So the read is
128
+ // answered by rankSkillsDeterministic() alone: no provider, no cache, no
129
+ // counters. The `braid.skills.ranking` setting and the `ranking` report
130
+ // survive because they are the seam a later tier plugs into.
131
+
132
+ /** The evidence label a row reports (see RelevanceReason in auto-inject.ts). */
133
+ export type RowRelevanceReason = RelevanceReason;
134
+
135
+ /** What a ranked row says about its own position. */
136
+ export interface RowRelevance {
137
+ /** The deterministic score behind the row. */
138
+ score: number;
139
+ reason: RowRelevanceReason;
140
+ }
141
+
142
+ /** Why the deterministic verdict is what it is, and who answered. */
143
+ export interface RankingReport {
144
+ /** Always "keyword": the deterministic scorer is the only tier. */
145
+ tier: "keyword";
146
+ basis: DeterministicRanking["basis"];
147
+ /**
148
+ * The configured `braid.skills.ranking` value, echoed so an operator can see
149
+ * what the policy asked for. Only "keyword" has an implementation today; the
150
+ * other two are reserved for a model tier that was measured and not kept, and
151
+ * they evaluate to the keyword path.
152
+ */
153
+ policy: SkillsRankingMode;
154
+ }
155
+
156
+ /**
157
+ * The current task text: the newest user message on the live branch.
158
+ *
159
+ * A tool call carries no prompt of its own and pi hands tools no turn prompt,
160
+ * but the branch does — and its newest user message is the task the agent is
161
+ * working on. A resumed session therefore ranks against its real task even
162
+ * though this process never saw the prompt arrive. Best-effort: no branch, an
163
+ * unreadable entry, or a non-text message yields "", and the caller then
164
+ * returns the catalog order rather than inventing a task.
165
+ */
166
+ export function currentTaskText(ctx: ExtensionContext): string {
167
+ try {
168
+ const sm = (ctx as { sessionManager?: { getBranch?: () => unknown[] } }).sessionManager;
169
+ const branch = sm?.getBranch?.() ?? [];
170
+ for (let i = branch.length - 1; i >= 0; i--) {
171
+ const entry = branch[i] as { type?: string; message?: { role?: string; content?: unknown } };
172
+ if (entry?.type !== "message" || entry.message?.role !== "user") continue;
173
+ const content = entry.message.content;
174
+ if (typeof content === "string") return content;
175
+ if (Array.isArray(content)) {
176
+ return content
177
+ .map((part) => part as { type?: string; text?: string })
178
+ .filter((part) => part?.type === "text" && typeof part.text === "string")
179
+ .map((part) => part.text as string)
180
+ .join("\n");
181
+ }
182
+ return "";
183
+ }
184
+ } catch {
185
+ // Best-effort: a stale ctx or an absent tree yields no text.
186
+ }
187
+ return "";
188
+ }
189
+
190
+ /** Reorder `skills` by `names`, appending everything the list did not mention. */
191
+ function orderByNames(skills: CachedSkill[], names: string[]): CachedSkill[] {
192
+ const byName = new Map(skills.map((s) => [s.name, s]));
193
+ const out: CachedSkill[] = [];
194
+ for (const name of names) {
195
+ const skill = byName.get(name);
196
+ if (!skill) continue;
197
+ byName.delete(name);
198
+ out.push(skill);
199
+ }
200
+ for (const skill of skills) if (byName.has(skill.name)) out.push(skill);
201
+ return out;
202
+ }
203
+
204
+ /** One catalog read's complete ranking answer. */
205
+ export interface RankingAnswer {
206
+ order: CachedSkill[];
207
+ /** Absent for `rank: "catalog"` — an unranked read has no tier or evidence. */
208
+ ranking?: RankingReport;
209
+ relevance: Map<string, RowRelevance>;
210
+ }
211
+
212
+ /**
213
+ * Resolve the catalog order for one read.
214
+ *
215
+ * `keyword` and `catalog` are the two answers: the deterministic ranking (with
216
+ * its verdict on how confident that ranking is), or the plain catalog. A
217
+ * missing task text is nobody's question, so the catalog order stands and the
218
+ * report says `no-task` rather than pretending to rank.
