THE REPOSITORY'S OWN DOCUMENT · OPERANT

Verification — the harness that proves work

operant's verification harness

Repository
agentic-harness/operant/docs/library
Commit
fbe90019
Date
2026-11-01
Length
54 lines · 1 min

Read the source on GitHub This page renders that file; it is not a summary of it.

Verification — the harness that proves work

Owns: operant's verification harness — recipe detection, phased execution, the evidence ledger, and the agent-facing verify_task tool. Not for: the guardrail/approval stack (security.md), the autonomous loop that consumes verification results (autonomous.md).

Source: crates/operant-core/src/verification.rs, crates/operant-core/src/tools/verification_tool.rs, crates/operant-core/src/estop.rs (the pause that halts scheduled runs).

The harness

verification.rs (hermes agent/verify/ parity: recipes.py + runner.py + environment.py) detects a project's verification recipe from its manifests — Rust (Cargo.toml), Node (package.json), Python, Go, Java, Makefile, docker-compose — then runs it phase by phase:

  1. bootstrap — install/prepare dependencies,
  2. build — compile,
  3. test — run the suite,
  4. start (optional) — boot the app and poll an HTTP readiness URL.

Every phase's result is recorded in the verification_events evidence ledger in the state database — an auditable trail of what was run and what it returned, not a boolean.

The agent tool

verify_task (tools/verification_tool.rs) exposes the harness to the model: it detects the recipe, runs the phases, and records every phase in the evidence ledger. The model can therefore prove its work instead of claiming it.

ESTOP — the operator's brake

estop.rs — operant pause writes a sentinel file at <data_dir>/ESTOP; operant resume removes it. While the sentinel exists, the cron scheduler skips dispatching due jobs, the kanban CLI dispatcher skips claiming tasks, and new gateway turns get a brief "Operant is paused" reply. Existing work finishes; NEW work does not start. A resumable pause for new work only.

Intention

The harness exists because "the agent says it is done" is not evidence. Recipe detection from manifests (rather than a configured command) keeps it zero-config across arbitrary projects; the phases mirror what a human run would do; the events ledger makes every verification citable later — the autonomous loop and the org layer both read it. The ESTOP is the operator's guarantee that a self-directed loop can always be stopped at a boundary rather than mid-write.

Other published documents in operant

Back to the operant case study 23 of this study's documents are published on the site; the repository holds the rest.