pi-braid 0.43.6 → 0.43.7

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.
package/CHANGELOG.md CHANGED
@@ -53,6 +53,50 @@ record; the tag scheme above applies unchanged.
53
53
 
54
54
  ---
55
55
 
56
+ ## Core — 0.43.7 (2026-10-30: an ignored skill is visible, and two rates say whether the obligation held)
57
+
58
+ - **A skill that applied and was ignored is now distinguishable from one that never applied.** The
59
+ subsystem could tell whether an agent read a skill and could not tell whether it should have. It now
60
+ records, per session and per skill, a verdict with the phrase that decided it and where the skill
61
+ declares it: `read`, `ignored`, or `not-applicable`, plus a census row naming the installed skills whose
62
+ files declare no condition at all. The record is additive by construction: a new record kind on the
63
+ existing store, refused by the adoption reader so the two cannot silently merge, and an existing
64
+ reader's output is byte-identical when the record is absent. `ignored` over `read + ignored` is the
65
+ rate nobody could compute before, and it is computable from the record alone.
66
+ - **The condition is the author's own declaration, not a coincidence.** Applicability reads the trigger
67
+ phrases the injector already matches, with the injector's own word-boundary rule, so the verdict can
68
+ never disagree with the whisper about whether a phrase is present; 1,098 phrase/text pairs over both
69
+ installed packs were checked for that agreement. The addition is polarity: "do not research this" is a
70
+ session where research did not apply. The negation rule is one cue, not a composed one, so a double
71
+ negative reads as negation and the ignored rate is a lower bound; that is pinned by a test with its
72
+ reason rather than left as a surprise.
73
+ - **A skill declares the tool family its body directs, and the session records what it did with it.** Three
74
+ of the shipped skills declare a family, five declare none because a body that governs no tool family
75
+ genuinely has none, and the contract reports both numbers rather than padding the count. The per-session
76
+ leg records `suggested`, `listed`, `read`, `called`, `applied`, where `applied` is ordered: a call made
77
+ before the read cannot be credited, and a read with no call still satisfies. An applicable skill with no
78
+ leg row is counted unsatisfied rather than dropped, which keeps the ignored population inside the
79
+ denominator.
80
+ - **The obligation rate ships with its own falsifier.** `allSatisfied` over the sessions that carry both a
81
+ verdict and a leg row, reported beside the read rate with that denominator stated rather than implied,
82
+ and a gap below five points marks the metric as non-discriminating. On a constructed population the two
83
+ rates separate three satisfied sessions from four that all read, so the number distinguishes what it
84
+ claims to.
85
+ - **The label vocabulary refuses both spellings on both surfaces.** The source gate's vocabulary was
86
+ uppercase while the spelling seats actually typed was lowercase, so 73 handles sat in 22 fixture files
87
+ the gate never saw, plus 23 glued spellings inside them and 11 more in two files the sweep could not
88
+ name. They are renamed to the fixtures' own subjects, the shared set now matches either case, and the
89
+ history gate's two redundant patterns are gone so one subject yields one problem line. A planted token
90
+ reddens the gate at a named line and column; a word character before a handle, two digits, four digits,
91
+ and an ISO calendar week are none of them hits.
92
+ - **A refusal was designed, measured, and killed by its own test.** The candidate enforcement was to refuse
93
+ the first call in a declared family once when the body was never read. Simulated over 703 sessions and
94
+ 107,463 tool calls with the production functions imported, 2 of 30 refused calls changed course (6.7%,
95
+ interval 1.8 to 21.3%), 0 of 30 on a stricter reading, and the gate could reach at most 8% of the calls
96
+ it governs. The best case in the data refutes the shape: the model read the body when the body mattered,
97
+ after the work it needed, without a gate. What ships is the measurement that makes that answer
98
+ possible; what does not ship is the gate.
99
+
56
100
  ## Core — 0.43.6 (2026-10-29: the rule says one thing, and a merge states what landed)
57
101
 
58
102
  - **The identity block contradicted itself, and the contradiction is gone.** One sentence forbade
package/build-info.json CHANGED
@@ -1,9 +1,9 @@
1
1
  {
2
2
  "name": "pi-braid",
3
- "version": "0.43.6",
4
- "sha": "e7ebbc542f5983f2270e96628def7a9b6d654610",
5
- "shaShort": "e7ebbc54",
3
+ "version": "0.43.7",
4
+ "sha": "202c7b26bdcfe036b330cde4d98470951553502f",
5
+ "shaShort": "202c7b26",
6
6
  "branch": "detached",
7
7
  "dirty": false,
8
- "packedAt": "2026-10-29T14:29:10.456Z"
8
+ "packedAt": "2026-10-30T08:48:55.796Z"
9
9
  }
@@ -0,0 +1,134 @@
1
+ /**
2
+ * Skill applicability — did a skill's own declared condition hold for the task?
