Evolution and lessons
This system was built by removing repeated friction rather than by starting with a large process framework.
Before → after
Evolution at a glance
Why GitHub became the center again
Project-management documents were useful for thinking, but they were not consistently consumed by the agents doing the real work.
GitHub already provided the main capabilities needed for durable engineering records:
- versioned source
- reviewable changes
- executable tests
- immutable commits
- CI evidence
- issue coordination
- release history
So the workflow stopped trying to maintain a second competing system of record.
Tela → Athena
A separate knowledge layer is still useful for rationale and cross-project history.
Tela was evaluated, but source-count constraints made it a poor fit for the intended scale. Athena is now used as the searchable knowledge layer.
The lesson is simple: the knowledge layer should help people and AI find context without becoming a second place where product state must be maintained.
Session continuity was the next bottleneck
Once engineering records and validation were clearer, the next major cost was moving context between tools:
- ChatGPT → Cursor prompt
- Cursor → ChatGPT result
- ChatGPT → new ChatGPT handoff
Engineering System 1.3.x introduced repository-scoped AI Work Packets and the Cursor /resume command to remove this repeated manual transfer.
Resume had to work before full adoption
A real release branch exposed an important edge case: active product work could have a valid Work Packet while Engineering System adoption was intentionally isolated in a different PR.
Treating missing local AGENTS.md or .engineering/* files as a hard resume failure would have mixed governance adoption into release work.
The resume flow was updated so a repository with adoption still pending can record ENGINEERING_SYSTEM_ADOPTION=ABSENT_OR_PENDING, continue under the current Engineering System rules, and leave adoption files untouched unless the active work explicitly includes them.
The lesson: continuity tooling must observe repository state without changing product scope merely to make itself work.
Mobility became an operational requirement
The workflow also had to work away from a desk.
A persistent development Linux server plus Cursor-native agent persist sessions made the implementation process independent of the physical client device. Cursor CLI can continue running on the server while the user reconnects from a phone over SSH and attaches to the persistent session.
Telegram completion messages removed another small but repeated cost: checking the terminal simply to learn whether a long task had finished.
Cursor persistence remains available as an optional compatibility path, but current default implementation stays in ChatGPT. Durable state belongs in the repository regardless of which worker is explicitly authorized.
Adoption itself became automation
By 1.3.x the lifecycle and session-continuity rules were strong, but applying the system to another repository still depended too much on an agent correctly interpreting templates.
Engineering System 1.4 added a canonical adoption standard and deterministic bootstrap.
The target behavior became:
This deliberately stops on ambiguous or destructive decisions instead of pretending that “automatic” means blind overwrite.
1.5 made ownership and operations explicit
Engineering System 1.5 tightened the places where an AI agent should not guess:
- adopted repositories pin an immutable canonical baseline and record whether shared or native CI owns affected validation
- native mode names the concrete workflow files that own that invariant
- merge-gate enforcement is recorded as verified, advisory, or unknown instead of being inferred from workflow existence
- production-oriented projects explicitly record runbook and incident-response posture
- material design-bearing changes route through a minimal design gate, while production-impacting failures route through the incident lifecycle
The goal is not more paperwork. It is to make hidden assumptions visible while keeping implementation context small.
1.5.1 made review feedback explicit
A real product PR exposed another gap: required CI could be green while a configured automated reviewer had already left an actionable finding in a non-blocking COMMENTED review.
Engineering System 1.5.1 therefore separates review state from review content. Before merge or terminal completion, actionable feedback must be fixed and revalidated or explicitly dispositioned with evidence. Advisory review status alone is never treated as proof that review is complete.
This keeps the gate lightweight: it does not require every comment to become an approval, only that actionable findings are not silently ignored.
Design principle
The system keeps evolving under one constraint:
Add automation only when it removes repeated manual work or materially improves correctness.
That rule keeps a solo-developer workflow from accumulating unnecessary process overhead.Last modified on September 29, 2026