DevFlow Toolkit

DevFlow is an installable command-line toolkit for managing specification-driven changes in existing software projects with the help of an AI coding agent.

It captures the project context, the project constitution and the approval history, so a specification can be reviewed and approved before anyone starts implementing it. DevFlow owns the workflow state; the selected agent harness is only the execution backend that drafts the specification.

This is the 0.3.17 MVP. It implements the discovery and specification stages only. It supports two specification contexts: Existing Project (the default) and Greenfield. It is not a complete SDD platform and does not include planning, implementation or review workflows.

Discovery decides whether a request is ready to become a specification. A specification is only generated when every decision is resolved or assumed, and every assumption has been explicitly accepted. Discovery is persisted, so an incomplete request can be resumed across sessions and agent harnesses. Discovery is also the requirement-gathering stage: ambiguity, authentication, permissions, moderation, data lifecycle, validation, ordering, administration, compatibility and out-of-scope decisions must be resolved there, not in spec.md.

The problem it solves

AI agents can produce useful specifications, but the result is often a throwaway chat response:

• it is not tied to the real codebase;
• it does not follow the project's own engineering rules;
• there is no record of who approved it or which exact version was approved;
• switching agents loses context and history.

DevFlow keeps the specification, the evidence used to write it, the constitution and the approval record inside the project. Approvals are bound to SHA-256 digests, and the workflow state can be rebuilt from an append-only event log.

Two specification contexts

DevFlow does not assume that a project already exists. The difference between the two modes is where the context comes from, not a permanent identity for the project:

Context	Source of context	Typical use
existing (default)	Analysed code, dependencies, conventions and the constitution, plus the declared architecture manifest when present	A change, fix, refactor or integration in a real codebase
greenfield	Requested requirements, a declared architecture manifest, stack, constraints and the constitution	The initial specification for a system that does not exist yet

• existing is the default and preserves the historical behaviour.
• greenfield never invents existing code. It does not scaffold a project, generate application code or implement tasks.
• Greenfield is a context for a specification execution, not a permanent project identity. After implementation exists, switch back to existing; the architecture manifest is kept and used as a constraint.
• DevFlow never switches modes silently. Use --project-mode explicitly when a run differs from the configured default.

Instruction precedence

Every generated context.md has two numbered sections: the Decision Context and the Project Evidence. The Decision Context is authoritative for decisions; the Project Evidence supports it. When instructions conflict, DevFlow applies this order, highest priority first:

1. Explicit decisions approved by the user.
2. The approved specification.
3. The constitution and the enabled technology profiles.
4. The declared architecture and verified project conventions.
5. General good practices.
6. Agent preferences.

A lower-priority instruction never overrides a higher-priority one. When two instructions at the same level conflict, the agent asks instead of resolving the conflict silently. The prompts state the output contract; DevFlow's Python code applies the enforcement.

Requirements

• Python 3.11 or newer.
• A POSIX operating system (Linux or macOS). Windows is not tested and is not supported in this release, because the workflow locks require fcntl.
• For agent-generated specifications: OpenCode or Google Antigravity (agy) installed and authenticated on PATH. OpenCode is verified end-to-end; Antigravity agent loading and print mode are verified locally.
• Git is not required. DevFlow does not initialize a repository, commit files or run destructive commands.

Installation

python -m venv .venv
source .venv/bin/activate
pip install devflow-toolkit

To install a specific built artifact instead:

pip install devflow_toolkit-0.3.17-py3-none-any.whl

Verify the installation:

devflow --help
devflow version

Initialization

Run devflow init inside the project you want to manage:

cd my-project
devflow init

DevFlow inspects the existing project, detects its stack and profiles, copies the applicable constitution documents, writes .devflow/config.yaml and generates the integration for the selected harness.

The stack is detected from pyproject.toml, requirements*.txt and package.json. When those manifests are absent or incomplete, DevFlow also reads the packages installed in a local virtual environment (.venv, venv or env). This keeps Django, FastAPI and the RAG/vector stack (langchain, llama-index, chromadb, qdrant-client, weaviate-client, pinecone-client, pgvector and sentence-transformers) detectable without versioning a dependency manifest.

