MiniMax-AI-Dev · SaladDay · Sep 20, 2026 · Sep 20, 2026 · Sep 20, 2026 · Sep 20, 2026
diff --git a/.env.example b/.env.example
@@ -5,7 +5,7 @@
 #
 # Then choose your target:
 #   make dev                                            # local hot-reload stack
-#   docker compose -f docker-compose.local.yml up -d    # single-host / LAN deploy
+#   docker compose up -d    # single-host / LAN deploy
 #
 # The production self-hosted compose (deploy/compose/compose.selfhost.yml)
 # has its own template at deploy/compose/.env.example for image-tag /
@@ -58,66 +58,19 @@ PARSAR_PG_PORT=15432
 PARSAR_HOST_IP=
 # Set to parsar:local after running: make docker-build PARSAR_IMAGE=parsar PARSAR_IMAGE_TAG=local
 PARSAR_SERVER_IMAGE=parsar:local
-# Shared token between parsar-server and the compose-resident local runtime.
-# install.sh generates this automatically for normal installs.
-PARSAR_SHARED_RUNTIME_TOKEN=
-# Internal WebSocket URL advertised to compose-resident daemon runtimes.
-PARSAR_AGENT_DAEMON_WS_URL=ws://parsar-server:8080/agent-daemon/ws
-
 # -----------------------------------------------------------------------------
 # Feishu Bot (optional — see docs/deploy/lan-deploy.md)
 # -----------------------------------------------------------------------------
 # Bot's own open_id (ou_xxx). Required for group chat @Bot to work.
 PARSAR_FEISHU_DEFAULT_BOT_OPEN_ID=
 
 # -----------------------------------------------------------------------------
-# Cloud sandbox — e2b.app (optional; powers the "Cloud isolation" agent mode)
-# -----------------------------------------------------------------------------
-# Both values are required to enable cloud isolation. With either empty the
-# server logs a warning at boot and cloud-isolation agents fail fast, while
-# local-device agents keep working.
-#
-# Build the template first (writes only gitignored build artifacts):
-#     make e2b-template
-# then paste the template id it reports here. Rebuilding the same template
-# name keeps the id, so this only changes on first creation.
-AGENT_DAEMON_SANDBOX_TEMPLATE=
-PARSAR_E2B_API_KEY=
-
-# Optional larger tier, selected per agent via agents.config.sandbox_size="xl".
-# An agent asking for a size with no template configured degrades to the
-# standard one with a warning rather than failing the acquire.
-AGENT_DAEMON_SANDBOX_TEMPLATE_XL=
-
-# Optional deployment-wide sandbox lifetime. Uses Go duration syntax (for
-# example 30m, 1h, or 24h) and defaults to 1h. An agent may override this with
-# agents.config.sandbox_ttl. Parsar does not impose a provider-specific maximum;
-# configure a value supported by your sandbox provider and account.
-AGENT_DAEMON_SANDBOX_TTL=
-
-# Deployment-wide default for periodic best-effort renewal. An agent may
-# override this with agents.config.sandbox_auto_renew. Renewal requires a TTL
-# longer than the five-minute scan interval. Enabled by default so a
-# continuously-used agent's sandbox is not reaped mid-conversation; set to
-# false to opt out (e.g. for a provider that does not support renewal).
-AGENT_DAEMON_SANDBOX_AUTO_RENEW=true
-
-# Deprecated compatibility setting. Used only when the duration above is
-# empty; prefer AGENT_DAEMON_SANDBOX_TTL for new deployments.
-AGENT_DAEMON_SANDBOX_TTL_HOURS=
-
-# Self-hosted / proxied e2b only. Leave empty for e2b.app.
-PARSAR_E2B_API_BASE_URL=
-PARSAR_E2B_SANDBOX_HOST=
-# PEM of a private CA, when the sandbox gateway serves a non-public cert.
-PARSAR_E2B_CA_CERT=
-
-# IMPORTANT for local testing: the daemon runs INSIDE the cloud sandbox and
-# dials back to PARSAR_PUBLIC_URL, so a loopback URL can never work — the
-# sandbox resolves 127.0.0.1 to itself and pairing times out. Expose the dev
-# server through a tunnel and set PARSAR_PUBLIC_URL to that hostname, e.g.
-#     ngrok http 18080
-#     PARSAR_PUBLIC_URL=https://<subdomain>.ngrok-free.dev
+# Core execution — configured independently from the product.
+# See docs/deploy/product-core.md. Docker/E2B and model credentials belong to Core.
+# Native development: absolute JSON file containing workspace/project/key-file bindings.
+PARSAR_CORE_WORKSPACES_FILE=
+# Compose deployment: directory mounted at /run/secrets/core.
+PARSAR_CORE_CONFIG_DIR=
 
 # -----------------------------------------------------------------------------
 # Network proxy (optional — only if your host needs a proxy for internet)

diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
diff --git a/Dockerfile b/Dockerfile
@@ -11,7 +11,7 @@
 #
 # The image carries NO real secrets and NO internal addresses.
 # Operators inject DATABASE_URL / PARSAR_MASTER_KEY / Feishu OIDC /
-# E2B credentials / etc. at runtime via environment variables or a
+# Core workspace credential paths / etc. at runtime via environment variables or a
 # YAML config file mounted at $PARSAR_CONFIG_FILE.
 #
 # Build:        make docker-build
@@ -80,6 +80,7 @@ RUN --mount=type=cache,target=/go/pkg/mod \
 
