clauderipple 0.8.9 → 0.9.0

package/docs/ARCHITECTURE.md

@@ -196,6 +196,15 @@ request went, and a `Pane` holds the request log.
   browser) are the app's own; the mod API (CLI 2.1.286 types) has no place for
   a plugin's button there. Its sites are the band, the status line, panes,
   transcript rows and slash commands.
+- **Pictures on the desktop are images, never frames.** The desktop app draws
+  each answer as a new DOM tree; an `Svg` becomes a new `<img>` of a data URL,
+  or with `isInteractive` a sandboxed `<iframe srcdoc>` that paints blank until
+  loaded. Only the frame blinks on a redraw: a plain `Svg` image does not, and
+  CSS `@keyframes` inside the SVG run in it (seen by the owner in the desktop
+  app on 2026-10-11 at commit 8d43b97, 리플이 moving in the band and the quick
+  settings with no flicker; Effortless animates its band the same way). A
+  redraw still starts the motion over, so the mod's loops all open on a rest
+  pose (`packages/mod/hooks/ui/ripple.ts`).
 - **A pane's open state is the engine's.** A module reload drops its pane
   without a `ui.close` the module hears, while `$.state` survives the reload: a
   flag kept there stuck the Log button on "Close log" (2026-10-05). The mod asks
@@ -205,9 +214,128 @@ request went, and a `Pane` holds the request log.
   path list other plugins may share, so only the entry whose manifest is named
   `clauderipple-status` is added or removed (`clauderipple mod on|off`, the
   Clients screen switch).
+- **Developing it, sessions load a copy at the last commit.** With
+  `CLAUDE_CODE_PLUGIN_DIR_WATCH=1` a session reloads a plugin folder whenever
+  it changes, and each reload leaves "ripple reloaded (N hooks: …)" in every
+  open transcript: pointed at the checkout, one session editing the mod
+  flooded all the others (2026-10-13). On this machine the entry is
+  `~/.clauderipple/mod`, which `scripts/sync-mod.mts` keeps at main's
+  `packages/mod` from the post-commit, post-merge and post-rewrite hooks
+  (`--install` writes them), writing only files that differ. Uncommitted work
+  never reaches everyday sessions; `clauderipple mod on` would point the entry
+  back at the checkout.
+
+### 3b. Shared mod settings and judge-only models (2026-10-10)
+
+The router is the single source of the mod's settings, persisted in
+`<homeDir()>/mod-settings.json`. `GET /api/mod-settings` returns
+`{settings:<full normalized Settings>, revision, stored:boolean}`. Before the first
+write it returns normalized defaults, revision 0 and `stored:false`, without creating
+storage. `PUT` accepts `{settings, revision?}`, normalizes with the mod's own pure
+`hooks/settings.ts:normalizeSettings` (type-only imports, no `claude-code` runtime),
+increments the revision under the account stores' `withFileLock` and `writeJsonAtomic`
+helpers (0600), and returns `{settings, revision, stored:true}`. A stale supplied
+revision returns 409 with the same full current snapshot, including `stored`.
+Existing stored settings are normalized on reads as well; unknown fields are dropped.
+Before normalization, the router refuses nonobjects, settings over 64 KiB of UTF-8
+JSON (413), and credential field names at any depth, including inside arrays (400).
+The final word of a field name decides (`apiKey`, `access_token`, `clientSecret`,
+`password`, etc.); `keepWarmMinTokens` is a count, not a credential, and is accepted.
+Corrupt storage is an error, never an empty store to overwrite.
+
+The mod reads this endpoint at `session.start` and every 3-second router poll, even
+when its band is hidden. A changed revision replaces the settings atom and redraws
+readers. Only when `stored:false` does it read legacy plugin-store `settings` and
