Rami HalabiSoftware developer · Stockholm

AI agent tooling

SoherAgent

A governed delivery harness for coding agents: an exact authorized batch, leased work, independent review and hard stops.

See the demo

When
Jul 2026
Status
Working extension suite; private source
Source
Private source, available for supervised review
Stack
  • TypeScript
  • Node.js
  • Pi SDK
  • Obsidian

Problem and idea

Coding agents work fast, but without limits they commit, deploy or widen the scope on their own.

The idea: the agent can only act on a batch that a person authorized exactly. Work runs under a lease, an independent reviewer checks it, and the harness stops before anything that was not approved.

How it works

  1. PlanIdea, specification and tickets
  2. AuthorizeThe person approves one exact batch
  3. DeliverLeased implementation with checkpoints
  4. ReviewAn independent reviewer checks the work
  5. StopHard stop before anything unauthorized

Demo

Five chaptered terminal recordings follow one imagined project. Recordings 1 and 2 are sanitized concept recreations of a proposed planning mode, 3 and 4 recreate reviewed current behavior, and 5 shows captured test results.

Proposed /plan concept, sanitized: after approval, Scene 13 opens a detailed, scrollable nine-section canonical idea with evidence, risks, decisions, and a continuation checkpoint.

Open detailed idea · Open transcript

Proposed /plan concept, sanitized: Scene 7 opens a detailed, scrollable fourteen-section specification with architecture, requirements, security, recovery, tests, delivery structure, and traceability.

Open detailed specification · Open transcript

Reviewed current /yolo behavior: exactly the V1-S1 Contract insight MVP batch is discovered, authorized, and armed.

Open transcript

Reviewed current /yolo behavior: cited extraction and grounded review run inside one leased batch; the engineering adoption kit remains unauthorized.

Open transcript

Actual captured results: 30 runtime assertions and 155/155 extension tests passed.

Open transcript

What I built

  • An extension suite of about 8,500 lines: a control plane with capability gates, deterministic phase routing and resumable sessions.
  • Cross-process leases, bounded parallel agents and hard stops before commits, deployment, destructive actions or scope changes.
  • Obsidian knowledge tools with bounded retrieval, path controls and approved changes only.

Stack

  • TypeScript
  • Node.js
  • Pi Extension SDK
  • Obsidian
  • Node test runner

Quality evidence

  • 155 of 155 extension tests pass (test output).
  • 30 of 30 runtime assertions pass in a public synthetic runner (result, runner).
  • Evidence captured on 24 July 2026 by importing the private modules into the public runner.

Current limits

  • Recordings 1 and 2 show a proposed planning mode, not implemented behavior.
  • The test evidence does not prove a full live conversation or a production deployment.
  • The source is private. I can arrange a supervised review.