3
+ *
4
+ * The suggestion whisper, the catalog ranking and the adoption counters all
5
+ * answer "might this skill help?". None answers the question the counters
6
+ * cannot: of the sessions that read nothing, which of them had a skill whose
7
+ * stated condition the task text actually met? A suggestion that was ignored
8
+ * and a skill that never applied looked identical in `suggested`/`loaded`.
9
+ * This module decides that difference and keeps its evidence.
10
+ *
11
+ * The condition is the skill's OWN declaration, never a heuristic invented
12
+ * here: the trigger phrases its file declares — frontmatter `trigger:` /
13
+ * `triggers:`, a `## trigger:` body heading, the description's `Triggers:`
14
+ * tail — the same entries `skill-triggers.ts` reads and the same provenance
15
+ * labels it records. The shipped pack's contract already treats that list as
16
+ * the machine-readable form of applicability ("a description that cannot fire
17
+ * is indistinguishable from one that fires"), so a skill file needs no new
18
+ * field for this module to see it. A file that declares no phrase yields no
19
+ * verdict at all; the caller reports it as undeclared rather than asserting
20
+ * "does not apply" on no evidence.
21
+ *
22
+ * Matching mirrors the injector on purpose. The occurrence boundary is the one
23
+ * `triggerMatches` uses (a phrase occurs where neither neighbour is a letter or
24
+ * digit, so "pane" is not an occurrence in "paneled"), because a verdict that
25
+ * disagreed with the whisper would be worse than no verdict: the suite pins
26
+ * the two answers equal over the installed catalog. What this module adds is
27
+ * polarity — a phrase inside a negated clause ("do not research this") is not
28
+ * the author's condition holding, so the clause is read and the phrase is
29
+ * dropped as evidence.
30
+ *
31
+ * What this module does NOT do: it decides nothing about delivery, slots,
32
+ * cooldowns, counters or ranking. It returns a classification and the caller
33
+ * records it. Nothing here reads the session, writes a file or sends a
34
+ * message.
35
+ *
36
+ * Cost: one pass over the task text per phrase of the cached catalog, with the
37
+ * native substring scan as the cheap reject before any boundary work. The
38
+ * caller holds the prompt in memory already (the injector matches against the
39
+ * same string), so no I/O is added per turn.
40
+ */
41
+ import type { TriggerSource } from "./skill-triggers.js";
42
+ /**
43
+ * A skill as the applicability read consumes it: its name and the phrase list
44
+ * the cache resolved from its file. `triggers` and `triggerSources` are the
45
+ * fields `setCachedSkills` fills, so this read never touches the disk — the
46
+ * phrases and their provenance are the ones the injector matched against.
47
+ */
48
+ export interface ConditionalSkill {
49
+ name: string;
50
+ /** Authored phrases (lowercased, deduped, capped) — empty when the file declares none. */
51
+ triggers?: string[];
52
+ /** Which authored source declared each phrase. */
53
+ triggerSources?: Record<string, TriggerSource>;
54
+ }
55
+ /** One authored phrase that matched, with the place in the file that declares it. */
56
+ export interface ApplicabilityEvidence {
57
+ /** The authored phrase, as the reader normalized it. */
58
+ phrase: string;
59
+ /**
60
+ * frontmatter | heading | description — absent only when the cached skill
61
+ * carried no provenance map (an unenriched cache, never a production one).
62
+ */
63
+ source?: TriggerSource;
64
+ }
65
+ /**
66
+ * One skill's verdict for a session.
67
+ *
68
+ * `ignored` is the evaluation's answer: the declared condition held and no
69
+ * detected read followed. `read` is the read-detector's upgrade of the same
70
+ * row. The third class, not-applicable, is never a row — it is a skill that
71
+ * declared a condition and did not meet the task text, so a reader derives it
72
+ * from the record's census (`evaluated` minus the verdict rows).
