English | 中文

Fault injection and chaos engineering plugin for DeepSeek Harness.

dsh-chaos is an independent third-party plugin for DeepSeek Harness. It intercepts the tools/execute / tools/post-execute and llm/stream waterfalls and injects latency, timeouts, HTTP-style errors, malformed tool results, dropped tool results, and LLM provider failures — so you can verify that retry, fallback, circuit-breaking, and agent-recovery paths actually work before a production incident tests them for you.

Injected failures reuse the same structured codes as real ones (TOOL_TIMEOUT, HTTP_429, …) so downstream policy routes on them identically; the error.info.name field carries a Chaos* class so telemetry can tell injected faults from genuine incidents.

From a dsh installation, install from the npm registry into any profile:

dsh plugin --profile web add @why-daydream/dsh-chaos          # latest
dsh plugin --profile web add @why-daydream/dsh-chaos@0.1.0    # pin a version
dsh --profile web --dump-config   # confirm `dsh-chaos` appears in the tree

The published package carries a prebuilt lib/ (via the dsh.bundle manifest declaration), so no source build runs on install.

To iterate against a local checkout, add it with a link:/file: spec instead:

dsh plugin --profile chaos-e2e add /path/to/dsh-chaos
dsh --profile chaos-e2e --dump-config   # confirm `dsh-chaos` appears in the tree

The plugin registers no tool and no prompt; it has zero effect when no rule matches — every rule is opt-in.

chaos:
  seed: 42              # optional: same config + same seed → same fault sequence
  rules:
    - tool: 'web_search'          # `*`-wildcard tool-name pattern; 10% fail with HTTP 429
      failureRate: 0.1
      error:
        status: 429
        probability: 1
    - tool: 'database_query'      # every call delayed 1–5s
      latency:
        min: 1000
        max: 5000
    - event: 'llm/request'        # 5% of model requests end with an HTTP 429 error finish
      error:
        status: 429
        probability: 0.05

Each rule names exactly one of tool (wildcard pattern) or event ('llm/request'). Faults:

Load-time validation is fail-loud: a rule naming both tool and event (or neither), a rule configuring no fault, a tool-only fault on an llm/request rule, a providerUnavailable on a tool rule, or latency.min > latency.max throws at plugin load.

pnpm install
pnpm run build      # tsc emits lib/
pnpm test           # vitest (no network)
pnpm run lint       # oxlint

Dev dependencies resolve to a local DeepSeek Harness checkout via file: links (../deepseek-harness), so build/test run against the same Harness APIs the plugin was developed against. Published @deepseek-ai/* peers are declared in peerDependencies with npm ranges.

This plugin depends on the tools/execute, tools/post-execute, and llm/stream Harness waterfalls. DeepSeek Harness is in developer preview and may introduce compatibility-breaking changes; verify the matrix above before upgrading either side.

• First-match-wins rule selection — a call is affected by only the first matching rule; overlapping rules do not compose per-fault.
• LLM latency and error are pre-stream — the llm/stream listener delays before returning the stream and can only end it with a terminal error finish; mid-stream chunk corruption (e.g. an injected tool-call-delta with broken JSON) is not yet supported.
• No chaos dashboard or metrics — this plugin injects faults but does not record or report them; a telemetry consumer of tools/result / session/event can attribute them via error.info.name (Chaos*).

MIT