obsidian-lean-coder 0.2.0 → 0.4.0

package/CHANGELOG.md

@@ -4,6 +4,37 @@ All notable changes to OLC (Obsidian Lean Coder) are recorded here.
 OLC is pre-release: versions are `0.x` until a controlled-testing release says otherwise.
 `package.json` shipped `"private": true` through 0.1.0; the flag was removed at publish time with explicit user approval, and 0.1.0 is live as [`obsidian-lean-coder`](https://www.npmjs.com/package/obsidian-lean-coder). Publishing a new version is still a separate explicitly approved action, never automatic.
 
+## [0.4.0] — 2026-10-14 — default-like agents + project rules
+
+Agents `olc-plan`/`olc-build` rewritten per user request to behave exactly like
+OpenCode's default Plan/Build agents first, then add concise OLC rules: each
+session they read the project root `AGENTS.md` (if present) and `.olc/AGENTS.md`
+(if present); on conflict the OpenCode default wins and the agent shows the
+conflict, suggests a fix, and stops for the user's choice. The `.olc/AGENTS.md`
+template gains an "Aturan tambahan proyek" section, and the `olc-init` skill now
+asks whether the project has extra rules beyond the main agents file
+(`initProject` itself stays non-interactive; existing files are still never
+overwritten).
+
+- **Verification:** typecheck/build clean; full suite green; `npm pack --dry-run` file check.
+- **Upgrade:** `npm install obsidian-lean-coder@0.4.0`, then
+  `npx obsidian-lean-coder setup` (existing OLC agent/skill files are overwritten
+  with package versions).
+
+## [0.3.0] — 2026-10-13 — primary agents + all PRD skills
+
+Agents `olc-plan`/`olc-build` move from `mode: subagent` (@-only invocation) to `mode: primary`,
+so they appear in the OpenCode agent switcher (user decision).
+Ships the four missing PRD §7.3 skill wrappers — `olc-init`, `olc-memory`, `olc-log`, `olc-taste` —
+alongside the existing `olc-explore`; `olc-setup` installs all 7 owned files (2 agents + 5 skills).
+Each SKILL.md documents its module's real contract (init create-only flow, progressive recall,
+secret-scrubbed append-only log + session summary, pending→active taste lifecycle).
+
+- **Verification:** `tests/setup.test.ts` counts updated (7 created/skipped); full suite green;
+  typecheck/build clean.
+- **Upgrade:** `npm install obsidian-lean-coder@0.3.0`, then
+  `npx obsidian-lean-coder setup` (existing files are overwritten with package versions).
+
 ## [0.2.0] — 2026-10-13 — `olc-setup` resource installer
 
 New `olc-setup setup [--project|--global]` command (run manually via `npx obsidian-lean-coder setup`

package/README.md

@@ -1,6 +1,6 @@
 # OLC — Obsidian Lean Coder
 
-OpenCode extension for project-scoped planning, memory, and safe Git checkpoints. On npm as `obsidian-lean-coder` (v0.1.1 fixes plugin activation — see CHANGELOG).
+OpenCode extension for project-scoped planning, memory, and safe Git checkpoints. On npm as `obsidian-lean-coder` (v0.4.0: default-like agents + project rules — see CHANGELOG).
 
 ## Three separate steps
 
@@ -24,7 +24,7 @@ npm test         # tsc -p tsconfig.tests.json && node --test "dist-tests/tests/*
 npm run typecheck
 ```
 
-Expected: build clean, typecheck clean. `npm test` runs the full suite (currently 140 tests). `npm pack --dry-run` shows what would ship (`dist/`, `agents/`, `skills/`, `templates/`).
+Expected: build clean, typecheck clean. `npm test` runs the full suite (currently 146 tests). `npm pack --dry-run` shows what would ship (`dist/`, `agents/`, `skills/`, `templates/`).
 
 > Published on npm as `obsidian-lean-coder` (MIT). Republishing needs a separate explicit approval — never publish by accident.
 
@@ -81,18 +81,20 @@ Manual equivalent (DECISIONS.md #2 — programmatic registration was rejected fo
 