73
+ */
74
+ export interface SkillApplicabilityVerdict {
75
+ name: string;
76
+ verdict: "read" | "ignored";
77
+ evidence: ApplicabilityEvidence[];
78
+ }
79
+ /** Max evidence rows carried per skill: enough to review, bounded like every other surface. */
80
+ export declare const MAX_APPLICABILITY_EVIDENCE = 4;
81
+ /**
82
+ * Word tokens a negation cue may sit before the phrase and still negate it.
83
+ * "do not research this" (0) and "no need to research" (2) are negated;
84
+ * "do not stop, then research it later" puts the cue past the clauseless
85
+ * distance and reads as the assertion it is. Four is the smallest window that
86
+ * covers the "no need to"/"there is no reason to" shapes without reaching
87
+ * across a clause.
88
+ */
89
+ export declare const NEGATION_GAP_TOKENS = 4;
90
+ /**
91
+ * Every index in `lowered` where `phrase` occurs as a whole phrase, using the
92
+ * matcher's boundary rule: neither neighbour may be a letter or digit. Phrase
93
+ * must already be lowercase. Pure.
94
+ */
95
+ export declare function phraseOccurrences(phrase: string, lowered: string): number[];
96
+ /**
97
+ * Index where the clause containing `at` begins: after the nearest sentence
98
+ * terminator. A clause boundary is what stops a negation in one statement from
99
+ * reaching a phrase in the next ("do not use the catalog; research the topic"
100
+ * asserts research). Pure.
101
+ */
102
+ export declare function clauseStart(lowered: string, at: number): number;
103
+ /**
104
+ * True when the clause text preceding a phrase carries a negation cue close
105
+ * enough to negate it. The LAST cue decides, so "not only did we not ask" is
106
+ * read by its nearest cue rather than its first. Pure.
107
+ */
108
+ export declare function isNegatedPrefix(prefix: string): boolean;
109
+ /**
110
+ * The evidence for one skill's declared condition holding on `taskText`.
111
+ *
112
+ * A phrase is evidence when at least one of its occurrences is asserted:
113
+ * negating one occurrence does not erase another ("do not skip research" and
114
+ * "research this" in the same text both count). Phrases are evidence at most
115
+ * MAX_APPLICABILITY_EVIDENCE times per skill, in authored order. An empty
116
+ * return means the condition did not hold. Pure.
117
+ */
118
+ export declare function skillApplicability(skill: ConditionalSkill, taskText: string): ApplicabilityEvidence[];
119
+ /** The whole catalog's answer for one task text. */
120
+ export interface ApplicabilityEvaluation {
121
+ /** Skills whose declared condition held, each still an `ignored` verdict. */
122
+ applicable: SkillApplicabilityVerdict[];
123
+ /** Installed skills that declare at least one condition phrase. */
124
+ evaluated: number;
125
+ /** Installed skills that declare none — no condition, so no verdict either way. */
126
+ undeclared: string[];
127
+ }
128
+ /**
129
+ * Classify every installed skill against one task text. Pure: the catalog
130
+ * arrives as the cache already resolved it, and the text is the prompt the
131
+ * injector matched against.
132
+ */
133
+ export declare function evaluateApplicability(skills: ConditionalSkill[], taskText: string): ApplicabilityEvaluation;
134
+ //# sourceMappingURL=applicability.d.ts.map
@@ -0,0 +1,161 @@
1
+ /**
2
+ * Skill applicability — did a skill's own declared condition hold for the task?
3
+ *
4
+ * The suggestion whisper, the catalog ranking and the adoption counters all
5
+ * answer "might this skill help?". None answers the question the counters
6
+ * cannot: of the sessions that read nothing, which of them had a skill whose
7
+ * stated condition the task text actually met? A suggestion that was ignored
8
+ * and a skill that never applied looked identical in `suggested`/`loaded`.
9
+ * This module decides that difference and keeps its evidence.
10
+ *
11
+ * The condition is the skill's OWN declaration, never a heuristic invented
12
+ * here: the trigger phrases its file declares — frontmatter `trigger:` /
13
+ * `triggers:`, a `## trigger:` body heading, the description's `Triggers:`
14
+ * tail — the same entries `skill-triggers.ts` reads and the same provenance
15
+ * labels it records. The shipped pack's contract already treats that list as
16
+ * the machine-readable form of applicability ("a description that cannot fire
17
+ * is indistinguishable from one that fires"), so a skill file needs no new
18
+ * field for this module to see it. A file that declares no phrase yields no
19
+ * verdict at all; the caller reports it as undeclared rather than asserting
20
+ * "does not apply" on no evidence.
