agent-work
agent-work is a forge-neutral Git lifecycle controller for AI-agent and human implementation tasks.
It makes parallel repository work safer without introducing a coordination service, database, daemon, or vendor-specific merge system. Projects define a small amount of policy; agent-work computes and enforces the mechanical lifecycle around implementation:
start task
→ implement bounded change
→ checkpoint owned files
→ validate exact commit
→ preserve submission
→ land against current integration head
→ verify reachability
→ close and audit
The tool is designed to keep agents focused on engineering. Agents should not have to remember when to commit, which validation receipt is still valid, whether a task is integrated, or whether cleanup is safe. The CLI derives those facts from Git and returns one deterministic next action.
Product principles
- Git is the authority. Durable work is represented by commits and refs, not chat history or mutable status prose.
- Forge-neutral. Local bare repositories, SSH remotes, self-hosted forges, and hosted forges use the same core workflow.
- No background platform. The initial product has no daemon, database, scheduler, dashboard, or forge API dependency.
- Fail closed. Undeclared scope, stale validation, unsafe cleanup, and ambiguous integration block rather than guessing.
- Idempotent transitions. Repeating a command after interruption must continue safely without duplicate commits or publication.
- Small integration units. One task packet represents one independently integrable batch.
- Project-owned validation. Consumer repositories define the commands that prove their own behavior.
- Compact agent protocol. Skills delegate lifecycle reasoning to the CLI and receive structured next actions.
Intended usage
A consumer repository contains a tracked configuration file and project-owned validation commands:
consumer-repository/
├── .agent-work.toml
└── tools/ci/
├── test-affected
└── test-integration
Install the stable wheel with an isolated tool installer:
pipx install sharedincc-agent-work
# or
uv tool install sharedincc-agent-work
Initialize and operate a repository:
agent-work init --json
agent-work doctor --json
agent-work profiles list --json
agent-work task create --id AUTH-17 --objective "Add token rotation" \
--commit-summary "add token rotation" --profile auth \
--output .agent-tasks/AUTH-17.json --json
agent-work workspace create .agent-tasks/AUTH-17.json --json
agent-work workspace inspect AUTH-17 --json
agent-work claim inspect AUTH-17 --json
agent-work claim recover --json
agent-work ready --json
agent-work run --json --output-profile agent
agent-work forecast --json
agent-work workspace retire AUTH-17 --json
agent-work operations list --json
agent-work operations inspect <operation-id> --json
agent-work operations resume <operation-id> --json
agent-work adopt --analyze --repository /path/to/consumer --json
agent-work adopt --write --analysis-sha256 <analysis-sha256> \
--repository /path/to/consumer --json
agent-work adopt --verify --repository /path/to/consumer --json
agent-work task create --dry-run --id AUTH-18 --objective "Preview task" \
--commit-summary "preview task" --allow "src/**" \
--output /tmp/AUTH-18.json --json
agent-work diagnose --json
agent-work audit --json
python tools/qualify_release.py --agent-work /path/to/installed/agent-work --json
python tools/qualify_release.py --agent-work /path/to/installed/agent-work --scenario synchronization --json
python tools/qualify_release.py --agent-work /path/to/installed/agent-work --scenario adversarial --json
python tools/qualify_release.py --agent-work /path/to/installed/agent-work --scenario coordination-recovery --json
python tools/qualify_release.py --agent-work /path/to/installed/agent-work \
--scenario lost-response --json
Install the canonical low-token skills with:
agent-work skills install --target all --json
agent-work skills verify --target all --json
Normal execution agents should not manually commit, merge, rebase, push, or retire task worktrees while the controller owns the task lifecycle.
Repository status
Version 2.0.0 provides the stable Git lifecycle, deterministic agent protocol, canonical task authoring, readiness, policy-controlled refresh, non-publishing integration forecast, managed workspaces, operation recovery, consumer adoption, and the completed packaged qualification matrix. The authoritative architecture, release procedure, and completed implementation plan are documented under docs/.
Documentation
- Architecture
- Lifecycle and state machine
- Configuration contract
- Security model
- Implementation plan
- Development guidelines
- Architecture decisions
Distribution identity
| Surface | Identity |
|---|---|
| Repository | agent-work |
| Python distribution | sharedincc-agent-work |
| Python import package | agent_work |
| Executable | agent-work |
| Runtime | Python 3.11–3.13, standard-library-only runtime dependencies |
The distribution name is intentionally distinct from the import package and executable.
Metadata
Release files for sharedincc-agent-work 2.3.4
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| sharedincc_agent_work-2.3.4.tar.gz | 286.1 kB | Without provenance | Details for sharedincc_agent_work-2.3.4.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| sharedincc_agent_work-2.3.4-py3-none-any.whl | Python 3 | none | any | Details for sharedincc_agent_work-2.3.4-py3-none-any.whl |
Total release size: 406.2 kB
Security for sharedincc-agent-work 2.3.4
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.