hotmilk 0.1.13 → 0.1.15

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "hotmilk",
-  "version": "0.1.13",
+  "version": "0.1.15",
   "description": "hotmilk — Pi package bundling gentle-pi, context-mode, graphify, subagents, and bundled extension toggles via hotmilk.json",
   "keywords": [
     "hotmilk",
@@ -38,45 +38,63 @@
   },
   "dependencies": {
     "@blackbelt-technology/pi-agent-dashboard": "^0.5.4",
-    "@blackbelt-technology/pi-flows": "^0.2.1",
+    "@blackbelt-technology/pi-flows": "^0.3.2",
     "@haispeed/pi-obsidian": "^0.1.1",
+    "@llblab/pi-actors": "^0.39.0",
+    "@mrclrchtr/supi-context": "^2.2.1",
+    "@plannotator/pi-extension": "^0.22.0",
     "@tomxprime/planning-with-files": "^1.1.0",
-    "context-mode": "^1.0.151",
-    "gentle-pi": "^0.4.5",
+    "context-mode": "^1.0.169",
+    "gentle-pi": "^0.13.0",
     "graphify-pi": "^0.3.0",
-    "pi-ask-user": "^0.11.2",
-    "pi-btw": "^0.4.0",
+    "latchkey": "^2.20.1",
+    "pi-ask-user": "^0.12.0",
+    "pi-autoresearch": "^1.6.2",
+    "pi-btw": "^0.4.1",
     "pi-caveman": "^1.0.7",
     "pi-docparser": "^3.0.1",
-    "pi-goal": "^0.1.5",
+    "pi-goal": "^0.1.7",
     "pi-kanagawa": "^1.0.0",
-    "pi-mcp-adapter": "^2.9.0",
+    "pi-mcp-adapter": "^2.11.0",
+    "pi-observational-memory": "^3.0.3",
+    "pi-prompt-template-model": "^0.10.0",
     "pi-red-green": "^0.2.2",
-    "pi-rtk-optimizer": "^0.8.2",
+    "pi-rtk-optimizer": "^0.9.0",
+    "pi-shazam": "^0.26.0",
     "pi-simplify": "^0.2.2",
-    "pi-subagents": "^0.28.0",
+    "pi-subagents": "^0.34.0",
     "pi-tetris": "^1.3.1",
-    "pi-web-access": "^0.10.7"
+    "pi-web-access": "^0.13.0"
   },
   "devDependencies": {
-    "@earendil-works/pi-agent-core": "^0.78.1",
-    "@earendil-works/pi-ai": "^0.78.1",
-    "@earendil-works/pi-coding-agent": "^0.78.1",
-    "@earendil-works/pi-tui": "^0.78.1",
-    "@types/node": "^22.19.19",
-    "typebox": "^1.2.0",
-    "vite-plus": "^0.1.24"
+    "@earendil-works/pi-agent-core": "^0.80.6",
+    "@earendil-works/pi-ai": "^0.80.6",
+    "@earendil-works/pi-coding-agent": "^0.80.6",
+    "@earendil-works/pi-tui": "^0.80.6",
+    "@types/node": "^22.20.1",
+    "typebox": "^1.3.6",
+    "vite-plus": "^0.2.4"
   },
   "peerDependencies": {
-    "@earendil-works/pi-agent-core": "^0.78.1",
-    "@earendil-works/pi-ai": "^0.78.1",
-    "@earendil-works/pi-coding-agent": "^0.78.1",
-    "@earendil-works/pi-tui": "^0.78.1",
-    "typebox": "^1.2.0"
+    "@earendil-works/pi-agent-core": "^0.80.6",
+    "@earendil-works/pi-ai": "^0.80.6",
+    "@earendil-works/pi-coding-agent": "^0.80.6",
+    "@earendil-works/pi-tui": "^0.80.6",
+    "typebox": "^1.3.6"
   },
   "overrides": {
+    "@earendil-works/pi-coding-agent": "^0.80.6",
     "@fastify/static": "9.1.3",
-    "@mozilla/readability": "0.6.0"
+    "@mariozechner/pi-coding-agent": "npm:@earendil-works/pi-coding-agent@^0.80.6",
+    "@mozilla/readability": "0.6.0",
+    "dompurify": "^3.4.11",
+    "esbuild": "0.28.1",
+    "hono": "^4.12.28",
+    "linkify-it": "^5.0.1",
+    "protobufjs": "^7.6.3",
+    "undici": "^7.28.0",
+    "vite": "^8.1.4",
+    "ws": "8.21.0"
   },
   "engines": {
     "node": "22"
@@ -100,6 +118,7 @@
       "node_modules/@haispeed/pi-obsidian/skills",
       "node_modules/pi-ask-user/skills",
       "node_modules/pi-web-access/skills",
+      "node_modules/pi-autoresearch/skills",
       "node_modules/@tomxprime/planning-with-files"
     ],
     "themes": [

