refactor(me/evolve): fix path/flag bugs + simplify Phase 0 & 1

docs/superpowers/specs/2026-05-27-evolve-skill-design.md

@@ -37,14 +37,14 @@ Claude Code 세션을 한참 돌리고 나면 같은 실수·우회·교정이 
 ## 사용법
 
 ```
-/me:evolve                          현재 세션 회고
-/me:evolve me:research              해당 스킬에 집중해 최근 세션 분석
-/me:evolve --session <id>           특정 세션 ID 회고
-/me:evolve --since 7d me:browse     7일치 + 스킬 필터
-/me:evolve --dry-run                제안만 보고 적용 안 함
+/me:evolve                          현재 세션 회고                    [구현됨]
+/me:evolve --session <id>           특정 세션 ID 회고                  [구현됨]
+/me:evolve --dry-run                제안만 보고 적용 안 함             [구현됨, 메인 에이전트가 해석]
+/me:evolve me:research              해당 스킬에 집중해 최근 세션 분석   [미구현]
+/me:evolve --since 7d me:browse     7일치 + 스킬 필터                  [미구현]
 ```
 
-기본값(인자 없음)은 현재 세션. 인자에 스킬명이 있으면 해당 스킬이 등장한 세션·구간으로 좁힌다.
+기본값(인자 없음)은 현재 세션. 스킬명·`--since` 필터는 설계 의도이며 build-index.ts에 아직 없다 (`parseArgs`는 `--session`과 jsonl 경로만 받음). 추가 시 build-index.ts·SKILL.md를 동기화할 것.
 
 ## 입력
 

plugins/me/skills/evolve/SKILL.md

@@ -41,15 +41,17 @@ The main agent does two things in Phase 0:
 
 2. **Build the index** — `build-index.ts` is read-only: auto-detects the transcript, extracts signals, prints JSON.
 
+Use the `Base directory for this skill: <path>` value injected at the top of this command prompt — do **not** rely on `${CLAUDE_PLUGIN_ROOT}` (it is not guaranteed to be set in the slash-command shell).
+
 ```bash
-bun "${CLAUDE_PLUGIN_ROOT}/skills/evolve/scripts/build-index.ts" [--session <id>]
+bun "<Base directory>/scripts/build-index.ts" [--session <id>]
 ```
 
-Exit codes: `0`=ok, `14`=transcript or project dir not found.
+Exit codes: `0`=ok, `2`=unknown flag (see §4 below — `--dry-run` must not be passed here), `14`=transcript or project dir not found.
 
 Capture stdout JSON into a variable. **Do not show it to the user** — pass it only to the next-step subagent.
 
-If `events` is empty, or both `summary.clusters` and `summary.signal_positions` are empty, print "no improvement signals found in this session" and exit.
+If `events` is empty, print "no improvement signals found in this session" and exit.
 
 ## Phase 1 — Subagent analysis (Agent)
 
@@ -60,7 +62,7 @@ Dispatch one subagent (`general-purpose`). The indexer hands user utterances ove
 The prompt must include all of:
 
 1. Spec path: `docs/superpowers/specs/2026-05-27-evolve-skill-design.md`
-2. The full index JSON from Phase 0 (only `summary` + `events`). **Read `summary` first** — `headline` (one-line state), `clusters` (same-kind events within ≤30 turns, ≥3 occurrences, with `t_range`/`n`/`example_t`), `signal_positions` (per-kind turn coordinates). Use this to decide which regions to inspect, then slice `events[]` for causal-chain analysis. summary is a simple heuristic — false positives are expected; skip meaningless clusters.
+2. The full index JSON from Phase 0 (`summary` + `events`). Read `summary.headline` (one-line state) and `summary.clusters` (same-kind events within ≤30 turns, ≥3 occurrences) first to spot dense regions, then walk `events[]` for causal-chain analysis. Clusters are a simple heuristic — skip meaningless ones.
 3. **Classification task for `kind: "user"` events** — label each one with exactly one of (refer by array index, e.g. `events[12]`):
    - **correction**: redirects/corrects the immediately preceding assistant action
    - **success**: positive feedback on the preceding assistant action
@@ -70,19 +72,13 @@ The prompt must include all of:
 
    Criterion is *the relation to the preceding assistant action (`event.prior`)*, not the words themselves. E.g. "stop and report" used as a directive verb is noise; "stop, that's wrong" reacting to a tool call is a correction.
 