219
+ */
220
+ export function resolveRanking(
221
+ ctx: ExtensionContext,
222
+ skills: CachedSkill[],
223
+ options: { rank: "relevance" | "catalog"; policy?: SkillsRankingMode },
224
+ ): RankingAnswer {
225
+ if (options.rank === "catalog") return { order: skills, relevance: new Map() };
226
+
227
+ const det = rankSkillsDeterministic(skills, currentTaskText(ctx));
228
+ const relevance = new Map<string, RowRelevance>();
229
+ for (const row of det.ranked) relevance.set(row.skill.name, rowOf(row));
230
+ const order =
231
+ det.basis === "no-task"
232
+ ? skills
233
+ : orderByNames(
234
+ skills,
235
+ det.ranked.map((row) => row.skill.name),
236
+ );
237
+ return {
238
+ order,
239
+ ranking: {
240
+ tier: "keyword",
241
+ basis: det.basis,
242
+ policy: options.policy ?? getSkillsRankingMode(),
243
+ },
244
+ relevance,
245
+ };
246
+ }
247
+
248
+ /** A ranked row as the response reports it (only rows with evidence get here). */
249
+ function rowOf(row: RankedSkill): RowRelevance {
250
+ return { score: row.score, reason: row.reason ?? "description" };
251
+ }
252
+
95
253
  // Extension factory:
96
254
 
97
255
  export default function skillsToolExtension(pi: ExtensionAPI): void {
@@ -101,10 +259,11 @@ export default function skillsToolExtension(pi: ExtensionAPI): void {
101
259
  name: "list_skills",
102
260
  label: "List Skills",
103
261
  description:
104
- "Discover skills on demand: paginated catalog (name, filePath, description, source); `read` a SKILL.md to load it.",
262
+ "Discover skills on demand: paginated catalog (name, filePath, description, source) plus each skill's reach (reachable, lines, bytes, chars, estTokens, sectionCount, section headings with line numbers), so the price of a read is visible before the read. Rows are ordered by relevance to the current task; a row with match evidence carries relevance {score, reason}, and the response carries ranking {tier, basis, policy}. Pass rank: 'catalog' for the plain catalog order. `read` a SKILL.md to load it.",
105
263
  parameters: Type.Object({
106
264
  action: Type.Literal("list", {
107
- description: "list = paginated catalog (name, filePath, description, source)",
265
+ description:
266
+ "list = paginated catalog with per-skill reach (reachable, lines, bytes, chars, estTokens, sectionCount, sections)",
108
267
  }),
109
268
  page: Type.Optional(
110
269
  Type.Number({
@@ -121,17 +280,36 @@ export default function skillsToolExtension(pi: ExtensionAPI): void {
121
280
  description: "Entries per page for list",
122
281
  }),
123
282
  ),
283
+ sectionsMax: Type.Optional(
284
+ Type.Number({
285
+ minimum: 0,
286
+ maximum: SECTIONS_MAX_LIMIT,
287
+ default: DEFAULT_SECTIONS_MAX,
288
+ description:
289
+ "Headings reported per skill; 0 omits the section list and keeps sectionCount",
290
+ }),
291
+ ),
292
+ rank: Type.Optional(
293
+ Type.Union([Type.Literal("relevance"), Type.Literal("catalog")], {
294
+ default: "relevance",
295
+ description:
296
+ "relevance (default) = order by fit to the current task; catalog = the plain catalog order, with no ranking fields",
297
+ }),
298
+ ),
124
299
  }),
125
- execute: async (_id, params, _signal, _onUpdate, _ctx) => {
300
+ execute: async (_id, params, signal, _onUpdate, ctx) => {
126
301
  const page = (params.page as number) ?? 1;
127
302
  const pageSize = (params.pageSize as number) ?? 10;
128
- const start = (page - 1) * pageSize;
303
+ const sectionsMax = clampSectionsMax(params.sectionsMax);
304
+ const rank = (params.rank as "relevance" | "catalog" | undefined) ?? "relevance";
129
305
  // Explicit-invocation-only skills stay out of the catalog; the cache
130
306
  // itself stays unfiltered for the read-detector that counts explicit
131
307
  // loads.
132
308
  const visible = cachedSkills.filter((s) => !s.disableModelInvocation);
133
- const slice = visible.slice(start, start + pageSize);
134
- const total = visible.length;
309
+ const answer = resolveRanking(ctx, visible, { rank });
310
+ const total = answer.order.length;
311
+ const start = (page - 1) * pageSize;
312
+ const slice = answer.order.slice(start, start + pageSize);
135
313
  const hasMore = start + pageSize < total;
136
314
 
137
315
  const skills = slice.map((s) => ({
@@ -139,16 +317,40 @@ export default function skillsToolExtension(pi: ExtensionAPI): void {
139
317
  filePath: s.filePath,
140
318
  description: (s.description ?? "").slice(0, 280),
141
319
  source: formatSource(s),
320
+ ...readSkillReach(s.filePath, sectionsMax),
321
+ // Evidence where it exists: a row with no match carries no relevance
322
+ // field. Every row paying ~60 bytes for "nothing matched" pushed the
323
+ // shipped page past its measured prompt budget (12.0KB -> 12.8KB), and
324
+ // the empty object carried no information the absence does not.
