Skip to content

The Orchestrated Coding Spec

Itervox is the reference implementation of Orchestrated Coding — a vendor-neutral spec for building real software with fleets of coding agents. Where single-agent guides tell you how to build one reliable agent, Orchestrated Coding specifies the system layer above: one owner of context, isolated workers, durable handoffs, and deterministic gates.

Itervox is also a conforming OpenAI Symphony runtime — Symphony specifies a single orchestrator daemon’s engine (poll → workspace → run → retry); Orchestrated Coding is the fleet layer above it. Itervox implements both, which is the cleanest proof that the layers compose.

Chaining agents naively makes reliability worse, not better — errors compound instead of cancelling. The measured failure research behind the spec (the MAST taxonomy) found that roughly 79% of multi-agent failures are organizational — specification, handoff, and verification problems — not model-capability problems. The spec is those organizational fixes, written down: 3 Foundations and 7 Disciplines, with conformance levels L0–L3.

Spec itemHow Itervox implements it
F1 — source of truth as lawWORKFLOW.md (schema-versioned, hard-fails on mismatch) + committable SOUL.md/INSTRUCTIONS.md per profile
F2 — shared context protocol.itervox/handoff/ deliverables assembled into every next prompt; “done” requires the handoff — a clean exit without one gets a synthesized, explicitly-marked handoff, committed on the issue branch
F3 — deterministic gatesmerge-bot requires green gh pr checks --required; refuses unarmed gates (repos with zero required checks) unless explicitly opted out; optional per-unit gate via hooks.after_run_required
D1 — orchestrator owns dispatchsingle-goroutine event loop, bounded concurrency, pidfile single-authority guard
D2 — isolation per unitone git worktree per issue; open PRs are continued on their branch, never duplicated
D3 — handoffs as durable artifactstimestamped Markdown deliverables, partial-on-failure marking, budgeted chronological assembly
D4 — dependency-ordered releaseBlockedBy gating with strict fail-safe: any blocker fetch error (including 404) leaves the dependency unknown ⇒ blocked
D5 — review + merge gatereviewer profiles + auto_review; daemon-guarded merge_pr with required checks and block labels
D6 — multiple harnessesper-profile claude/codex backends; automatic switch on rate limits
D7 — control loopsten event triggers wired to agent profiles, with a durable, restart-surviving automation queue
Recoverybounded retries with backoff, stall detection, partial-work preservation, input_required human-in-the-loop resume, corrupt-state quarantine

See the Automations, Agent Profiles, and Dependency Management guides for the operational detail behind each row.

The spec’s rule is strict: a system may claim a level only if it satisfies every requirement at that level and below. Itervox conforms at L3 — every L1–L3 MUST verified against code with named regression tests run under Go’s race detector.

That claim is deliberately evidence-linked rather than asserted:

  • IMPLEMENTATIONS.md — the audited conformance map in the spec repo
  • v0.2.0/gaps_must.md — the MUST-level gaps the spec’s own audit found in Itervox, and the named tests that closed them

The honest part of that story: the audit caught real MUST-level gaps in the spec’s own reference implementation — fail-open dependency handling, best-effort-only handoffs, an unvalidated merge gate. They were closed with tests before the L3 badge went on. That’s the spec’s strict rule working as intended.

The spec is the point — Itervox is one way to build it. If you’re building an orchestrator (or already run one), self-assess against the conformance checklist and add your system to the list. An evidenced L1 is more useful to readers than an aspirational L3.