-4. Candidate-file mapping table (use alongside interrupt/repeat/error events):
+4. Target-file selection — pick by *what kind of knowledge is missing*, not by signal pattern:
 
-   | Signal pattern | 1st choice | 2nd |
-   |---|---|---|
-   | many corrections → skill not invoked | that SKILL.md (description, triggers) | nearest AGENTS.md |
-   | correction → rule violated *after* skill invoked | that SKILL.md (body, Red Flags) | — |
-   | repeat + eventually finds X | nearest AGENTS.md (Key Files) | CLAUDE.md |
-   | correction → project rule/convention violated | nearest CLAUDE.md | AGENTS.md |
-   | success → a skill clearly worked well | that frequently-invoked SKILL.md (reinforce) | — |
-   | interrupt + prior action clearly wrong | that SKILL.md | AGENTS.md |
-   | repeated error → same tool failing | that SKILL.md (usage) | CLAUDE.md |
+   - **that SKILL.md** — skill triggers, body rules, Red Flags. Use when a skill should have run but didn't, or ran but violated its own contract.
+   - **nearest AGENTS.md** — repo navigation, file locations, "where to look" knowledge. Use when the agent searched/read repeatedly before finding something.
+   - **nearest CLAUDE.md** — project conventions and rules that govern *all* work in this tree.
 
-   **"Nearest" resolution**: "that SKILL.md" = the SKILL.md inferred from `events[]` items with `kind:"skill"` or from tool calls in `prior`. "Nearest AGENTS.md/CLAUDE.md" = walk up from that SKILL.md's directory; first one found wins (fallback to repo root). If no skill can be inferred, default to repo-root CLAUDE.md.
+   **"Nearest" resolution**: "that SKILL.md" = inferred from `events[]` items with `kind:"skill"` or tool calls in `prior`. "Nearest AGENTS.md/CLAUDE.md" = walk up from that SKILL.md's directory; first hit wins (fallback to repo-root CLAUDE.md). If no skill can be inferred, default to repo-root CLAUDE.md.
 
 5. Output schema — one JSON block only, no other text. `event_index` is the integer position in the index's `events[]` array:
 
@@ -102,8 +98,7 @@ The prompt must include all of:
          "patch": "<unified diff applicable with `git apply`>",
          "rationale": "1-2 sentences"
        }
-     ],
-     "skipped": [{"event_index": 5, "reason": "low confidence"}]
+     ]
    }
    ```
 
@@ -133,7 +128,7 @@ After parsing the subagent's returned JSON:
    - **edit**: let the user edit the patch, then proceed as `y`.
    - **skip / n**: move on to the next proposal.
 
-4. If `--dry-run` is passed, skip Phase 2 entirely and just print the proposal list.
+4. If `--dry-run` is passed, skip Phase 2 entirely and just print the proposal list. `--dry-run` is consumed by the main agent only — never forward it to `build-index.ts` (the script will exit 2 on unknown flags).
 
 5. Finalize: print the list of applied commit SHAs and the upstream-suggestions path (if any).
 

plugins/me/skills/evolve/scripts/build-index.ts

@@ -8,20 +8,20 @@ import { basename, join, resolve } from "node:path";
 import { homedir } from "node:os";
 
 // ── 타입 ───────────────────────────────────────────────
-type EventKind = "user" | "skill" | "interrupt" | "error" | "agent" | "large_out" | "repeat";
+type EventKind = "user" | "skill" | "interrupt" | "error" | "agent" | "repeat";
 
 interface Event {
   t: number;
   kind: EventKind;
   text?: string;
   prior?: string[];
   name?: string;
+  args?: string;
   by?: "user" | "assistant";
   tool?: string;
   desc?: string;
   sub?: string;
   model?: string;
-  bytes?: number;
   pattern?: string;
   n?: number;
 }
@@ -36,7 +36,6 @@ interface Cluster {
 interface Summary {
   headline: string;
   clusters: Cluster[];
-  signal_positions: Partial<Record<EventKind, number[]>>;
 }
 
 interface SessionIndex {
@@ -51,7 +50,6 @@ interface SessionIndex {
 interface ToolResultPayload {
   content: string;
   isError: boolean;
-  size: number;
 }
 
 interface Turn {
@@ -106,7 +104,7 @@ function loadTurns(jsonlPath: string): LoadedTranscript {
         for (const c of content) {
           if (c.type === "tool_result") {
             const norm = normalizeToolResultContent(c.content);
-            t.toolResults.push({ content: norm, isError: c.is_error === true, size: norm.length });
+            t.toolResults.push({ content: norm, isError: c.is_error === true });
           }
         }
       }
@@ -118,7 +116,7 @@ function loadTurns(jsonlPath: string): LoadedTranscript {
   return { turns, sessionTitle };
 }
 