Common options:

devflow init --harness opencode --mode non-interactive
devflow init --harness antigravity --mode interactive
devflow init --profile django-backend --name my-service
devflow init --project-mode greenfield
devflow init --language es
devflow init --no-agent
devflow init --yes

init --language <code> sets the language used for specification prose. The value is stored as documentation.language in .devflow/config.yaml and defaults to en. Common language names such as Spanish are normalized to their code (es). Required headings, section numbers, FR-XXX/AC-XXX identifiers, paths, commands and digests remain unchanged so structural validation still works.

init --project-mode greenfield creates .devflow/architecture.yaml from a template when it does not exist. It never overwrites an existing manifest. Edit the manifest before running a Greenfield specification; until then, spec fails with a clear structural error.

init never overwrites an existing constitution document, configuration file, specification, approval or unmanaged integration file. Running it again refreshes missing constitution files and managed integration files without destroying project work.

If an explicitly selected harness is not installed, init still records the selection but prints a warning. Subsequent spec runs fall back to deterministic static analysis until the executable is available.

When an interactive spec run asks you to choose a harness and none of the supported executables is installed, the command fails with a clear error instead of silently switching to deterministic analysis. Pass --harness none or --no-agent to opt into deterministic output explicitly.

Configuration

.devflow/config.yaml stores project settings. It references the constitution documents; it does not duplicate their contents.

project:
  name: my-service
  type: backend
  mode: existing
profiles:
  - django-backend
constitution:
  base: true
  profile_specific: true
  path: .devflow/constitution
  files:
    - constitution.md
    - code-standards.md
    - architecture.md
    - django-standards.md
specifications:
  path: .devflow/specs
plans:
  path: .devflow/plans
templates:
  path: .devflow/templates
agent:
  harness: opencode
  execution:
    mode: non-interactive
  permissions:
    allow_project_read: true
    allow_workflow_write: true
    allow_source_code_changes: false

constitution.files accepts plain file names inside the constitution directory. Absolute paths, .. and nested paths are rejected.

Existing configurations without project.mode load as existing.

Declared architecture

The declared architecture lives in .devflow/architecture.yaml, not in config.yaml. It is a reviewable, versionable design decision. Example:

version: 1
style: clean-architecture
summary: >
  REST API for contact management with domain separation.
modules:
  - name: contacts
    responsibility: Contact registration and queries
    depends_on: [shared]
  - name: shared
    responsibility: Cross-cutting utilities
    depends_on: []
layers:
  - name: domain
    may_depend_on: []
  - name: application
    may_depend_on: [domain]
  - name: infrastructure
    may_depend_on: [domain, application]
  - name: presentation
    may_depend_on: [application]
stack:
  language: python
  framework: django-rest-framework
  database: postgresql
constraints:
  containerization: docker
  authentication: jwt
  testing: pytest
quality_attributes:
  - testability
  - modularity
fitness_functions:
  - domain must not import infrastructure

style is free text. Clean Architecture, Hexagonal Architecture, Microservices, Modular Monolith and Event-Driven are examples and presets, not a closed list. Custom architectures are valid as long as they are described in the manifest.

DevFlow validates structure and internal consistency only:

• required fields and non-empty values;
• duplicate module or layer names;
• references to unknown modules or layers;
• dependency cycles, reported as warnings;
• invalid YAML, unsafe paths, symlinks and oversized manifests.

It does not certify architectural quality. Contradictions and open decisions must be recorded as specification questions.

init may create the template. spec only reads the manifest and never modifies it. The manifest digest is recorded in context.md and in spec.md, including an appendix with the declared manifest, so each approval identifies the architecture version used. Greenfield specifications require the manifest; existing-mode specifications use it as an architectural constraint when it is present.

Harnesses

Harness	Status in 0.3.17
OpenCode	Verified end-to-end; interactive Discovery uses the question tool
Antigravity	Custom agents and scoped permissions generated; agy loading and print mode verified
Codex	Integration files are generated, but not verified for this release
Claude Code	Integration files are generated, but not verified for this release
Kiro	Integration files are generated, but not verified for this release
none	Deterministic static analysis only; no agent is invoked

Do not treat an unverified adapter as a compatibility guarantee. The generated files and command lines for Codex, Claude Code and Kiro may change as those integrations are tested.

For interactive Codex runs, DevFlow enables Codex's experimental request_user_input tool so Discovery questions can be answered by selecting an option instead of typing a free-form reply. Claude Code uses its native AskUserQuestion dialog. Both adapters fall back to numbered text options if the installed harness does not expose a structured question tool.

