Broaden to a tool-agnostic loop landscape (Claude Code as reference implementation)

[espi]

Routine self-improvements

• Applied (auto-gated): none. Human-authored change.
• Suggested: none.

What this is (PR 1 of 2)

Your question was: what alternatives to Claude Code exist for running loops, and how to integrate them without loss of quality. This PR does the integration; PR 2 (coming next) adds the maintenance structure to keep it current.

Sourced from a 3-agent research survey (peer CLI harnesses / orchestration platforms / cross-tool portability). The framing: principles-first, Claude Code as the reference implementation — the loop primitives are converging cross-tool, so the discipline (three hard stops + verification) is what's portable, not the vendor.

Changes

• Primer §4 → "Beyond Claude Code — the same loop on other harnesses": a cross-tool capability matrix (Codex CLI, Goose, Gemini CLI, opencode, Cursor, Amp/Aider/Cline, Devin/Factory, Gas Town/City, Claude Agent SDK, LangGraph/ADK/CrewAI/AG2) scored on goal-stop / three-hard-stops / scheduling vs. a single Claude Code loop. Plus three durable facts:
  • the validator-judge stop is now cross-tool (both Codex and Claude Code /goal);
  • "durable execution" is the field's biggest hype-vs-substance gap — LangGraph/CrewAI/ADK checkpoints are recovery points, not crash-survival, without Temporal/Diagrid;
  • portability is real for MCP, contested for SKILL.md.
• Primer §6 + guardrails/README.md → tool-agnostic enforcement: the cleanest place for hard stops knowledge: second update pass (2026-06-09) #1/knowledge: 2026-06-22 update pass (7-day delta from Jun 15) #3 is the LLM gateway every agent routes through (LiteLLM max_budget_per_session + fail_closed_budget_enforcement; OpenRouter 402 windows; Portkey/Helicone), plus in-harness libs (AgentGuard, LoopGain). This strengthens the repo's knowledge: second update pass (2026-06-09) #1 rule beyond Claude rather than diluting it.
• sources.md → new "Alternative harnesses & cross-tool landscape" section with per-tool entries + honest confidence, and two re-verify items added to the standing backlog.
• update-knowledge research angles expanded (step 3) so peer harnesses + cross-tool standards stay maintained by the weekly pass.

How quality is preserved (the "without loss of quality" part)

The failure mode to avoid was becoming a shallow 12-tool listicle. Guards against that:

• Claude Code stays the single deeply-runnable reference; other tools are covered at capability-matrix depth, not 12 parallel guides.
• Most new tool-claims entered at Medium/sourced, not primer-High — vendor guardrail docs were thin/403'd, and the honest confidence is recorded.
• Coverage serves the discipline (three hard stops + verification), which is the actual product; breadth is in service of that, not a goal.

Two re-verify items you should know about

• SKILL.md cross-tool execution is contested. My two research passes disagreed — one found the format widely read, the other found no CLI tool confirmed to execute SKILL.md, with AGENTS.md the more common file. Recorded as with-testing, not drop-in.
• Agent Skills governance may be Anthropic-spec + community, not LF/AAIF-governed (the LF appears to govern MCP/AGENTS.md/goose). Earlier KB passes may have overstated it. Flagged to re-verify.

Not merging — human reviews and merges.

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01Nsk6ryTqYCAuE6dRW8ZnHJ

---

Generated by Claude Code