Axstack

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

align → spec → tickets → implement → review → watch

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

What you can do

| Layer | Skill | What it does | | --- | --- | --- | | Plan | axstack-align | Settle scope through questions, a design lens sketch, and a four-family arena 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. | | Verify | axstack-review | Review a PR or bounded codebase at an exact revision. | | Verify | axstack-improve | Find evidenced codebase improvements without editing code. | | Verify | axstack-audit | Measure a run's outcomes and evidence gaps. | | Operate | axstack-watch | Observe or maintain an existing PR within its authority. | | Operate | axstack-cleanup | Retire eligible completed agent resources. | | Operate | axstack-relay | Send an explicit message or authorized notification. | | Understand | axstack-research | Answer one bounded question with sources. | | Understand | axstack-explain | Explain a system and separate known behavior from gaps. |

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.

Why Axstack

| Failure mode | How Axstack responds | | --- | --- | | Wrong thing built | Align rounds clarify the request; a four-family arena compares approaches for hard choices. | | Nobody really reviewed it | Strict TDD checks behavior first; with the mixed preset, cross-provider review checks the exact revision. | | Design rot | The design lens sketches boundaries before a build; Improve surfaces evidenced changes later. | | Agents left a mess | Orca makes delegation visible, one writer owns each PR, cleanup stays bounded, and a human merges. |

Quick start

[!NOTE] You need Bun >=1.3.14, Git, the GitHub CLI (gh) with gh stack, and a running Orca with the orca-cli, orchestration, and orca-linear guides. The agents selected by your preset must also be available in Orca.

Install the CLI and skills. This example targets Codex:

bun add --global axstack
axstack check --harness codex
axstack install --harness codex --preset mixed --yes

Open an Orca chat and ask for the phase you need:

$axstack-align Help me scope account recovery.
$axstack-implement Build the task we agreed on.
$axstack-review Review this pull request: <PR URL>
$axstack-review Find issues in <paths> at <commit SHA>.
$axstack-watch Monitor this PR without making changes: <PR URL>
$axstack-watch Watch every PR raised by this chat until all merge or close

Codex skills default to the shared ~/.agents/skills root. The owned AGENTS.md block stays under $CODEX_HOME (default ~/.codex). A default install retires only unchanged manifest-owned legacy skills; --skills-dir sets an explicit target without automatic migration.

Use --harness claude, opencode, or antigravity with axstack check and axstack install, or provide explicit skill and instruction paths. Installation adds an owned instruction block and preserves unrelated content. It does not enable automations or prove that every configured model is available.

See installation for source installs, custom paths, upgrades, conflicts, and uninstalling.

How work stays controlled

• The current chat drives scope, coordination, and publication. Delegation uses visible Orca orchestration via the orca CLI, not harness-native subagent tools. Separate worktrees keep one writer on each candidate.
• Peer PRs receive two independent reviews. 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. The human merges by default.

Choose one explicit preset (28 roles each): mixed (recommended), codex-only, or claude-only. Mixed supports cross-provider implementation review; single-provider presets have workflow limits and are not automatic fallbacks when a model is unavailable. See workflow and routing details.

In mixed and claude-only, requirements research, web research, and the optional monitor use Claude Sonnet 5.5 high.

Optional PR automation

Manual review works without a schedule. Own open PRs in chat-run mode use a harness-native monitoring or scheduled wake to resume the driver chat every 10 minutes; the existing Orca */10 observer is fallback only when the harness has no such capability. Each wake checks feedback, base, CI, and human approval; delegated work still uses Orca. Stop the chosen wake when all watched PRs merge or close, the user cancels, or it expires.

An optional native Orca review manager handles recurring peer review; the review automation never merges for you. Its activation is opt-in and needs live host validation. See PR-manager setup and safety.

Some notes

• Orca is the only supported active runtime. Axstack adds no daemon or runtime database.
• A human merges by default.
• This is an early project; expect the workflows to evolve.

Documentation

• Installation and configuration
• Workflows, review policy, and model routing
• PR scope and sizing
• Releases

License

MIT © 2026 Axat Bhardwaj.