Harness

Lightweight D agent runtime with persisted actor graphs, Mixr routing, and a local desk UI.

Explore the docs »

Harness is D only, enables tgc (thread-local GC) by default, persists its node graph on disk, and uses Mixr for model routing.

Phase	Ships
v0	graph.json, per-node meta.json + chat.jsonl, orchestrator metathread, CLI
v0.5	Mixr on-device + suppress lists; /api/health, OpenAI-compat stub, /api/provider/* for t3code
v1	Mixr router, wait-graph (warn default), HTTP desk (harness serve)
v2	Temporal layout engine — scrubber, fade, heatmap, scoped bookmarks
desk PM	Status colors for awaiting_user / failed; Mark completed + Archive (POST /api/hide)
desk views	Linear / Grid / Tree — discussion vs workers separation; fork drill-in for full node transcript
plan queue	Cross-cutting plan-stack dock (GET /api/plan-queue) — ETA-sized speculative waits; not a 4th view mode

Contents

• Changelog
• Build
• Quick start
• Disk layout
• Harness configuration
• Architecture
• Built with

Changelog

See CHANGELOG.adoc.

(back to top)

Build

Requires DUB and LDC/DMD. Thread-local GC via dependency on dlang-supplemental/tgc (Tgc_default → --DRT-gcopt=gc:tgc embedded in the binary).

cd $env:code\github.com\dev-centr\harness
dub build
dub test

(back to top)

Quick start

.\harness.exe init .\my-chat
.\harness.exe spawn .\my-chat --parent=coordinator --type=task --title=docs-sync
.\harness.exe serve .\my-chat --port=8765

Open http://127.0.0.1:8765 for the desk UI. Toggle Linear, Grid, or Tree in the shell bar — discussion chat stays in the main panel; task workers live in the workers region (or the tree canvas in Tree mode). The timeline scrubber and plan-stack dock work in every mode (start plan-stackd on :17358 for live queues).

(back to top)

Disk layout

Matches actor-model-agentic-ui:

$CHAT_ROOT/
  graph.json
  orchestrator/meta.jsonl
  nodes/{id}/meta.json
  nodes/{id}/chat.jsonl
  timeline/bookmarks.jsonl

(back to top)

Harness configuration

Set HARNESS_NAME = harness in $CODE_ROOT/harness.md.

Parallel / swarm close-out (see agent-rules general/parallel-git-closeout.md):

GIT_CLOSEOUT = coordinator-batch

Task nodes spawn with gitCloseout=coordinator-batch and a system brief: commit locally, do not push; the coordinator batches push after the wave.

(back to top)

Architecture

• Long-lived: harness serve — supervisor + HTTP desk (future: harnessd RPC for thin CLI clients).
• One-shot: CLI subcommands touch disk directly (git-shaped); no Python, no embedded interpreter per invoke.
• tgc: per-thread heaps; collections do not stop-the-world sibling threads — fits actor swarms + @nogc workers.
• Discovery vs routing: Open Provider Registry / UniProvider finds endpoints; Mixr chooses models. Prefer OPR manifests over hard-coding a single local runner brand.

(back to top)

Built with

Runtime

• D — native implementation
• tgc — thread-local garbage collection

Interface

• Browser-native HTML, CSS, and JavaScript — local desk UI

(back to top)