Skip to main content

VS Code-native AI engineering workflow system for microservices, libraries, and backend systems.

Project description

vstack

PyPI version Python version Verify status Security checks Runtime: PyYAML License: MIT GitHub Discussions

The VS Code-native AI workflow system for backend engineering.

vstack installs structured agents, skills, instructions, and prompts into .github/ so GitHub Copilot Agent Mode can run repeatable backend workflows with clear role boundaries.

It provides six delivery roles for end-to-end software work: product, architect, designer, engineer, tester, and release, coordinated by planner.

Best for

  • Backend and API teams using GitHub Copilot Agent Mode in VS Code
  • Repositories that want consistent planning, implementation, verification, and release flow
  • Teams that want reusable AI workflows instead of one-off prompt crafting

What you get

  • Fixed role model: six delivery roles plus a planner coordinator
  • Template-driven install model from src/vstack/_templates/
  • Backend-first verification, security, and release discipline
  • One runtime dependency: PyYAML

Building blocks

Artifact type Purpose Typical invocation
Agents Main operating interface for role-based work @product, @tester
Skills Reusable task procedures /verify, /security
Instructions Baseline policy and repository guardrails auto-loaded by context
Prompts Reusable prompt artifacts where direct prompting is useful explicit prompt use

Prompt catalog

Prompts are .prompt.md files installed to .github/prompts/. Invoke them via the VS Code command palette (Chat: Run Prompt File) or the Copilot Chat attach button.

Prompt Purpose
api-design-review Review an API design or OpenAPI spec for correctness
architecture-risk Identify architectural risks and mitigation priorities
code-review Review a change for bugs, regressions, and missing tests
dependency-audit Audit dependencies for vulnerabilities and licence risks
incident-timeline Build an evidence-based incident timeline and post-mortem
migration-safety Review DB migration safety, rollback, and zero-downtime
release-readiness Evaluate release readiness from reports and open blockers

Quickstart — fresh install

Install with pipx, then install vstack artifacts into your repository:

# Install the CLI once, globally
pipx install vstack

# Move to your repository root and run install — no --target needed
cd /path/to/your/project
vstack install        # seeds .vstack/config.yaml and generates .github/ in the current directory
vstack validate       # confirm no errors

When you omit --target, vstack uses the current working directory. The equivalent explicit form is vstack install --target /path/to/your/project.

Run a first task in Copilot Agent Mode:

@tester /verify Check this repository and summarize findings

Expected result:

  • vstack validate reports no unresolved template tokens
  • Agent command returns a concrete verification summary for your repository

Quick upgrade

Patch or minor version (e.g. v3.1 → v3.2, same major)

Docs paths never change within a major version. Only .github/ artifacts are updated.

pipx upgrade vstack

cd /path/to/your/project
vstack init           # idempotent — safe to run in CI

Major version (e.g. v2 → v3)

Docs paths may change on a major version bump. Run vstack migrate before vstack init.

pipx upgrade vstack

cd /path/to/your/project
vstack migrate        # moves docs files to their new paths (auto-detects installed version)
vstack init           # regenerates .github/ artifacts

# Only if you see "Legacy manifest schema detected" in the output above:
vstack manifest upgrade
vstack init

Preview the docs moves without touching any files:

vstack migrate --dry-run

For upgrades spanning multiple major versions (e.g. v1 → v3), vstack migrate chains all intermediate steps automatically. Use --from and --to to specify the range explicitly if auto-detection from the manifest fails:

vstack migrate --from 1 --to 3
vstack init

Force reinstall (overwrite local edits)

vstack install --force                       # overwrite all managed artifacts
vstack install --force-name agent/engineer   # overwrite one specific artifact

Why this helps

  • Consistent role boundaries for planning, implementation, validation, and release
  • Reusable skills and instructions instead of ad hoc prompts
  • Better release hygiene with documented workflows and CI alignment

Core commands

vstack --version
vstack validate

# Run from your repository root (--target defaults to the current directory)
vstack install
vstack init
vstack migrate
vstack manifest verify
vstack manifest status
vstack manifest upgrade

# Or specify a path explicitly
vstack install --target /path/to/your/project

Common usage patterns

Repository-scoped install (recommended for teams):

# Move to your repository root and install there
cd /path/to/your/project
vstack install

# Or specify a path explicitly from any directory
vstack install --target /path/to/your/project

Profile-wide install (optional defaults for all projects):

vstack install --global

vstack install is the first-run command: it seeds .vstack/config.yaml in your project (never overwrites), then generates .github/ artifacts from templates. vstack init re-runs generation idempotently — safe to use in CI after upgrading vstack.

