[HaoZeke]

Manual discovery links

What and why

The README names llms.txt, llms-full.txt, AGENTS.md, skills/, and server.json.
The manual index, which is the published site home, did not name any of them, so a reader of the site had no path to those files.

The index now has a Discovery section with those links.
Client setup stays on the Model Context Protocol page, which the section links.
llms-full.txt follows the Org manual, and the metadata test requires the same names on the index that it already requires on the README.

Before

docs/orgmode/index.org ended its front matter with the license and went on to the docs build:

  MIT License. See [[https://github.com/HaoZeke/eb-stack/blob/main/LICENSE][LICENSE]] in the repository.

* Build these docs

After

* Discovery

  [[https://github.com/HaoZeke/eb-stack/blob/main/llms.txt][llms.txt]] indexes
  this manual.
  [[https://github.com/HaoZeke/eb-stack/blob/main/llms-full.txt][llms-full.txt]]
  repeats every page in one file.
  [[https://github.com/HaoZeke/eb-stack/blob/main/AGENTS.md][AGENTS.md]] lists
  the build, test, lint, and docs commands. The skills for inspect, plan,
  bump, a stack SBOM, pull-request overlap, and the stdio server are under
  [[https://github.com/HaoZeke/eb-stack/tree/main/skills][skills/]]. The Model
  Context Protocol manifest is
  [[https://github.com/HaoZeke/eb-stack/blob/main/server.json][server.json]].
  Client configuration is the [[file:reference/mcp.org][server page]].

The built docs/build/index.html contains llms.txt and server.json.

Tests

• Formatting: cargo +stable fmt --all --check exited 0.
• Lints: cargo +stable clippy --locked --all-targets -- -D warnings exited 0.
• Full suite: cargo +stable test --locked --all-targets exited 0, with 93 suites, 1503 passed, 0 failed, 10 ignored.
• Packaging: cargo +stable package --locked exited 0.
• Hooks: uvx prek run -a exited 0.
• Documentation: pixi run -e docs mkrst, scripts/gen-rust-api.sh, scripts/sphinx-build-docs.sh, and scripts/check-doc-links.sh exited 0, and lychee 0.24.2 reported 0 errors.
• Generated text: python3 scripts/gen-llms-txt.py --check exited 0.

Repro

cargo +stable test --locked --test public_metadata -- --exact discovery_metadata_uses_real_index_terms
python3 scripts/gen-llms-txt.py --check

---

Note

Medium Risk
Build-order graph walking changes what gets scheduled when --installed is used, and PR-overlap fetch scripts may include truncated file lists; both affect operational planning workflows but not auth or data security.

Overview
This PR adds several stack and workflow improvements plus a large documentation pass, not only the manual Discovery links described in the PR note.

stack order --installed (and eb_stack_order installed) treats modules from a listing file or Lmod tree like stack gap --installed: listed modules and their installed dependency subtrees are omitted from the build order, with installed=N on stdout before order=. --filter-deps still runs first. The Rust layer adds build_order_against_installed / build_graph_against_installed and wires them through workflows::stack_order.

Both examples/*/fetch-prs.sh scripts now page open PRs 100 at a time without stopping at 500, page changed files through GitHub’s 3000-file cap, and keep PRs with file_list_truncated when the list is truncated.

stack arch --probe-sources with two --arch values compares only that pair: a miss is arch-missing when the other family fetched the slot, unreachable when neither did (documented in the how-to and CLI reference).

The manual and README expand the MCP catalog (51 tools, command↔tool table in mcp.org, test that every CLI subcommand appears there), refresh the CLI command map (stack gap, campaign finding rows, stack oom usage block), and add Discovery links on the index (llms.txt, AGENTS.md, skills/, server.json).

^{Reviewed by Cursor Bugbot for commit f85acb2. Bugbot is set up for automated code reviews on this repo. Configure here.}