21
+ *
22
+ * Matching mirrors the injector on purpose. The occurrence boundary is the one
23
+ * `triggerMatches` uses (a phrase occurs where neither neighbour is a letter or
24
+ * digit, so "pane" is not an occurrence in "paneled"), because a verdict that
25
+ * disagreed with the whisper would be worse than no verdict: the suite pins
26
+ * the two answers equal over the installed catalog. What this module adds is
27
+ * polarity — a phrase inside a negated clause ("do not research this") is not
28
+ * the author's condition holding, so the clause is read and the phrase is
29
+ * dropped as evidence.
30
+ *
31
+ * What this module does NOT do: it decides nothing about delivery, slots,
32
+ * cooldowns, counters or ranking. It returns a classification and the caller
33
+ * records it. Nothing here reads the session, writes a file or sends a
34
+ * message.
35
+ *
36
+ * Cost: one pass over the task text per phrase of the cached catalog, with the
37
+ * native substring scan as the cheap reject before any boundary work. The
38
+ * caller holds the prompt in memory already (the injector matches against the
39
+ * same string), so no I/O is added per turn.
40
+ */
41
+ /** Max evidence rows carried per skill: enough to review, bounded like every other surface. */
42
+ export const MAX_APPLICABILITY_EVIDENCE = 4;
43
+ /**
44
+ * Word tokens a negation cue may sit before the phrase and still negate it.
45
+ * "do not research this" (0) and "no need to research" (2) are negated;
46
+ * "do not stop, then research it later" puts the cue past the clauseless
47
+ * distance and reads as the assertion it is. Four is the smallest window that
48
+ * covers the "no need to"/"there is no reason to" shapes without reaching
49
+ * across a clause.
50
+ */
51
+ export const NEGATION_GAP_TOKENS = 4;
52
+ /**
53
+ * Negation cues, matched in the clause text BEFORE a phrase. The `-n't`
54
+ * alternative is spelled out rather than listed per verb so a contraction the
55
+ * list forgot ("won't", "shouldn't") still counts.
56
+ */
57
+ const NEGATION_CUE = /\b(?:not|never|no|nor|without|avoid|avoiding|skip|skipping)\b|\b[a-z]+n['\u2019]t\b/gi;
58
+ /** True when `char` continues a word (the boundary rule `triggerMatches` uses). */
59
+ function isWordChar(char) {
60
+ return char !== undefined && /[a-z0-9]/.test(char);
61
+ }
62
+ /**
63
+ * Every index in `lowered` where `phrase` occurs as a whole phrase, using the
64
+ * matcher's boundary rule: neither neighbour may be a letter or digit. Phrase
65
+ * must already be lowercase. Pure.
66
+ */
67
+ export function phraseOccurrences(phrase, lowered) {
68
+ if (!phrase)
69
+ return [];
70
+ const out = [];
71
+ let from = 0;
72
+ for (;;) {
73
+ const at = lowered.indexOf(phrase, from);
74
+ if (at < 0)
75
+ break;
76
+ if (!isWordChar(lowered[at - 1]) && !isWordChar(lowered[at + phrase.length]))
77
+ out.push(at);
78
+ from = at + 1;
79
+ }
80
+ return out;
81
+ }
82
+ /**
83
+ * Index where the clause containing `at` begins: after the nearest sentence
84
+ * terminator. A clause boundary is what stops a negation in one statement from
85
+ * reaching a phrase in the next ("do not use the catalog; research the topic"
86
+ * asserts research). Pure.
87
+ */
88
+ export function clauseStart(lowered, at) {
89
+ for (let i = at - 1; i >= 0; i--) {
90
+ const char = lowered[i];
91
+ if (char === "." || char === "!" || char === "?" || char === ";" || char === "\n")
92
+ return i + 1;
93
+ }
94
+ return 0;
95
+ }
96
+ /**
97
+ * True when the clause text preceding a phrase carries a negation cue close
98
+ * enough to negate it. The LAST cue decides, so "not only did we not ask" is
99
+ * read by its nearest cue rather than its first. Pure.
100
+ */
101
+ export function isNegatedPrefix(prefix) {
102
+ const cue = new RegExp(NEGATION_CUE.source, "gi");
103
+ let last = null;
104
+ let match;
105
+ while ((match = cue.exec(prefix)))
106
+ last = match;
107
+ if (!last)
108
+ return false;
109
+ const gap = prefix.slice(last.index + last[0].length);
110
+ return (gap.match(/[a-z0-9]+/g) ?? []).length <= NEGATION_GAP_TOKENS;
111
+ }
112
+ /**
113
+ * The evidence for one skill's declared condition holding on `taskText`.
