Provider Portability Begins With Repository-Owned Workflow

A repository-owned core connects to interchangeable provider modules.
In this article
  1. Separate the operating model from the host
  2. Portable does not mean identical
  3. Keep the safety floor outside the model
  4. Preserve durable state
  5. Keep context provider-neutral
  6. Evaluate providers against the workflow
  7. paqad-ai’s portability position
  8. Source notes

Last updated September 18, 2026

3 min read

AI coding providers change quickly.

Models improve. Pricing changes. Features appear in one tool and later in another. Teams adopt several tools at once
because different roles prefer different interfaces.

If the development process lives inside one assistant’s private prompt, switching providers means rebuilding the way
the team works.

Real portability begins in the repository.

Separate the operating model from the host

The team-owned operating model includes:

  • rules and architecture boundaries;
  • risk routing;
  • planning and specification fields;
  • acceptance evidence;
  • role definitions;
  • deterministic commands;
  • human decision categories;
  • workflow state;
  • documentation ownership;
  • audit and evidence formats.

These should exist independently of a provider.

The host-specific adapter should explain how one tool loads, executes, records, or enforces that model.

Portable does not mean identical

Providers have different control surfaces.

One may support lifecycle hooks before tool calls or file writes.

Another may load repository instructions and run commands but not block the same event.

Another may record state only after the action.

Some integrations are fully interactive. Others support background agents, worktrees, or scheduled runs.

An honest portability layer maintains a capability matrix:

  • session-start loading;
  • task and role invocation;
  • hook points;
  • blocking gates;
  • checkpoint support;
  • state persistence;
  • structured output;
  • permissions;
  • external safety gates.

The final receipt should distinguish enforced controls from advisory instructions.

Keep the safety floor outside the model

Important controls should not depend only on the provider following a sentence.

Put deterministic safety floors in:

  • repository scripts;
  • CI;
  • schemas;
  • protected branches;
  • registry policy;
  • test runners;
  • architecture linters;
  • approval and publication systems.

The agent can invoke these checks. The checks produce their own results.

This makes the workflow more portable and more trustworthy.

Preserve durable state

A provider switch during a feature should not require replaying a chat.

Store:

  • workflow identifier and active step;
  • approved plan and specification;
  • changed artifacts;
  • check outputs;
  • review findings;
  • pending decision;
  • human answers;
  • next action.

The next provider can read a structured handoff and continue within the same boundary.

Keep context provider-neutral

Write rules in repository language, not a tool’s marketing vocabulary.

Use small entry files for each provider that point to shared knowledge. Adapt commands and configuration where needed.
Do not duplicate the whole rule set into several files that drift separately.

Review provider-specific instructions after updates. The model may change, but the team still owns the expected process.

Evaluate providers against the workflow

Do not choose only by benchmark score.

Evaluate:

  • task success in your repository;
  • adherence to scope and rules;
  • quality of structured handoffs;
  • evidence fidelity;
  • permission controls;
  • review and correction cost;
  • context and token use;
  • recovery behaviour;
  • provider-specific gaps.

One provider can be better for a task and weaker for a control the risk lane requires.

paqad-ai’s portability position

paqad-ai’s current first-party documentation says it connects shared workflows and repository knowledge to several
coding-agent platforms. It generates thin provider-specific entry files and keeps shared rules, specifications,
workflows, decisions, and state in the project.

This should be treated as a product mechanism claim, not independent proof that every provider behaves equally.

The mature position is:

Own the workflow. Adapt to the host. Report the differences.

The team can change tools without losing the operating memory that makes its software theirs.

Install paqad-ai: https://github.com/Eliyce/paqad-ai/

Source notes

  • Eliyce/paqad-ai, first-party compatibility and product mechanism.
  • OpenAI, Harness engineering, first-party repository-owned knowledge
    and mechanical control account.
  • Supplied paqad feature-development explanation for provider enforcement distinctions.

Recognise this in your own team?

See how a change travels from request to live in one enforced process, then tell us about your team.