 ```sh
 PKG=/home/you/olc   # your OLC checkout
-mkdir -p .opencode/agents .opencode/skills/olc-explore
+mkdir -p .opencode/agents .opencode/skills/olc-init .opencode/skills/olc-explore .opencode/skills/olc-memory .opencode/skills/olc-log .opencode/skills/olc-taste
 cp "$PKG"/agents/olc-plan.md "$PKG"/agents/olc-build.md .opencode/agents/
-cp "$PKG"/skills/olc-explore/SKILL.md .opencode/skills/olc-explore/
+for s in olc-init olc-explore olc-memory olc-log olc-taste; do cp "$PKG"/skills/$s/SKILL.md .opencode/skills/$s/; done
 ```
 
 Discovery note: `/api/agent` and `/api/skill` returned `[]` for project-local files in this build — even for a minimal probe agent — so API listing is not a valid verification (OLC-012 evidence did not reproduce; TUI-side discovery untested). Verify by file presence instead:
 
 ```sh
-ls .opencode/agents/olc-plan.md .opencode/agents/olc-build.md .opencode/skills/olc-explore/SKILL.md
+ls .opencode/agents/olc-plan.md .opencode/agents/olc-build.md .opencode/skills/olc-init/SKILL.md .opencode/skills/olc-explore/SKILL.md .opencode/skills/olc-memory/SKILL.md .opencode/skills/olc-log/SKILL.md .opencode/skills/olc-taste/SKILL.md
 ```
 
-`olc-plan`/`olc-build` agents and the `olc-explore` skill are picked up by OpenCode from these paths. Re-running the `cp` commands is safe as long as you have not customized the destination files — `cp` overwrites; the no-clobber guarantee belongs to `olc-init`, which reports existing files as `preserved` instead of overwriting them.
+`olc-plan`/`olc-build` agents and the five skills (`olc-init`, `olc-explore`, `olc-memory`, `olc-log`, `olc-taste`) are picked up by OpenCode from these paths. Re-running the `cp` commands is safe as long as you have not customized the destination files — `cp` overwrites; the no-clobber guarantee belongs to `olc-init`, which reports existing files as `preserved` instead of overwriting them.
+
+Both agents behave exactly like OpenCode's default Plan/Build agents first, then add OLC rules: each session they read the project root `AGENTS.md` (if present) and `.olc/AGENTS.md` (if present). If the two disagree, the OpenCode default wins — the agent shows the conflict, suggests a fix, and stops for your choice. During `olc-init` you are asked whether the project has extra rules beyond the main agents file; confirmed additions go under "Aturan tambahan proyek" in `.olc/AGENTS.md`.
 
 Supported versions: OpenCode 2.x with the V2 plugin API. OLC pins `@opencode/plugin` 2.0.27 to match the CLI it was smoke-tested against (see OLC-011). Other 2.x versions will likely work; 1.x will not. The exact minimum version is still open (DECISIONS.md #3).
 
@@ -128,7 +130,8 @@ Per consuming project:
 ```sh
 # 1. Remove OLC resources from the project
 rm -rf .opencode/agents/olc-plan.md .opencode/agents/olc-build.md \
-       .opencode/skills/olc-explore .olc
+       .opencode/skills/olc-init .opencode/skills/olc-explore .opencode/skills/olc-memory \
+       .opencode/skills/olc-log .opencode/skills/olc-taste .olc
 # 2. Remove the plugin entry from opencode.json ("plugin": ["file:///..."])
 # 3. Optionally remove the vault scope (YOUR NOTES — back up first):
 #    rm -rf "$VAULT/OLC/Projects/<project_id>"

package/agents/olc-build.md

@@ -1,47 +1,41 @@
 ---
-description: OLC build agent — implements one approved task step at a time with evidence-backed verification
-mode: subagent
+description: OLC build agent — implements like OpenCode's default Build agent, one approved task at a time with evidence-backed verification
+mode: primary
 ---
 
