Review agent-made plans, docs, and diagrams without losing the thread.
Your agent ships a plan. You comment in chat. It “fixes” something. Three revisions later, nobody remembers what changed or what you accepted. Tweakloop is a local review workspace for that loop: comment on the exact paragraph or shape, turn feedback into claimable agent work, publish immutable revisions, and accept or reopen — nothing is done until you say so.
v0.1 alpha — the core loop works today for HTML, Markdown, and Excalidraw on your machine. What's ready and what's not →
Quickstart · Why · How it works · Agent setup · Docs
If this matches how you work with agents, star the repo for updates.
You need Node.js 24+,
pnpm 10.12+, and
just.
git clone https://github.com/Excoriate/tweakloop.git
cd tweakloop
pnpm install --frozen-lockfile
just open examples/plan.htmlThis builds the project, starts a workspace-local daemon, snapshots the example as revision R1, and opens a one-time authenticated review URL. You should see the document, its revision selector, Comment mode, and the collaboration rail.
Headless or CI (no browser opener)
just build
node dist/cli/index.js open examples/plan.html --no-browserOpen the printed URL yourself.
Better than scattered agent chat and ad-hoc file edits when you need accountability across review rounds:
- Revisions, not a mutable preview. Every publication is a new revision you can compare. An older result stays reviewable after the source file on disk changes.
- Feedback becomes work. A comment can stay conversational or become typed, claimable work tied to the exact artifact and revision that produced it.
- Agents are interchangeable. Codex, Claude Code, Cursor, OpenCode, or any CLI-capable process can use the same protocol. Tweakloop does not host or special-case a model.
- “Done” is not “accepted.” Agent activity, claimed work, a returned revision, readiness for review, and your decision remain separate facts.
For: developers who use coding agents daily and review plans, architecture docs, or diagrams before trusting them — and want that review to survive past the current chat thread.
Not for: replacing Git, hosting a model, cloud sync, or running arbitrary repo mutations from the review shell. Tweakloop coordinates and records the review loop; it does not execute implementation work on your codebase.
| 🔁 Stage | 👤 You | 🤖 Agent | 📋 Output |
|---|---|---|---|
| Open | Publish an artifact for review | — | Revision R1 |
| Review | Select exact text or shape; leave typed feedback | — | Comments / intents |
| Claim | — | Takes ownership of the work durably | Claim receipt |
| Revise | — | Publishes a child revision | Revision R2+ |
| Decide | Accept or reopen | — | Human decision recorded |
Loop at a glance (ASCII)
Open R1 ──▶ Review ──▶ Claim ──▶ Revise R2 ──▶ Decide
▲ │
└──────────────── reopen ────────────────────┘
- Open. Tweakloop registers the artifact and publishes its bytes as revision R1.
- Review. You select an exact section, paragraph, or whiteboard element and describe the required change, constraint, or decision.
- Claim. The assigned agent receives that precise work and claims it durably. Retries and handoffs cannot silently create a second owner.
- Revise. The agent publishes R2 as a child of R1 and reports what it addressed.
- Decide. You compare the result and accept it or reopen the work for another pass.
| 📄 Artifact | ✨ What Tweakloop preserves | 👉 Start here |
|---|---|---|
| HTML | semantic anchors, interactive rendering, immutable revisions | examples/plan.html |
| Markdown | heading ancestry, stable block anchors, safe rendering | examples/markdown-collaboration.md |
| Excalidraw | semantic nodes, edges, groups, managed drafts, immutable publication | examples/engineering-whiteboard.excalidraw |
HTML plans can also embed managed whiteboards. The browser keeps Documents, Tasks, Comments, and Chat in one session instead of scattering the review across unrelated tools.
Install the public skill, then ask your agent to run the workflow:
npx skills add Excoriate/tweakloop --skill tweakloopUse the Tweakloop skill to draft a plan for this change, open it for my review, address the feedback I submit, and return the revised artifact for acceptance.
Choose the agent and project scope when prompted. The skill runs with that agent's permissions. Full workflow: agent skill · CLI reference
- Loopback-only daemon; review shell and artifact content on separate origins (artifact content has no shell credential or mutation route).
- Append-only local event log plus content-addressed bytes; human accept/reopen requires the authenticated browser session.
- Trust boundary is the local OS user — see SECURITY.md for the full model.
More: failure model · design principles
v0.1 alpha. The core review loop runs end to end for HTML, Markdown, and Excalidraw — typed feedback, atomic agent claims, immutable child revisions, live browser updates, and explicit human accept/reopen. CI runs the Chromium end-to-end suite on every pull request and
mainpush.
Known gaps (read before adopting deeply)
- First-class evidence objects and verification records are not implemented yet; completion summaries, semantic checks, diffs, and human decisions are available today.
- Automatic cross-revision anchor re-resolution and explicit orphan facts remain incomplete.
- Owned-daemon restart and recovery paths are exercised, but arbitrary SIGKILL, power-loss, and cross-platform crash consistency are not claimed.
- Chromium is the primary browser test target; Firefox, WebKit, and screen-reader verification are still open.
Exact docs-to-code ledger:
docs/architecture/16-implementation-status.md
Read CONTRIBUTING.md for setup, just check, end-to-end tests, and architecture
expectations. Before modeling new behavior, read the
design principles and
ubiquitous language.
Report vulnerabilities via GitHub Security Advisories.
- Start with the documentation map
- Understand the product and non-goals
- complete CLI reference
- Read the architecture
Tweakloop was partly inspired by Lavish AXI, which made precise feedback on agent-generated HTML feel immediate. Tweakloop extends that idea into a durable, agent-neutral review workflow with immutable revisions and explicit human decisions.
MIT © Alex Torres.