Axstack

Engineering workflows for AI agents, from a first question to a reviewed PR.

align → spec → tickets → implement → review → watch

Axstack gives your current T3 thread a way to scope work, build it with tests, and review the exact result. T3 Code provides worktrees, agent threads, and visible coordination. You can start at the phase you need.

Why Axstack

| Failure mode | How Axstack responds | | --- | --- | | Wrong thing built | Align rounds clarify the request; every brainstorm runs a light arena across configured families. Use judges only for Rung 2 hard-to-reverse choices. | | Nobody really reviewed it | Strict TDD checks behavior first; mixed authored review is cross-provider, and the first mixed peer review pairs Sol high with Opus medium in isolated sessions at the exact revision. Peer follow-ups use one reviewer. | | Design rot | The design lens sketches boundaries before a build; Improve surfaces evidenced changes later. | | Agents left a mess | T3 makes delegation visible, one writer owns each isolated candidate, cleanup stays bounded, and watched own PRs merge under guarded rules. |

Prerequisites

Axstack supports only Linux hosts. You need Bun >=1.3.14, Git, the GitHub CLI (gh), and T3 Code stable 0.1.0 or newer on PATH. Nightly and other prerelease builds are unsupported. The providers selected by your preset must be available inside T3.

Choose one explicit preset (each contains all role IDs): mixed (recommended), codex-only, or claude-only. Mixed supports cross-provider implementation review; single-provider presets have workflow limits and are not automatic cross-class or cross-provider fallbacks when a model is unavailable. See workflow and routing details.

Quick start

Install the CLI and skills, then check your host. This example targets Codex:

bun add --global axstack
export PATH="$HOME/.bun/bin:$PATH"
axstack install --harness codex --preset mixed --yes
axstack check --harness codex

Reload the harness's skills in T3 after installation. A green check verifies local capabilities and installed files; it does not prove live provider readiness, schedule activation or mobile delivery. See installation for other harnesses, custom paths, upgrades, conflicts and uninstalling.

First task

Open your repository as a T3 project and start a driver thread. Codex invokes skills with $skill, and Claude uses /skill. For a small repair, send this request (replace $ with / in Claude):

$axstack-implement Fix the empty-state message in the existing list view within one PR, preserve list behavior, and verify the displayed result.

Small, bounded changes can begin with your request or an existing issue; substantial work needs an approved spec and matching tickets before implementation. Research, explanation, and peer review can start directly. The driver returns a candidate revision, test evidence and unverified boundaries. You make decisions at holds and substantial-spec approval. Follow getting started for the full first-task journey and guides for features, reviews, watches, debugging and releases.

What you can do (20 skills)

| Layer | Skill | What it does | | --- | --- | --- | | Plan | axstack-align | Settle scope through questions, a design lens sketch, and inline brainstorm validation. | | Plan | axstack-brainstorm | Validate an approach with independent candidates; judges for hard choices. | | Plan | axstack-spec | Write and approve an observable specification. | | Plan | axstack-tickets | Break approved scope into executable tasks. | | Build | axstack-implement | Build with strict TDD and an author → review → repair loop. | | Build | axstack-debug | Diagnose a bug with a failing check and hand off a bounded repair. | | Build | axstack-perf | Route measured performance work to Debug, Improve, or Implement. | | Verify | axstack-review | Review a PR or bounded codebase at an exact revision. | | Verify | axstack-verify | Create, prove and maintain a repository verification skill. | | Verify | axstack-improve | Find evidenced codebase improvements without editing code. | | Verify | axstack-audit | Measure a run's outcomes and evidence gaps. | | Verify | axstack-correct | Report repeated mistakes and propose stronger checks when invoked by the user. | | Operate | axstack-watch | Observe or maintain an existing PR within its authority. | | Operate | axstack-cleanup | Retire eligible completed agent resources. | | Operate | axstack-publish | Publish one checked Explain article in the explicitly authorized private or public mode. | | Understand | axstack-research | Answer one bounded question with sources. | | Understand | axstack-scout | Gather codebase evidence for an agent across several areas. | | Understand | axstack-explain | Explain a system and separate known behavior from gaps. | | Understand | axstack-anatomy | Build an isometric physical-object figure with the vendored anatomy kit. |