OpenCode and Antigravity receive an explicit instruction to call their structured question tool for every pending Discovery question and for every assumption acceptance. OpenCode's tool is question; Antigravity's tool is ask_question. The instruction is injected only for those harnesses, so the Codex and Claude Code prompts are unchanged.

For Antigravity, DevFlow generates AGENTS.md and custom agents under .agents/agents/. Interactive runs use the writable devflow-spec agent with agy --settings <json> --agent devflow-spec --prompt-interactive <prompt>. Non-interactive runs use the read-only devflow-spec-readonly agent with agy --settings <json> --agent devflow-spec-readonly --print --output-format text <prompt>. Project reads are allowed, writes are scoped to .devflow/specs in interactive mode, and commands, web access and MCP servers are denied. The alias agy is accepted for --harness.

To configure Discovery in Spanish, pass --language es to init, or to a single spec run. The conversation and the stored question text use the selected language.

For OpenCode, DevFlow generates AGENTS.md, opencode.json and the agents under .opencode/agents/. The project-level opencode.json can also select the model used by the agent, for example:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "provider/model-name"
}

DevFlow creates opencode.json only when it does not exist and never replaces an existing OpenCode configuration.

DevFlow records the files it generates for the selected harness in a managed block inside .gitignore, for example CLAUDE.md, AGENTS.md, opencode.json and the DevFlow agents. Existing ignore entries are preserved, and the block is shared across harnesses, so switching adapters keeps earlier generated files ignored. Unmanaged files that DevFlow did not create are never added. DevFlow does not ignore .devflow/ or the constitution.

Permissions

allow_workflow_write controls writes to DevFlow artifacts. allow_spec_write is accepted as a compatibility alias. allow_source_code_changes is disabled by default.

DevFlow translates these settings into technical controls where the harness exposes them. For OpenCode, the generated agents deny edit outside .devflow/specs/, deny shell execution unless source changes are explicitly allowed, and disable web access and subagents. Prompt text is not treated as the only permission boundary.

When a harness cannot enforce a restriction technically, DevFlow prints a warning instead of claiming that the restriction is guaranteed.

For Claude Code, interactive runs use Manual mode (--permission-mode manual) so Discovery can open the AskUserQuestion dialog. Reads and writes under .devflow/specs/ are pre-approved, while shell, web and task tools remain denied. Non-interactive runs keep the locked dontAsk mode and do not allow the question tool.

Discovery

Discovery is separate from specification. The project context decides where evidence comes from; Discovery decides whether the request is ready to become a specification.

Every new devflow spec request creates a persistent record at .devflow/specs/<id>-<slug>/discovery.yaml. It is plain YAML and can be reviewed, edited and committed:

version: 1
spec_id: "001"
change: "001-add-a-contact-form"
request: "Add a contact form"
project_mode: existing
status: awaiting_clarification
created_at: "2026-11-02T10:00:00Z"
updated_at: "2026-11-02T10:00:00Z"
questions:
  - id: Q1
    category: functional_scope
    materiality: functional
    question: What should happen when the contact form is submitted?
    reason: Persistence, confirmation and email delivery change scope.
    blocking: true
    status: pending
    options: []
    answer: null
    answered_by: null
    answered_at: null
  - id: Q2
    category: architecture_boundary
    materiality: technical
    question: Which form mechanism should the page use?
    reason: The project already uses Django Forms.
    blocking: false
    status: resolved
    resolution_source: evidence
    evidence:
      - "profiles/django-backend"
      - "src/contact/forms.py:12"
    answer: Reuse Django Forms.
    options: []
assumptions:
  - id: A1
    question_id: Q3
    value: Use `/contacto/` as the URL path.
    reason: Internal routing must be visible before approval.
    source: agent_default
    status: pending_acceptance
    accepted_by: null
    accepted_at: null

Every decision has one outcome: resolved, assumed or pending. A question presents a pending decision; an assumption is the acceptance ledger for an assumed decision. Each decision has a category, a materiality (critical, functional, technical or cosmetic) and an explicit blocking flag.

• pending is valid only for blocking or material decisions.
• resolved by the user records answer, answered_by and answered_at.
• resolved by evidence records resolution_source: evidence and at least one evidence entry.
• assumed requires a linked assumption with explicit human acceptance. Critical and security decisions can never be assumed.
• answered is a legacy alias that DevFlow reads as resolved.