 COPY internal ./internal
 COPY catalog ./catalog
+COPY packages/agents-client ./packages/agents-client
 COPY server ./server
 
 # Build all three binaries in one RUN so the layer represents one
@@ -129,8 +130,6 @@ RUN --mount=type=cache,target=/go/pkg/mod \
 #   - ca-certificates + tini ship pre-built.
 #   - Server-side Skills.sh installs require npx and git.
 #   - Operators can `docker exec -it ... bash` to debug.
-#   - The opencode local runner may shell out (rg, basic core utils);
-#     keeping a real userland avoids surprises.
 ###############################################################################
 FROM --platform=$TARGETPLATFORM ${RUNTIME_BASE} AS runtime
 
@@ -209,10 +208,6 @@ EXPOSE 8080
 HEALTHCHECK --interval=10s --timeout=3s --start-period=15s --retries=3 \
   CMD wget -qO- --tries=1 --timeout=3 http://127.0.0.1:8080/healthz >/dev/null || exit 1
 
-# tini reaps zombie subprocesses spawned by the opencode local runner
-# (or any future fork-exec path). Without it the server PID-1 would
-# leak zombies on every prompt that shells out and an SRE poking the
-# image weeks later would find a process table full of `<defunct>`
-# entries with no obvious culprit.
+# tini reaps subprocesses used by product asset import tools.
 ENTRYPOINT ["/usr/bin/tini", "--"]
 CMD ["/usr/local/bin/parsar-server"]
diff --git a/Makefile b/Makefile
@@ -18,7 +18,7 @@ endif
 PARSAR_IMAGE     ?= parsar
 PARSAR_IMAGE_TAG ?= dev
 
-.PHONY: help setup node-deps dev dev-db check check-setup check-sqlc check-go check-store check-web check-cli check-hygiene test test-fast test-go test-web typecheck-web lint-web-design lint-web test-cli typecheck reset-dev clean-dev paths migrate-dev sqlc-generate server web cli devgateway http-runner-once http-runner-loop dev-all smoke e2e-http-agent e2e-feishu-gateway dev-server-up dev-server-down dev-server-log bootstrap docker-build docker-build-no-cache openapi e2b-template e2b-template-binaries
+.PHONY: help setup node-deps dev dev-db check check-setup check-sqlc check-go check-store check-web check-cli check-hygiene test test-fast test-go test-web typecheck-web lint-web-design lint-web test-cli typecheck reset-dev clean-dev paths migrate-dev sqlc-generate server web cli devgateway dev-all smoke e2e-feishu-gateway dev-server-up dev-server-down dev-server-log bootstrap docker-build docker-build-no-cache openapi e2b-template e2b-template-binaries
 
 help:
 	@printf '%s\n' \
@@ -211,12 +211,6 @@ cli:
 devgateway:
 	cd server && go run ./cmd/devgateway --help
 
-http-runner-once:
-	cd server && go run ./cmd/httprunner --once
-
-http-runner-loop:
-	cd server && go run ./cmd/httprunner --interval $${PARSAR_HTTP_RUNNER_INTERVAL:-2s} --max-runs $${PARSAR_HTTP_RUNNER_MAX_RUNS:-100}
-
 dev-all:
 	./scripts/dev-all.sh
 
@@ -226,9 +220,6 @@ dev-all:
 smoke:
 	./scripts/smoke.sh
 
-e2e-http-agent:
-	./scripts/e2e-http-agent.sh
-
 e2e-feishu-gateway:
 	./scripts/e2e-feishu-gateway.sh
 

diff --git a/README.md b/README.md
@@ -18,17 +18,14 @@
 
 Parsar is a team-first platform for dispatching, managing, and auditing AI coding agents. Send tasks from the tools your team already uses — chat, web UI, API — and get results back where they started: a thread, a PR, a webhook.
 
-Supported agent runtimes:
-
-- **Claude Code**
-- **Piagent**
-- **Codex**
-- More to come — the runtime layer is pluggable.
+Agent execution goes through the independently deployed Agents API Core, whose
+qualified profiles include **Claude Code**, **Codex**, and **MiniMax Code**. Parsar
+keeps its own database for members, business assets, permissions and collaboration.
 
 ### Why Parsar
 
 - **Team-first.** Shared queues, run history, and permissions — not single-player agent loops.
-- **Pluggable runtimes.** Claude Code today, Codex tomorrow, your in-house agent next week.
+- **Independent execution.** Core owns models, runtime environments and execution; Parsar uses its public Agents API.
 - **Pluggable surfaces.** Feishu / Lark ships today; Slack, Discord, and webhooks on the roadmap.
 - **Auditable.** Every run is persisted: prompt, diff, logs, exit code.
 - **Self-hosted.** Your code, your secrets, your machine. No telemetry.
@@ -54,8 +51,10 @@ make docker-build
 ./install.sh --image parsar:dev
 ```
 
-The default Compose stack starts PostgreSQL, the Parsar web control plane,
-and a shared agent runtime together.
+The default Compose stack starts PostgreSQL and the Parsar web control plane.
+Deploy Core separately and configure a distinct Core project credential for each
+workspace using the [product integration guide](docs/deploy/product-core.md).
+Asset management works independently; Agent execution requires this connection.
 
 ## Contributing
 

diff --git a/apps/web/src/components/admin/DevicePicker.tsx b/apps/web/src/components/admin/DevicePicker.tsx