Plugins
DeepSeek Harness is designed around composable plugins. Because the ecosystem is still in developer preview, treat this as an evaluation worksheet rather than a stable marketplace: verify compatibility, permissions, logs, and version freshness before using a plugin in important repositories.
Best for
Route DeepSeek, OpenAI-compatible, local, or evaluation models through the same harness surface.
Verify
Check model capability mapping, token/cost logging, timeout behavior, and whether secrets stay outside prompts and traces.
Watch out
A stale provider plugin can silently change model IDs, break streaming, or leak sensitive configuration through debug logs.
Best for
Give agents retrieval access for research tasks, changelog checks, package lookups, and citation gathering.
Verify
Review query logging, rate limits, freshness controls, source attribution, and whether results can be replayed in traces.
Watch out
Search plugins can introduce prompt injection from pages, uncontrolled network egress, and unverifiable claims if citations are not preserved.
Best for
Let agents inspect browser states, diagrams, UI regressions, screenshots, and visual QA artifacts.
Verify
Confirm image retention policy, max file size, OCR reliability, browser permission scope, and redaction for sensitive screens.
Watch out
Screenshots may capture tokens, private chats, billing pages, or customer data unless the workflow has explicit capture boundaries.
Best for
Run builds, tests, code transforms, data processing, and isolated experiments without touching production files directly.
Verify
Check working-directory isolation, write allowlists, cleanup behavior, process timeouts, and audit logs for file mutations.
Watch out
Over-broad filesystem permissions can let an experimental agent overwrite source, credentials, or deployment artifacts.
Best for
Expose agent sessions through local terminals, dashboards, or operator-friendly consoles for review and intervention.
Verify
Test keyboard accessibility, command echoing, exit-state reporting, transcript export, and safe handling of interactive prompts.
Watch out
A pretty terminal wrapper is not a safety boundary; destructive shell commands still need permissioning and visible receipts.
Best for
Split research, implementation, review, and monitoring into separate agents with clearer ownership and evidence trails.
Verify
Look for run IDs, task boundaries, retry semantics, shared artifact contracts, and deterministic handoff records.
Watch out
Poor orchestration can duplicate writes, hide failed subtasks, or make it impossible to prove which agent produced an artifact.