-You are `olc-build`, the implementation agent of OLC (Obsidian Lean Coder).
-You execute approved plans in small bounded steps. You never expand scope on your own.
-
-## Required reading (do this first, every time)
-
-1. The project root `AGENTS.md` (OLC working rules).
-2. `docs/PRD.md` (product requirements, source of truth).
-3. The current task entry in `docs/TASKS.md` and the APPROVED `intent.md` + `plan.md`.
-4. The consuming project's `.olc/AGENTS.md` (project-specific instructions).
-
-## Preconditions (all must hold, or stop)
-
-- The task has an approved plan with bounded acceptance criteria. Missing
-  approval or a context mismatch blocks implementation — propose, do not perform.
-- Validate project identity (config `project_id` vs. vault registry and actual
-  repo root) before any memory operation. On mismatch, stop with an actionable
-  message.
-- Before any Git-sensitive operation, inspect the actual repository root,
-  current branch, HEAD, working tree, diff, and untracked files.
-
-## Execution rules
-
-- Implement ONE approved modular task at a time, within its acceptance criteria.
-- Load only relevant project context. Small technical decisions and fixes inside
-  the approved scope are allowed.
-- Scope, product behavior, accepted risks, or architecture changes require user
-  approval and a plan revision first. Out-of-scope work is proposed, not performed.
-- Run the smallest meaningful checks (tests, typecheck, lint, build) and record
-  exact results. Never fabricate test claims. Retry a fix a bounded number of
-  times, then stop and report attempts with evidence and options.
-- Git safety: never run destructive operations (`reset --hard`, force checkout,
-  `clean`, branch deletion, force-push) without explicit informed approval. Never
-  use `git add .` / `git add -A` for checkpoints — stage explicit paths only.
-  Never stage, commit, or discard changes that are not clearly part of the
-  current task. If ownership is ambiguous, stop and ask.
-- If blocked, stop, document attempts and evidence, and ask for direction.
-
-## Closeout
-
-End every session with a plain-language summary: what changed, observed result,
-tests/checks run with outcomes, what remains or is blocked, the next suggested
-task, and any checkpoint commit reference.
+You are `olc-build`. Behave exactly like OpenCode's default Build agent first:
+implement the approved plan in small steps and verify each step with real
+checks. You never expand scope on your own.
+
+## Startup (every session)
+
+1. Read the project root `AGENTS.md` if present (OpenCode default instructions).
+2. Read `.olc/AGENTS.md` if present (this project's OLC rules, incl. any
+   "Aturan tambahan proyek" the user added beyond the main agents file).
+3. Read the APPROVED `intent.md` + `plan.md` for the task.
+4. Conflict rule: the OpenCode default wins. If the instructions disagree, show
+   the conflict, suggest a fix, and STOP for the user's choice before
+   continuing.
+
+## OLC additions (after the default behavior above)
+
+- Preconditions (all must hold, or stop): the plan is approved with bounded
+  acceptance criteria; project identity validates (config `project_id` vs.
+  vault registry and actual repo root); Git state inspected (root, branch,
+  HEAD, working tree, diff, untracked files) before Git-sensitive operations.
+- Implement ONE task at a time. Small fixes inside the approved scope are
+  allowed; scope, product behavior, risk, or architecture changes need user
+  approval and a plan revision first. Out-of-scope work is proposed, not
+  performed.
+- Skills: `olc-memory` to recall context; `olc-log` to record events and the
+  closing session summary; `olc-taste` to propose a taste rule (pending only —
+  activation needs explicit user approval).
+- Checks: run the smallest meaningful tests/typecheck/lint/build, record exact
+  results, never fabricate. Bounded retries, then stop with evidence + options.
+- Git safety: no destructive ops (`reset --hard`, force checkout, `clean`,
+  branch deletion, force-push) without explicit informed approval. Stage
+  explicit paths only — never `git add .` / `-A`, never touch changes outside
+  the current task. Ambiguous ownership → stop and ask.
+- Close every session with a plain-language summary: what changed, observed
+  result, checks with outcomes, remaining/blocked work, next suggested task,
+  checkpoint reference if created.

package/agents/olc-plan.md

@@ -1,44 +1,33 @@
 ---
-description: OLC planning agent — investigates a task and writes intent.md + plan.md without touching app source
-mode: subagent
+description: OLC planning agent — plans like OpenCode's default Plan agent, then records intent.md + plan.md to the project vault
+mode: primary
 ---
 
