dsh-audit-log

Who mutated my data flow? — Runtime data-flow audit log for DeepSeek Harness plugins, with per-plugin, per-fiber attribution.

English | 简体中文

dsh-audit-log answers the question every plugin ecosystem eventually asks: when a field in your session data
disappears or gets rewritten, which plugin did it, in what order, and what exactly changed?

It is a read-only observer: it never blocks, never rewrites, and never stores payload values — only structural
fingerprints (types, keys, lengths). Secrets never enter the log.

Why

Cordis (the framework under DeepSeek Harness) shares the same args array across every listener of a
waterfall dispatch, and the internal/get / internal/set interception points are public. Any plugin can
silently rewrite data flowing through the system. With hundreds of plugins, "which plugin washed out my
field" is otherwise nearly impossible to answer.

dsh-audit-log makes it answerable:

• Dispatch-level diff — every audited dispatch snapshots argument shapes before and after, and records
  the mutations in between.
• Per-listener window attribution — each listener is wrapped at registration; mutations are attributed
  to the exact plugin + fiber that performed them (confidence: "window"), not a heuristic.

Install

# into your web profile
cd ~/.dsh/profiles/web
pnpm add dsh-audit-log

Then add dsh-audit-log to dsh.profile.bundles in package.json (or use dsh plugin --profile web add dsh-audit-log),
and restart your dsh instance.

The plugin registers ctx.auditLog and starts recording immediately. No configuration required for basic use.

Usage

Query the audit trail from any plugin:

const records = await ctx.auditLog.query({
  events: ['message/send'],
  mutationsOnly: true,   // only dispatches that changed something
  fromSeq: 100,
  limit: 50,
})

// Each mutation carries exact attribution (when attribute_by_window is on):
//   { listenerIndex, package, fiber, confidence: 'window' }
const culprit = records[0].mutations[0].attribution?.package

// Raw per-listener windows:
const windows = ctx.auditLog.queryWindows({ event: 'message/send' })

Record shape

{
  "v": 1, "ts": "2026-08-19T08:00:00.000Z", "seq": 42,
  "mode": "waterfall", "event": "message/send",
  "listeners": [{ "order": 0, "package": "my-plugin", "fiber": 3 }],
  "before": [ { "type": "object", "keys": ["content"] } ],
  "after":  [ { "type": "object", "keys": ["content"] } ],
  "mutations": [{
    "argIndex": 0, "path": "arg[0].content",
    "kind": "replace", "beforeLength": 56, "afterLength": 36,
    "attribution": { "listenerIndex": 1, "package": "spam-filter", "fiber": 7, "confidence": "window" }
  }]
}

Configuration

Via the profile patch layer (cordis.patch.yml), or $DSH_HOME/settings.yaml namespace audit-log:

- id: audit-log
  config:
    events: ["message/send", "before/*"]
    package_blocklist: ["my-noisy-plugin"]

How it works

dispatch → internal/dispatch (prepend) → fingerprint(args) → listeners run → fingerprint(args) → diff → store

1. Dispatch-level diff (P0/P1) — Cordis emits internal/dispatch synchronously before public listeners
   run. The service mounts that one hook (prepend: true, global: true), snapshots argument shapes, lets the
   dispatch proceed, then diffs and stores. Shape-only fingerprints (maxDepth 3, maxKeys 20) keep the
   overhead negligible and payload values out of the log.
2. Per-listener window attribution (P2) — intercepts internal/listener (bail) at registration time and
   wraps every non-internal listener. Each wrapped call snapshots the shared args before and after that
   listener, so mutations are attributed to the exact plugin + fiber (confidence: "window").
   Framework-internal events are never wrapped — no recursion, no self-noise.

Limitations

• Shape-only fingerprints: equal-length string swaps ('by-c' → 'by-d') and in-place number edits are
  invisible to the diff. This is a deliberate privacy/cost trade-off; a value-aware deep mode is future work.
• Sync-path timing: the microtask resume is exact for synchronous emit; for async listeners in
  serial/parallel/waterfall, the per-listener window wrapper (P2) covers the gap precisely, while the
  dispatch-level diff remains a coarse overview.

Relationship to other plugins

Observability tools in the ecosystem come in layers: some show what plugins put into the model prompt
(context layer), while dsh-audit-log shows what plugins changed in the event data flow (runtime data
layer). They are complementary — install both.

Development

npx @dsh-io/dsh-dev check    # validate manifest, YAML, build
npx @dsh-io/dsh-dev dev      # run under dsh web with file watching

Tests (no framework needed):

node --test tests/

License

MIT