A specification may only be generated when no decision is pending and every assumption has been explicitly accepted. DevFlow validates the record and confirms completeness; the agent never declares Discovery complete on its own.

Discovery uses a per-round question budget and batches independent decisions into one interaction. The defaults live in .devflow/config.yaml:

discovery:
  question_budget:
    initial: 6
    follow_up: 3
    maximum: 12
  interaction:
    batch_size: 6
    max_rounds: 3
  assumptions:
    enabled: true

The budget limits questions but never permits ignoring a critical decision. If material decisions remain after the budget is exhausted, Discovery stays incomplete and asks for intervention.

The deterministic planner recognizes contact pages, forms, comments, authentication, persistence, deployments and architecture boundaries in English and Spanish. Generic data vocabulary such as datos, data or database does not by itself prove that the feature scope is understood: when it is the only match, the planner adds a generic blocking scope question instead of marking Discovery complete. Interactive runs always produce at least one question when no specific question matches, so the selected harness is invoked instead of being skipped with an empty Discovery. Non-interactive runs keep the deterministic behavior for concrete requests such as Add a health endpoint.

For user-generated content, Discovery must also resolve who writes it, who may create, edit or delete it, moderation, authentication mechanism, validation, ordering, lifecycle on parent deletion, admin registration, compatibility and explicit out-of-scope items.

An assumption proposed by DevFlow uses source: agent_default and starts as pending_acceptance. It never becomes accepted automatically. A human must set status: accepted with accepted_by and accepted_at, or reject it and record a different decision.

While any decision is pending or any assumption remains unaccepted:

• discovery.yaml and context.md are written;
• no spec.md is generated;
• devflow approve --stage spec refuses the change;
• the workflow status is awaiting_clarification.

context.md is the single generated context file. Its ## 1. Decision Context section is generated deterministically by DevFlow from the project facts, conventions, approved decisions, evidence, pending decisions and constraints. Its ## 2. Project Evidence section contains the request, static analysis, relevant code, constitution scope and declared architecture. The file is capped and carries the Discovery and architecture digests, so a new session does not need to rediscover the whole project.

Answer the questions in the file, or run an interactive harness, then resume:

devflow spec --continue 006

--continue never regenerates or overwrites an approved specification. It reuses a current, valid specification when its Discovery digest still matches. If an answer changes before approval, DevFlow marks the old specification stale, preserves it under history/, and generates a new specification bound to the new Discovery digest.

The deterministic --no-agent path follows the same rules. It generates Discovery, records it, and stops when any question or assumption remains unresolved. It never invents an answer to continue.

Changes created with 0.2.0 that have no discovery.yaml remain approvable as legacy changes, with an explicit warning. New changes require Discovery.

A generated specification is valid only when section 7 is exactly None.. If the specification agent discovers a decision that Discovery did not cover, DevFlow rejects the generated document, archives it, appends the question to discovery.yaml, and returns the change to awaiting_clarification. The question must be answered before a new specification is generated. devflow approve --stage spec rejects a specification that still lists open questions.

Workflow

1. Initialize the project with devflow init.
2. Create a change and run Discovery.
3. Answer every question and accept or reject assumptions.
4. Generate the specification with devflow spec --continue <id>.
5. Review spec.md. It must have None. in section 7; unresolved questions belong in Discovery.
6. Record the approval.
7. Inspect the state, artifacts and approval digests.
8. If artifacts exist without workflow state, diagnose with devflow doctor <id> and repair with devflow recover <id>.

Available commands in 0.3.17:

Command	Status
devflow init	Implemented
devflow spec	Implemented: Discovery, specification, and --continue
devflow approve <id> --stage spec	Implemented
devflow status <id>	Implemented
devflow doctor <id>	Implemented: read-only diagnostics
devflow recover <id>	Implemented: register existing artifacts
devflow version	Implemented
devflow plan	Not implemented
devflow approve <id> --stage plan	Rejected
devflow implement	Not implemented
devflow review	Not implemented
devflow approve <id> --stage review	Rejected

There is no separate execute command. approve is a human authorization and never changes code. When the implement stage is added, it will be the technical execution command, and it will remain separate from approval.

