mirror of
https://github.com/deepseek-ai/deepseek-harness
synced 2026-08-15 21:04:50 +00:00
Address the two remaining review warnings on PR #106. - tools/post-execute: when a downstream listener/policy returns `block`, return early without loading or attaching workspace instructions. The registry turns a block into a final isError result, so reconciling off the original successful result leaked instructions from a rejected call and advanced nested/baseline tracking off a touch that never happened. - Disable workspaceContext in the Code Mode examples: fs tools run as run_code sub-dispatches and code-mode.ts drops sub-call additionalContext, so dynamic AGENTS.md updates are silently discarded there. Update the block regression test to assert no context is attached, and add a waterfall case proving accept still surfaces the discovered instructions.
prompt/ — prompt and request-context extensions
Product packages that contribute model-facing prompt or request-context behavior without being core agent/session/tool primitives. These packages usually consume existing seams such as agent/session-prefix, agent/request, tools/post-execute, or system-prompt/assemble; they do not own the loop and do not provide LLM adapters, execution backends, or UI front doors.
| Package | Role | ctx key |
|---|---|---|
workspace-context/ |
AGENTS.md/CLAUDE.md workspace context loader |
(listens on agent/session-prefix + tools/post-execute) |
workspace-context lives here because it adds workspace guidance to the model request without owning a core service. Its decision record explains the per-agent/session isolation and lifecycle split.