Skip to content

spillwavesolutions/okf-agent-graph

v0.8.1MIT

OKF Agent Graph Engineering Runtime (AGER): author multi-agent loop graphs as OKF and reverse-engineer existing agent stacks. Multi-host bindings and write isolation.

Changelog

Unreleased

0.8.1 — 2026-08-27

  • Nested --out links (#16). ager_capture.py / ager_reverse_engineer.py accept --bundle-root (or --link-prefix) so a draft written to knowledge/agent-graph emits /agent-graph/tools/read.md instead of /tools/read.md. PKC/OKF doctors that validate the shared knowledge/ tree no longer see broken edges.
  • Scan noise floor (#17). Generic keyword hits on the same phrase (Supervisor / orchestrator role) collapse to one patterns/ node with multiple evidence_paths. Tool findings without a real identifier (Tool: { ... }, tools = [blob]) are dropped; named tools (web_search) collapse to a single node.

0.8.0 — 2026-08-27

  • Plugin discovery for AGKC (#14). ager_scan.py now finds Claude Code, Grok Build, Codex, Cursor, and Agent Plugins 1.0 surfaces before the keyword scan: .claude-plugin/plugin.json, .claude/agents/*.md, .claude/skills/*/SKILL.md, plus the undotted claude/, grok/, codex/ twins, .codex-plugin, .grok-plugin, and root plugin.json. Frontmatter name / description / tools map a subagent to WorkerAgent or JudgeAgent, and a SKILL.md that is the orchestrator to OrchestratorAgent. Capture emits those as typed concepts under agents/, tools/, and graphs/ instead of patterns/ prose.

0.7.1 — 2026-08-24

0.7.0 — 2026-08-24

Added

  • Noun ownership. AGER now ships JSON schemas for the full agent/harness set (51 types), including AgentNode, Workflow, Harness, SharedState, ToolCapability, loop/runtime/ops/eval types, with x-impact for the graph engine. WriteEvent stays in second-brain-core / the graph engine write journal; AGER does not own it.
  • README lists every AGER noun.

Changed

  • Catalog / ContextPack stay in okf-plugin. Authoring AgentNode is no longer an okf-graph-eng concern.

0.6.4

  • WikiTicket SDD (worklog) is the tracking system for this plugin.

0.6.3 — 2026-08-17

  • Cursor host. .cursor-plugin/plugin.json (Cursor Plugins) plus .cursor/rules/second-brain.mdc. Docs: docs/CURSOR.md. docs/GROK_BOT.md now covers Grok Bot spawning Cursor cloud agents.

0.6.2 — 2026-08-16

  • Local ContextPack: scripts/ager_pack.py, /ager-pack, $ager-pack.
  • Default budget = 1/4 of SECOND_BRAIN_WINDOW_TOKENS (128000 → 32000).
  • Override: --max-tokens / SECOND_BRAIN_PACK_MAX_TOKENS.
  • Fail-closed: over-budget pack does not --write.
  • Bodies off unless that node is the pack root. Neighbors keep title / type / path / description.
  • Inbound typed edges are visible from the seed (same shape as PKC inbound index).
  • Catalog index.md files and write-events/ are not concepts.
  • Node clip (--max-nodes / --tiny) is not a token budget.
  • Implements part of okf-plugin#55.

0.6.1 — 2026-08-16

  • Required identity on every knowledge write: --author or SECOND_BRAIN_IDENTITY.
  • ager_common.resolve_author() fail-closes without a claim. Capture, reverse-engineer, and init stamp author and emit a WriteEvent.
  • Scan / validate / print-only paths do not require identity.

0.6.0 — 2026-08-15

  • Multi-host bindings + write isolation. Root Agent Plugins 1.0 plugin.json, Grok Bot / Deep Agents / isolation / onboarding docs, host wrappers, vendored scripts/brain_session.py, and ager-session.
  • Concurrent writers read main and write brain/<actor>/<session-id>. Isolation tests use fictional lumenfield-detector / northstar-console only.
  • LoopPolicy / KnowledgeBind: retrieve from main; persist loop/graph writes only in the session worktree.

0.5.1 — 2026-08-13

  • Register AGER types in schemas/okf-concepts/registry.json so a mixed second-brain okf-graph.py validate recognizes them (fallback: shared BaseConcept, required type + title only). Field-depth validation stays in ager-validate.py.
  • Land the v0.4.0 post-release wiki ledger (.work/published.json), freeze the dated design/code-walkthrough snapshots, ignore .work/wiki-checkout/, and correct Codex dependency guidance (okf-plugin v0.3.2 has no Codex manifest).
  • Refresh docs/.index/publish-manifest.json so worklog ia-render --check matches the current spec and user-guide bodies.

0.5.0 — 2026-08-09

  • Add AGKC reverse engineering path: scan existing agent codebases and capture draft AGER knowledge.
  • New scripts: ager_scan.py, ager_capture.py, ager_reverse_engineer.py (stdlib only).
  • New skills/commands: ager-scan, ager-reverse-engineer (Claude /…, Codex $…).
  • New agent: agent-graph-re-orchestrator.
  • Detect LangGraph, LangChain, CrewAI, LlamaIndex, OpenAI/Claude APIs, Claude Agent SDK, Deep Agents, AutoGen, Semantic Kernel, MCP/JSON-RPC, Bedrock AgentCore, Azure AI Agents, Vertex Agent Engine, Firecracker/gVisor/Kata/E2B sandboxes.
  • Extract prompts/system prompts, tools, schemas, graphs, loops, orchestration, hyperscaler + hardened runtime signals with provenance.
  • Fixture agent repo + tests/test_ager_scan.py regression coverage.
  • Docs: docs/REVERSE_ENGINEERING.md.

0.4.0 — 2026-08-06

  • Add a complete canonical scaffold and atomic ager-init.py generator.
  • Add dependency-free AGER structural validation with strict CI checks.
  • Make the sample bundle pass strict AGER and OKF validation without warnings.
  • Add native Codex plugin packaging while preserving Claude Code and Grok Build support.
  • Initialize WikiTicket SDD for plans, tickets, roadmap, PR, and wiki synchronization.

0.3.0 — 2026-08-04

  • Initial AGER plugin with skills, commands, specification, and worked sample graph.