devswha · devswha · Jul 29, 2026 · Jul 24, 2026 · Jul 24, 2026 · Jul 24, 2026
diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
@@ -10,7 +10,7 @@
     {
       "name": "patina",
       "source": "./",
-      "version": "6.3.1",
+      "version": "6.3.2",
       "description": "Strip the AI packaging, keep the meaning. Detect and rewrite AI writing patterns across KO/EN/ZH/JA with MPS/fidelity checks. Runs the root SKILL.md as the /patina skill.",
       "homepage": "https://github.com/devswha/patina",
       "repository": "https://github.com/devswha/patina",

diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
@@ -1,7 +1,7 @@
 {
   "$schema": "https://www.schemastore.org/claude-code-plugin-manifest.json",
   "name": "patina",
-  "version": "6.3.1",
+  "version": "6.3.2",
   "description": "Detect and rewrite AI writing patterns in Korean, English, Chinese, and Japanese so text reads as if a human wrote it. Meaning-preservation (MPS) verified, audit-friendly. The root SKILL.md is loaded as the /patina skill.",
   "author": {
     "name": "devswha",

diff --git a/.env.example b/.env.example
@@ -1,38 +1,25 @@
-# Local-only runtime / Discord config
-# Copy to .env and fill with your own values.
-
-DISCORD_CHANNEL=your-discord-channel-id
-DISCORD_GUILD=your-discord-guild-id
-DISCORD_ALLOWED_USERS=comma-separated-user-ids
-
-PATINA_AGENT_ID=patina
-PLANNER_AGENT_ID=planner
-GENERATOR_AGENT_ID=generator
-EVALUATOR_AGENT_ID=evaluator
-
-RUNTIME_ENFORCE_ALLOWLIST=true
-# ↑ Default flipped to true. Bootstrap will refuse to start unless
-# DISCORD_ALLOWED_USERS is set. Set this back to false ONLY if you understand
-# that any user who can post in the bound channel can drive the runtime.
-PATINA_RUNTIME_CLI=your-runtime-cli-binary
-
-# Optional: override if your old bot token lives elsewhere
-# CLAWHIP_CONFIG=/home/you/.clawhip/config.toml
-# Optional: direct token override if your runtime does not reuse the migrated bot token
-# RUNTIME_DISCORD_TOKEN=
-
-# Optional: isolated marketing bot profile (separate OpenClaw state/config)
-# MARKETING_RUNTIME_PROFILE=marketing
-# MARKETING_AGENT_ID=patina-marketing
-# MARKETING_SOURCE_AGENT_ID=patina
-# MARKETING_DISCORD_GUILD=your-discord-guild-id
-# MARKETING_DISCORD_CHANNEL=your-marketing-channel-id
-# MARKETING_DISCORD_ALLOWED_USERS=comma-separated-user-ids
-# MARKETING_DISCORD_TOKEN=
-# MARKETING_GATEWAY_PORT=18889
-# MARKETING_WORKSPACE=/home/you/.openclaw-marketing/workspace
-# MARKETING_ENFORCE_ALLOWLIST=true   # default; bootstrap refuses without MARKETING_DISCORD_ALLOWED_USERS
-# MARKETING_RESTART_GATEWAY=false
+# patina environment reference — copy to .env and fill only what you use.
+# .env is gitignored; never commit real values.
+
+# ---------------------------------------------------------------------------
+# CLI / local LLM provider keys (openai-http backend, quality:live harness)
+# Lookup order: PATINA_API_KEY first, then the provider-specific key
+# (src/auth.js HTTP_KEY_ENV_VARS). `--provider <name>` presets live in
+# src/providers.js: openai, gemini, groq, kimi, moonshot, together.
+# ---------------------------------------------------------------------------
+# PATINA_API_KEY=            # generic key, wins over provider-specific vars
+# OPENAI_API_KEY=            # default openai-http backend (gpt-5.5)
+# GEMINI_API_KEY=            # --provider gemini (gemini-2.5-pro)
+# GROQ_API_KEY=              # --provider groq
+# TOGETHER_API_KEY=          # --provider together
+# KIMI_API_KEY=              # --provider kimi (kimi-k2.5, api.moonshot.ai)
+# MOONSHOT_API_KEY=          # --provider moonshot (same endpoint as kimi)
+# PATINA_API_KEY_FILE=       # or read any key from a file (recommended for CI)
+
+# Local subscription CLIs (claude / codex / gemini / kimi) need no key here —
+# a logged-in seat is the credential. See also PATINA_LIVE_* and
+# PATINA_LIVE_JUDGE_* (fixed-judge scoring) in tests/quality/README.md.
+
 # ---------------------------------------------------------------------------
 # Hosted playground / API service (Vercel `/api/rewrite`)
 # Server-side only. The browser never sees these. All tiers fail closed:
@@ -46,9 +33,15 @@ PATINA_RUNTIME_CLI=your-runtime-cli-binary
 