-// ── tool_use 요약 (user.prior, large_out.tool에 사용) ──
+// ── tool_use 요약 (user.prior에 사용) ──
 function summarizeToolUse(tu: { name: string; input: any }): string {
   const name = tu.name;
   let arg = "";
@@ -149,13 +147,25 @@ function priorAssistantActions(turns: Turn[], currentIdx: number): string[] {
 
 // ── 슬래시 커맨드 검출 ──────────────────────────────────
 const SLASH_CMD_TAG = /<command-name>\/([a-z0-9:_-]+)<\/command-name>/i;
-const SLASH_CMD_PREFIX = /^\/([a-z0-9:_-]+)\b/i;
+const SLASH_CMD_PREFIX = /^\/([a-z0-9:_-]+)\b(.*)$/i;
+const SLASH_CMD_ARGS_TAG = /<command-args>([\s\S]*?)<\/command-args>/i;
 
-function detectSlashCommand(userText: string): string | undefined {
+interface DetectedSlash {
+  name: string;
+  args?: string;
+}
+
+function detectSlashCommand(userText: string): DetectedSlash | undefined {
   const tag = userText.match(SLASH_CMD_TAG);
-  if (tag) return tag[1];
+  if (tag) {
+    const argsTag = userText.match(SLASH_CMD_ARGS_TAG);
+    const args = argsTag?.[1]?.trim();
+    return args ? { name: tag[1], args: args.slice(0, 200) } : { name: tag[1] };
+  }
   const prefix = userText.trim().match(SLASH_CMD_PREFIX);
-  return prefix?.[1];
+  if (!prefix) return undefined;
+  const args = prefix[2]?.trim();
+  return args ? { name: prefix[1], args: args.slice(0, 200) } : { name: prefix[1] };
 }
 
 const PSEUDO_USER_PREFIXES = [
@@ -173,13 +183,11 @@ function isPseudoUser(userText: string): boolean {
 }
 
 // ── events 빌드 ────────────────────────────────────────
-const LARGE_OUTPUT_THRESHOLD = 10 * 1024;
-
 function buildEvents(turns: Turn[]): Event[] {
   const events: Event[] = [];
   const interruptClaimed = new Set<number>();
 
-  // user / skill / interrupt(user) / error / agent / large_out 를 순회 중에 emit
+  // user / skill / interrupt(user) / error / agent 를 순회 중에 emit
   for (let i = 0; i < turns.length; i++) {
     const t = turns[i];
 
@@ -188,9 +196,11 @@ function buildEvents(turns: Turn[]): Event[] {
       // 단 사용자가 직접 친 슬래시 커맨드 (prefix form)는 user event도 함께 의미 있을 수 있지만
       // 분류 노이즈를 줄이기 위해 skill만 emit.
       if (t.userText) {
-        const slashName = detectSlashCommand(t.userText);
-        if (slashName) {
-          events.push({ t: t.index, kind: "skill", name: slashName });
+        const slash = detectSlashCommand(t.userText);
+        if (slash) {
+          const ev: Event = { t: t.index, kind: "skill", name: slash.name };
+          if (slash.args) ev.args = slash.args;
+          events.push(ev);
         } else if (!isPseudoUser(t.userText)) {
           events.push({
             t: t.index,
@@ -205,16 +215,12 @@ function buildEvents(turns: Turn[]): Event[] {
         events.push({ t: t.index, kind: "interrupt", by: t.interruptedBy ?? "user" });
         interruptClaimed.add(t.index);
       }
-      // tool_result 안의 error / large_out
+      // tool_result 안의 error
       for (const tr of t.toolResults) {
+        if (!tr.isError) continue;
         const prevAssist = [...turns.slice(0, i)].reverse().find((a) => a.type === "assistant" && a.toolUses.length > 0);
         const toolName = prevAssist?.toolUses[prevAssist.toolUses.length - 1]?.name ?? "unknown";
-        if (tr.isError) {
-          events.push({ t: t.index, kind: "error", tool: toolName, text: tr.content.slice(0, 200) });
-        }
-        if (tr.size > LARGE_OUTPUT_THRESHOLD) {
-          events.push({ t: t.index, kind: "large_out", tool: toolName, bytes: tr.size });
-        }
+        events.push({ t: t.index, kind: "error", tool: toolName, text: tr.content.slice(0, 200) });
       }
     } else {
       // assistant
@@ -297,15 +303,6 @@ function buildClusters(events: Event[]): Cluster[] {
   return clusters.sort((a, b) => a.t_range[0] - b.t_range[0]);
 }
 
