Horus

Agent packet

A compact, honesty-framed briefing built from an investigation — designed to fit a coding agent's context window.

horus packet produces a grounded briefing for a coding agent: the headline cause, the evidence behind it, the honesty caveats, suggested next steps, relevant files, lower-priority areas based on current evidence, and any relevant remembered context — tightly capped to fit a context window. It is a projection of a Horus investigation, not a new kind of analysis: the same deterministic engine produces the report, and the packet re-shapes it for an agent.

When to use it

Use horus investigate when a human is reading the report. Use horus packet when an agent is — it strips the report down to what an agent needs and labels exactly how much to trust it.

Two input modes

The argument is auto-detected:

  • A saved investigation id (a UUID) — Horus loads the persisted report and projects it. No re-query of your systems.
  • A hint (free text) — Horus runs a fresh investigation, then projects it. Building a packet from a hint needs a source-intelligence connector, exactly like horus investigate.
bash
horus packet "checkout latency spike"
horus packet <savedId>

JSON output

Pass --json to emit the packet as structured JSON — clean for piping into an agent. Without it, Horus renders the packet as Markdown. The two carry the same content; only the format differs.

bash
horus packet <savedId> --json

The honesty header

Every packet leads with an honesty header so an agent knows how much to trust the briefing before it acts. It surfaces:

  • Headline confidence — the confidence, honesty band, and a working-hypothesis flag when a ceiling capped it.
  • Integrity caveats — low-confidence/fuzzy seed, headline not seed-linked, gap confidence-ceiling, single-source (single-provider) ceiling, blind spots, and a 'Runtime-only (source intelligence unavailable)' degraded note.
  • To raise confidence — the gaps' next-source actions.
  • Per-source contribution status — contributed / empty / failed / not-configured.
  • Freshness caveats — remembered context flagged when it may be stale.
  • Routing — where the investigation was sent (incident vs source-impact).

Honesty is computed, not narrated

The caveats and routing come from the same deterministic signals the report already carries — index metadata, evidence timestamps, the router. The packet surfaces them; it does not generate prose about them.

Remembered context from memory

A packet is enriched with what Horus already knows: the memory items relevant to the packet's project and seed, recalled from the local memory store. This recall is best-effort and fail-closed — a missing project or an unavailable store simply yields no memory section and never blocks delivery — and the recalled items are context only, never live evidence and never part of scoring.

Agent presets

--for <agent> selects a presentation preset — claude, cursor, or generic. A preset only tightens the caps (and may drop timestamps) to suit the target. It never changes which evidence or caveats are computed, so the substance of the packet is identical across presets.

bash
horus packet "queue backlog" --for claude --json

Next steps

See the CLI reference for every horus packet option, Memory for the remembered-context model, and Self-routing for how Suggested next steps are produced.

PreviousMemoryNextSecurity