Skip to content

Repository files navigation

AI-assisted Software Engineering Workflow Framework

Warning

Work in progress — pilot, not fully validated. Use with care, verify all outputs, and protect sensitive data. You are responsible for its use and resulting decisions or changes.

Provider-neutral, evidence-driven playbooks, roles, skills, and execution contracts for AI-assisted software engineering.

The framework turns a work item into a traceable workflow with investigation, design, implementation, independent review, validation, durable context, and an honest, human-readable handoff.

Every terminal handoff reports the workflow outcome—whether the selected graph completed—and the engineering outcome— what value the run delivered to the work item—as separate fields.

Core promise

AI-assisted engineering where every material conclusion is traceable from evidence to claim, decision, and action. Workers and provider adapters support that reasoning chain; they are not the product's central promise.

Introduction

This framework is a tool for engineers, not a replacement for them. The user sets the goal and scope, provides context, interprets results, adjusts the workflow, makes decisions, grants approval, and remains solely responsible for how the tool is used and for the resulting changes.

The project is an evolving pilot with known gaps and many opportunities for improvement. It favors quality over quantity: more workers, tokens, or effort do not automatically produce a better result. Role quality should remain consistent across execution profiles; profiles change the evidence and coordination depth, not the standard expected from a role.

Three further principles are essential:

  • AI output is evidence to assess, not authority to trust; uncertainty and unknowns must remain visible.
  • Human approval, independent review, and executable validation are control points, not optional ceremony.
  • Scope, permissions, privacy, and security remain explicit; the workflow must not expose sensitive data or make irreversible changes without authorization.

What it covers

Use it for work with meaningful uncertainty, dependencies, risk, or coordination needs, including:

  • Jira features, improvements, bugs, and TechOps issues;
  • Sentry production failures;
  • vulnerability and scanner findings; and
  • special workflows that need distinct stages or gates.

Do not use the full worker graph for a trivial, well-bounded change. Use the smallest role and skill set that provides enough evidence and validation.

Quick start

  1. Complete SETUP.md for the framework and execution repository; configure a provider runtime view when needed.
  2. Choose the playbook that matches the primary evidence and goal in PLAYBOOK_CATALOG.md.
  3. Copy the matching canonical run template from templates/ into a session started in the execution repository. Fill only the required first-run and scenario fields; do not invent a second prompt format.
  4. Start with planning for read-only discovery, diagnosis, design, and an implementation plan. Create the plan only after required fan-in passes.
  5. After explicit approval, run remediation through implementation, Code Review, validation, and handoff.

First-use view

The user-facing model is intentionally small:

Question Answer
What playbook do I use? Choose the most specialized playbook for the work item's primary evidence and goal in PLAYBOOK_CATALOG.md.
What do I provide? Work-item ID or URL, execution repository, lifecycle, profile, and relevant context.
What happens first? The run initializes or recovers the work record, activates the selected worker graph, and completes required fan-in before claiming success.
What may I approve? Scope and design approvals are conditional. Implementation approval is required before remediation. Release approval is required for deployment, cutover, or another external operational write. See the human control model.
Where do results go? The execution repository's .thoughts/<WORK-ITEM-ID>/work_record.md; implementation_plan.md and its optional portable handoff are created after planning reaches ready_for_implementation.

Users normally make only two execution choices:

Choice Meaning
Lifecycle planning is read-only; remediation may implement an explicitly approved plan.
Profile standard or deep; it selects the required worker graph and independent coverage.

The selected playbook and provider role policy derive workers, skills, tools, models, and reasoning effort. Actual model and effort values belong in the worker ledger, not in a first-use run prompt.

The execution repository's .codex/agents/ runtime view is optional. When it is absent, resolve provider definitions from the bundled framework/plugin or the selected work-graph binding. Versioned evaluations still require a resolved, reproducible provider configuration source; never inherit unverified Coordinator settings.

The shared rules are in contracts/workflow_execution.md. Vocabulary, portable-handoff rules, and non-normative execution guidance are linked from that core and loaded only when needed. The evidence-to-action reasoning model is contracts/claims.md; OPERATING_GUIDE.md explains normal operation.

Ask an AI agent to prepare a run prompt

If you are unsure which fields to provide, use the request and example in SETUP.md from a session started in the execution repository. The request explicitly supplies the absolute framework checkout path, execution repository, primary code repository when different, work-item ID or URL, selected playbook, and canonical template. It instructs the agent to fill the existing template, mark missing information as Unknown, and avoid executing or modifying the workflow.