By default, vstack install preserves existing unmanaged files and local edits to tracked files by comparing the current file contents with the SHA-256 checksum recorded in .vstack/vstack.json. Use --adopt-name <name> to start tracking one existing unmanaged file without overwriting it. vstack uninstall also preserves locally modified tracked files unless you explicitly pass --force or --force-name <name>. Use vstack manifest status --target ... (or vstack status --target ...) to see what still matches the manifest. If a legacy manifest schema is detected, run vstack manifest upgrade --target ... first.

To skip artifact types or individual artifacts you do not need, edit .vstack/config.yaml:

exclude:
  skills:
    - terraform
    - helm
  instructions: all   # skip the entire type

If you already have agents, skills, or other files in .github/, run a dry-run first to see what would be preserved before committing:

# Run from your repository root
vstack install --dry-run

The summary lists preserved files as type/name selectors (e.g. agent/engineer). Resolve each conflict with --force-name type/name to overwrite, --adopt-name type/name to take ownership without overwriting, or --force to overwrite everything.

Reading .vstack/config.yaml

  • Lines starting with # are comments, explanation, or example configuration and are not active.
  • Only uncommented YAML keys are active configuration.
  • To enable an example block, remove # from that block and keep valid YAML indentation.
  • After any config change, run vstack init to apply it to generated .github/ artifacts.

Workflow modes

vstack supports three workflow modes via .vstack/config.yaml:

workflow:
  mode: agentic  # default

After changing workflow.mode, regenerate artifacts:

vstack init
Mode Behavior Planner file Worker handoff buttons
agentic Planner orchestrates stage progression using subagents generated omitted
manual User progresses stage-by-stage manually not generated shown
hybrid Both planner orchestration and manual handoffs are available generated shown

Execution semantics:

  • workflow.stages order is the canonical progression order.
  • agentic is stage-sequential by default: planner advances one stage at a time in configured order.
  • Set depends_on to unlock a DAG topology and let planner run independent branches in parallel.

Parallel stages with depends_on: By default each stage implicitly depends on the previous one. Add depends_on to declare explicit predecessors. The canonical vstack DAG (seeded automatically by vstack install):

workflow:
  mode: agentic
  version: 1
  stages:
    - role: product
      gate: required
      hitl: always
    - role: architect
      gate: required
      hitl: always
      depends_on: [product]
    - role: designer
      gate: optional
      hitl: on-change
      depends_on: [product]        # runs in parallel with architect
    - role: engineer
      gate: required
      hitl: always
      depends_on: [architect, designer]   # waits for both
    - role: tester
      gate: required
      hitl: always
      depends_on: [engineer]
    - role: release
      gate: required
      hitl: always
      depends_on: [tester]

depends_on: [] marks a root stage. Absent depends_on falls back to sequential. Circular dependencies are caught at install/init time.

Handoff target semantics:

  • handoffs.prompt is the transition prompt text.
  • If handoffs.agent is omitted, the target defaults to the next role in workflow.stages.
  • You can set handoffs.agent explicitly to override that default target in manual/hybrid.
  • In agentic, worker handoff buttons are hidden; planner controls progression.

Mode quickstart in Copilot Agent Mode:

In agentic mode, @planner is the primary entry point. Start every session with @planner and let it drive all stage transitions automatically.

Mode Start here First prompt example
agentic @planner @planner Run the workflow for this repository change.
manual @product @product Define requirements for this change.
hybrid @planner or @product @planner Run the workflow for this repository change.