Start, resume and generate a specification

Start a new change:

devflow spec "Add a GET /patients endpoint that returns JSON"

The command:

1. loads and validates the configuration;
2. loads the configured profiles and constitution;
3. builds the context for the effective project mode (existing by default, or greenfield);
4. creates a persistent discovery.yaml with classified questions and proposed assumptions;
5. writes the Decision Context and the project evidence or declared architecture to context.md;
6. stops without spec.md when any question or assumption remains unresolved;
7. in an interactive run, stops after Discovery is complete and prints the next command, so you can choose the harness for the specification;
8. on --continue (or with --auto) invokes the selected harness with the specification prompt, validates the returned specification, and binds it to the Discovery digest;
9. writes spec.md and registers the Discovery and specification digests in the event log and the workflow projection.

An interactive devflow spec therefore uses two harness invocations: one for Discovery and one for the specification. The first run stops at the Discovery checkpoint. The second run resumes with the harness you choose:

devflow spec "Add a GET /patients endpoint"
devflow spec --continue 006 . --harness opencode

When no --harness is passed, an interactive devflow spec asks which installed harness to use before starting, even if .devflow/config.yaml already names one. The configured harness is offered as the default. Pass --harness <name> to skip the prompt, or --no-agent to use deterministic static analysis.

Use --auto to run both phases in a single invocation, as earlier releases did:

devflow spec "Add a GET /patients endpoint" --auto

Resume an existing change:

devflow spec --continue 006

--continue is not a new specification. It reuses the stored request and project mode, loads Discovery, and stops until no decision is pending and every assumption is accepted. It reuses a current validated specification when the Discovery digest still matches; otherwise it archives the old specification and generates a new one. If an interrupted run left a complete discovery.yaml without workflow.yaml, --continue creates the missing state and registers Discovery before generating the specification.

For an existing project, DevFlow inspects relevant code and dependencies. For Greenfield, it uses the declared architecture and does not search for code or invent code references.

Useful options:

devflow spec "Description" --harness opencode --mode non-interactive
devflow spec "Description" --auto
devflow spec "Description" --no-agent
devflow spec "Description" --language es
devflow spec "Change in a real project" --project-mode existing
devflow spec "Create a new service" --project-mode greenfield
devflow spec --continue 006 --harness opencode

--auto keeps the older one-shot flow: Discovery and the specification run in the same invocation. Without it, an interactive run stops at the Discovery checkpoint.

spec --language <code> overrides the configured language for that run without changing .devflow/config.yaml. Interactive agent Discovery asks existing questions and writes new questions in that language, even when the stored question text was generated in English by the deterministic planner; the stored question, reason and option text is translated in place. Specifications written by an agent use the same language. The deterministic --no-agent planner still writes its own stored question text in English.

If the harness is unavailable, fails, or returns an invalid document, DevFlow writes a deterministic draft and reports the reason. The command remains usable in CI and on machines without an AI harness, but the draft still requires human review. A request that is not ready never receives a silent deterministic specification: Discovery stops instead.

Check status

devflow status 001
devflow status 001-add-a-get-patients-endpoint

status shows the workflow state, registered artifacts and digests, approvals, stale artifacts and orphan files. It reconciles workflow.yaml from events.jsonl before reporting.

Diagnose and recover a change

If context.md, discovery.yaml or spec.md were written outside devflow spec, the change can exist on disk without workflow.yaml, events.jsonl or runs/. In that case status reports that no workflow state was found.

Diagnose the change without changing anything:

devflow doctor 001

doctor checks the workflow, Discovery and specification, reports orphan artifacts, and exits non-zero when the change is unhealthy.

If Discovery is complete but the specification is missing (for example, an interrupted interactive run), doctor prints the resume command:

devflow spec --continue 001

--continue creates the missing workflow state, registers the existing Discovery digest and generates the specification without repeating Discovery.

If discovery.yaml is complete and spec.md is structurally valid, register the existing artifacts without regenerating them:

devflow recover 001

recover creates the missing workflow state and registers Discovery and the specification with source: existing. It preserves the artifact bytes and their digests, does not modify spec.md, and does not approve the change. Running it again after recovery is a no-op.

Recovery does not bypass review. After recovering, inspect devflow status 001, then approve explicitly:

devflow approve 001 --stage spec

If Discovery is incomplete or spec.md is invalid, recover refuses to register the artifacts and names the problem. Resolve it with devflow spec --continue 001 first.