package/prompts/prompt-eval.md

@@ -0,0 +1,241 @@
+---
+description: Harden a prompt or skill with fresh subagent evaluation loops
+argument-hint: "<path to SKILL.md, AGENTS.md section, or prompt file>"
+---
+
+Run a parent-orchestrated prompt hardening loop for the target below.
+
+Target:
+
+$@
+
+Keep this parent session as loop controller and sole editor. Child subagents must not spawn subagents, edit files, or inherit authoring context.
+
+## Core rule: fresh executor
+
+A prompt's quality is opaque to its author. Dispatch a **new** fresh executor every iteration. Self re-reading never qualifies.
+
+Always set on executor and instructor dispatches:
+
+- `context: "fresh"`
+- `skill: false`
+- `clarify: false`
+- explicit `model` when the default resolver fails
+
+**NOT fresh**: `context: "fork"`, same session, or reusing a prior executor's thread.
+
+If `subagent` is unavailable, ask the user to open a separate session or report `prompt-eval skipped: no fresh agent available`. Never substitute self re-reading.
+
+## Step 0 — Static coherence check (parent inline)
+
+Before any dispatch, compare what the target **advertises** vs what the body **delivers**. Fix divergences first.
+
+| Target type                   | "Advertises" source          | Check                                       |
+| ----------------------------- | ---------------------------- | ------------------------------------------- |
+| Skill file                    | Frontmatter `description`    | Triggers/use-cases match body coverage?     |
+| CLAUDE.md / AGENTS.md section | Section heading              | Heading promise matches actionable content? |
+| Task prompt / system prompt   | Opening instruction or title | Stated goal matches the actual rules?       |
+
+Skipping this causes false positives — the fresh executor silently re-interprets the body.
+
+## Step 1 — Baseline (parent inline)
+
+1. Read the target file and freeze one stakes tier:
+
+| Stakes                            | Scenarios             | `[critical]` tags | Stop condition       | Iteration cap |
+| --------------------------------- | --------------------- | ----------------- | -------------------- | ------------- |
+| **High** (core skill, automation) | 3 (1 median + 2 edge) | ≥ 2               | 3 consecutive clears | none          |
+| **Standard** (default)            | 2 (1 median + 1 edge) | ≥ 1               | 2 consecutive clears | 5             |
+| **Low** (small, low-blast)        | 2 (1 median + 1 edge) | ≥ 1               | ship at 80%          | 3             |
+
+2. Design scenarios and requirement checklists. Do not change checklists after seeing results.
+
+Checklist rules:
+
+- 3–7 items per scenario.
+- At least one `[critical]` item — without it, success judgment is vacuous.
+- Translate user complaints into observable criteria (e.g. "shallow" → "names at least one concrete bug with line reference").
+- `[critical]` failure = whole scenario fails (binary ×), regardless of other items.
+- Non-critical items: ○ = 1, partial = 0.5, × = 0. Accuracy = sum / total.
+
+3. Paste the **full target prompt text** into every executor task (not a path).
+
+## Step 2 — Evaluation loop
+
+Repeat until convergence, iteration cap, or divergence. Each iteration:
+
+### 2a. Fresh executor (one per scenario)
+
+For each scenario, launch:
+
+```text
+subagent({
+  agent: "assistant",
+  context: "fresh",
+  skill: false,
+  clarify: false,
+  model: "<explicit-model-id>",
+  task: "<executor contract below with full pasted prompt + one scenario + its checklist>"
+})
+```
+
+Executor task contract:
+
+```markdown
+You are an executor reading <target prompt name> as a blank slate.
+You have no prior exposure to this prompt; treat it as newly encountered.
+
+## Target Prompt
+
+<paste the full target prompt here>
+
+## Scenario
+
+<one paragraph describing the situation>
+
+## Requirement Checklist
+
+1. [critical] <item>
+2. <item>
+   ...
+
+## Task
+
+1. Execute the scenario following the target prompt and produce the deliverable.
+2. On completion, respond with the report below.
+
+## Report Structure
+
+- Deliverable: <the artifact or run summary>
+- Requirement status: for each item, ○ / × / partial (with reason)
+- Ambiguities: places where wording was open to interpretation (bullets)
+- Discretionary fills: decisions the instructions did not specify (bullets)
+- Retries: how many times you redid a decision and why
+```
+
+Parallel scenarios: use `subagent({ tasks: [...] })` with models from the **same provider** only.
+
+### 2b. Instructor scoring (fresh reviewer)
+
+Launch one fresh reviewer per iteration (or one per scenario when scenarios are large):
+
+```text
+subagent({
+  agent: "reviewer",
+  context: "fresh",
+  skill: false,
+  clarify: false,
+  model: "<explicit-model-id>",
+  task: "Score executor reports using the rubric below. Return: per-scenario Success/Accuracy/Steps/Duration/Retries table, new ambiguities, new discretionary fills, and the single highest-impact ambiguity to fix next. Qualitative signals outweigh raw speed."
+})
+```
+
+Instructor rubric:
+
+| Axis                    | Source                | How to measure                                                                                                                                        |
+| ----------------------- | --------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------- |
+| **Success/Failure**     | Instructor (reviewer) | ○ only if every `[critical]` met. Binary.                                                                                                             |
