A workflow-first control plane for heterogeneous executable capabilities with eval-driven promotion and retirement. RCOS wires together three existing open-source systems — Pi (agent specs + execution), DSH (cognitive control + procedural compiler), and Archon (durable orchestration) — into one compounding loop. Nothing is vendored: Pi, DSH, and Archon are referenced as external dependencies you point at, never copies inside this repo.
Status: early public scaffold. The prototype/ reference implementation and
benchmarks/ Fam-R suite are landing incrementally; Fam-C / Fam-N and the paper
draft are staged TODOs with schemas defined. See paper/outline.md.
Single-shot agents plateau: they do not accumulate skill. RCOS treats durable
workflows as the primary abstraction and capabilities (reusable, versioned,
eval-gated units of behavior — prompts, scripts, model nodes, agent teams) as
the asset that compounds. A router selects capabilities per task, an evaluator
scores every execution, and a promotion gate admits a capability to the shared
registry only after repeated shipped results — retiring capabilities whose
performance decays. The claimed result is longitudinal: RCOS starts equal to or
worse than fixed harnesses, then overtakes them as experience accumulates. That
claim is tested by the compounding experiment in benchmarks/, not asserted here.
DSH General is the cognitive control surface; the DSH Workflow Manager plus the workflow-engineer role compiles intent into procedures; Archon DAGs execute those procedures durably; Pi AgentSpecs (plus code/model tiers) execute steps in isolation; scoped memory plus per-run artifact directories isolate state; and a registry plus promotion policy gates what becomes permanent. Full detail: ARCHITECTURE.md.
flowchart LR
U[User intent] --> G[DSH General<br/>cognitive control]
G --> WM[DSH Workflow Manager<br/>+ workflow-engineer<br/>procedural compiler]
WM --> DAG[Archon DAGs<br/>durable orchestration]
DAG --> EX[Execution tiers<br/>Pi AgentSpecs · code · model]
EX --> EV[Evaluators<br/>ship / fix / blocked]
EV -->|ship × N| REG[(Capability registry<br/>promotion gate)]
REG -->|reuse| WM
EV -->|decay| RET[Retirement]
EX --> MEM[(Scoped memory<br/>+ run artifact dirs)]
git clone https://github.com/Foshowithit/rcos.git
cd rcos
cp .env.example .env # fill in YOUR Pi/DSH/Archon checkout paths — no secrets
cat INSTALL.md # prereqs + first dry run (~5 min)Prereqs: git, gh, python3, and (where needed) node/pnpm. Full steps in
INSTALL.md. There is deliberately no pip install rcos yet — the
prototype is interfaces plus a pilot, not a package.
| Path | What lives there |
|---|---|
ARCHITECTURE.md |
Component map: control, compiler, orchestration, execution, memory, promotion |
INSTALL.md |
Prereqs, env-var wiring to your own Pi/DSH/Archon checkouts, dry-run smoke |
.env.example |
Placeholder-only paths/URLs. Never commit a filled .env |
prototype/ |
capability-registry.json schema, promotion-gate wiring, router reuse hook, per-task logger. Houses the Fam-R pilot when it lands |
benchmarks/ |
Fam-R: 15 file-ops / repo-scan / fix-loop tasks, each with prompt + checker + EVAL.json gate. Fam-C + Fam-N staged with schema |
paper/ |
Two-stage publication plan: v1 outline, conservative abstract, 12-section skeleton stubs |
.github/ |
PR template, bug + capability-proposal issue templates, minimal CI (markdown lint, JSON schema check, pilot smoke) |
CONTRIBUTING.md |
How to contribute, commit hygiene, security rules |
CODE_OF_CONDUCT.md |
Conduct-lite: be decent, no harassment, maintainers may remove |
- Fam-R (reproducible): 15 deterministic tasks — file ops, repo scans,
fix-loops — each with a prompt, a script checker, and an
EVAL.jsongate. Runnable today; see benchmarks/README.md. - Fam-C (compounding, staged): repeated-task families measuring reuse rate and cost-per-task decay over time. Schema defined, tasks TODO.
- Fam-N (novelty/generalization, staged): held-out tasks measuring transfer of promoted capabilities. Schema defined, tasks TODO.
cost-per-task ^
| fixed harness ───────────── (flat)
| RCOS ╲ ╲___
| ╲_______________________________ (decays as reuse compounds)
+-------------------------------------------------> tasks completed
The real figure is produced by running Fam-C after the pilot lands; until then
this is a hypothesis sketch, not a result. Generating it is tracked as a TODO in
benchmarks/.
Small, reviewable PRs. Security hygiene is a merge requirement: no keys, tokens, credentials, or vault contents — CI greps for them. See CONTRIBUTING.md.
Be decent. No harassment, no hate, no spam. Assume good faith; disagree with the work, not the person. Maintainers may warn, then remove, anyone who persistently degrades the commons. Full short text: CODE_OF_CONDUCT.md.
MIT. See LICENSE. This license choice is deliberate: MIT is permissive and compatible with the Pi / DSH / Archon open-source ecosystem this project builds on.
RCOS stands on existing open-source work and does not fork or vendor it:
- Pi — AgentSpecs, model/code tiers,
isolated execution. Referenced via
PI_CHECKOUTenv var, never copied. - DSH (DeepSeek Harness) — General cognitive control, Workflow Manager,
workflow-engineer / router / builder / evaluator / verifier roles, promotion
policy. Referenced via
DSH_CHECKOUT. - Archon — durable DAG orchestration, workflow library, eval gates.
Referenced via
ARCHON_CHECKOUT/ARCHON_WORKFLOWS.
Related ideas we learned from (independent projects, no affiliation):
FlowEvo, Voyager (skill-library compounding in Minecraft),
AWM / Agentic Workflow Memory. The RCOS novelty claim is narrower than any
of these: heterogeneous executable capabilities under a workflow-first control
plane with eval-driven promotion and retirement, tested by a longitudinal
overtaking experiment. Prior-art comparison belongs in the paper
(paper/outline.md § related work).