 # Free tier: the service's own provider key (rate-limited by IP) + HMAC secret
 # that signs quota/subject KV keys (raw keys/licenses are never stored).
+#
+# Model choice (owner-approved 2026-07-26): gemini-3.6-flash. The prior default
+# gpt-4.1-mini measured an AI-score improvement of only 0.2 across 22 fixtures
+# — it returns the input nearly unchanged, so free traffic received no real
+# rewrite while still scoring high on meaning preservation. gemini-3.6-flash
+# measured 13.0. See docs/operations/serving-engine-cost-20260725.md.
 # PATINA_FREE_API_KEY=your-provider-key
-# PATINA_FREE_PROVIDER=openai
-# PATINA_FREE_MODEL=gpt-4.1-mini
+# PATINA_FREE_PROVIDER=gemini
+# PATINA_FREE_MODEL=gemini-3.6-flash
 # PATINA_QUOTA_HMAC_SECRET=your-long-random-secret
 
 # --- Pro tier ($9.99/mo USD, Lemon Squeezy license-gated) -------------------
@@ -72,10 +65,42 @@ PATINA_RUNTIME_CLI=your-runtime-cli-binary
 # Pro provider/model — REQUIRED in production (missing values fail pro requests
 # closed with 503; no silent fallback to the free provider/model). Outside
 # production they fall back to PATINA_FREE_PROVIDER/MODEL, then the preset.
-# Both must be on the PROVIDER_PRESETS allowlist. Pro ships on Claude Sonnet 5.
+# Both must be on the PROVIDER_PRESETS allowlist.
+#
+# Model choice (owner-approved 2026-07-26; evidence re-established 2026-07-27):
+# gemini-3.6-flash, replacing claude-sonnet-5. Rerun on 22 live-quality
+# fixtures after fixing a broken fidelity rubric and a harness prompt that
+# omitted the persona (docs/operations/serving-engine-cost-20260725.md): the
+# two engines both score 20/22 with MPS 90.4 vs 91.5. Quality is level; the
+# case is cost and latency — $0.030 vs $0.156 per rewrite, 8.3s vs 27.7s.
+# PATINA_PRO_PROVIDER=gemini
+# PATINA_PRO_MODEL=gemini-3.6-flash
+#
+# Previous pin, kept for rollback:
 # PATINA_PRO_PROVIDER=claude
 # PATINA_PRO_MODEL=claude-sonnet-5
 
+# --- Prompt caching (Anthropic only, opt-in) --------------------------------
+# The rewrite prompt is ~99.7% fixed prefix (pattern catalog + profile + voice)
+# with only the user text after the data fence, so it caches almost perfectly.
+# Anthropic's OpenAI-compat endpoint ignores cache_control, so this flag routes
+# paid calls through the native /v1/messages adapter with a cached prefix.
+#
+# Measured 2026-07-25 (ko web prompt, claude-sonnet-5): 34,254 cacheable prefix
+# tokens; second identical call read 34,254/34,254 from cache and halved latency
+# (25.2s -> 12.0s). Cost per request: $0.126 uncached, $0.151 on a cache miss
+# (+20%, the 5m write surcharge), $0.033 on a hit (-74%).
+#
+# Break-even hit rate is 21.7%, i.e. roughly 3 requests/hour per language
+# variant (the 5-minute ephemeral TTL refreshes on each hit). Below that this
+# flag is a ~20% input surcharge; above it, savings approach 74%.
+# PATINA_ANTHROPIC_NATIVE_CACHE=1
+#
+# Thinking stays at the provider default (ON for sonnet-5) because a measured
+# A/B showed thinking-off rewrites amputate content. Set to 0 only for
+# experiments, never to cut cost on the paid path.
+# PATINA_ANTHROPIC_THINKING=1
+
 # Pro limits (per-license). Defaults shown.
 # PATINA_PRO_MAX_CHARS=20000
 # PATINA_PRO_REQ_PER_DAY=200

diff --git a/.patina.default.yaml b/.patina.default.yaml
@@ -1,4 +1,4 @@
-version: "6.3.1"
+version: "6.3.2"
 language: ko              # Korean (default) -- auto-loads all ko-*.md patterns
                           # language: en      -- English -- auto-loads all en-*.md patterns
                           # language: zh      -- Chinese -- auto-loads all zh-*.md patterns