325
+ ...(answer.relevance.has(s.name) ? { relevance: answer.relevance.get(s.name) } : {}),
142
326
  }));
143
327
 
144
328
  return {
145
329
  content: [
146
330
  {
147
331
  type: "text" as const,
148
- text: JSON.stringify({ page, pageSize, total, hasMore, skills }, null, 2),
332
+ text: JSON.stringify(
333
+ {
334
+ page,
335
+ pageSize,
336
+ total,
337
+ hasMore,
338
+ ...(answer.ranking ? { ranking: answer.ranking } : {}),
339
+ skills,
340
+ },
341
+ null,
342
+ 2,
343
+ ),
149
344
  },
150
345
  ],
151
- details: { total, page, pageSize, hasMore, returned: skills.length },
346
+ details: {
347
+ total,
348
+ page,
349
+ pageSize,
350
+ hasMore,
351
+ returned: skills.length,
352
+ ...(answer.ranking ? { tier: answer.ranking.tier, basis: answer.ranking.basis } : {}),
353
+ },
152
354
  };
153
355
  },
154
356
  });
@@ -0,0 +1,256 @@
1
+ /**
2
+ * What a SKILL.md costs before it is read.
3
+ *
4
+ * `list_skills` is a choice between reads, and a row that names a path but
5
+ * prices nothing leaves that choice blind: a 650-byte pointer and a 25KB
6
+ * reference look identical until one of them is already in context. The fields
7
+ * here are the price - line count, byte count, character count, a token
8
+ * estimate, and the file's ATX headings with line numbers, which is the
9
+ * cheapest useful table of contents for "read section 88 and stop".
10
+ *
11
+ * Counts are the whole file, frontmatter included: the question a reader is
12
+ * asking is what the `read` call will cost, not what the body costs. `chars`
13
+ * is `String.length`, so it counts UTF-16 code units and not code points: text
14
+ * outside the BMP counts twice, and `estTokens` inherits that over-count. A
15
+ * byte-priced budget is what this row feeds, `chars` is the name callers
16
+ * already read, and the doc block is where the difference is stated.
17
+ * Lines end at a newline; a last line without a terminating newline still
18
+ * counts, because the read pays for it.
19
+ *
20
+ * One read per change. The cache is keyed on mtime + size, the same
21
+ * invalidation `skill-triggers.ts` uses: a page render stats every row, and
22
+ * only a first-seen or changed file is opened and parsed. Failures are not
23
+ * cached, because a chmod that took away read access leaves mtime untouched -
24
+ * a cached failure would outlive the repair.
25
+ *
26
+ * Headings are parsed fence-aware. Skill files carry long fenced blocks of
27
+ * shell, TypeScript and markdown samples whose comment lines start with `#`.
28
+ * Over the 22 skills of the shipped pack, 293 lines are shaped like headings
29
+ * and 248 are headings: 45 phantom entries, 1.18x. The ratio is not the
30
+ * argument, the entries are. Those 45 lines are shell and TypeScript
31
+ * commentary, and a fence-blind parse advertises them as document structure,
32
+ * sending a reader to a line that holds no section.
33
+ *
34
+ * Two fence edges are known limits, not open bugs. A fence opener indented
35
+ * four spaces is not a fence - the fence and ATX rules both stop at three
36
+ * spaces - so `#` lines inside such a block are reported as headings. An
37
+ * unclosed fence swallows every later heading, which is CommonMark's rule too
38
+ * and costs a truncated file its table of contents. Neither occurs in the
39
+ * shipped pack. Fixing either is a behaviour change with its own evidence, so
40
+ * this block is where they are recorded rather than where they are repaired.
41
+ *
42
+ * The default page is a budget, not a ceiling. A row carries at most
43
+ * SECTIONS_MAX_LIMIT headings and `list_skills` serves at most 50 rows per
44
+ * page, so one call can ask for tens of kilobytes: `pageSize: 50` with
45
+ * `sectionsMax: 64` over the shipped pack is 35852 bytes, nearly three times
46
+ * the 12KB default page. That is the caller's choice, priced here so it is not
47
+ * a surprise; the budget below binds on the default arguments only.
