v2026.10.15 — Check a Half-Remembered Reference and Find Related Papers

Two research commands that resolve partial references and find similar papers, a shorter AGENTS.md hook, prompt-first orchestration docs, and a security fix for aiwg serve.

AIWG 2026.10.15 - Check a half-remembered reference and find related papers

AIWG 2026.10.15 adds two research commands: one turns a partial reference into a verified paper identity, and one finds papers similar to the ones you already track. It also shrinks the AGENTS.md file your assistant reads, rewrites the handoff docs around example prompts, and ships a security fix for `aiwg serve`.

Turn a partial reference into a verified paper

You have a title fragment, a pasted link, or a bibliography line, and you want to know what paper it means. Ask your assistant to check it, for example: "What paper does this reference mean: <paste the title or link>?" Behind that question, your assistant runs:

`aiwg corpus resolve "<title, URL, arXiv ID, DOI, or bibliography line>"`

The command recognizes arXiv, DOI, OpenReview, ACL Anthology and Semantic Scholar identifiers, looks titles up on Semantic Scholar, arXiv and CrossRef, and either resolves the reference or tells you it is ambiguous or unresolved instead of guessing. When several papers look alike, it lists all of them. It also tells you whether the paper is already in your corpus. You can check a whole list at once from a file, one reference per line.

Existing research searches work again too: Semantic Scholar and arXiv requests had been going to the wrong addresses, and now reach the right APIs.

Find papers similar to the ones you track

Once your corpus has papers in it, you can ask for more like them: "Find papers similar to the recent entries in my corpus." Your assistant runs:

`aiwg corpus expand --since YYYY-MM-DD`

The command asks Semantic Scholar for papers similar to each seed, removes ones you already track or rejected, and ranks the rest by how many seeds pointed at them. It only suggests by default: nothing is recorded until you add `--write`, and a scope is always required, so a bare call never scans your whole corpus.

A smaller AGENTS.md that leaves your content alone

AGENTS.md is now a short hook instead of a long rulebook. It points your assistant at WORKSPACE.md and AIWG.md, keeps the short core rules inline, and names one place per tool where the rest of the rules live. Your own content outside the AIWG block is untouched. Run `aiwg use all` to refresh your setup.

Handoff docs that start with what to say

The orchestration guide and the worker-dispatch pages now open each task with an example prompt and what the assistant does with it. Start from the prompt that matches your situation instead of assembling flags from a table. The commands themselves are unchanged. The full walkthrough is in the orchestration guide.

A security fix, and fewer surprises at release time

This release updates the static file server behind `aiwg serve` so it no longer decodes request paths twice, and serves dashboard files for reads only. Update to pick it up.

It also moves three packaging checks into every main-branch run. They ran only when cutting a release, which is how a broken wrapper format slipped through last time. Problems like that now fail on the change that caused them, so releases carry fewer last-minute fixes.

If you are upgrading

  • Run `aiwg use all` to refresh your AGENTS.md, rule files and workflow skills.
  • If you run `aiwg serve`, update for the static file server fix.
  • The two research commands are new and need no setup. `expand` records nothing until you pass `--write`.

Full technical detail is in the CHANGELOG entry.