mirror of
https://github.com/boshu2/agentops.git
synced 2026-09-14 15:08:13 +08:00
59 lines
2.5 KiB
Markdown
59 lines
2.5 KiB
Markdown
|
# AgentOps 4.0: The Cathedral Cut
|
||
|
|
|
||
|
|
AgentOps 3.2 tried to make validation inescapable by owning the path around it.
|
||
|
|
That turned a useful trust boundary into a control plane: retries, queues, work
|
||
|
|
claims, Git admission, landing, closure, budgets, and delivery accumulated around
|
||
|
|
one fresh-context review. The result was correct-looking gridlock. Validation
|
||
|
|
often took longer than implementation, and moving `main` repeatedly invalidated
|
||
|
|
expensive proof.
|
||
|
|
|
||
|
|
Version 4.0 removes that control plane.
|
||
|
|
|
||
|
|
```text
|
||
|
|
RPI -> Plan -> Implement -> fresh Validate -> durable verdict -> report and stop
|
||
|
|
```
|
||
|
|
|
||
|
|
## What AgentOps owns
|
||
|
|
|
||
|
|
- one behavior with explicit normal and edge acceptance;
|
||
|
|
- one bounded implementation experiment;
|
||
|
|
- deterministic identity for the exact candidate content;
|
||
|
|
- one independent semantic judgment from fresh context;
|
||
|
|
- one durable, content-addressed verdict with checked and unchecked scope.
|
||
|
|
|
||
|
|
## What the caller owns
|
||
|
|
|
||
|
|
The caller decides whether to revise, retry, re-plan, schedule, assign, commit,
|
||
|
|
push, merge, release, or abandon the candidate. Existing repository Git and CI
|
||
|
|
policy remains authoritative. Optional councils, genies, NTM, Agent Mail, Gas
|
||
|
|
City, and swarm adapters may help, but none can change the core sequence or
|
||
|
|
strengthen a verdict.
|
||
|
|
|
||
|
|
## Skills
|
||
|
|
|
||
|
|
The load-bearing core is `rpi`, `plan`, `implement`, and `validate`. `learn` is an
|
||
|
|
optional later analysis of verdict collections. Discovery, behavior-first
|
||
|
|
planning, and goal design were folded into Plan. Crank and controller/delivery
|
||
|
|
skills were removed. Canonical mortem names are `premortem` and `postmortem`.
|
||
|
|
|
||
|
|
The remaining 50 skills are specialist strategies and optional adapters. Their
|
||
|
|
metadata is the source for the generated registry, router, graph, counts, and
|
||
|
|
runtime projections. The release audit reads every canonical `SKILL.md` and
|
||
|
|
runs the strict deep auditor over the entire corpus.
|
||
|
|
|
||
|
|
## CLI
|
||
|
|
|
||
|
|
The retained `ao` CLI supplies ordinary deterministic checks, inspection,
|
||
|
|
generic provenance records, goal measurement, session handoff, and skill
|
||
|
|
discovery/linking. Removed lifecycle commands have inert, non-mutating
|
||
|
|
tombstones for this release. Semantic validation lives in the Validate skill;
|
||
|
|
`ao gate check` proves deterministic repository facts only.
|
||
|
|
|
||
|
|
## Installation
|
||
|
|
|
||
|
|
New installs use one canonical checkout plus `ao skills link`. The command links
|
||
|
|
that checkout into the portable skill root and every detected runtime without
|
||
|
|
overwriting foreign skills. Plugin caches and copied mirrors are no longer the
|
||
|
|
recommended path. See [Install and day-2 operations](install-day2-ops.md) and
|
||
|
|
[the migration guide](MIGRATION.md).
|