48
+ */
49
+
50
+ import { readFileSync, statSync } from "node:fs";
51
+ import { resolve } from "node:path";
52
+
53
+ /**
54
+ * Characters per token. The same divisor the pre-send guard uses
55
+ * (`packages/providers/presend-guard.ts`, `CHARS_PER_TOKEN`) and the one
56
+ * pi-core's `estimateTokens` applies, so a reach estimate and a payload
57
+ * estimate are comparable numbers rather than two private conventions.
58
+ * Declared here instead of imported: providers is a standalone package with no
59
+ * import path into core.
60
+ */
61
+ export const CHARS_PER_TOKEN = 4;
62
+
63
+ /**
64
+ * Headings reported per skill when the caller does not ask for another cap.
65
+ *
66
+ * Six is a prompt-budget number, not a taste one: over the shipped pack the
67
+ * heaviest ten-row page of the pretty-printed catalog measures 11984 bytes at
68
+ * six headings, 304 bytes under the 12KB bound, and 13568 at eight, where it
69
+ * crosses. Callers who want the full contents pass `sectionsMax` (up to
70
+ * SECTIONS_MAX_LIMIT).
71
+ */
72
+ export const DEFAULT_SECTIONS_MAX = 6;
73
+
74
+ /** A heading is a table-of-contents entry, not a paragraph. */
75
+ const SECTION_TEXT_MAX = 80;
76
+
77
+ /**
78
+ * Upper bound on the per-skill heading cap - a request above it is a bug.
79
+ * `list_skills` declares this same value as its schema maximum, so the ceiling
80
+ * a caller meets through the tool and the ceiling `clampSectionsMax` enforces
81
+ * cannot drift into two numbers.
82
+ */
83
+ export const SECTIONS_MAX_LIMIT = 64;
84
+
85
+ /**
86
+ * ATX heading: up to three spaces of indent (four would be an indented code
87
+ * block), one to six `#`, then a space or the end of the line - so `#region`
88
+ * in a TypeScript sample is not a heading.
89
+ */
90
+ const ATX_HEADING = /^ {0,3}(#{1,6})(?:[ \t]+(.*))?$/;
91
+
92
+ /**
93
+ * A heading's optional closing sequence: `## Title ##`, trailing whitespace
94
+ * allowed - `## Title ## ` closes the same way, and the spaces after the run
95
+ * are not part of the text.
96
+ */
97
+ const CLOSING_HASHES = /[ \t]+#+[ \t]*$/;
98
+
99
+ /** A fence line: up to three spaces of indent, then three or more backticks or tildes. */
100
+ const FENCE = /^ {0,3}(`{3,}|~{3,})(.*)$/;
101
+
102
+ export interface SkillSection {
103
+ /** 1-indexed line the heading sits on. */
104
+ line: number;
105
+ /** Heading text with its leading `#`s stripped, trimmed, capped. */
106
+ text: string;
107
+ }
108
+
109
+ export interface SkillReach {
110
+ reachable: true;
111
+ lines: number;
112
+ bytes: number;
113
+ /** `String.length`: UTF-16 code units, so astral-plane text counts twice. */
114
+ chars: number;
115
+ estTokens: number;
116
+ /** Total headings in the file, independent of the per-call report cap. */
117
+ sectionCount: number;
118
+ /** Omitted when the caller asked for no section list. */
119
+ sections?: SkillSection[];
120
+ }
121
+
122
+ export interface UnreachableReach {
123
+ reachable: false;
124
+ /** Stat size when stat() succeeded and the read failed; absent otherwise. */
125
+ bytes?: number;
126
+ }
127
+
128
+ export type Reach = SkillReach | UnreachableReach;
129
+
130
+ interface ReachCacheEntry {
131
+ mtimeMs: number;
132
+ size: number;
133
+ reach: SkillReach;
134
+ }
135
+
136
+ const cache = new Map<string, ReachCacheEntry>();
137
+
138
+ /**
139
+ * Lines in a file's text: a trailing newline ends the last line rather than
140
+ * opening a blank one, an empty file has no lines, and a final line with no
141
+ * terminating newline counts, because the partial line is content the read
142
+ * pays for.
