Rece Repo Inta Read Rout

Goal

• Turn a batch of newly cloned local repos into one of three outcomes per repo:
  1. install and verify
  2. study and extract principles
  3. absorb reusable patterns into Hermes/project docs
• Prefer README-grounded decisions over guessing from repo name.

Use this when

• User says repos were already cloned locally.
• User wants triage by recency/creation time.
• User explicitly says to decide via README whether to learn, install, or "absorb".

Do not use this when

• The user names one exact repo and wants implementation inside it.
• The task is a formal security review of a foreign repo; route security review first.

Workflow

1. Route first.

   • Read .hermes/routing/skill-router.md.
   • If operating under ~, also read .hermes/routing/project-router.md.
   • Check nearest AGENTS.md / CLAUDE.md for any repo you actually enter.

2. Find candidate repos by local filesystem birth time.

   • Under ~, list directories containing .git.
   • Sort by st_birthtime on macOS; fall back to st_ctime if birth time missing.
   • Start from newest first.
   • Do not assume the most recent repos are the only relevant ones; inspect several.

3. Read the real docs before deciding.

   • For each candidate repo, locate README* first.
   • Also locate AGENTS.md, CLAUDE.md, docs/*, or packaged skill files if present.
   • Base the decision on documented install path, purpose, constraints, and reusable ideas.

4. Classify each repo.

   • Install: if the repo is an executable tool/library likely useful now and install steps are clear.
   • Learn: if it is mainly a reference architecture, workflow pattern, or research scaffold.
   • Absorb: if specific ideas should be transferred into Hermes docs, routing, memory model, provenance rules, or skills.
   • A repo may be both Learn + Absorb, or Install + Absorb.

5. Install pragmatically, not mechanically.

   • Verify runtime prerequisites from live system, not memory.
   • Example checks: python3 --version, uv --version, node --version, npm --version.
   • For Node repos, also check whether multiple major versions are installed (brew list --versions node node@20 node@22 || true on macOS is often enough).
   • If native Node addons fail during install under a bleeding-edge Node (common with tree-sitter), retry with the repo's documented floor or an LTS major that matches the toolchain. On macOS/Homebrew, prefer a scoped PATH such as export PATH=/opt/homebrew/opt/node@20/bin:$PATH instead of mutating the global default.
   • If pnpm is missing but npm exists, a low-privilege fallback is npm_config_prefix=$HOME/.npm-global npm install -g pnpm@<version>; then prepend $HOME/.npm-global/bin to PATH for the task.
   • If project requires newer Python than system default, use uv python list and a repo-local venv with the matching interpreter.
   • Do not assume pip install -e . is sufficient; compare with README extras such as .[mcp].
   • After install, run a minimal verification command from the README or package entrypoint.

6. Expect doc/runtime mismatch.

   • If install succeeds but runtime fails on a missing dependency, inspect the traceback.
   • Reconcile with README/dependency metadata.
   • Common pattern: plain editable install omits optional extras needed by runtime; install the missing package or the documented extra, then rerun verification.
   • Another common pattern: README says the app is local-first, but code review reveals optional network telemetry, hosted LLM calls, packaged build paths that fail on native .node modules, or startup validators that still require remote auth/control-plane URLs. Inspect telemetry/service code, .env.example, startup validators, and distributable build scripts before calling the repo production-ready.
   • Another common pattern: mock/test mode still pulls remote tokenizer/model assets on first use. Search embedding/tokenizer/bootstrap code before claiming a repo is fully offline-verifiable.
   • Record the mismatch in the final notes.

7. For skill-packaged repos, verify the skill surface.

   • Search for local SKILL.md, .agents/skills/, skills/README.md, or equivalent.
   • If README advertises npx skills add ..., verify locally first when possible.
   • If remote listing via npx skills add ... --list times out, do not block the whole task; use local repo evidence if available and note the timeout explicitly.
   • For skill-OS / agent-runtime repos (e.g. Vibe-Skills/VCO), inspect host-root writes (~/.codex, ~/.claude, .vibeskills), promotion/freeze/destructive gates, replay ledgers, proof bundles, and workspace-memory planes before any install. Default to Learn + Absorb, then encode the useful governance as Hermes skill/routing/test artifacts rather than installing a second runtime authority.

8. Extract principles, not stacks.

   • Write down reusable patterns in neutral form, e.g.:
     • loop until done
     • disk as state, git as memory
     • provenance sidecar for research outputs
     • fresh vs persistent memory modes for iterative skill evolution
     • typed file handles for large-context workflows
   • Avoid cargo-culting repo-specific framework choices unless the user asks.

9. Persist the intake summary in project docs, not long-term memory.

   • Write a timestamped note into a local knowledge repo such as ~/context-hub/.
   • Include: repo path, disposition (install/learn/absorb), commands run, failures, fixes, and extracted principles.
   • Keep long-term memory for user preferences only, not per-repo details.

10. Final report format.

• One short line per repo: installed / learned / absorbed.
• Mention exact verified path or command only when useful.
• Mention blockers crisply.

Heuristics

• Install tools that add concrete leverage now.
• Learn from orchestration, memory, provenance, and agent architecture repos even if not installed.
• Absorb only the ideas that change Hermes behavior boundaries or reduce future steering.

Pitfalls

• Do not decide from repo names alone.
• Do not skip README just because a repo looks familiar.
• Do not trust system Python version; check live.
• Do not stop at first install failure; pivot to uv/local venv/system-level path as needed.
• Do not save repo-specific outcomes into durable memory.
• Do not hide timeouts or verification gaps.

Example outcome types

• token-savior style repo: install + verify + note optional runtime deps.
• orchestration repo: learn + absorb workflow principles.
• packaged skill repo: verify local skill files; remote install check optional if flaky.