Feat/native web admin

.dockerignore

@@ -0,0 +1,16 @@
+.git
+.codex
+.config
+node_modules
+out
+out-server
+out-admin
+dist
+coverage
+backup
+logs
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+*.log
+docs/screenshots

.gitignore

@@ -3,6 +3,8 @@ node_modules/
 
 # Build outputs
 out/
+out-server/
+out-admin/
 dist/
 build/entitlements.mac.plist
 
@@ -39,16 +41,25 @@ Thumbs.db
 *.swp
 *.swo
 .codex/
+.playwright-mcp/
 CLAUDE.md
 
 # Environment
 .env
 .env.local
 .env.*.local
 
+# Python
+__pycache__/
+*.pyc
+*.pyo
+
 # Test
 coverage/
 
+# Manual live test scripts (local debugging only)
+tests/manual/*-live.mjs
+
 # Misc
 *.pem
 *.p12
@@ -58,3 +69,10 @@ src/renderer/next-app/
 
 # Backup
 backup/
+
+# Local migration artifacts and development data
+.local/
+.local-artifacts/
+
+# Local live debug logs
+.tmp-live-*.ndjson

Dockerfile

@@ -0,0 +1,113 @@
+ARG NODE_IMAGE=node:22.21.1
+
+FROM ${NODE_IMAGE} AS deps
+WORKDIR /app
+COPY package.json package-lock.json ./
+RUN npm ci --ignore-scripts
+
+FROM ${NODE_IMAGE} AS build
+WORKDIR /app
+COPY --from=deps /app/node_modules ./node_modules
+COPY . .
+RUN npm run build:server
+
+FROM ${NODE_IMAGE} AS runtime
+WORKDIR /app
+ENV NODE_ENV=production
+ENV CHAT2API_HOST=0.0.0.0
+ENV CHAT2API_PORT=8080
+ENV CHAT2API_DATA_DIR=/data
+ENV CHAT2API_COMPACTION_DETECTION=auto
+ENV CHAT2API_QWEN_AI_COMPACTION_THINKING=auto
+# Compaction input uses live model limits first; these values are deployment
+# controls for an explicit override, optional metadata cap, or a
+# catalogue-without-limits fallback. Zero leaves live metadata uncapped.
+ENV CHAT2API_QWEN_AI_COMPACTION_INPUT_TOKEN_BUDGET=0
+ENV CHAT2API_QWEN_AI_COMPACTION_METADATA_MAX_INPUT_TOKENS=0
+ENV CHAT2API_QWEN_AI_COMPACTION_FALLBACK_INPUT_TOKENS=12000
+ENV CHAT2API_QWEN_AI_COMPACTION_PROMPT_TOKEN_RESERVE=512
+ENV CHAT2API_QWEN_AI_COMPACTION_CHUNK_DELAY_MS=0
+ENV CHAT2API_QWEN_AI_COMPACTION_MAX_REDUCTION_ROUNDS=6
+# Zero means use the complete active account pool discovered at runtime.
+ENV CHAT2API_QWEN_AI_COMPACTION_MAX_ACCOUNT_ATTEMPTS=0
+# Limit simultaneous recovery candidates only; account rotation still uses
+# the complete active pool unless the deployment sets an attempt cap.
+ENV CHAT2API_QWEN_AI_COMPACTION_FAILOVER_WAVE_SIZE=2
+ENV CHAT2API_QWEN_AI_MAX_ACCOUNT_FAILOVERS=0
+# Keep the adaptive pacing floor aligned with the validated multi-account
+# deployment; upstream 429/risk responses still control account cooldowns.
+ENV CHAT2API_QWEN_AI_AUTO_TUNE_MIN_GLOBAL_INTERVAL_MS=1000
+# Repair active Qwen AI accounts that have a JWT but no Web session cookie.
+# Sign-ins are serialized and globally paused when Qwen returns risk control.
+ENV CHAT2API_QWEN_AI_SESSION_REPAIR_ENABLED=true
+ENV CHAT2API_QWEN_AI_SESSION_REPAIR_INTERVAL_MS=25000
+ENV CHAT2API_QWEN_AI_SESSION_REPAIR_RESCAN_MS=60000
+ENV CHAT2API_QWEN_AI_SESSION_REPAIR_RISK_COOLDOWN_MS=180000
+ENV CHAT2API_QWEN_AI_SESSION_REPAIR_FAILURE_RETRY_MS=300000
+ENV CHAT2API_QWEN_AI_SESSION_REPAIR_CREDENTIAL_RETRY_MS=21600000
+# Docker deployments allow long active generations within the cumulative
+# request deadline while separately bounding streams that stop producing data.
+# Queue admission has its own timer but still shares the route deadline.
+ENV CHAT2API_QWEN_AI_QUEUE_TIMEOUT_MS=120000
+# Keep one effective governor slot available for ordinary client requests
+# while a context-compaction map/reduce is active.
+ENV CHAT2API_QWEN_AI_COMPACTION_RESERVED_SLOTS=1