+`setupDone` and migrate them with a revision-checked PUT; a racing first writer wins
+instead of being overwritten. Afterwards settings are never read from or saved to the
+plugin store; unrelated keys such as picks, band visibility and handoff choices remain.
+Panel/setup saves PUT the last seen revision. On 409 the mod takes the server copy,
+reapplies only the user's changed paths (including sibling rule/provider fields), and
+retries once. A panel keeps its opening baseline so a poll while it is open cannot
+turn an unrelated remote edit into a local overwrite. Offline changes remain active
+locally with the existing router-not-answering notice and retry on the next poll,
+rebased over that poll's server copy. Revision, pending edits and migration state live
+in session atoms and survive module hot reloads; settings HTTP work is serialized.
+Language resolution and the separate Jev key file are unchanged. Verified with
+`test/admin.test.ts`, `test/mod/settings.test.ts` and the mod's 2.1.293 engine tests.
+
+`GET /api/mod-options` returns `agents` (owned manifest entries with existing agent
+files), all configured `models` (`model`, `provider`, `name`, `judgeOnly`), `providers`
+(`name`, `type`, `usable`), `zen:{configured,model}`, and `scores`. ChatGPT usability
+shares `/api/status`'s quota refresh and account states, including purchased credits;
+AA scores reuse `model-scores.ts` and retain its attribution. No provider key is exposed.
+Since 2026-10-11 (the worker ladder, docs/specs/2026-10-11-worker-ladder.md) each agent also
+carries `efforts` (its model's ladder from `effortLevels`, mod efforts only) and `billing`
+(`free | payg | subscription`, `hooks/ladder.ts:billingOf`: ChatGPT, Claude login, Antigravity
+and the OpenCode Go preset are subscriptions, Ollama/LM Studio, a hand-entered local URL and
+`:free`/`-free` models are free, every other preset or API is pay-per-token); providers carry
+`billing`; `claudeEfforts` is Anthropic's ladder; `presets` are the four ladder presets computed
+by the mod's own `hooks/ladder.ts` from the usable workers plus Claude's models, so the dashboard
+and the session's quick settings fill rungs alike.
+
+The worker ladder replaced Pin as the default rule mode (`{mode:'ladder', rungs:{light, standard,
+hard, plan}}`, each rung `{target, effort, backup?}`). The judge adds `kind: plan|do`; level picks
+Light/Standard/Hard, a plan at standard or deep goes to Plan, and `sure < 0.5` moves Light→Standard
+→Hard (Plan does not move). An unavailable rung tries its backup, then one rung up (its model, then
+its backup), then the session model (`general-purpose` with no model; refused only when Anthropic is
+switched off), never down. Stored pins normalize to ladders on every read (missing levels climb,
+efforts start at the default ladder's for the same model or low/medium/high, calibrated as the pin
+was; Plan is Hard's model one effort up); Auto stays Auto, shown as experimental. `fallback` and
+`preference` are kept; only Auto still reads `preference`. Every pick writes one
+`ClaudeRipple worker pick:` debug line. Tests: `test/mod/ladder.test.ts`, `test/admin.test.ts`.
+
+`model-policy.ts:isJudgeOnlyModel` is the single judge-only rule: ids beginning with
+`jev` are not chat models (the OpenCode Zen note in `presets.ts`: `/systemone` only).
+Desktop/CLI picker injection (including aliases/routes targeting Jev), derived worker
+aliases, generated agents, AA auto-worker score candidates, OpenAI model lists and
+Codex catalogs exclude them; the next enabled agent sync removes previously
+owned Jev files, never hand-written ones. Provider configuration and mod options keep
+them available as judge candidates. Tests: `test/admin.test.ts`, `test/model-policy.test.ts`.
 