114
+ *
115
+ * A phrase is evidence when at least one of its occurrences is asserted:
116
+ * negating one occurrence does not erase another ("do not skip research" and
117
+ * "research this" in the same text both count). Phrases are evidence at most
118
+ * MAX_APPLICABILITY_EVIDENCE times per skill, in authored order. An empty
119
+ * return means the condition did not hold. Pure.
120
+ */
121
+ export function skillApplicability(skill, taskText) {
122
+ const lowered = taskText.toLowerCase();
123
+ const evidence = [];
124
+ for (const authored of skill.triggers ?? []) {
125
+ if (evidence.length >= MAX_APPLICABILITY_EVIDENCE)
126
+ break;
127
+ const phrase = authored.toLowerCase();
128
+ // Cheap reject first: most phrases are absent from most prompts.
129
+ if (!phrase || !lowered.includes(phrase))
130
+ continue;
131
+ const asserted = phraseOccurrences(phrase, lowered).some((at) => !isNegatedPrefix(lowered.slice(clauseStart(lowered, at), at)));
132
+ if (!asserted)
133
+ continue;
134
+ const source = skill.triggerSources?.[authored] ?? skill.triggerSources?.[phrase];
135
+ evidence.push(source ? { phrase, source } : { phrase });
136
+ }
137
+ return evidence;
138
+ }
139
+ /**
140
+ * Classify every installed skill against one task text. Pure: the catalog
141
+ * arrives as the cache already resolved it, and the text is the prompt the
142
+ * injector matched against.
143
+ */
144
+ export function evaluateApplicability(skills, taskText) {
145
+ const applicable = [];
146
+ const undeclared = [];
147
+ let evaluated = 0;
148
+ const text = taskText ?? "";
149
+ for (const skill of skills) {
150
+ if (!skill.triggers?.length) {
151
+ undeclared.push(skill.name);
152
+ continue;
153
+ }
154
+ evaluated++;
155
+ const evidence = skillApplicability(skill, text);
156
+ if (evidence.length)
157
+ applicable.push({ name: skill.name, verdict: "ignored", evidence });
158
+ }
159
+ return { applicable, evaluated, undeclared };
160
+ }
161
+ //# sourceMappingURL=applicability.js.map
@@ -189,6 +189,8 @@ export interface CachedSkill {
189
189
  triggers?: string[];
190
190
  /** Which source authored each phrase (the row's evidence label). */
191
191
  triggerSources?: Record<string, TriggerSource>;
192
+ /** Tool families the file's `governs:` declaration claims (skill-families.ts). */
193
+ families?: string[];
192
194
  }
193
195
  /** Called by index.ts's setCachedSkills to keep the auto-inject cache in sync. */
194
196
  export declare function updateSkillCache(skills: CachedSkill[]): void;
@@ -497,6 +499,14 @@ interface DiscoveryState {
497
499
  * discovery mode. Returns the current record for skill_stats.
498
500
  */
499
501
  export declare function resolveDiscoveryMode(ctx: ExtensionContext): DiscoveryState;
502
+ /**
503
+ * Record the skills a delivered suggestion offered this session. Called only
504
+ * after the send was accepted, so the row counts arrivals like the `suggested`
505
+ * counter does. Never throws: a record is not worth a broken turn.
506
+ */
507
+ export declare function noteObligationSuggested(ctx: ExtensionContext, names: string[]): void;
508
+ /** Record the skills a catalog page listed (called from the list_skills tool). */
509
+ export declare function noteObligationListed(ctx: ExtensionContext, names: string[]): void;
500
510
  /**
501
511
  * Register the auto-inject handler on before_agent_start, the read-tool load
502
512
  * detector and the drift trigger on tool_result.
