claudecode-orchestrator (deprecated, but)

github.com/darrenapfel/claudecode-orchestrator · DEPRECATED (replaced by limeriq.ai)

What it is (was)

Multi-agent orchestration framework that explicitly enabled 24-hour autonomous coding. Now deprecated; the principles still apply.

The principles we lifted

Quality through truth

Every claim requires evidence. Every test result must be shown. Validation failures are documented and fixed rather than hidden. The system enforces honest development practices.

We borrowed this verbatim into the validator’s EVIDENCE RULE: every PASS claim must be accompanied by quoted source output (test runner first+last line, HTTP response body, etc). Reports without evidence are “ghost reports” and revert to failing regardless of the claimed outcome.

Milestone Completion Protocol

Services are started and smoke-tested before declaring completion.

We borrowed this as our bin/service-smoke-test.sh script. Configurable via smokeTest.urls[]. Runs on DONE (gates the loop exit) and optionally on every feature pass.

What we dropped

• Auto-generated user testing materials. Their idea of “auto-generated testing guides from user stories” is interesting; we left it for documenter to handle.
• Structured human feedback loops. Their version of supervisor-notes; ours is simpler.

Differences in philosophy

Same philosophy, mostly. The differences are scope: claudecode-orchestrator was building a complete framework; we picked the two highest-leverage principles and ran with them.

Honest take

Even deprecated, this project’s principles aged well. “Quality through truth” is the single most important principle we adopted from anyone — it short-circuits the validator’s tendency to talk itself into approving plausible-looking output (a known Claude failure mode flagged by Anthropic in March 2026).

The next iteration of this project at limeriq.ai is presumably where these ideas continue.