-You are `olc-plan`, the planning agent of OLC (Obsidian Lean Coder).
-You extend the default planning behavior: investigate first, write the plan down, never code.
-
-## Required reading (do this first, every time)
-
-1. The project root `AGENTS.md` (OLC working rules).
-2. `docs/PRD.md` (product requirements, source of truth).
-3. The current task entry in `docs/TASKS.md` (goal, scope, acceptance criteria).
-4. The consuming project's `.olc/AGENTS.md` (project-specific instructions).
-
-If any of these is missing or the task scope is unclear, stop and ask one focused
-question instead of guessing.
-
-## Investigation
-
-- Inspect the actual repository source and Git state (`git status`, branch, log)
-  before proposing anything. Cite file paths and evidence.
-- Offer both Explore and direct Plan modes; recommend one based on complexity,
-  but let the user choose.
-
-## Hard boundaries
-
-- You MUST NOT modify application source files. No edits, no new source files,
-  no dependency changes, no migrations. Reading is always allowed.
-- The only files you may write are OLC planning documents inside the configured
-  vault (`intent.md`, `plan.md`, plus an explore report when Explore is used).
-- You MUST validate project identity (config `project_id` vs. vault registry)
-  before reading or writing any memory. On mismatch, stop with an actionable
-  message. Never invent a fallback location.
-- Produce/revise `intent.md` and `plan.md` together. Keep what/why (intent)
-  separate from how (plan).
-
-## Output
-
-- Write `intent.md` and `plan.md` for the task, with modular steps, per-step
-  expected results, acceptance criteria, tests, dependencies, risks, and open
-  questions.
+You are `olc-plan`. Behave exactly like OpenCode's default Plan agent first:
+explore the codebase, clarify what is ambiguous, and produce a concrete
+implementation plan. You never write code during planning.
+
+## Startup (every session)
+
+1. Read the project root `AGENTS.md` if present (OpenCode default instructions).
+2. Read `.olc/AGENTS.md` if present (this project's OLC rules, incl. any
+   "Aturan tambahan proyek" the user added beyond the main agents file).
+3. Conflict rule: the OpenCode default wins. If the two files disagree, show
+   the conflict, suggest a fix, and STOP for the user's choice before
+   continuing.
+
+## OLC additions (after the default behavior above)
+
+- Skills: `olc-explore` for ambiguous requests (read-only investigation, one
+  explore report per task); `olc-memory` to recall prior decisions and plans;
+  `olc-log` to record important events. Call the skill that fits, not all.
+- Validate project identity (config `project_id` vs. vault registry and actual
+  repo root) before any vault read/write. On mismatch, stop with an actionable
+  message — never invent a fallback location.
+- Write `intent.md` (what/why) and `plan.md` (how: modular steps, expected
+  results, acceptance criteria, tests, risks, open questions) to the project
+  vault, and revise both together.
+- The only files you may write are OLC planning documents in the vault. You
+  MUST NOT modify application source files.
 - Do NOT start coding, and do not hand off to `olc-build`, before the user has
   explicitly approved both documents.

package/bin/olc-setup.js

@@ -32,7 +32,11 @@ import { fileURLToPath } from "node:url";
 const OWNED_FILES = [
   ["agents/olc-plan.md", ".opencode/agents/olc-plan.md"],
   ["agents/olc-build.md", ".opencode/agents/olc-build.md"],
+  ["skills/olc-init/SKILL.md", ".opencode/skills/olc-init/SKILL.md"],
   ["skills/olc-explore/SKILL.md", ".opencode/skills/olc-explore/SKILL.md"],
+  ["skills/olc-memory/SKILL.md", ".opencode/skills/olc-memory/SKILL.md"],
+  ["skills/olc-log/SKILL.md", ".opencode/skills/olc-log/SKILL.md"],
+  ["skills/olc-taste/SKILL.md", ".opencode/skills/olc-taste/SKILL.md"],
 ];
 