@@ -112,8 +112,10 @@ import { getExtensionState } from "../../../lib/state/extension-state.js";
112
112
  import { BRAID_INJECTED, BRAID_SOURCE } from "../../../lib/constants/injection-details.js";
113
113
  import { getSkillsUsageFilePath } from "../../../lib/config/braid-paths.js";
114
114
  import { SESSION_COMPACT, AGENT_SETTLED } from "../../../lib/core/event-names.js";
115
- import { getSessionProfile, recordSessionSkillLoad, recordSessionStart, } from "./session-records.js";
115
+ import { getSessionProfile, recordApplicabilityCensus, recordApplicabilityVerdict, recordObligationLegs, recordSessionSkillLoad, recordSessionStart, } from "./session-records.js";
116
+ import { evaluateApplicability } from "./applicability.js";
116
117
  import { triggerMatches } from "./skill-triggers.js";
118
+ import { toolFamilyToken } from "./skill-families.js";
117
119
  import { advanceDrift, driftDirectoryFor, driftPayload, initialDrift, noteDriftSurface, recordDriftDrop, recordDriftFire, skillReadRoot, } from "./loading-guidance.js";
118
120
  const log = createLogger("skills-auto-inject");
119
121
  // Constants
@@ -238,6 +240,12 @@ const KEYWORD_STOP_WORDS = new Set([
238
240
  "need",
239
241
  "look",
240
242
  ]);
243
+ function initialObligations() {
244
+ return { legs: new Map(), called: new Set(), written: new Map() };
245
+ }
246
+ function initialApplicability() {
247
+ return { censusWritten: false, verdicts: new Map(), evaluated: 0, undeclared: [], taskChars: 0 };
248
+ }
241
249
  const STATE_KEY = "skills-auto-inject";
242
250
  function getState(ctx) {
243
251
  return getExtensionState(ctx, STATE_KEY, () => ({
@@ -245,6 +253,8 @@ function getState(ctx) {
245
253
  currentTurn: 0,
246
254
  hintSent: false,
247
255
  drift: initialDrift(),
256
+ applicability: initialApplicability(),
257
+ obligations: initialObligations(),
248
258
  }));
249
259
  }
250
260
  // Shared module-level variable rather than a per-module copy: the import
@@ -910,6 +920,159 @@ export function resolveDiscoveryMode(ctx) {
910
920
  }
911
921
  return state;
912
922
  }
923
+ // Applicability verdict (recorded, never enforced)
924
+ /**
925
+ * Classify the catalog against this turn's task text and record what changed.
926
+ *
927
+ * Two writes, both append-only:
928
+ * - the session's census row, once, retried until it lands. It is what makes
929
+ * "not-applicable" derivable (evaluated minus the verdict rows) and what
930
+ * names the skills whose files state no condition at all.
931
+ * - one verdict row per skill whose declared condition held, and never twice
932
+ * for the same (session, skill): the set only grows because the verdict is
933
+ * per-session — a skill whose condition held on any turn applied to the
934
+ * session.
935
+ *
936
+ * The row shape is the joint the sibling compliance record expects: it joins on
937
+ * (sessionId, skill) and adds its own `suggested`/`listed`/`read`/`applied`
938
+ * beside this verdict, so `ignored / (read + ignored)` becomes computable. This
939
+ * module writes no `applied` leg: a tool belong to a skill's family only once
940
+ * the skill declares that family, and no skill does yet.
941
+ *
942
+ * Reads only — it never touches `injected`, a counter or a send. That is the
943
+ * whole point of the ordering in before_agent_start: this step cannot move the
944
+ * suggestion, and the suggestion cannot move the verdict.
945
+ */
946
+ function stepApplicability(ctx, state, skills, prompt) {
947
+ const evaluation = evaluateApplicability(skills, prompt);
948
+ const app = state.applicability;
949
+ const fresh = evaluation.applicable.filter((verdict) => !app.verdicts.has(verdict.name));
950
+ for (const verdict of fresh)
951
+ app.verdicts.set(verdict.name, verdict);
952
+ app.evaluated = evaluation.evaluated;
953
+ app.undeclared = evaluation.undeclared;
954
+ app.taskChars = prompt.length;
955
+ if (!app.censusWritten) {
956
+ const wrote = recordApplicabilityCensus(ctx, {
957
+ taskChars: app.taskChars,
958
+ evaluated: app.evaluated,
959
+ undeclared: app.undeclared,
960
+ });
961
+ // A session without an id, or a refused store: retry on the next turn.
962
+ if (wrote)
963
+ app.censusWritten = true;
964
+ }
965
+ if (!fresh.length)
966
+ return;
967
+ log.debug(`[skills-auto-inject] applicability: ${fresh.length} newly applicable, ` +
968
+ `${evaluation.evaluated} evaluated, ${evaluation.undeclared.length} undeclared`);
969
+ for (const verdict of fresh) {
970
+ recordApplicabilityVerdict(ctx, {
971
+ skill: verdict.name,
972
+ applicability: verdict.verdict,
973
+ evidence: verdict.evidence,
974
+ });
975
+ // The obligation row lands with the verdict, not with the first leg: a
976
+ // session that is merely told which body it owes, and then does nothing,
977
+ // is the population the rate exists to count. Without this row that
978
+ // session would be mistaken for one the instrument never saw.
