Skip to content

Point agents at Alexandria feedback from instructions, descriptions and results - #428

Merged
nickscamara merged 3 commits into
mainfrom
nsc/alexandria-feedback-hints
Sep 22, 2026
Merged

nickscamara merged 3 commits into
mainfrom
nsc/alexandria-feedback-hints

Conversation

@nickscamara

@nickscamara nickscamara commented Sep 22, 2026 •

Copy link
Copy Markdown
Member

Why

Agents using Alexandria purely through MCP were never told to send feedback: the firecrawl_find_tools and firecrawl_scrape descriptions and the server instructions did not mention it, and firecrawl_feedback only described the Alexandria mode inside its own schema. Search, by contrast, surfaces its feedback loop in the tool description and through the response id.

These commits were pushed to alexandria-mcp after #405 had already been squash-merged, so they are re-applied here on top of main.

Summary

The pointer now appears in three places:

  • Server instructions (FULL_PROFILE_INSTRUCTIONS) and the Alexandria block of the full firecrawl_search description: after an Alexandria task, whether a capability ran or discovery found nothing for the website, call firecrawl_feedback once per website with endpoint: "alexandria".
  • firecrawl_scrape (Alexandria mode) and firecrawl_find_tools descriptions: one sentence each pointing at firecrawl_feedback and the feedbackTool result field.
  • A top-level feedbackTool object on every Alexandria execution result (inline and retained) and every firecrawl_find_tools result, carrying the tool name and an argument skeleton.

The hint is omitted for Firecrawl-internal calls (bash, terms/*) and whenever firecrawl_feedback is not registered (FIRECRAWL_NO_ENDPOINT_FEEDBACK or keyless startup), so it never points at a tool that does not exist. The search-only marketplace profile is unchanged.

Test plan

  • New tests/mcp-alexandria-feedback-hints.test.mjs: instructions and descriptions, hints on execution and discovery results, no hint on bash, hint on retained results, no hint when disabled by environment.
  • npm test: 114/114 (includes the agent metadata policy check on the new wording).
  • npm run lint: only the 3 preexisting errors in src/alexandria.ts.

Made with Cursor


Summary by cubic

Adds feedback pointers so agents using Alexandria through MCP learn to report how the catalogue served the website.

  • Appends guidance to the server instructions and the firecrawl_scrape, firecrawl_find_tools, and firecrawl_search descriptions.
  • Attaches a feedbackTool object with the tool name and an argument skeleton to every Alexandria execution and discovery result, including retained results.
  • Omits the hint for Firecrawl-internal calls like bash and when firecrawl_feedback isn't registered (FIRECRAWL_NO_ENDPOINT_FEEDBACK or keyless startup).
  • Bumps the package version to 3.25.1.

Written for commit 48d5e0a. Summary will update on new commits.

Review in cubic

nickscamara and others added 3 commits September 21, 2026 18:29
…nd results

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
@nickscamara
nickscamara merged commit 4d0b56c into main Sep 22, 2026
2 checks passed
rakshith48 added a commit that referenced this pull request Sep 22, 2026
Resolves conflicts in the profile instructions, the scrape tool schema,
the feedback schema and the CHANGELOG, and threads the new Alexandria
surface: firecrawl_find_tools declares threadId (its schema is strict),
the Alexandria-mode refinement on firecrawl_scrape allows threadId
alongside alexandria, and result stamping keeps compact JSON compact
now that main emits compact results. Two Alexandria tests that
deep-compare payloads split off the thread ID first.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant