boshu2__agentops/docs/one-session-mode.md

AgentOps — One-Session Mode (maximally-Claude)

Run the AgentOps factory loop in a single Claude session, no external infrastructure — the assured work-loop with nothing but Claude + a beads tracker (br) + git. This is the distilled core of AgentOps, and the native Claude golden image: the same laws scale out unchanged to the orchestrated factory (NTM swarms of Claude/Codex/AGY). It realizes the ag-eli0 distillation from the other direction — instead of carving orchestration out, rebuild the core up from one session and prove it runs assured.

Canonical reference implementation + proof: the control-plane bootstrap at ~/dev/control-plane — n=1 green, independently validated, zero Dolt, zero claude -p.

What it is

The control plane is the main Claude agent + br (SQLite + JSONL + git) + subagents. No daemon, no message bus, no Dolt. One operator, one session. Three roles, mapped to Claude-native primitives:

Role	Primitive
State store	br (SQLite + JSONL + git)
Orchestrator	the main agent + a tick.sh helper
Worker	an Agent subagent (fresh context), loads the AgentOps image (skills/corpus)
Validator	a separate-context Agent subagent (author ≠ judge)

The loop (one tick, idempotent)

next → claim → worker subagent → separate-context validator subagent → orchestrator closes on PASS (carrying an evidence ref) → br sync + git commit → repeat.

• Empty ready set = NO_READY (stop scheduling). No open/in-progress beads = CONVERGED.

The disciplines (what this encodes as standards)

• The seam = different context, not different vendor. Author ≠ judge holds in an all-Claude system because worker and validator are separate Agent invocations.
• Single writer. Only the orchestrator closes a bead; workers and validators propose.
• Evidence-gated, verified against reality. Validators execute/inspect the real artifact, not assert from a glance. Close only on VERDICT: PASS.
• Contested/unverified FAIL → fresh tie-break validator. Record both verdicts; the orchestrator never self-overrules. (Earned: a real false-FAIL was caught this way.)
• git is the durable ledger. Committed issues.jsonl is the provenance; br close --reason carries the evidence ref. No Dolt.
• No claude -p. In-harness subagents only (Max sub). Idempotent ticks — read state first.

The fractal (why this scales)

Native Claude image (this) → Codex image → AGY image → orchestrated NTM swarms managing swarms. Same contract at every scale; harden a discipline in the native image once and it propagates to the big factory. Native ⊂ orchestrated, same laws.

Quickstart

1. br init --prefix <p> in a git repo — the state store.
2. Map work to br beads with dependencies (the DAG).
3. Run the tick loop: orchestrator = main agent + tick helper; dispatch worker then a separate-context validator per bead; close on PASS with an evidence ref.
4. Every close cites evidence; the commit is the ledger.

See the reference implementation for the exact contracts: control-plane/{SOURCE-OF-TRUTH,ARCHITECTURE,ORCHESTRATOR,WORKER,VALIDATOR,LEARNINGS,N1-PROOF}.md + tick.sh + status.sh.

Relation to ag-eli0

ag-eli0 distills AgentOps to its core (the mind) by relocating orchestration to MTO. One-session mode is the same destination, validated — it rebuilds the core up from a single session and proves the assured loop runs with only Claude + br + git. The disciplines above are the standards that core enforces.