143
+ */
144
+ export function countLines(text: string): number {
145
+ if (text === "") return 0;
146
+ let newlines = 0;
147
+ for (let i = 0; i < text.length; i++) if (text[i] === "\n") newlines++;
148
+ return text.endsWith("\n") ? newlines : newlines + 1;
149
+ }
150
+
151
+ /**
152
+ * Headings in document order, skipping fenced blocks. A fence opens on a run
153
+ * of backticks or tildes (with or without an info string) and closes only on a
154
+ * run of the same character at least as long, so a `~~~` sample inside a
155
+ * backtick fence is content, not a delimiter. A backtick run followed by more
156
+ * backticks (` ```a`` `) is not a fence at all, per CommonMark.
157
+ */
158
+ function scanHeadings(text: string): SkillSection[] {
159
+ const headings: SkillSection[] = [];
160
+ let fence: { marker: string; length: number } | null = null;
161
+ const lines = text.split("\n");
162
+ for (let i = 0; i < lines.length; i++) {
163
+ const line = lines[i].endsWith("\r") ? lines[i].slice(0, -1) : lines[i];
164
+ const fenceLine = FENCE.exec(line);
165
+ if (fence) {
166
+ if (
167
+ fenceLine &&
168
+ fenceLine[1][0] === fence.marker &&
169
+ fenceLine[1].length >= fence.length &&
170
+ fenceLine[2].trim() === ""
171
+ ) {
172
+ fence = null;
173
+ }
174
+ continue;
175
+ }
176
+ if (fenceLine && !(fenceLine[1][0] === "`" && fenceLine[2].includes("`"))) {
177
+ fence = { marker: fenceLine[1][0], length: fenceLine[1].length };
178
+ continue;
179
+ }
180
+ const heading = ATX_HEADING.exec(line);
181
+ if (!heading) continue;
182
+ const text = (heading[2] ?? "").replace(CLOSING_HASHES, "").trim();
183
+ headings.push({
184
+ line: i + 1,
185
+ text: text.length > SECTION_TEXT_MAX ? text.slice(0, SECTION_TEXT_MAX) : text,
186
+ });
187
+ }
188
+ return headings;
189
+ }
190
+
191
+ /** Clamp a caller-supplied heading cap into range; 0 means "no list". */
192
+ export function clampSectionsMax(value: unknown): number {
193
+ const n = typeof value === "number" && Number.isFinite(value) ? Math.floor(value) : NaN;
194
+ if (Number.isNaN(n)) return DEFAULT_SECTIONS_MAX;
195
+ if (n <= 0) return 0;
196
+ return Math.min(n, SECTIONS_MAX_LIMIT);
197
+ }
198
+
199
+ /**
200
+ * Reach for one SKILL.md. A file that cannot be stat'ed or read answers
201
+ * `reachable: false` instead of throwing: the catalog has to render, and one
202
+ * unreadable skill must not take the page down with it.
203
+ *
204
+ * `sectionsMax` caps the returned heading list; the cached parse keeps the
205
+ * whole list, so two callers asking for different caps share one read.
206
+ */
207
+ export function readSkillReach(
208
+ filePath: string,
209
+ sectionsMax: number = DEFAULT_SECTIONS_MAX,
210
+ ): Reach {
211
+ const key = resolve(filePath);
212
+ let size: number;
213
+ let mtimeMs: number;
214
+ try {
215
+ const stat = statSync(key);
216
+ size = stat.size;
217
+ mtimeMs = stat.mtimeMs;
218
+ } catch {
219
+ cache.delete(key);
220
+ return { reachable: false };
221
+ }
222
+ let entry = cache.get(key);
223
+ if (!entry || entry.mtimeMs !== mtimeMs || entry.size !== size) {
224
+ let text: string;
225
+ try {
226
+ text = readFileSync(key, "utf-8");
227
+ } catch {
228
+ cache.delete(key);
229
+ return { reachable: false, bytes: size };
230
+ }
231
+ const headings = scanHeadings(text);
232
+ entry = {
233
+ mtimeMs,
234
+ size,
235
+ reach: {
236
+ reachable: true,
237
+ lines: countLines(text),
238
+ bytes: size,
239
+ chars: text.length,
240
+ estTokens: Math.ceil(text.length / CHARS_PER_TOKEN),
241
+ sectionCount: headings.length,
242
+ sections: headings,
243
+ },
244
+ };
245
+ cache.set(key, entry);
246
+ }
247
+ const { sections: all = [], ...scalars } = entry.reach;
248
+ const cap = clampSectionsMax(sectionsMax);
249
+ if (cap === 0) return scalars;
250
+ return { ...scalars, sections: cap >= all.length ? all : all.slice(0, cap) };
251
+ }
252
+
253
+ /** Test seam: forget every memoized reach measurement. */
254
+ export function clearReachCache(): void {
255
+ cache.clear();
256
+ }