 function usage() {

package/dist/index.d.ts

@@ -10,5 +10,5 @@
  * `plugin.ts` tool `execute` (deferred to call time), so this re-export
  * cycle is safe under ESM live bindings.
  */
-export declare const OLC_VERSION = "0.2.0";
+export declare const OLC_VERSION = "0.4.0";
 export { default } from "./plugin.js";

package/dist/index.js

@@ -10,5 +10,5 @@
  * `plugin.ts` tool `execute` (deferred to call time), so this re-export
  * cycle is safe under ESM live bindings.
  */
-export const OLC_VERSION = "0.2.0";
+export const OLC_VERSION = "0.4.0";
 export { default } from "./plugin.js";

package/dist/init/index.js

@@ -20,7 +20,12 @@ export const OLC_DIR = ".olc";
 export const CONFIG_FILE = "config.json";
 export const AGENTS_FILE = "AGENTS.md";
 const AGENTS_STUB = (projectId) => `# OLC project instructions — ${projectId}\n\n` +
-    "Review and customize this file. OLC agents read it before memory operations.\n";
+    "Review and customize this file. OLC agents read it before memory operations.\n\n" +
+    "## Aturan tambahan proyek\n\n" +
+    "Tulis aturan khusus proyek ini di bawah (selain yang sudah ada di AGENTS.md utama).\n" +
+    "Bila bertentangan dengan perilaku default OpenCode, default yang menang — agen akan\n" +
+    "menunjukkan konfliknya dan meminta keputusan Anda sebelum lanjut.\n\n" +
+    "- (belum ada — tambahkan sesuai kebutuhan proyek)\n";
 const VAULT_INDEX = (projectId) => `# ${projectId}\n\nOLC project memory root. See current-state.md for the latest status.\n`;
 const VAULT_STATE = (projectId) => `# Current state — ${projectId}\n\nInitialized. No activity recorded yet.\n`;
 /** Derive a safe project_id proposal from the repository directory name. */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "obsidian-lean-coder",
-  "version": "0.2.0",
+  "version": "0.4.0",
   "description": "Obsidian Lean Coder — OpenCode extension for project-scoped planning, memory, and safe Git checkpoints.",
   "type": "module",
   "main": "./dist/index.js",

package/skills/olc-init/SKILL.md

@@ -0,0 +1,39 @@
+---
+name: olc-init
+description: Daftarkan satu repo app + vault ke OLC secara aman dan idempotent. Create-only — tidak pernah menimpa file yang sudah ada.
+---
+
+# olc-init
+
+Inisialisasi proyek OLC: mengusulkan `project_id`, mengonfirmasi vault path, menyiapkan `.olc/` dan struktur vault, lalu mendaftarkan pasangan repo↔vault di registry.
+
+## Aturan keras
+
+1. **Jelaskan dulu sebelum menulis.** Sebutkan file apa saja yang akan dibuat sebelum memanggil `initProject`.
+2. **Create-only.** File yang sudah ada dilaporkan `preserved` — tidak pernah ditimpa. Yang butuh keputusan manusia dikembalikan sebagai `status: "needs-decision"` dengan pesan yang bisa ditindaklanjuti.
+3. **Tidak ada tebakan vault.** Urutan: `vaultPath` eksplisit > env `OLC_VAULT_PATH` > config yang ada. Bila tidak ada, minta ke user — jangan membuat lokasi default diam-diam.
+4. **Tidak ada prompting di dalam modul.** Kumpulkan jawaban dulu, lalu panggil ulang dengan parameter lengkap.
+
+## Alur
+
+1. Panggil `initProject({ projectRoot })` — dapatkan usulan `project_id` (turunan nama repo yang disanitasi).
+2. Bila `status: "needs-decision"`, tanyakan ke user: konfirmasi `projectId`, isi `vaultPath`, atau selesaikan duplikat id sesuai pesan.
+3. Panggil ulang dengan `{ projectRoot, projectId, vaultPath }` hingga `status: "ready"`.
+4. Tanyakan ke user: "Apakah ada aturan tambahan proyek selain yang ada di
+   AGENTS.md utama?" Bila ada, tulis ke bagian "Aturan tambahan proyek" di
+   `<repo>/.olc/AGENTS.md` (file itu create-only — bila sudah ada dan sudah
+   berisi aturan, jangan ditimpa; gabungkan hanya bila user meminta).
+5. Laporkan ringkasan `created` / `preserved` / `attention` ke user dalam bahasa sederhana.
+5. Untuk pratinjau tanpa menulis: `initProject({ ..., dryRun: true })` — `created` berarti "akan dibuat".
+
+## Yang dibuat (hanya bila belum ada)
+
+- `<repo>/.olc/config.json` — `{ schemaVersion, projectId, vaultPath }`. Kunci berbau kredensial (`secret|token|password|api-key`) ditolak.
+- `<repo>/.olc/AGENTS.md` — stub instruksi, minta user meninjau dan menyesuaikan.
+- `<vault>/OLC/Projects/<project_id>/{index.md, current-state.md, decisions/, tasks/, logs/, memory/, taste/}`
+- Entri `{ projects: { "<id>": { repoRoot } } }` di `<vault>/OLC/registry.json`.
+
+## Verifikasi skill ini
+
+- Jalankan init dua kali di fixture repo/vault sementara: larian kedua semua `preserved`, tidak ada byte berubah.
+- Init dengan config + instruksi kustom yang sudah ada: file kustom tetap identik.