Approve the specification

devflow approve 001 --stage spec

Approval is explicit and binds the decision to the current artifact digests. For a 0.3.17 change it binds spec.md, discovery.yaml and, when a manifest applies, the architecture manifest digest. Repeating the same approval is a no-op. If an artifact changes after validation, DevFlow marks it stale and refuses approval until it is regenerated and validated again.

Discovery must be complete before approval. If it is not, approval fails and names the pending questions and assumptions. Approval also rejects a specification whose section 7 still lists open questions. Legacy 0.2.0 changes without discovery.yaml remain approvable with a warning.

Only the spec stage is available in 0.3.17. --stage plan and --stage review are rejected. There are no plan, implement or review commands in this release.

Generated files

my-project/
├── .devflow/
│   ├── config.yaml
│   ├── architecture.yaml           # declared architecture (Greenfield)
│   ├── .specs.lock                 # serializes specification id assignment
│   ├── constitution/
│   │   ├── constitution.md
│   │   ├── code-standards.md
│   │   ├── architecture.md
│   │   └── django-standards.md     # only when the profile is enabled
│   └── specs/
│       └── 001-add-patients-endpoint/
│           ├── discovery.yaml      # persistent questions and answers
│           ├── context.md          # Decision Context + Project Evidence
│           ├── spec.md             # only when Discovery is complete
│           ├── workflow.yaml       # projection of events.jsonl
│           ├── events.jsonl        # append-only source of truth
│           └── .lock               # per-change advisory lock
├── AGENTS.md
├── opencode.json                   # when OpenCode is selected
└── .opencode/agents/               # when OpenCode is selected

All managed files are identifiable: Markdown files contain <!-- devflow:managed -->, and JSON files contain "devflowManaged": true. Unmanaged files with the same names are left untouched and reported.

How approvals and history work

events.jsonl is the source of truth. workflow.yaml is only a projection: if it is deleted, lags behind or disagrees with the event log, DevFlow rebuilds or rejects it rather than trusting it.

generated -> validated -> approved
     |            |
     +-> rejected +-> stale

• An approval stores the artifact kind and its SHA-256 digest.
• Editing an approved artifact makes it stale.
• A stale, unvalidated or orphan artifact cannot be approved.
• Semantic validation rejects malformed or inconsistent histories.
• Artifact and state writes are atomic; a failed write does not leave a partially written managed file.
• Writes are restricted to the project and the configured directories. Symlinks that escape the project are rejected or ignored.
• Specification directory allocation is serialized with a lock, so two concurrent runs do not receive the same numeric id.

The event log is local and append-only, not cryptographically signed. A user with full filesystem access can modify it; DevFlow detects structural and semantic inconsistency, not deliberate tampering by an administrator.

Updating DevFlow

pip install --upgrade devflow-toolkit
devflow version

Then run devflow init again in each project. The update:

• restores missing constitution documents;
• refreshes managed integration files;
• preserves existing constitution documents, specifications, approvals and event logs.

Review .devflow/config.yaml after a major upgrade, because new options may have been added. DevFlow does not silently rewrite user-modified configuration.

Known limitations

• Only the specification stage is implemented. Planning, implementation and review are not available.
• Greenfield mode produces a specification from a declared architecture. It does not scaffold a project, generate application code, create Docker or cloud infrastructure, or implement tasks.
• DevFlow validates the structure of the architecture manifest. It does not certify that the architecture is correct, complete or appropriate.
• Only OpenCode has been verified end-to-end. Codex, Claude Code and Kiro adapters are unverified scaffolds for future releases.
• Interactive harness mode exists, but the verified release path is OpenCode non-interactive mode. Interactive mode requires a terminal.
• Without an available harness, output is produced by deterministic static analysis. It is a starting draft, not a substitute for human review.
• Static analysis is heuristic: it finds likely affected files, not a complete impact map.
• DevFlow is POSIX-only in this release.
• The package does not provide a hosted service, telemetry or remote synchronization. All state stays in the project.

Development

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pytest
ruff check
python -m build
python -m twine check dist/*

The test suite covers the core, CLI, state replay and harness adapters. Release verification additionally installs the built wheel and sdist in clean virtual environments and runs the consumer workflow from outside the repository.