979
+ legsFor(state, verdict.name);
980
+ appendObligationRow(ctx, state, verdict.name);
981
+ }
982
+ }
983
+ /**
984
+ * Upgrade an applicable skill to `read` and append its verdict row. A read of a
985
+ * skill whose declared condition never held gets no row: the three classes are
986
+ * about skills that applied, and a read is not a fourth one.
987
+ */
988
+ function markApplicabilityRead(ctx, state, name) {
989
+ const app = state.applicability;
990
+ const held = app.verdicts.get(name);
991
+ if (!held || held.verdict === "read")
992
+ return;
993
+ app.verdicts.set(name, { ...held, verdict: "read" });
994
+ recordApplicabilityVerdict(ctx, {
995
+ skill: name,
996
+ applicability: "read",
997
+ evidence: held.evidence,
998
+ });
999
+ }
1000
+ // Obligation rows (the applied leg)
1001
+ /**
1002
+ * The session's legs for one skill, created on first sight. The object is the
1003
+ * snapshot every row carries, so a row states all five legs as of the moment
1004
+ * the leg changed rather than the single delta that triggered the write.
1005
+ */
1006
+ function legsFor(state, name) {
1007
+ let legs = state.obligations.legs.get(name);
1008
+ if (!legs) {
1009
+ legs = {
1010
+ skill: name,
1011
+ suggested: false,
1012
+ listed: false,
1013
+ read: false,
1014
+ called: false,
1015
+ applied: false,
1016
+ };
1017
+ state.obligations.legs.set(name, legs);
1018
+ }
1019
+ return legs;
1020
+ }
1021
+ /**
1022
+ * Append the skill's legs when they changed since the last row. One row per
1023
+ * distinct state rather than one per event: a session that reads the same body
1024
+ * three times, or lists it on every turn, writes one line. A refused write
1025
+ * leaves the signature unset, so the next leg change retries.
1026
+ */
1027
+ function appendObligationRow(ctx, state, name) {
1028
+ const legs = state.obligations.legs.get(name);
1029
+ if (!legs)
1030
+ return;
1031
+ const signature = [legs.suggested, legs.listed, legs.read, legs.called, legs.applied]
1032
+ .map((flag) => (flag ? "1" : "0"))
1033
+ .join("");
1034
+ if (state.obligations.written.get(name) === signature)
1035
+ return;
1036
+ if (!recordObligationLegs(ctx, legs))
1037
+ return;
1038
+ state.obligations.written.set(name, signature);
1039
+ }
1040
+ /** Note one observed leg for a skill and record the row it changed. */
1041
+ function noteObligationLeg(ctx, state, name, leg) {
1042
+ if (!name)
1043
+ return;
1044
+ const legs = legsFor(state, name);
1045
+ if (legs[leg])
1046
+ return;
1047
+ legs[leg] = true;
1048
+ appendObligationRow(ctx, state, name);
1049
+ }
1050
+ /**
1051
+ * Record the skills a delivered suggestion offered this session. Called only
1052
+ * after the send was accepted, so the row counts arrivals like the `suggested`
1053
+ * counter does. Never throws: a record is not worth a broken turn.
1054
+ */
1055
+ export function noteObligationSuggested(ctx, names) {
1056
+ try {
1057
+ const state = getState(ctx);
1058
+ for (const name of names)
1059
+ noteObligationLeg(ctx, state, name, "suggested");
1060
+ }
1061
+ catch (err) {
1062
+ log.debug(`[skills-auto-inject] obligation suggested leg failed (best-effort): ${err}`);
1063
+ }
1064
+ }
1065
+ /** Record the skills a catalog page listed (called from the list_skills tool). */
1066
+ export function noteObligationListed(ctx, names) {
1067
+ try {
1068
+ const state = getState(ctx);
1069
+ for (const name of names)
1070
+ noteObligationLeg(ctx, state, name, "listed");
1071
+ }
1072
+ catch (err) {
1073
+ log.debug(`[skills-auto-inject] obligation listed leg failed (best-effort): ${err}`);
1074
+ }
1075
+ }
913
1076
  /**
914
1077
  * Count a SKILL.md read (the list→read load path) and mark the matching
915
1078
  * injection loaded (releases its slot and extends its cooldown: the skill is
@@ -932,9 +1095,44 @@ function detectSkillLoad(ev, ctx, state) {
932
1095
  if (held)
933
1096
  held.loadedTurn = state.currentTurn;
934
1097
  recordSkillLoad(ctx, name);
1098
+ // The load counters say a read happened; this says whether the session was
1099
+ // one where the skill applied. Independent writes, independent failures.