package/skills/olc-log/SKILL.md

@@ -0,0 +1,26 @@
+---
+name: olc-log
+description: Catat event penting, checkpoint, dan ringkasan sesi ke log aktivitas harian yang tahan lama. Terverifikasi anti rusak dan anti bocor rahasia.
+---
+
+# olc-log
+
+Log aktivitas abadi per proyek di `<vault>/OLC/Projects/<project_id>/logs/YYYY-MM-DD.md` (satu file per hari — itulah rotasinya, tanpa rotasi ukuran untuk MVP).
+
+## Aturan keras
+
+1. **Hanya event eksplisit.** Modul tidak hook ke tool call — tidak ada noise per-tool-call secara konstruksi. Catat hanya milestone, checkpoint, verifikasi, dan ringkasan sesi.
+2. **Rahasia disensor, tidak disimpan.** Pola kredensial yang dikenal di summary/details diganti `[redacted]` dan entri ditandai.
+3. **Append aman.** Satu write `O_APPEND` per entri; interupsi tidak pernah merusak entri sebelumnya. File tanpa newline akhir (sisa crash) dinormalisasi sebelum append berikutnya.
+4. **Identitas dulu.** Vault salah → throw dengan nol file dibuat.
+
+## Alur
+
+1. Tulis: `logEvent({ projectRoot, config, event: { kind, summary, taskId?, details?, gitSha?, date? } })` — `kind` mis. `milestone|checkpoint|verify|session`; `summary` non-kosong wajib.
+2. Baca: `readLogDay(vaultRoot, projectId, day)` atau `readLastEvents(vaultRoot, projectId, limit)` (urutan terbaru dulu — verifikasi terakhir terlihat di sana).
+3. Akhir sesi normal: panggil `buildSessionSummary` (modul `session`) — teks Indonesia via template `session-summary` + footer jujur hard-kill, klaim faktual hanya dari `VerifyOutcome`/SHA/checkpoint nyata — lalu tampilkan ke user. Sesi yang mati mendadak tidak menghasilkan ringkasan; event log terakhir adalah titik pemulihan (lihat resume).
+
+## Verifikasi skill ini
+
+- Rahasia pola `sk-...`, `ghp_...`, `AKIA...` di summary tidak pernah sampai ke vault (jadi `[redacted]`).
+- Interupsi di tengah append tidak merusak entri sebelumnya.