 ## 4. Provider adapters
 
+### Artificial Analysis model scores (2026-10-10)
+
+`scripts/aa-dataset.mjs` builds `aa-models.json` on the orphan GitHub `data` branch;
+`.github/workflows/aa-data.yml` checks daily at 03:37 UTC, replacing its single commit
+only for substantive changes or after a week. Source: [AA API v2](https://artificialanalysis.ai/api/v2/data/llms/models)
+(`x-api-key`, free limit 1,000 requests/day; `AA_API_KEY` stays in the environment).
+The API supplies slug/name/creator/release date, `evaluations.artificial_analysis_intelligence_index`
+and input/output USD-per-million prices, **not** cost per task. The model page's Next.js
+flight record supplies `intelligenceIndex`, `intelligenceIndexIsEstimated` and
+`intelligenceIndexCostPerTask.cost.total` (USD, already including input/cache/output);
+`intelligenceIndexCost.total` is a different, whole-benchmark total and is never used.
+Measured [GPT-6.1 Sol High](https://artificialanalysis.ai/models/gpt-6-1-sol-high):
+index **50.2378** (≈50.24), **$0.319144 per task** (≈$0.3191), not estimated.
+
+Track families with at least one measured release in the last **365 days**, including
+all their older/unmeasured variants. Group by creator and model name with reasoning
+parentheses removed; retain unrelated Preview/Beta/date labels. Strip only recognized
+reasoning endings from slugs. Effort uses the page suffix/name parentheses, then slug:
+`max/xhigh/high/medium/low/minimal/non-reasoning`; Non-reasoning takes precedence over
+an accompanying High/Low label. Generic Reasoning or the unsuffixed default maps to
+max in a multi-variant family; a single-variant model has null effort. Successful
+`checkedAt` values under **7 days** are reused; failed pages retain prior values/time
+(or null cost/time), and embedded defaults are harvested opportunistically by exact
+slug. Page requests are sequential, ≥1.5 seconds apart with one backed-off 429/5xx
+retry. Fewer than half of attempted pages parsing their requested slug aborts publication.
+The extraction marker `canonicalIntelligenceIndexTokenCount` is a private page-format
+anchor, not an output metric: no token counts are stored. `--limit N` caps page fetches
+for a local smoke run; use `--output <path>` and never publish capped output.
+
+Auto routing and field rules were removed 2026-10-12 (along with `GET /api/model-scores`);
+routing is the four workers (docs/specs/2026-10-12-worker-tiers.md). `model-scores.ts`
+remains as the admin's internal catalog for the mod's options, with normalized
+AA bases (decimal dots → hyphens, provider namespace/Claude snapshot date removed),
+variant intelligence/cost, attribution and generation time. The picker's five effort
+names are preserved; AA minimal/non-reasoning and single variants return null rather
+than being mislabeled low. Unknown models remain present with `matched:false`.
+Scores are metadata, never routing policy or the user's subscription bill. The public
+raw JSON is fetched with a 5-second timeout, kept 24 hours in memory and
+`<homeDir()>/aa-models.json` (mtime records retrieval), shared by concurrent lookups;
+on failure the last cache is kept and retried after one hour. Displays/republication
+must preserve **Data: Artificial Analysis (artificialanalysis.ai)** attribution.
+
 - **The Claude model list is Anthropic's own (2026-10-08).** The native `anthropic`
   provider's models came from a constant in admin.ts, so each release after it was cut
   was missing from the provider form, the mapping table and the declared list until a
@@ -335,8 +463,12 @@ request went, and a `Pane` holds the request log.
     visibility: "list"|"hide", context_window, supported_reasoning_levels:
     [{effort}] }] }` — what the Codex CLI's picker shows. **The server filters by
     `client_version`:** `0.146.0` omitted gpt-6-astra/sol/luna, `0.155.0`
-    listed them. The adapter (`catalog.ts`) asks with the higher of a floor and
-    `~/.codex/models_cache.json`'s `client_version`, drops `hide` entries, keeps
+    listed them; `0.158.0` omitted gpt-6.1-sol and `0.159.0` listed it
+    (2026-10-10, also listed at `0.162.1`, `1.0.0`, `99.0.0`). A floor alone
+    trails every release — a Windows user with no Codex CLI never saw 6.1 Sol
+    (#50) — so the adapter (`catalog.ts`) asks with the highest of the floor,
+    `~/.codex/models_cache.json`'s `client_version` and the newest `@openai/codex`
+    on the npm registry (kept a day), drops `hide` entries, keeps
     the list an hour, and falls back to a measured list in the same file. The
     GUI's ChatGPT form fetches it on open; new models arrive unticked. Model ids
     unknown to the backend answer `400 "… model is not supported when using

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "clauderipple",
-  "version": "0.8.9",
+  "version": "0.9.0",
   "description": "Use GPT (any ChatGPT plan with Codex), DeepSeek and 400+ other models in the Claude Desktop Code tab and Claude Code while staying signed in to your Claude subscription — no third-party gateway mode. Several ChatGPT accounts that take over when one runs out, in Codex too.",
   "license": "GPL-3.0-only",
   "workspaces": [
@@ -10,11 +10,15 @@
     "node": ">=24"
   },
   "scripts": {
-    "typecheck": "tsc -p tsconfig.json --noEmit",
+    "typecheck": "tsc -p tsconfig.json --noEmit && tsc -p packages/remote/tsconfig.web.json --noEmit && tsc -p packages/remote/relay/tsconfig.json --noEmit",
     "dist": "npm --workspace @clauderipple/app run dist",
     "release": "npm --workspace @clauderipple/app run release",
     "router": "node packages/router/src/index.ts",
-    "test": "node --test packages/router/test/*.test.ts packages/cli/test/*.test.ts",
+    "remote": "node --disable-warning=ExperimentalWarning packages/remote/src/main.ts",
+    "relay:build": "node --disable-warning=ExperimentalWarning packages/remote/relay/build-web.ts",
+    "relay:dev": "npm run relay:build && wrangler dev -c packages/remote/relay/wrangler.toml --ip 127.0.0.1 --port 8798",
+    "relay:deploy": "npm run relay:build && wrangler deploy -c packages/remote/relay/wrangler.toml",
+    "test": "node --test packages/router/test/*.test.ts packages/cli/test/*.test.ts test/mod/*.test.ts packages/remote/test/*.test.ts",
     "cli": "node packages/cli/src/index.ts",
     "prepack": "node scripts/build-npm.mjs",
     "build": "node scripts/build-npm.mjs"