spice-harness 0.22.0

Spice Harness: an agent harness / fleet operations console for coding repositories.

Spice Harness

Spice Harness is an agent harness / fleet operations console.

Simultaneous Production, Integration, and Control Environment.

spice is an installed, repo-native harness for operating coding agents. It treats the agent transcript as the source of truth and the repository filesystem as the steering channel; supervision, task routing, git pressure, live feedback, and hygiene gates are derived from those two surfaces.

It is built for agents moving fast in parallel: every correction is durable, every task boundary is observable, and the gate catches structural drift before it lands.

spice is building itself, but it was not created in a vacuum: the loop was born from a harsher polyglot environment where many languages, conventions, and agent lanes had to survive contact with one another.

_{Operator steering arrives in the live stream; an assistant ACK retires the exact inbox key from the durable filesystem queue.}

What it does

• Semantic ACKs: steering is not considered handled until the agent acknowledges the durable key in assistant prose.
• Task allocation: spice task next owns work selection; task boundaries own git synchronization and review phases.
• Conscience: curated maxims judge assistant prose while work is still in flight, then route violations back as ordinary steering.
• Constitution: pre-commit and spice study ... enforce repository shape, file/routine limits, env policy, reachability, assertion density, private internals, and commit-message rules.
• Serve UI: spice serve exposes lanes, teams, live transcripts, steering, attachments, task routing, and browser-visible diagnostics.
• Observer UI: spice watch <session-dir>... follows existing Codex and Claude transcripts without initializing or modifying their directories.

See docs/overview.md for the operating model and docs/interface.md for the serve UI.

Posture: a single-operator console

spice serve is a single-operator console, and that is a deliberate identity — not a limitation to grow out of:

• SQLite, localhost, one shared token. The server is a stdlib process backed by SQLite, bound to 127.0.0.1 by default; when it is reached beyond loopback it is gated by a single shared --auth-token. There are no accounts, sessions, or per-user identities.
• The growth vector is remote reach for one operator, not multi-user auth. Reaching your own fleet from elsewhere is a transport choice — an SSH tunnel or a tailnet bind over the same one-token surface (see single-operator remote reach).
• Multi-user auth is an explicit non-goal. Do not grow a multi-operator team product out of the stdlib server. Many humans may steer one lane, but only through the same durable filesystem queue — never privileged per-user channels (see no-privileged-channel).

Start Small

Spice Harness is a progressive-disclosure product: watch, then gates, then steer, then fleet. Start by observing existing agent sessions with no repository changes; add constitution gates when the team wants enforceable hygiene; bind one agent when direct intervention becomes necessary; move to the task-backed fleet only when work needs multiple coordinated lanes. The full prerequisite and graduation path is the entry ladder.

Commands

Surface	Command
Watch existing agent sessions	spice watch <session-dir>...
Install constitution gates only	spice init --gates
Prepare steering and fleet surfaces	spice init / spice doctor
Open a manually steered lane	spice agent ensure / spice serve
Run through the agent wrapper	spice agent run -- <cmd>
Pull allocator work	spice task next
Rehydrate context	spice session briefing
Open the operator UI	spice serve
Observe foreign sessions read-only	spice watch <session-dir>...
Run studies and gates	spice study ... / git pre-commit hook

Configuration lives in CONFIG.md. The design contract lives in DESIGN.md. Wrapper command behavior is detailed in docs/cli/wrapper-commands.md. Stability expectations for extensions and command coupling live in STABILITY.md.

Install

uv tool install -e /path/to/spice-main
# or, for the released package:
uv tool install spice-harness

# RTK is required for the agent shell (version 0.42.4 or newer):
brew install rtk
# or: cargo install --git https://github.com/rtk-ai/rtk

cd /path/to/your/repo
spice init
spice doctor

For repository hygiene without the task plane, shell wrapper, or agent skill, install the standalone constitution tier instead:

spice init --gates

This installs the pre-commit constitution (including sticky-flex limits, regression-only magic-number ratchets, taste policy, and configured extensions) plus commit-message hygiene. It does not install the fleet-specific reference guard or materialize agent files. Commit normally to run the gates, or invoke the staged gate directly with spice dev pre-commit.

The default install is a uv tool. Operators who deploy from a main tree should use the editable form so the installed spice command resolves to that tree; that editable main tree is the server deployment. Other worktrees remain operated trees and do not supply their own runtime.

Graceful degradation

RTK is a required companion for the agent shell: spice agent run delegates command selection to rtk rewrite, and spice doctor verifies the supported protocol before agents work. The local judge and speech synthesis are degradable companions; when either is unavailable, transcript capture, steering, tasks, and the constitution keep working while maxim feedback or audio narration is skipped. Runtime, verification, and protocol details are in CONFIG.md.

Release

Release workflow is documented in docs/release.md. Most users only need to know that releases are cut from clean synchronized worktrees through the repository's mounted spice release command.

Status

Work in progress toward a standalone, releasable product. The loop described here is real, exercised daily, and guarded by the same constitution that spice init installs elsewhere.