+| **Accuracy**            | Instructor            | ○/partial/× per item → percentage                                                                                                                     |
+| **Step count**          | Environment metadata  | `tool_uses` or tool-use blocks. Text-only targets: count review comments, checklist items, or code blocks — state the unit once in the iteration log. |
+| **Duration**            | Environment metadata  | `duration_ms` or wall-clock                                                                                                                           |
+| **Retry count**         | Self-report           | Times the agent redid a decision                                                                                                                      |
+| **Ambiguities**         | Self-report           | Primary signal for improvement                                                                                                                        |
+| **Discretionary fills** | Self-report           | Surfaces implicit requirements                                                                                                                        |
+
+Weighting: ambiguities and discretionary fills are primary. Quantitative metrics are supporting. Chasing time alone makes prompts brittle.
+
+On failure: append one line to Ambiguities naming which `[critical]` item dropped.
+
+Step-count skew (even at 100% accuracy): one scenario at 3–5× the others → prompt lacks self-containment for that case. Fix with inline guidance or a minimal complete example.
+
+Reviewers must not edit files. They score from executor self-reports plus environment metadata when available.
+
+### 2c. Parent applies ONE delta
+
+1. Identify the most impactful ambiguity from the instructor report.
+2. State which checklist item the edit satisfies.
+3. Apply the minimum edit to the target file — one semantic theme only.
+4. Record the iteration:
+
+```
+## Iteration N
+
+### Changes (delta from previous)
+- <one-line edit description>
+
+### Results
+| Scenario | Success | Accuracy | Steps | Duration | Retries |
+|---|---|---|---|---|---|
+| A | ○ | 90% | 4 | 20s | 0 |
+| B | × | 60% | 9 | 41s | 2 |
+
+### Ambiguities (new this iteration)
+- <Scenario B>: [critical] item N was × — <reason>
+- <Scenario A>: (none new)
+
+### Discretionary fills (new this iteration)
+- <Scenario B>: <what was filled>
+
+### Next edit
+- <one-line minimal edit>
+
+(Convergence: X consecutive clears / Y iterations until stop)
+```
+
+Never leave an axis blank (write "0", not "—"). "None new" is explicit signal.
+
+Delta rules: one semantic theme per iteration. A cluster of 2–3 related micro-edits is one theme; unrelated edits go to the next iteration.
+
+### 2d. Re-evaluate
+
+Dispatch **new** fresh executors. Never reuse prior executor threads.
+
+## Stopping
+
+**Converged** when for N consecutive iterations (N = 2 standard, 3 high-stakes):
+
+- New ambiguities: 0
+- Accuracy gain: ≤ +3 points
+- Step count change: within ±10%
+- Duration change: within ±15%
+
+**Diverged**: after 3+ iterations ambiguities aren't decreasing → stop patching, rewrite.
+
+**Resource cutoff**: ship at 80% when improvement cost exceeds prompt importance.
+
+Before declaring done on any tier, run one hold-out scenario not used earlier. Accuracy drop ≥ 15 points → overfitting; return to scenario design.
+
+## Completion summary
+
+Report:
+
+- stakes tier and scenarios used
+- iterations run and deltas applied
+- final per-scenario scores
+- hold-out result (if run)
+- plateau reason (converged / cap / diverged / skipped)
+- remaining ambiguities worth accepting or deferring
+
+## Red flags
+
+| Rationalization                                  | Reality                                                          |
+| ------------------------------------------------ | ---------------------------------------------------------------- |
+| "Re-reading it myself is equivalent."            | Structurally impossible to objectively view text you just wrote. |
+| "One scenario is enough."                        | Overfits. Minimum 2.                                             |
+| "Zero ambiguities once, so done."                | May be coincidence. Require consecutive clears.                  |
+| "Fix multiple things in one pass."               | Lose attribution. One theme per iteration.                       |
+| "Metrics look good, ignore qualitative."         | Faster ≠ better. Qualitative is primary.                         |
+| "Reuse the same fresh agent."                    | It learned. Dispatch anew every time.                            |
+| "Split every micro-edit into its own iteration." | Over-splitting. One semantic unit = one iteration.               |
+| "Rewrite from scratch."                          | Valid only after 3+ stalled iterations.                          |
+
+## Subagent troubleshooting
+
+| Symptom                     | Cause                    | Fix                        |
+| --------------------------- | ------------------------ | -------------------------- |
+| "No API key found"          | Default model resolution | Specify `model` explicitly |
+| Agent reads wrong files     | Skill/cwd pollution      | Add `skill: false`         |
+| Parallel timeout (exit 143) | Cross-provider mixing    | Same-provider models only  |
+| Subagent hangs              | Provider instability     | Try different model id     |
+
+Structural-audit mode: for coherence/clarity checks only (not empirical behavior), mark the request with _"structural-audit mode: check textual coherence only, do not execute."_ Cannot count toward consecutive-clear convergence.