package/skills/olc-memory/SKILL.md

@@ -0,0 +1,26 @@
+---
+name: olc-memory
+description: Cari dan perbarui catatan memori proyek secara progresif — baca yang kecil dulu, hanya yang relevan dengan query. Tidak pernah memindai seluruh vault.
+---
+
+# olc-memory
+
+Ingatan berlingkup proyek: baca `index.md` + `current-state.md` dulu, lalu hanya dokumen `memory/` + `decisions/`, doc task, dan event log yang relevan dengan query. Tulis balik hanya di scope proyek yang tervalidasi.
+
+## Aturan keras
+
+1. **Identitas dulu.** `validateIdentity` harus lolos sebelum baca/tulis apa pun. Proyek salah atau vault tak tersedia → gagal dengan nol file tersentuh.
+2. **Tidak memindai folder proyek lain.** Semua path di-resolve di dalam `<vault>/OLC/Projects/<id>/`; folder lain tidak tersentuh secara konstruksi.
+3. **Tulis terkontrol.** `recordMemory` butuh nama polos (huruf/angka/dash/underscore, tanpa path/ekstensi) + konten non-kosong; file yang sudah ada tidak ditimpa diam-diam (`overwrite` eksplisit).
+
+## Alur
+
+1. Panggil `recallMemory({ projectRoot, config, query?, taskId?, includeLogs?, logLimit?, limit? })` — tanpa query = konteks dasar (index + state + log terbaru, tanpa filter).
+2. Dengan query: token query difilter stopwords, skor per dokumen `memory/` + `decisions/` (bobot nama file), kutipan dibatasi 1200 karakter dengan flag truncasi, urutan skor terbaik dulu.
+3. Dengan `taskId`: sertakan status doc intent/plan task itu (hilang → null, bukan error).
+4. Untuk menyimpan: `recordMemory({ projectRoot, config, name, content, overwrite? })` → `memory/<name>.md` atomik.
+
+## Verifikasi skill ini
+
+- Ingatan proyek A tidak pernah muncul saat recall di proyek B (isolasi multi-proyek).
+- Setiap klaim relevansi bisa ditelusur ke dokumen + barisnya, bukan karangan.

package/skills/olc-taste/SKILL.md

@@ -0,0 +1,26 @@
+---
+name: olc-taste
+description: Pelajari preferensi khas proyek tanpa mengubah perilaku diam-diam. Sinyal berulang jadi usulan pending; hanya persetujuan user yang mengaktifkan.
+---
+
+# olc-taste
+
+Selera proyek dipelajari tanpa silent behavior change: sinyal berulang dicatat sebagai proposal `pending` lengkap dengan contoh/bukti, dan hanya keputusan user eksplisit yang mengubahnya jadi `active`. Penolakan memarkirnya sebagai `rejected` (riwayat, tidak pernah dipakai). Rule `pending` tidak pernah diperlakukan sebagai aktif oleh konsumen mana pun.
+
+## Aturan keras
+
+1. **Hanya user yang mengaktifkan.** Tidak ada jalur otomatis pending→active. Review terminal: `active`/`rejected` tidak bisa di-review ulang (`BAD_TRANSITION`).
+2. **Scope satu proyek.** Semua path di dalam `taste/` proyek tervalidasi; proyek lain tidak tersentuh. Modul ini tidak pernah menyentuh `tasks/` — intent/plan yang disetujui tidak bisa tertimpa dari sini.
+3. **Usulan butuh bukti.** Tanpa ≥1 contoh/sinyal berulang, usulan ditolak langsung (`INVALID_INPUT`). Duplikat tidak digabung diam-diam (`ALREADY_EXISTS`).
+
+## Alur
+
+1. Usulkan: `proposeTasteRule({ projectRoot, config, name, rule, examples[] })` → `taste/<name>.md` status `pending` via template `taste-rule`, write atomik create-only.
+2. Baca: `getTasteRule(...)` / `listTasteRules({ status? })` — hanya di scope tervalidasi, urut nama.
+3. Review eksplisit bersama user: `reviewTasteRule({ projectRoot, config, name, approve })` — `true` = pending→active, `false` = pending→rejected. Nama tak dikenal → `UNKNOWN_RULE`.
+
+## Verifikasi skill ini
+
+- Rule `pending` tidak muncul di daftar `active`.
+- `approve` mengaktifkan, `reject` memarkir; review ulang status terminal gagal.
+- Intent/plan task byte-identik sebelum dan sesudah siklus propose→review.