What planner does: reads workflow.stages and depends_on, invokes each role agent as a subagent at the right time, runs independent branches in parallel when depends_on permits, pauses at each gate for human approval, and reports a structured stage outcome after each step. Valid role names: product, architect, designer, engineer, tester, release.

Agentic runbook (copy/paste):

@planner Run the workflow for this repository change.
@planner Show current stage status, ready stages, blocked stages, and next action.
@planner Continue with all ready stages in parallel where workflow.depends_on allows it.
@planner Pause at required HITL gates and ask for approval before advancing.
@planner Finalize with a release-readiness summary and list changed artifacts.

Troubleshooting: why planner is not running stages in parallel

  • Check workflow.mode in .vstack/config.yaml: parallel orchestration requires agentic (or planner-led hybrid).
  • Validate stage dependencies: run vstack validate to catch invalid roles, self-dependencies, and cycles.
  • Inspect depends_on shape: a stage runs only when all listed predecessors are ready or skipped.
  • Check implicit sequential fallback: if depends_on is omitted, the stage depends on the previous stage.
  • Check optional stage behavior: gate: optional can be skipped when unaffected, reducing apparent parallel fan-out.
  • Check blockers in stage report: any blocked predecessor prevents dependent stages from becoming ready.
  • Check mixed execution path: in hybrid, mixing manual handoffs and planner in one session can mask parallel readiness.
  • Confirm you started with planner: in agentic, begin with @planner, not a worker role agent.

Usage guidance:

  • Use agentic when you want one deterministic orchestration path.
  • Use manual when your team prefers explicit user-controlled stage transitions.
  • Use hybrid only when your team intentionally wants both options.

Hybrid operating rule:

  • Choose one path per session (planner-led or manual handoffs) and stay on it.
  • Mixing both paths in one session increases the chance of duplicate stage transitions.

Hybrid warning:

  • In hybrid, users can click handoff buttons while a planner-led flow is also available.
  • This can cause unintended progression jumps or duplicated transitions if your process assumes one strict path.

Fast troubleshooting

  • Command not found after install: ensure your pipx binary path is in PATH
  • Validation error: rerun vstack install from your repository root and then vstack validate
  • Agent results look generic: explicitly invoke a role (for example @tester) before a skill

Full documentation

For complete documentation (including architecture details, workflow diagrams, and contributor guides), use GitHub:

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

vstack-3.3.0.tar.gz (214.2 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

vstack-3.3.0-py3-none-any.whl (318.8 kB view details)

Uploaded Python 3

File details

Details for the file vstack-3.3.0.tar.gz.

File metadata

  • Download URL: vstack-3.3.0.tar.gz
  • Upload date:
  • Size: 214.2 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.13

File hashes

Hashes for vstack-3.3.0.tar.gz
Algorithm Hash digest
SHA256 e3ef932d875c3dfd005a7924385c29c24382b8385f2c139f580b946b6a63e828
MD5 f7927041fb20e5a14d3f284b3d209b26
BLAKE2b-256 e6362c31d49a9962a7a78c65abdf59a1089d8b9b274b2bf75d4f59f4183e7ff8

See more details on using hashes here.

Provenance

The following attestation bundles were made for vstack-3.3.0.tar.gz:

Publisher: publish.yml on eschaar/vstack

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file vstack-3.3.0-py3-none-any.whl.

File metadata

  • Download URL: vstack-3.3.0-py3-none-any.whl
  • Upload date:
  • Size: 318.8 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.13

File hashes

Hashes for vstack-3.3.0-py3-none-any.whl
Algorithm Hash digest
SHA256 28a827337474e47e9fd012b18344dc8a043b6f209cdc99a3699bfe013844bb7a
MD5 2be2cdb27ae41b6e7703ab7081532223
BLAKE2b-256 f982a9dc0c6f590249e2232fadf6d6c76ffeee4f5f859c64471b5627e5c7d3f8

See more details on using hashes here.

Provenance

The following attestation bundles were made for vstack-3.3.0-py3-none-any.whl:

Publisher: publish.yml on eschaar/vstack

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page