v1.1.0-rc.1 — the original AED canon
Pre-release
Release candidate. main still carries v1.0.0 — pin to this tag, not to a branch.
What this is
v1.0.0 published six edit-level style rules. That was a narrow slice: it described how a single line should read and contained none of the doctrine AED is actually named for.
This release publishes that doctrine for the first time. The author's original notes — written 2024-06-03, with process_manager_conventions.md revised 2025-11-24 — are here verbatim, including his own work-in-progress markers and one section still marked TBD. Nothing was rewritten or smoothed over.
Reading order
| # | Chapter | What it settles |
|---|---|---|
| 01 | Why models need this | Token windows — why naming carries so much weight for a model |
| 02 | Naming conventions | list_of_, boolean-as-question, attributes as short statements |
| 03 | Process managers | The "when" grammar |
| 04 | Feature stories | Personas, operations, authorization levels |
| 05 | Edit-level style (CONVENTIONS.md) |
The six v1.0 rules |
| 06 | Control flow | CF-1…CF-11 |
| 07 | How the workflow runs | Plan in batches, let the agent run, walk away |
Plus quick_reference.md, runnable examples/, and evidence/.
New in this release
- CF-1 … CF-11 — control flow whose syntax can't read like a statement (
case, loops, guards,unless, ternaries, chains, rescues, compound conditions, fibers,?/!, macros). Built against a census of two live production codebases, so the rules target real frequency rather than theory. Thresholds are still open — the questions are listed at the end of the chapter. - Evidence — a small-model comprehension benchmark: AED-style Crystal scored 60/60 against 54/60 for conventional compressed style, blind-answered by Claude Haiku and blind-graded. The report states its own limits up front (n=10, one small model, single run, model grader, and the pairs were authored by the same party that authored the conventions). It is a directional signal, not proof.
- Single-file bundle — the whole canon in one file, for one-request fetches and context windows.
Getting it
# everything, one file
curl -sL https://raw.githubusercontent.com/AgentC-Consulting/aed-conventions/v1.1.0-rc.1/dist/aed-v1.1.0-rc.1.md
# one chapter
curl -sL https://raw.githubusercontent.com/AgentC-Consulting/aed-conventions/v1.1.0-rc.1/02_naming_conventions.md
# the whole repo
git clone --branch v1.1.0-rc.1 --depth 1 https://github.com/AgentC-Consulting/aed-conventions.git
Also mirrored as a Hugging Face dataset: https://huggingface.co/datasets/agentc-consulting/aed-conventions
Compatibility
CONVENTIONS.md keeps its filename and its #rule-1 … #rule-6 anchors. They were published in v1.0.0 and are cited from outside this repository, so nothing that already links to them breaks.
Why a candidate, not a final
Two things are genuinely unsettled: the CF thresholds above, and CONVENTIONS.md rule 4, which offers expected_state as an improved name when chapter 02 considers it still under-specified. Chapter 02 is the authority; reconciling the example is tracked for final.
Versioning
This stays on the v1 line. The question was raised at publication — an expansion this large arguably justifies an honest v2.0.0, since it changes what the repository is rather than adding to it. Settled 2026-07-29: it is a big expansion, but it belongs to the same frontier foundation v1.0.0 started, filled in rather than replaced. The final tag on this work will be v1.1.0.
v2.0.0 is reserved for the next generation of improvements — thinking that supersedes this foundation rather than completing it. Read a major bump here as a genuine change of generation, not merely a big release.
Arguments welcome in Discussions.
Verify the signature: gpg --import agentc-signing-pubkey.asc && git tag -v v1.1.0-rc.1