feat(alibaba): add token-plan routing and cloud discovery

.github/workflows/docs-pages.yml

@@ -14,14 +14,15 @@ permissions:
 env:
   FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
 
+concurrency:
+  group: pages
+  cancel-in-progress: false
+
 jobs:
-  deploy:
+  build:
     # Only run on the original repository, not on forks
     if: github.repository == 'Pythoughts-labs/pythinker-code'
     runs-on: ubuntu-latest
-    environment:
-      name: github-pages
-      url: ${{ steps.deploy.outputs.page_url }}
     steps:
       - name: Checkout repository
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # pinned from v6.0.2
@@ -76,6 +77,16 @@ jobs:
         with:
           path: docs/.vitepress/dist
 
+  deploy:
+    needs: build
+    runs-on: ubuntu-latest
+    permissions:
+      pages: write
+      id-token: write
+    environment:
+      name: github-pages
+      url: ${{ steps.deploy.outputs.page_url }}
+    steps:
       - name: Deploy to GitHub Pages
         id: deploy
         uses: actions/deploy-pages@cd2ce8fcbc39b97be8ca5fce6e763baed58fa128 # pinned from v5.0.0

.gitignore

@@ -78,3 +78,6 @@ blackbox/
 coverage.xml
 htmlcov/
 *.scratchpad.lock
+
+.playwright-mcp/
+.playwright/

CHANGELOG.md

@@ -19,6 +19,7 @@ GitHub Releases page; `0.8.0` is the new starting line.
 - **Deep-audit remediation: security, correctness, and multi-instance robustness.** Permission gate: awk programs that shell out via `print | "cmd"` / `getline` are now classified as mutating AND destructive (previously only `system(`/`>` and only mutating), and `xargs -L N` no longer hides its payload from classification. Glob resolves symlinks before its workspace-boundary check (an in-workspace symlink could previously list outside content); progress-note titles are ANSI-sanitized like every other transcript field. Grep content lines are parsed with unambiguous field separators, so paths like `utf-8-codec.py` are no longer mangled with `-n=false` and sensitive-file attribution is exact. Multi-line edits on CRLF files work again (LF-joined old strings are CRLF-translated when needed). `/import` preserves paths byte-for-byte (only a standalone leading/trailing `--force` is treated as the flag). Post-compaction file reminders include `--add-dir` files. Double-interrupt can no longer orphan the interruption-marker write (unanswered tool_calls). Background web replay falls back to full history (not empty) when the watermark stat fails, and a malformed Agent resume id returns a clean "Agent not found". OAuth: login fails loud when the token response lacks a `refresh_token`; a refresh response without `expires_in` carries the previous lifetime forward instead of refreshing every tick; the device-id file can no longer be read empty mid-creation. A failed `theme="auto"` background probe can be retried by re-selecting auto via `/theme`. Multi-instance: sessions now take a per-session writer lock (a second `pythinker -r <id>`/web worker on the same session is refused instead of interleaving turns), the shared `pythinker.json` index uses a locked read-modify-write (no more lost work-dir registrations), JSONL appenders repair torn final lines after a crash, forks materialize atomically, project-memory mutations abort on read failure instead of wiping the file, the journal is capped at 100 recaps, inbox approve/reject claims candidates atomically, and recall re-arms when another instance writes new memory. Subagents: a failed summary continuation no longer discards a completed agent's work, hallucinated subagent types fail fast with the valid-type list (before any RunAgents child launches), background failures carry an `Agent ID:` + resume hint, and a crash inside the runner's own error handling is logged instead of silently lost.
 - **Breaking (CLI flags): `pythinker web` / `pythinker vis` host short flag is now `-H`.** `-h` is a help alias on both subcommands (matching the root CLI); previously `-h <ip>` bound the host. Scripts using `-h 0.0.0.0` now print help and exit 0 without starting a server — switch to `-H <ip>` or `--host <ip>`. Part of the security/correctness audit (which also confined Grep to the workspace, gated non-HTTPS provider URLs in the web config API to loopback, and stopped saving OpenAI keys on 401/403).
 - **Thinking effort moved to a single top-right label on the input border.** The input box border is now one static frame grey at every effort level instead of recoloring the whole bar cold→hot, and the effort is no longer duplicated in the footer line. It's shown once, as a small label flushed to the right of the input's top border — a level-colored dot (slate→blue→teal→amber→orange→red as `off→max`) plus the muted level word — so the dial stays glanceable without tinting the typing area or cluttering the footer. The label is hidden entirely for native-thinking models (`always_thinking`, no user dial) and non-thinking models, and the rule auto-shortens by the label width so the line never wraps.
+- **Alibaba login: correct plan-key routing, full live model discovery, and `qwen3.7-plus` default.** Subscription "plan" keys — `sk-sp-`, `sk-tok-`, and `sk-ws-` — now route to the shared international Token Plan endpoint (`token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1`), fixing a bug where `sk-sp-` keys were sent to the Coding Plan host (`coding-intl`) and every chat then failed with `401 invalid_api_key`. Generic `sk-` keys remain pay-as-you-go Cloud (DashScope) keys. Model discovery no longer filters the live `/models` response down to a hardcoded allowlist — every chat model the endpoint returns now appears (known models keep curated capabilities/context; the rest get sensible defaults), while text-to-image models (Wan, Qwen-Image) are excluded since they can't serve as the agent's LLM. The catalog therefore reflects the key's tier: a Token Plan key surfaces the multi-vendor set (Qwen, DeepSeek, GLM, Kimi, MiniMax), a pay-as-you-go key surfaces that account's Qwen line. The first-login default model is now `qwen3.7-plus` (falling back to the first available model when an endpoint doesn't offer it). The usage panel detects endpoints with no quota API and shows a clear note pointing to the Model Studio console alongside Pythinker's local token tally instead of probing a 404 quota endpoint. The built-in catalog remains the offline fallback when discovery is unavailable. For the Token Plan, whose usage policy forbids automated balance polling and exposes no quota API, `/usage` shows Pythinker's local token tally plus a note pointing to the Model Studio console (My Subscriptions / Usage Analysis) for the Credits balance.
 - **Qwen models treated as native-thinking across both plans.** Qwen3.x/3.7 (e.g. `qwen3.7-max`, `qwen3.6-plus`, the Qwen3 Coder models) now carry the `always_thinking` capability on both the Alibaba Model Studio and OpenCode Go plans, matching GLM/MiniMax: reasoning is built in and always on, with no user effort dial and no top-border effort label. Reasoning still flows over the Anthropic `thinking` block that both Anthropic-compatible routes accept.
 - **TUI enhancements: adaptive theme, layout, and agent prompt overhaul.** Adaptive terminal-background probe + color-depth blending; reference-CLI layout and palette refinements; unified todo-list renderer; white running-task titles with consistent diff palette; elapsed/tokens/t-s metadata on the background status line; transcript-row bullet fix; renderer guards and markdown fence table unwrapping. All default agent prompts restructured with explicit Mission / Hard Constraints / Workflow / Output Contract sections. Background manager and subagent runner hardened with stale-record reconciliation and resume contract enforcement. Automatic turn recaps disabled by default.