package/prompts/tidy.md

@@ -18,12 +18,14 @@ All of these must hold:
 
 If you would hesitate to ship it by itself, it is not tidying.
 
+**Extract boundary:** one extract moves a small cohesive unit with no logic change. Restructuring most of a large function is normal refactor work, not tidying.
+
 ## Rules
 
 1. **Name the tidying** when you describe the change
 2. **One tidying at a time**
 3. **Stay in scope** — only that cleanup
-4. **Tests pass** after the change
+4. **Tests pass** after the change — run the relevant test command and report the result; if you cannot run tests, say why
 
 ## Agent instructions
 

package/prompts/translate.md

@@ -1,5 +1,5 @@
 ---
-description: Translate text to Japanese
+description: Translate text between English and Japanese (auto-detect direction)
 argument-hint: "<text to translate>"
 ---
 
@@ -8,6 +8,8 @@ You are a professional translator fluent in both English and Japanese. Your task
 Guidelines:
 
 - Maintain the original intent and nuance of the text
+- Match the source's force and register: prohibitions stay prohibitions, advisories stay advisories, polite requests stay polite; do not soften or strengthen
+- Apply register matching independently to each sentence or clause
 - Use natural, fluent Japanese expressions
 - Preserve technical terms when appropriate
 - Keep formatting and special characters intact

package/skills/make-docs/SKILL.md

@@ -0,0 +1,43 @@
+---
+name: make-docs
+description: "Greenfield: create docs/* (9 files), README.md from interview or repo scan. If docs/ or README.md already exists, stop — use doc sync instead. [Triggers: /make-docs, create docs, init docs, generate AGENTS.md]"
+disable-model-invocation: true
+---
+
+# /make-docs — Documentation bootstrap
+
+**前提**: `docs/` も `README.md` も無いリポジトリだけ。既にあるなら **update-docs 系に切り替え**。このスキルは初回スキャフォールド用で、勝手に上書きしない。
+
+## モード選択（最初に確定・途中で混ぜない）
+
+```
+リポジトリ無し / 仕様だけ → Interactive（2〜3 問ずつ）
+リポジトリあり かつ docs と AGENTS が無い → Auto-scan（下の bash）→ 足りない所だけ短命インタビュー
+docs または AGENTS が既にある → **STOP**（同期ワークフローへ）
+```
+
+## Auto-scan（プロジェクトルートで実行）
+
+```bash
+find . -type f -not -path '*/node_modules/*' -not -path '*/.git/*' -not -path '*/dist/*' -not -path '*/.next/*' -not -path '*/build/*' -not -path '*/target/*' | head -100
+cat README.md AGENTS.md 2>/dev/null
+cat package.json 2>/dev/null || cat Cargo.toml 2>/dev/null || cat pyproject.toml 2>/dev/null
+```
+
+## 現在のプロジェクトのファイルを読み内容に合わせて作成、またはインタラクティブに作成
+
+## progresive disclosure
+
+- `./docs/requirements.md` (要件定義)
+- `./docs/design.md` (詳細設計、スキーマ設計)
+- `./docs/directory.md` (ディレクトリ構成)
+- `./docs/tech.md` (技術スタック)
+- `./docs/tasks.md` (タスク管理、進捗管理)
+- `./docs/test.md` (テスト設計、テストガイド)
+- `./docs/security.md` (セキュリティ設計)
+- `./docs/problems.md` (注意点、落とし穴)
+- `./docs/references.md` (参考文献、サンプルコード)
+
+## テーブル, mermaidダイアグラム
+
+- 必要に応じてtasks.md と連動した `ROADMAP.md` も作成