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -12,6 +12,29 @@ All notable changes to patina. Dates are release dates (YYYY-MM-DD).
 Semver rationale: patch | minor | major — explain whether this changes patterns, schemas, CLI behavior, or docs only.
 ```
 
+## 6.3.2 — 2026-07-29
+
+**Hosted-rewrite fixes (chat-log pastes no longer 422), a fidelity-gate scoring fix, backend compatibility, and opt-in LLM plumbing. 6.4 stays reserved for the payment launch.**
+
+Semver rationale: patch — bug fixes on the hosted rewrite surface, scoring, and CLI backends, plus strictly opt-in or allowlist-only additions; no default contract, tier limit, or pinned model changes. The v6.4 payment-launch hold (tag/publish prohibitions for 6.4.x) is untouched.
+
+### Fixed
+
+- **Chat-log/timeline pastes no longer fail the web rewrite**: clock times (`16:47`, `16:47 – 16:50`, `9:05:30`) hit the numeric-operator check as `digit:digit`, so the number-safety gate 422-rejected the whole paste before scoring — even for an identity rewrite — and the playground showed a misleading generic "check the mode/key" failure. Clock times are now first-class exact `time:` claims (leading-zero hour normalized); a rewrite that drops or drifts a time still fails closed as `numeric_claim_changed`, and ratios/scores (`1:2`, `3:1`) plus invalid times (`25:30`) keep the old fail-closed behavior. `number_safety_failed` also gets its own localized (en/ko/zh/ja) error copy in the playground.
+- **Fidelity gate no longer punishes the rewrite it exists to grade**: the fidelity judge rubric penalized register normalization that the rewrite prompt itself mandates; live-quality fixtures went from 9/22 to 20/22 after the rubric fix plus prompt parity between the CLI and web surfaces.
+- **Rewrite prompt forbids invented claims**: the strict rewrite prompt now explicitly prohibits adding claims the source does not state, closing a fabrication path the meaning floors could only catch after the fact.
+- **`gemini-cli` backend disables MCP servers** on its invocations, so a user's local MCP config can never inject tools into a patina rewrite call.
+- **`kimi-cli` ≥ 0.28 compatibility**: the backend now parses the stream-json assistant output of newer Kimi Code releases (with the legacy plain-text fallback retained).
+- **SSE streams request usage accounting** (`include_usage`), so streamed rewrites report token usage like buffered calls.
+- **Playground approval status line is screen-reader-only**: the visible "Unapproved — checks have not passed" line under every streaming message read as an alarming warning during normal in-flight rewrites; it is now visually clipped while the `role="status"` live region and localized copy stay intact for assistive tech.
+
+### Added
+
+- **`gemini-3.6-flash` allowlisted** on the web BYOK surface and eligible for `PATINA_FREE_MODEL` / `PATINA_PRO_MODEL`, after a 22-fixture live-quality comparison (better AI-score improvement and fewer meaning-loss fixtures than the prior Pro pin at ~1/5 the cost). All pinned defaults are unchanged; this only widens the allowlist.
+- **Native Anthropic adapter with prompt caching (opt-in)** plus thinking control for OpenAI-compatible providers via `extraBody` pass-through — both dormant unless explicitly enabled.
+- **Free-tier observability**: the pro monitor now also watches the free tier (the tier that actually has users), with the same closed outcome schema.
+- **Live-quality harness upgrades** (dev tooling): fixed-judge override, `--judge-backend` / `--backend` subscription-CLI seats, per-call usage/latency capture, and `--repeat` for variance-aware sweeps.
+
 ## 6.3.1 — 2026-07-07
 
 **Launch polish for the hosted playground and README — no CLI/engine changes.**

diff --git a/README.md b/README.md
@@ -20,7 +20,7 @@
   <a href="https://opensource.org/licenses/MIT"><img alt="License: MIT" src="https://img.shields.io/badge/License-MIT-yellow.svg"></a>
   <a href="#quick-start"><img alt="Skill: Claude Code | Codex | Cursor | OpenCode" src="https://img.shields.io/badge/Skill-Claude%20Code%20%7C%20Codex%20%7C%20Cursor%20%7C%20OpenCode-blueviolet"></a>
   <a href="https://github.com/devswha/patina"><img alt="Languages: KO | EN | ZH | JA" src="https://img.shields.io/badge/Languages-KO%20%7C%20EN%20%7C%20ZH%20%7C%20JA-green"></a>
-  <a href="CHANGELOG.md"><img alt="Version 6.3.1" src="https://img.shields.io/badge/version-6.3.1-blue"></a>
+  <a href="CHANGELOG.md"><img alt="Version 6.3.2" src="https://img.shields.io/badge/version-6.3.2-blue"></a>
 </p>
 
 <p align="center">
@@ -204,7 +204,7 @@ If meaning drifts, the change is retried or rolled back. Deterministic analysis
 