Review the generated prompt before using it. The user remains responsible for the selected playbook, scope, lifecycle, profile, permissions, and approvals.

Architecture

User-facing flow

Work item
  -> choose playbook
    -> fill canonical run template
      -> planning
        -> approve implementation plan
          -> remediation
            -> Code Review, validation, and handoff

Execution flow

Coordinator
  -> worker graph
    -> roles + skills + tools
      -> provider adapter and model policy
        -> evidence, artifacts, validation, and work record

The execution flow is the framework's internal model. Users primarily interact with the user-facing flow and the canonical run template.

Building block Purpose
Framework Shared investigation-first engineering method
Contract Common lifecycle, worker, claims, evidence, decisions, gates, and handoff semantics
Strategy Coordination and parallelization approach
Role Reusable responsibility and reasoning boundary
Skill Reusable provider-neutral capability
Playbook Scenario-specific stages, dependencies, and outputs
Provider adapter Mapping to Codex or another execution platform
Work record Durable facts, decisions, errors, evidence, and next steps

Current playbooks

Playbook Use for
Technical Spike Bounded technical questions and existing Spike reviews
Feature Delivery Jira features and improvements
TechOps Issue Remediation Support- and operations-reported Jira issues
Sentry Issue Remediation Production issues backed by Sentry evidence
Vulnerability Investigation Scanner findings, advisories, CVEs, and security risk

Exercise state and worker graphs are maintained in PLAYBOOK_CATALOG.md. Add another playbook only when the existing stages, gates, and artifacts cannot express the scenario cleanly.

Guides and examples

Use this reading path:

  1. Setup — clone, configure, and start a run.
  2. Operating Guide — architecture, responsibilities, lifecycle, and usage rules.
  3. Playbook Catalog — choose a scenario and see its worker graph.
  4. Templates — start a run with the canonical prompt and create the durable work record.
  5. Examples — follow a safe, generic scenario guide for each current playbook.
  6. Contributing — extend the framework without duplicating contracts, roles, skills, or provider behavior.

Repository map

.agents/plugins/  Codex marketplace metadata
.codex-plugin/    Codex plugin manifest
contracts/       shared execution and reasoning semantics
examples/        safe, generic scenario guides
frameworks/      reusable engineering method
  experimental/  deferred, opt-in methods excluded from normal runs
integrations/    external evidence and work-item sources
playbooks/       scenario workflows
providers/       platform adapters and agent definitions
roles/           reusable responsibilities
scripts/         package preflight, run preparation, and deterministic framework validation
skills/          provider-neutral capabilities and the explicit Codex launcher
strategies/      coordination approaches
templates/       canonical prompts and durable work artifacts
tests/           redacted regression fixtures

The Codex plugin is a thin package over this repository; it does not duplicate framework behavior. See the launcher setup for package layout, installation, and update requirements; skills/run/SKILL.md for execution controls; and providers/codex.md plus the model and effort policy for provider behavior.

Quality and evolution

Run the framework validator after changes:

python3 scripts/validate_library.py
python3 scripts/validate_library.py /path/to/.thoughts/WORK-ITEM/work_record.md

The optional path validates a terminal work record's identity, playbook-selection evidence, repository revisions, and evidence-to-action references.

Document facts and limitations in the work record. Exercise changes against real work items, then simplify. See CONTRIBUTING.md and ROADMAP.md.

Evaluation and benchmarking are deliberately deferred from the normal workflow while the pilot stabilizes. Ordinary runs keep only core execution, provenance, outcome, and artifact controls. Add the evaluation addendum only when a prompt explicitly declares an evaluation or benchmark; it must not be used to retrofit telemetry that the run did not observe. See the experimental evaluation guide and evaluation_work_record_addendum.md.

Versioning

Versioned framework documents use independent semantic versions. Change a document's version when its contract or required behavior changes; related documents do not need matching versions. Git revisions identify the exact framework snapshot used by a run.

Status

This is an evolving pilot foundation, not a guarantee of correct code or production readiness. A playbook is not considered delivery-validated until its required implementation, Code Review, validation, fan-in, and runtime closure have been exercised successfully.

About

AI-assisted Software Engineering Workflow Framework

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Contributors

Languages