Skip to main content
Origin
13 May 2026Docs · 01

← Docs

Concepts

What's actually in the box.

devrel-origin is a Python CLI that operates on your project repo the way git does. devrel init scaffolds a .devrel/ directory; from then on, every command reads and writes inside it.

§1

The .devrel/ directory.

Eight files and folders. Four are version-controlled (the editorial contract + config). Three are runtime (kb, deliverables, state.db). One is local secrets (.env, gitignored).

  • .devrel/config.tomlProduct identity (name, URL, github_repo), model routing per agent, budget caps.
  • .devrel/voice.mdThree to five short samples of your best published content. The editorial north star.
  • .devrel/style.mdReadability targets, structural rules, header conventions, code-block voice.
  • .devrel/slop-blocklist.mdPhrases that must never appear. The Kai pipeline checks every revision against this.
  • .devrel/kb/Markdown harvested from your docs site, README, or pasted by hand. Source of grounded evidence.
  • .devrel/deliverables/Generated content lands here, partitioned by ISO week. Each piece ships with a trace JSON.
  • .devrel/state.dbSQLite. Run history, OKR rollups, schedule registrations, agent state.
  • .devrel/.envANTHROPIC_API_KEY or OPENROUTER_API_KEY. chmod 600. Written by devrel auth.
§2

Fifteen agents, three pillars.

Hub-and-spoke. Atlas orchestrates; fourteen specialists execute. Each agent is purpose-built for one job, has its own prompt under prompts/, and persists structured output to state.db so downstream agents have receipts.

Health pillar

  • AtlasOrchestrator. Brokers context, surfaces decisions, reports OKRs.
  • WatchdogPre-flight system health, budget, integration status.
  • SentinelPre-publish cross-agent brand and ICP audit.
  • ArgusPost-publish content performance analyst.

DevRel pillar

  • SageCommunity manager. GitHub issue triage + churn risk.
  • EchoSocial listener. Reddit, HN, X mentions + sentiment.
  • IrisFeedback synthesizer. Chunked theme extraction.
  • NovaGrowth strategist. Experiments, funnels, cohorts.
  • CyraCRO auditor. PostHog funnel drop-off + A/B hypotheses.
  • KaiContent creator. Revision-looped tutorials.
  • VoxVideo producer. Screen-recorded tutorials + TTS.
  • DexDocumentation generator. AST-based reference + arch docs.

Sales pillar

  • RexCompetitive intel. Parallel search + Apollo enrichment.
  • PaxSales enablement. Outreach sequences, battle cards.
  • MoxCampaign marketing. Cross-channel content packages.
§3

Seven editorial stages.

Every piece of content Kai produces passes through this pipeline before it lands in deliverables/. Each stage has explicit pass/fail criteria; failures trigger a force-rewrite loop, not a soft skip. The full revision trace is persisted next to the deliverable as <slug>-trace.json.

  1. 01Developmental edit — argument shape, missing context, scope drift.
  2. 02Line edit — sentence cadence, transitions, redundancy.
  3. 03Copy edit — grammar, mechanics, fact-checks against source.
  4. 04Anti-slop — removes hedges, AI tics, and cliché phrasing.
  5. 05Reader-persona test — does it land for the stated audience?
  6. 06Readability check — Flesch-Kincaid + structural scan.
  7. 07Brand audit — voice, messaging, ICP alignment against the editorial contract.
§4

Repo-grounded generation.

Kai refuses to generate when no KB, official-docs, or repository evidence is available, returning status="insufficient_evidence" with an evidence-gaps list instead of producing ungrounded content. The forbidden-claims list bans inventing SDK methods, endpoints, install commands, file paths, benchmarks, or issue numbers.

Code blocks get validated. Bash, sh, zsh, Python, JS/TS, YAML syntax-check via parser; failures surface in the trace JSON and block ship until the rewrite loop produces working code.