Explanations default to checked inline T3 pages above a short reply. Use Anatomy only on an explicit isometric or physical-object figure request. Every Explain visual is authored and independently reviewed; plain prose stays proportional. Architecture diagrams add scout grounding, conditional research, specialist authorship and review by a different provider than the actual author.

How work stays controlled

• The current T3 thread drives scope, coordination, and publication. Delegation uses visible T3 orchestration via the t3-code MCP; harness-native subagent tools are forbidden. Separate worktrees keep one writer on each candidate. Useful isolated packages can feed one PR with one integration writer; dispatch independent ready work before waiting under Parallel-first delivery. See the T3 runtime boundary.
• Peer PRs receive two independent reviews on the first review and one reviewer on follow-ups. Authored changes receive a reviewer selected from the author's configured pairing. Reviews bind to exact revisions.
• Agents keep accepted decisions and evidence for resume. Missing authority, unavailable models, and serious risks surface as holds. For own PRs, automatic merge is the default under the watch predicate.

Optional PR automation

Manual review works without a schedule. Every verified own-PR publication arms or joins the driver's chat-run watch, subject to explicit stop-after-publication or observation requests. See Host operations for wake activation and cadence. See Chat-run PR watch for the lifecycle and quiet cadence.

An optional native T3 review manager runs finite peer-review passes hourly (everyMs:3600000); the review automation never merges for you. Activation needs live host validation. See PR-manager setup and safety.

Automatic merge boundaries

Eligible own PRs use guarded automatic merge; see merge boundaries, exclusions and accepted risks.

Some notes

• T3 Code is the only supported active runtime. Axstack adds no daemon or runtime database. The exception is exactly one opt-in systemd user timer: axstack-single-server removes duplicate t3 serve process trees for one base dir while keeping the ownership-lock holder, otherwise the recorded server. It runs a stateless script, keeps no workflow state and launches no agent; see Single T3 server.
• The recorded owning watch thread merges eligible own PRs; excluded PRs use a merge card.
• This is an early project; expect the workflows to evolve.

Documentation

| Page | Use it for | | --- | --- | | Getting started | Install, understand check, and run a first bounded task. | | Concepts | Driver threads, phases, roles, presets, holds, evidence and provider bindings. | | Guides | Step-by-step journeys for features, fixes, reviews, watches, debugging and releases. | | Workflows | Routing, review policy, merge boundaries, exclusions and accepted risks. | | Installation | CLI commands, harness paths, configuration, upgrades and troubleshooting. | | Host operations | Configure your host, provider runtimes, remote access and automations. | | Skill writing | Contribute self-contained skills and contract tests. |

See PR scope and sizing for candidate shape and Releases for release history.

Development

bun install --frozen-lockfile
bun run lint
bun run typecheck
git config core.hooksPath .husky

The hook is opt-in per clone via git config core.hooksPath .husky and requires bun install. Without dev tools it prints a setup hint and skips the checks. The hook runs lint and typecheck only. Biome formatting is optional for editors or manual use; commits do not enforce formatting. TypeScript checks JavaScript in src/ and bin/ without emitting files. All tooling is development-only and is excluded from packed assets by package.json.files. Regenerate bun.lock with the minimum Bun version in package.json.engines (currently 1.3.14); newer Bun versions can write an incompatible lockfile format.

License

MIT © 2026 Axat Bhardwaj.

The opt-in saved New-chat default script uses subscription capacity and preserves project overrides; its five-minute timer is never enabled by installation. Role-labelled title instructions need T3 >=0.1.3.