 ```yaml
 # .patina.default.yaml
-version: "6.3.1"
+version: "6.3.2"
 language: ko              # ko | en | zh | ja
 profile: default
 output: rewrite           # rewrite | diff | audit | score

diff --git a/README_JA.md b/README_JA.md
@@ -6,7 +6,7 @@
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
 [![Skill](https://img.shields.io/badge/Skill-Claude%20Code%20%7C%20Codex%20%7C%20Cursor%20%7C%20OpenCode-blueviolet)](#クイックスタート)
 [![Multi-language](https://img.shields.io/badge/Languages-KO%20%7C%20EN%20%7C%20ZH%20%7C%20JA-green)](https://github.com/devswha/patina)
-[![Version](https://img.shields.io/badge/version-6.3.1-blue)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-6.3.2-blue)](CHANGELOG.md)
 
 > **AIっぽさだけを落として、意味はそのまま。**
 
@@ -186,7 +186,7 @@ Input
 
 ```yaml
 # .patina.default.yaml
-version: "6.3.1"
+version: "6.3.2"
 language: ko              # ko | en | zh | ja
 profile: default
 output: rewrite           # rewrite | diff | audit | score

diff --git a/README_KR.md b/README_KR.md
@@ -6,7 +6,7 @@
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
 [![Skill](https://img.shields.io/badge/Skill-Claude%20Code%20%7C%20Codex%20%7C%20Cursor%20%7C%20OpenCode-blueviolet)](#빠른-시작)
 [![Multi-language](https://img.shields.io/badge/Languages-KO%20%7C%20EN%20%7C%20ZH%20%7C%20JA-green)](https://github.com/devswha/patina)
-[![Version](https://img.shields.io/badge/version-6.3.1-blue)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-6.3.2-blue)](CHANGELOG.md)
 
 <p align="center">
   <a href="https://patina.vibetip.help/"><b>브라우저에서 바로 써보기 — 설치 없음</b></a>
@@ -186,7 +186,7 @@ Input
 
 ```yaml
 # .patina.default.yaml
-version: "6.3.1"
+version: "6.3.2"
 language: ko              # ko | en | zh | ja
 profile: default
 output: rewrite           # rewrite | diff | audit | score

diff --git a/README_ZH.md b/README_ZH.md
@@ -6,7 +6,7 @@
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
 [![Skill](https://img.shields.io/badge/Skill-Claude%20Code%20%7C%20Codex%20%7C%20Cursor%20%7C%20OpenCode-blueviolet)](#快速开始)
 [![Multi-language](https://img.shields.io/badge/Languages-KO%20%7C%20EN%20%7C%20ZH%20%7C%20JA-green)](https://github.com/devswha/patina)
-[![Version](https://img.shields.io/badge/version-6.3.1-blue)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-6.3.2-blue)](CHANGELOG.md)
 
 <p align="center">
   <strong>去掉 AI 味，保留原意。</strong>
@@ -188,7 +188,7 @@ Input
 
 ```yaml
 # .patina.default.yaml
-version: "6.3.1"
+version: "6.3.2"
 language: ko              # ko | en | zh | ja
 profile: default
 output: rewrite           # rewrite | diff | audit | score

diff --git a/SKILL.md b/SKILL.md
@@ -1,6 +1,6 @@
 ---
 name: patina
-version: 6.3.1
+version: "6.3.2"
 description: Detect and rewrite AI writing patterns in Korean, English, Chinese, and Japanese text so it reads as if a human wrote it. Meaning-preservation (MPS) verified.
 allowed-tools:
   - Read