Auto-approve sandbox Shell only for host-bound task execution, stage native results within their owning session, and preserve active lease state on response stream loss. The repeated failures came from interactive approval in unattended runs and transport cleanup invalidating execution state; focused regressions cover both boundaries.
Use verified task names until Agent generates a specific conversation title.
Repair legacy names without changing manual titles or history ordering.
Bound page repairs to 30 entries and a shared two-second deadline, and
re-read titles after task access validation to avoid stale reuse receipts.
Validation: bun run check; 40 focused tests across six files.
Move project and host queries into native MCP with host-owned OAuth/context,
while retaining memory, skills, actions and publication plugins. Add nearby-model
queries and explicit chat/task/learning model defaults.
Adopt Responses HTTP/SSE and native steer/queue inbox delivery, preserving scoped
session ownership, pending input edits and investigation communication.
Disable local-provider discovery: Server /health allowed native vLLM to probe
/v1/models on port 8000. A native before/after regression prevents recurrence.
Validation: 170 focused Bun tests; service/plugin typechecks; native discovery,
provider-discovery and input-inbox smoke checks; no paid model inference.
Generate chat, task and learning models from agent-models.json.
Use WaterNetwork fast/expert models with 272k context and validate
upstream names before readiness. Add an isolated model check command
with signal-safe cleanup and plugin, Skill and MCP management docs.
Verified in isolation: type checks, 26 focused tests and both model
catalog checks; no model inference requests.
Reuse OpenCode 2 HTTP/stdio clients with per-session read-only grants,
native approval and sanitized owner-scoped MCP status. Narrow the
environment inherited by trusted local MCP processes.
Validated in an isolated checkout with bun run check and 20 focused tests.
Upgrade the runtime and native plugin API to OpenCode 2.0.24. Restore
session-scoped learning, credential renewal, typed project queries and
controlled frontend actions from the legacy Agent.
Replace host Shell execution with an approved Landlock/seccomp executor
and persistent per-session workspaces. Add typed Bocha search and Tianditu
geocoding bridges using the owning session credential. Require a working
kernel sandbox before startup and persist workspace files in deployments.
Validation completed before this commit: TypeScript checks, 150 Bun tests,
Docker build and UID 10001 sandbox execution, native Shell approval and
session recovery. Tianditu succeeded; Bocha currently reports exhausted
account quota. Unrelated investigation-report contract work is excluded.
jiang