deepseek-harness tested by Argusic Runner
Attempt 1 of 3 · Runs with mocks · Argusic Score 88.0 of 100
all runs of deepseek-harnessai-agentscordisdshdsh-plugin
Session recording
Download the original recording · fingerprint 3842570120024808...
What the test found
The repository installs, builds, launches (dsh --help, dsh --version, and dsh web serving the authenticated Web UI with real JS/CSS bundles), completes a full agent turn with real tool execution against a mock LLM endpoint, and passes 17,257 of its 17,383 unit tests — rising to all but one (an IPv6-only test this container cannot support) when run under umask 022 with the one flaky-test fix applied.
No DEEPSEEK_API_KEY available, so the agent loop was verified against a schema-correct mock of the OpenAI-compatible /chat/completions SSE endpoint (matching WireChunk/WireDelta/WireUsage in packages/llm/llm-deepseek/src/types.ts), injected via DEEPSEEK_BASE_URL. The mock was sufficient: the real write tool executed and notes.txt was verified on disk with exact expected content, plus the tool result 'Created file' observed in the fed-back messages. Real credentials are NOT required to install, build, launch, or run the test suite. The 17,257-test unit suite and the Web UI launch are real (unmocked) verification; only the model boundary was mocked. The repo also ships its own keyless tiers I did not run for time (test:snapshot replay of 75 recorded sessions, test:web browser snapshots). Only one repo file was modified: packages/spill/spill-local/tests/spill-local.spec.ts. IPv6 could not be enabled (kernel disable_ipv6=1, /proc/sys read-only, unprivileged user, no sudo).
Argusic installed deepseek-harness in 2.5 minutes and launched it; verification reached Verified with mock services. 4 of 5 recorded errors were worked through (see the timeline below).
Runs with mocksall runs of deepseek-harness
This summary is drawn from the agent's recorded report for this run. Every figure traces to the log and recording above; nothing here is authored.
Error and fix timeline
No errors were recorded for this run.