blumeops/docs/how-to/agent-change-process.md

title	modified	tags
Agent Change Process	2026-02-20	how-to ai

Agent Change Process

How to classify and execute infrastructure changes, especially when working with AI agents that may lose context across sessions.

Change Classification

Before starting work, classify the change:

Class	Scope	Process
C0	Quick fix, single-file, obvious	Read ai-docs, implement directly
C1	Moderate, potential hidden complexity	Mikado method, single session, single PR
C2	Complex, multi-session	Mikado method, documentation-driven, single PR

C0 — Quick Fix

1. Run mise run ai-docs to load context
2. Implement the change directly
3. Commit, push, create PR

Examples: fix a typo, bump a version, add a simple config value.

C1 — Guided Change (Single Session)

Use the Mikado method within a single session:

1. Run mise run ai-docs to load context
2. Attempt the change on a feature branch, amending a single commit as you iterate
3. If it works: push and create PR
4. If it fails: revert the broken change (git revert), then:
   • Amend or add a commit with documentation updates noting what prerequisite was discovered
   • Update frontmatter: add requires: [prerequisite-card] to the goal card
   • Work the leaf nodes (prerequisites with no further dependencies) first
   • Repeat until the goal succeeds

Single feature branch, squash-merge when complete. GitOps may require pushing to test — if a pushed commit breaks, revert it promptly.

C2 — Documented Change (Multi-Session)

Like C1 but designed to survive agent context loss across sessions:

1. Goal card: Create a how-to doc in docs/how-to/ describing the desired end state
   • Add status: active to frontmatter
2. Attempt the change — GitOps may require pushing code to test (e.g., ArgoCD sync). When the attempt fails:
   • First, reset the failed code changes (the branch should not carry broken code forward)
   • Then, create/update prerequisite cards as how-to docs with status: active
   • Add requires: [prerequisite-stem, ...] to the goal card's frontmatter
   • Commit only the doc updates (the documentation IS the Mikado graph)
3. Work leaf nodes first — cards with status: active and no unmet requires
4. Re-attempt the goal after leaf nodes are resolved — code from the attempt comes back here
5. New agent sessions pick up state by running mise run docs-mikado
6. When a card's change succeeds, remove status: active (or the entire field) from its frontmatter

Documentation IS the Mikado graph. Each card captures what was learned from failed attempts, so the next agent session doesn't repeat mistakes.

Handling failed attempts

When an attempt fails and you discover prerequisites, the branch must be cleaned up before documenting what you learned:

1. Reset to before the code attempt (git reset --hard)
2. Commit the new prerequisite cards and frontmatter updates
3. If you already committed docs mixed with code, cherry-pick the doc commits onto a clean base rather than reverting (avoids noisy add/revert history)

The branch between attempts should contain only documentation. Code returns when prerequisites are satisfied and the next attempt succeeds.

Build artifacts

Mikado resets apply to branch code, not build artifacts. Container images in the registry are independent of branch lifecycle:

• Registry images are build outputs cached in zot — tagged with commit SHAs, so each build is unique and traceable.
• Automatic builds trigger when container changes merge to main. Use mise run container-build-and-release for manual dispatch.
• If a build succeeds but deployment fails, the image is fine; the problem is elsewhere. Document what you learned and try again.
• If a build fails in CI, no image is pushed. Fix the nix/dockerfile and re-merge or re-dispatch.

Card Conventions

Frontmatter

---
title: Deploy Authentik
status: active          # omit when complete
requires:               # explicit dependencies
  - configure-postgres
  - setup-redis
tags:
  - how-to
---

• status: active marks in-progress work; remove when done (this is the ONLY way a card is marked complete)
• requires lists card stems (filenames without .md) that must be completed first. Keep requires permanently even after prerequisites are done — it documents the dependency graph history
• required-by is NOT stored — it's computed by docs-mikado

Writing Cards

• Mikado cards are not plans. Plans are designed upfront; Mikado cards are discovered through failed attempts. Don't put Mikado prerequisite cards in docs/how-to/plans/.
• Cards live in a topic subdirectory under docs/how-to/ (e.g., docs/how-to/authentik/ for the deploy-authentik chain). The goal card may live in plans/ if it started as a plan.
• Keep cards brief (<30 seconds to read)
• Link to other cards rather than inlining their content
• Document what was learned from failures, not just what to do

Git Discipline

• Single feature branch per C1/C2 change
• Create a PR early — open a draft PR after the first doc commit so the user can review the Mikado graph as it evolves between iterations.
• Push after every iteration — after completing a leaf node or documenting a failed attempt, push to origin. This is the save point for multi-session work.
• Amend a single working commit as you iterate; keep the branch history clean
• GitOps requires pushing to test — if a pushed commit breaks, revert it promptly
• Commit doc updates noting what was learned from failures

Tools

Command	Purpose
mise run docs-mikado	List all active Mikado chains
mise run docs-mikado <card>	Show dependency chain for a goal card
mise run docs-mikado <card> --all	Include completed cards in full

Related

• ai-assistance-guide — General AI agent conventions
• exploring-the-docs — Documentation structure overview