1100
+ markApplicabilityRead(ctx, state, name);
1101
+ noteObligationLeg(ctx, state, name, "read");
935
1102
  log.debug(`[skills-auto-inject] counted actual load of "${name}" via read`);
936
1103
  return name;
937
1104
  }
1105
+ /**
1106
+ * Record a call into a declared family (the applied leg).
1107
+ *
1108
+ * The FIRST call per (session, skill) decides it: a read already on record
1109
+ * means the body came first and the call applies what it says; no read means
1110
+ * the call came first, and no later read can turn that into satisfaction — the
1111
+ * ordering is a fact about what happened, not a state that improves.
1112
+ *
1113
+ * A failed call still counts: a refusal would fire before the tool ran and
1114
+ * could not know the outcome. Skills that declare no family are never reached
1115
+ * here, which is most of them.
1116
+ */
1117
+ function stepFamilyCalls(ctx, state, event) {
1118
+ const toolName = event?.toolName;
1119
+ if (!toolName)
1120
+ return;
1121
+ const family = toolFamilyToken(toolName);
1122
+ if (!family)
1123
+ return;
1124
+ for (const skill of _skillsCache) {
1125
+ if (!skill.families?.length || !skill.families.includes(family))
1126
+ continue;
1127
+ if (state.obligations.called.has(skill.name))
1128
+ continue;
1129
+ const legs = legsFor(state, skill.name);
1130
+ state.obligations.called.add(skill.name);
1131
+ legs.called = true;
1132
+ legs.applied = legs.read;
1133
+ appendObligationRow(ctx, state, skill.name);
1134
+ }
1135
+ }
938
1136
  /**
939
1137
  * Fire the drift line when the work has moved into a new work area (the rule
940
1138
  * and its bounds live in loading-guidance.ts, beside the floor text the line
@@ -993,6 +1191,10 @@ export function registerAutoInject(pi) {
993
1191
  const state = getState(ctx);
994
1192
  if (detectSkillLoad(ev, ctx, state))
995
1193
  noteDriftSurface(state.drift);
1194
+ // A call in a declared family is the applied leg. It runs after the load
1195
+ // detector so a SKILL.md read of the very tool call in flight is already
1196
+ // on record when the family is looked up.
1197
+ stepFamilyCalls(ctx, state, ev);
996
1198
  stepDrift(pi, ctx, state, ev);
997
1199
  }
998
1200
  catch {
@@ -1044,6 +1246,17 @@ export function registerAutoInject(pi) {
1044
1246
  if (!prompt.trim()) {
1045
1247
  return; // No user input to match against
1046
1248
  }
1249
+ // Applicability is recorded BEFORE every scoring early-return below.
1250
+ // Most sessions match no skill, and those are exactly the sessions the
1251
+ // adoption counters cannot explain: the returns would erase them. The
1252
+ // step reads the catalog and the prompt only, so the suggestion decision
1253
+ // below is untouched by it (and its failure cannot cost a suggestion).
1254
+ try {
1255
+ stepApplicability(ctx, state, skills, prompt);
1256
+ }
1257
+ catch (err) {
1258
+ log.debug(`[skills-auto-inject] applicability step failed (best-effort): ${err}`);
1259
+ }
1047
1260
  // One-shot profile hint: the first non-empty prompt of a session whose
1048
1261
  // configured profile carries a hint gets exactly one steer — the hint
1049
1262
  // text itself, byte-for-byte. Profiles without a hint and sessions
@@ -1125,6 +1338,8 @@ export function registerAutoInject(pi) {
1125
1338
  (i.loadedTurn !== undefined && state.currentTurn - i.loadedTurn < LOADED_COOLDOWN_TURNS));
1126
1339
  // E3: per-skill suggestion counter — counts deliveries, not intent.
1127
1340
  recordSkillSuggestion(ctx, skillsToInject.map((s) => s.name));
1341
+ // The same delivery is an obligation leg: the body was offered here.
1342
+ noteObligationSuggested(ctx, skillsToInject.map((s) => s.name));
1128
1343
  }
1129
1344
  catch (err) {
1130
1345
  log.debug(`[skills-auto-inject] Error in before_agent_start: ${err}`);