Ecosystem Positioning

May 27, 2026 ยท View on GitHub

This page describes how lessonweaver relates to adjacent tools and standards. It is factual, not promotional: each entry states what lessonweaver can do in that relationship and what it does not do or replace.

lessonweaver is an offline, deterministic loop that turns reviewed trace evidence into governed, reusable operational guidance. It does not orchestrate agents, run models, or inject anything without human review.

For category-level boundaries (observability vs. evals vs. memory vs. frameworks), see comparisons.

AGENTS.md

  • Can: Export reviewed skill content as Markdown that a maintainer can paste into AGENTS.md (today via the markdown or runtime export formats).
  • Does not: Parse, manage, or auto-write AGENTS.md files. A dedicated agents-md export format is planned (#48).

GitHub Copilot custom instructions

  • Can: Export Copilot-compatible instruction fragments via export-skill --format copilot.
  • Does not: Manage .github/copilot-instructions.md or .github/instructions/ for you, and does not commit instructions automatically. You review the fragment and decide where it goes.

Claude Code (CLAUDE.md, .claude/rules/, SKILL.md)

  • Can: Export Claude-style skill fragments via export-skill --format claude.
  • Does not: Replace Claude Code's native configuration, install skills, or integrate with Claude Code hooks. Output formats may evolve.

OpenAI Agents SDK / Codex

  • Can: Produce reviewed skill content that can be injected into agent instructions before a run. A codex_directory export format is reserved in the ExportFormat enum.
  • Does not: Manage agent configuration, SDK setup, or hosted Agent Builder. A runnable adapter example is planned (#21).

LlamaIndex applications

  • Can: Produce runtime prompt snippets (export-skill --format runtime, or SkillLoader.load_for_task) for injection into a system prompt before an agent/chat-engine call.
  • Does not: Integrate with LlamaIndex retrieval pipelines or add a LlamaIndex dependency. A runnable adapter example is planned (#20).

Pipecat voice agents

  • Can: Mine post-call transcripts (any tool that emits the documented trace format) into lesson candidates, and produce short snippets suitable for latency-sensitive contexts.
  • Does not: Replace Pipecat's pipeline or real-time features, and adds no Pipecat dependency. Voice integration docs are planned (#22).

Evals frameworks (Braintrust, promptfoo, etc.)

  • Can: Recommend eval specs and guardrail rules as an action type for a reviewed lesson. Dedicated eval/guardrail exporters are planned (#47).
  • Does not: Run evals, score model output, or integrate with an eval runner. lessonweaver complements evals; it does not execute them.

Observability / tracing tools

  • Can: Consume traces these tools already produce, once normalized into the documented trace format.
  • Does not: Record, store, or visualize live telemetry. Observability tells you what happened; lessonweaver turns reviewed evidence into guidance.

Sibling repository tools

  • Can: Consume structured findings from sibling tools (for example a repository checker such as dgenio/vibeguard) when their output is mapped to the trace format. See interoperability.
  • Does not: Take a hard dependency on any specific sibling tool. All such integrations are optional and live in examples/ or docs, never in core.