+# Managed reasoning, answers, and tool arguments remain private until terminal
+# validation so a failed provider branch cannot leak before account recovery.
+ENV CHAT2API_QWEN_AI_BUFFER_MANAGED_STREAMS=true
+# Start document offload before a large Qwen Web request reaches its model context.
+# This is a transport target, not a local client request limit; zero disables it.
+ENV CHAT2API_QWEN_AI_REQUEST_MAX_BYTES=92160
+# Bound inline Hermes routing summaries while complete tool documentation stays
+# in the account-scoped reference attachment. Zero omits inline descriptions.
+ENV CHAT2API_QWEN_AI_HERMES_ROUTING_SUMMARY_MAX_CODE_POINTS=240
+# Managed-branch and upstream-busy recovery counts are deployment controls.
+# Their request deadlines remain authoritative; zero disables each path.
+ENV CHAT2API_QWEN_AI_RETRY_COUNT=1
+ENV CHAT2API_QWEN_AI_BUSY_RETRY_COUNT=1
+# A transport reset can continue the same Qwen response without resubmitting
+# the prompt. Deployments can tune or disable this bounded recovery budget.
+ENV CHAT2API_QWEN_AI_STREAM_RESUME_ATTEMPTS=3
+ENV CHAT2API_QWEN_AI_STREAM_RESUME_DELAY_MS=1000
+# Response-id resumes and managed workflow continuations share this
+# no-progress budget; it pauses while a replacement stream is active.
+ENV CHAT2API_QWEN_AI_RECOVERY_BUDGET_MS=180000
+# Default to one same-chat semantic correction; the deployment controls the count.
+ENV CHAT2API_QWEN_AI_WORKFLOW_CONTINUATION_ATTEMPTS=1
+# Semantic continuation branches also share an absolute wall-clock deadline.
+ENV CHAT2API_QWEN_AI_WORKFLOW_RECOVERY_TIMEOUT_MS=840000
+# Busy-chat admission is bounded separately from the long generation timeout.
+# Leave the generic retry-count override unset so ordinary semantic workflow
+# continuations retain deadline mode; Responses tool-result continuations use
+# the dedicated override below.
+ENV CHAT2API_QWEN_AI_CHAT_IN_PROGRESS_RETRY_BUDGET_MS=300000
+ENV CHAT2API_QWEN_AI_CHAT_IN_PROGRESS_RETRY_DELAY_MS=1000
+# Retained Responses tool-result continuations fail fast into same-account
+# full replay; semantic workflow continuations use the generic policy above.
+ENV CHAT2API_QWEN_AI_RESPONSES_CONTINUATION_RETRY_ATTEMPTS=0
+ENV CHAT2API_VALIDATED_SSE_MAX_HOLD_MS=60000
+ENV CHAT2API_SSE_KEEPALIVE_INTERVAL_MS=15000
+# Keep the HTTP listener alive long enough for the longest configured request
+# to finish when Docker sends SIGTERM during an update.
+ENV CHAT2API_SHUTDOWN_DRAIN_TIMEOUT_MS=540000
+# Keep the default cumulative deadline below typical downstream transport
+# limits so a structured terminal response has time to propagate.
+ENV QWEN_AI_REQUEST_TIMEOUT_MS=840000
+# Zero disables only the additional post-admission response cap. The
+# cumulative QWEN_AI_REQUEST_TIMEOUT_MS deadline still bounds the full request.
+ENV QWEN_AI_RESPONSE_TIMEOUT_MS=0
+ENV QWEN_AI_STREAM_IDLE_TIMEOUT_MS=180000
+# Bound each account's document parse stage independently so a stalled parse
+# can move to another account while the cumulative request deadline remains.
+ENV QWEN_AI_FILE_PARSE_POLL_INTERVAL_MS=2000
+ENV QWEN_AI_FILE_PARSE_TIMEOUT_MS=120000
+ENV QWEN_AI_OSS_STS_REFRESH_INTERVAL_MS=240000
+COPY package.json package-lock.json ./
+RUN npm ci --omit=dev --ignore-scripts && npm cache clean --force
+COPY --from=build /app/out-server ./out-server
+COPY --from=build /app/out-admin ./out-admin
+COPY --from=build /app/sha3_wasm_bg.7b9ca65ddd.wasm ./sha3_wasm_bg.7b9ca65ddd.wasm
+RUN mkdir -p /data
+VOLUME ["/data"]
+EXPOSE 8080
+CMD ["node", "out-server/server/index.js"]