feat: 發布 v1.0.0 訪談 reasoner 基線

.gitignore

@@ -16,9 +16,10 @@ build/
 
 # Environment
 .env
+.env*
+!.env.example
 .env.local
 .env.*.local
-!.env.example
 
 # SQLite databases (local data)
 *.db
@@ -47,6 +48,7 @@ Thumbs.db
 # Never commit actual interview transcripts or extracted SOUL/VOICE/SKILL files
 artifacts/
 transcripts/
+exports/
 *.transcript.md
 SOUL-*.md
 VOICE-*.md
@@ -67,6 +69,25 @@ STATE-*.md
 
 # Internal working dir (not for public)
 concept-for-chatgpt-diagram.md
+GEMINI.md
+docs/design/
+.private/
+private/
 
 # Claude Code local settings (operator-specific MCP refs, not for public)
 .claude/
+
+# === Moat protocol: v2 methodology + internal roadmap, withdrawn 2026-05-18 ===
+/src/virtualme/data/question-pool-v2.yaml
+/src/virtualme/data/domain-packs-v2.yaml
+/specs/09-interview-engine-v2.md
+/docs/research/virtualme-domain-pack-8-fields.md
+/src/virtualme/interview/v2_schema.py
+/src/virtualme/interview/v2_loader.py
+/tests/unit/test_interview_v2_loader.py
+/tests/unit/test_domain_packs_v2.py
+/docs/TRUNK.md
+
+# === Moat protocol: methodology implementations stay private（含 Version D 腳本 + 個人 PII）===
+/src/virtualme/interview/bait_reaction.py
+/tests/unit/test_bait_reaction.py

README.en.md

@@ -2,6 +2,21 @@
 
 > Extract a person into an AI agent — through an 8-week interview, not a form.
 
+**Current release: v1.1.0** — baseline interview + coverage tracking + persona markdown export, plus Constitution v1.1 (six Stability & Restraint Principles) and matching M1 hard gates.
+
+### v1.1.0 Highlights
+
+- **[Constitution v1.1](specs/11-constitution.md)** — ratified by seven-agent council on 2026-05-20. Codifies the project's "caution, restraint, reverence" stance previously scattered across `docs/TRUNK.md` / `specs/05` / milestones into six principles: P1 State-Trait Separation / P2 Contradiction Preservation / P3 Reflective Restraint / P4 Multi-Session Validation / P5 Self-Correction & Agency / P6 Provenance, Confidence & Temporal Decay.
+- **Interview reasoning engine refactor** — L0 transport idempotency fail-closed + L1 read-only TurnState + L2 `turn_reasoner.decide_and_reply()` + Guardrail + feature flag (`reasoning_turn_enabled`) whitelist rollout.
+- **User-initiated persona archive export + download link**
+- **M1 hard gate detectors (4 principles) + 115 contract tests**:
+  - P3 — `SkipStopReason` enum + Guardrail metadata + `reflection_note` no-leak
+  - P5 — `hedge_validator` (8 forbidden patterns / 12 hedge markers) + `unlike_me` regression
+  - P1 — `stability_gate.is_eligible_for_core_truths()` (STATE never enters Core Truths)
+  - P4 — `multi_session_validator.can_be_validated()` (single-session can never be validated)
+
+> M2 will wire detectors into `build_snapshot_bundle` / export pipeline. This release ships detectors + contract tests that lock the invariants. See `specs/11-constitution.md` §M2/M3.
+
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
 [![Python 3.11+](https://img.shields.io/badge/Python-3.11+-blue.svg)](https://www.python.org/)
 [![中文](https://img.shields.io/badge/Lang-中文-red.svg)](README.md)

README.md

@@ -2,6 +2,22 @@
 
 > 把一個人**萃取**成 AI 代理人——用 8 週訪談，不用填表。
 
+**Current release: v1.1.0** — baseline interview + coverage tracking + persona markdown export 之上，補上 Constitution v1.1（六條 Stability & Restraint Principles）與對應的 M1 hard gates。
+
+### v1.1.0 Highlights
+
+- **[Constitution v1.1](specs/11-constitution.md)** — 七位一體 council 2026-05-20 ratified；把先前散落於 `docs/TRUNK.md` / `specs/05` / milestone 的「謹慎、克制、有敬畏」立場 codify 為六條：P1 State-Trait Separation / P2 Contradiction Preservation / P3 Reflective Restraint / P4 Multi-Session Validation / P5 Self-Correction & Agency / P6 Provenance, Confidence & Temporal Decay
+- **訪談 reasoning engine 重構** — L0 transport idempotency fail-closed + L1 TurnState 只讀狀態物件 + L2 `turn_reasoner.decide_and_reply()` + Guardrail + feature flag (`reasoning_turn_enabled`) whitelist rollout
+- **使用者自助匯出人格檔 + 下載連結**
+- **Production demo map** — [`docs/architecture-demo-flow.md`](docs/architecture-demo-flow.md) documents the deployed LINE Bot / VPS architecture and a short demo script.
+- **M1 hard gate detectors（4 條）+ 115 contract tests**：
+  - P3 — `SkipStopReason` enum + Guardrail metadata + reflection_note no-leak
+  - P5 — `hedge_validator`（8 forbidden patterns / 12 hedge markers）+ unlike_me regression
+  - P1 — `stability_gate.is_eligible_for_core_truths()`（STATE 不進 Core Truths）
+  - P4 — `multi_session_validator.can_be_validated()`（single-session 不得 validated）
+
+> M2 將把 detector wire 進 build_snapshot_bundle / export pipeline；本版只交付 detector + contract test 鎖住 invariant。詳見 `specs/11-constitution.md` §M2/M3。
+
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
 [![Python 3.11+](https://img.shields.io/badge/Python-3.11+-blue.svg)](https://www.python.org/)
 [![English](https://img.shields.io/badge/Lang-English-red.svg)](README.en.md)
@@ -89,6 +105,11 @@ VirtualMe 把這個發現延伸成可上線的 pipeline：
 
 匯出時會再次 scrub anchor 內容中的 PII；`interviewee_id`、資料夾名與 archive metadata 不會被改名，請不要用 email / 真名當 interviewee id。
 
+v1.0.0 的輸出包含兩種層級：
+
+- **Raw archive**：`python -m virtualme.export` 產生 8 個 dimension markdown、入口檔與 manifest。
+- **Review draft**：可依 anchors 人工整理成 `SOUL.md` / `VOICE.md` / `SKILL.md` / `PEOPLE.md` / `HISTORY.md` / `JOURNAL.md` / `BOUNDARIES.md` / `STATE.md` 八份可讀人格檔，用於「像 / 不像 / 缺例子」review。
+
 加上一個可用的 agent endpoint，可以：
 - 起草給客戶 / 候選人 / 同事的訊息
 - 用你的語氣回覆公開貼文
@@ -159,6 +180,15 @@ python scripts/init_db.py --path ./data/virtualme.db
 python -m virtualme.cli --interviewee yourself
 ```
 
+### LINE dogfood
+
+v1.0.0 的 LINE path 可用 feature flag 開啟 supervised dogfood。新路徑會在每輪訪談後寫入 anchors，讓「進度」查詢顯示真實 coverage。
+
+```env
+REASONING_TURN_ENABLED=true
+REASONING_TEST_USER_IDS=<your-line-user-id>
+```
+
 ### 本機 demo flow
 
 訪談累積資料後，可以先輸出 markdown archive，再準備一次手動 blind test：