-function buildSignalPositions(events: Event[]): Partial<Record<EventKind, number[]>> {
-  const out: Partial<Record<EventKind, number[]>> = {};
-  for (const kind of SUMMARY_KINDS) {
-    const turns = events.filter((e) => e.kind === kind).map((e) => e.t);
-    if (turns.length > 0) out[kind] = turns;
-  }
-  return out;
-}
-
 function buildHeadline(turns: number, events: Event[], clusters: Cluster[]): string {
   const counts: Record<string, number> = {};
   for (const e of events) counts[e.kind] = (counts[e.kind] ?? 0) + 1;
@@ -322,7 +319,6 @@ function buildSummary(turns: number, events: Event[]): Summary {
   return {
     headline: buildHeadline(turns, events, clusters),
     clusters,
-    signal_positions: buildSignalPositions(events),
   };
 }
 

tests/me/evolve-build-index.bats

@@ -43,7 +43,7 @@ RICH_FIXTURE="${PROJECT_ROOT}/tests/fixtures/evolve/rich-signals-session.jsonl"
     [ "$status" -eq 0 ]
     # 인덱서는 분류 안 함 — user_correction/success_pattern 같은 kind는 절대 emit 안 함
     echo "$output" | jq -e '
-      [.events[].kind] | all(. as $k | ["user","skill","interrupt","error","agent","large_out","repeat"] | index($k) != null)
+      [.events[].kind] | all(. as $k | ["user","skill","interrupt","error","agent","repeat"] | index($k) != null)
     '
 }
 
@@ -60,12 +60,13 @@ RICH_FIXTURE="${PROJECT_ROOT}/tests/fixtures/evolve/rich-signals-session.jsonl"
     echo "$output" | jq -e '(keys - ["session_id","session_title","turns","summary","events"]) == []'
 }
 
-@test "evolve build-index: summary has headline, clusters, signal_positions" {
+@test "evolve build-index: summary has headline and clusters" {
     run bun "$INDEXER" "$FIXTURE"
     [ "$status" -eq 0 ]
     echo "$output" | jq -e '.summary.headline | type == "string"'
     echo "$output" | jq -e '.summary.clusters | type == "array"'
-    echo "$output" | jq -e '.summary.signal_positions | type == "object"'
+    # signal_positions intentionally removed — clusters + events[] cover its role
+    echo "$output" | jq -e '.summary | has("signal_positions") | not'
 }
 
 @test "evolve build-index: detects slash-command in <command-name> tag form" {
@@ -119,24 +120,11 @@ RICH_FIXTURE="${PROJECT_ROOT}/tests/fixtures/evolve/rich-signals-session.jsonl"
     '
 }
 
-@test "evolve build-index: detects large_out event above 10KB" {
+@test "evolve build-index: large_out event kind is no longer emitted" {
     run bun "$INDEXER" "$RICH_FIXTURE"
     [ "$status" -eq 0 ]
-    echo "$output" | jq -e '
-      [.events[] | select(.kind == "large_out")] as $l
-      | ($l | length >= 1) and ($l[0].bytes > 10240)
-    '
-}
-
-@test "evolve build-index: signal_positions kinds appear in events" {
-    run bun "$INDEXER" "$RICH_FIXTURE"
-    [ "$status" -eq 0 ]
-    # 각 signal_positions 키는 실제 events 의 kind 와 일치해야 함
-    echo "$output" | jq -e '
-      (.summary.signal_positions | keys) as $sk
-      | ([.events[].kind] | unique) as $ek
-      | ($sk | all(. as $k | $ek | index($k) != null))
-    '
+    # large_out was removed because Phase 1's mapping table never consumed it.
+    echo "$output" | jq -e '[.events[] | select(.kind == "large_out")] | length == 0'
 }
 
 @test "evolve build-index: false-positive guard — body words do NOT trigger semantic kinds" {
@@ -148,5 +136,5 @@ EOF
     [ "$status" -eq 0 ]
     # user event 1건, 분류 kind 일체 없음
     echo "$output" | jq -e '[.events[] | select(.kind == "user")] | length == 1'
-    echo "$output" | jq -e '[.events[].kind] | all(. as $k | ["user","skill","interrupt","error","agent","large_out","repeat"] | index($k) != null)'
+    echo "$output" | jq -e '[.events[].kind] | all(. as $k | ["user","skill","interrupt","error","agent","repeat"] | index($k) != null)'
 }