Skip to content

feat(copilot): build the docs vfs from a generated manifest, rescope search_docs#5948

Open
j15z wants to merge 12 commits into
stagingfrom
feat/enhance-search-agent
Open

feat(copilot): build the docs vfs from a generated manifest, rescope search_docs#5948
j15z wants to merge 12 commits into
stagingfrom
feat/enhance-search-agent

Conversation

@j15z

@j15z j15z commented Jul 25, 2026

Copy link
Copy Markdown
Collaborator

Summary

Mounts Sim's public documentation as a read-only docs/ tree in the copilot VFS, and replaces search_documentation with a path-scoped search_docs.

The corpus is navigated like a filesystem. glob matches a checked-in manifest generated from apps/docs/content/docs/en with no network access; read fetches the page live from docs.sim.ai; grep takes exactly one page, since each is a separate request. Manifest entries double as both the docs/ VFS path and the docs.sim.ai URL path, so a read needs no lookup table. Like uploads/, the tree is opt-in — only a pattern or path that explicitly starts with docs/ ever sees it, and only the search agent's prompt advertises it.

search_docs adds an optional path scope (one page or a section) and returns the docs/ path with every hit, so search chains into read. Unscoped searches exclude academy/ and api-reference/, which are indexed but not mounted, and results whose source no longer maps to a live path are dropped — the index is rebuilt on release, so it lags the manifest.

@docs chat tagging is disabled: its query was the raw user message, and its mention UI was already unreachable.

search_documentation is kept in the catalog as a hidden, deprecated entry routed to Sim, so the id still resolves during a mixed-version deploy. It is in no agent's allow list. Remove it, along with the alias in baseServerToolRegistry, once search_docs is live in prod on both sides.

Type of Change

  • New feature

Testing

Tested manually against the running app. tsc clean, 947 copilot tests, biome clean, and all 11 CI audits pass (boundaries, api-validation:strict, utils, zustand-v5, react-query, client-boundary, bare-icons, icon-paths, realtime-prune, skills, agent-stream-docs).

Reviewers should focus on:

  • docs-search.ts scope mapping. source_document stores the en-relative mdx path while VFS paths mirror the public URL, so a section overview is docs/workflows.mdx in the VFS but workflows/index.mdx on disk. Both directions are tested, including a round-trip over every manifest entry.
  • sync-docs-manifest.ts + the CI check. Adding, renaming, or deleting a page without regenerating fails the build; editing page content does not.
  • search-docs-alias.test.ts. Dispatch gates on catalog membership before consulting the handler registry, so the alias needs the hidden catalog entry to be reachable at all. The test pins every link in that chain.

Checklist

  • Code follows project style guidelines
  • Self-reviewed my changes
  • Tests added/updated and passing
  • No new warnings introduced
  • I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

Companion PRs

  • simstudioai/mothership#371

Merge sim first. The alias covers an older Mothership calling this build; the reverse — a new Mothership calling an older Sim — cannot be fixed from either side.

j15z and others added 8 commits July 24, 2026 17:13
…ocs; serve openapi.json publicly

- search_docs server tool: same vector search over docs_embeddings plus an
  optional docs/documentation/... VFS path prefix mapped onto a
  source_document scope (covers both <tail>.mdx and <tail>/... layouts);
  unscoped searches exclude academy/ and api-reference/ rows so the scope
  is exactly the Documentation tab
- @docs chat context repointed to the new tool; display label updated
- apps/docs now serves /openapi.json so the mothership can build its
  docs/api-reference/<tag>.json VFS views from the deployed spec
- generated tool catalog/schemas regenerated from the mothership contract

Companion: simstudioai/mothership feat/enhance-search-agent

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…tool chips

read("docs/documentation/workflows/index.mdx") now renders "Read
Workflows/index" instead of the leaf-only fallback ("Read Index").

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
"Searched docs" becomes 'Searched docs for "<query>"' (toolTitle/title
preferred, query fallback, truncated at 60 chars). Also adds the missing
browser_list_sessions display title the catalog regen surfaced.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A 400 from the chat POST's ChatMessageSchema parse returned the zod
issues to the client but logged nothing, making "message disappears on
send" undiagnosable from server logs.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A workspace_resource tag emitted with id:"" resolved through `??` chains
(which keep empty strings), got persisted onto the chat's resources, and
then every send attached it and failed the chat request's id min-length
validation — the message vanished on send, forever. Resolve chip ids by
first non-empty candidate and drop id-less resources when building
request attachments.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…search_docs

Replaces the mothership's runtime docs corpus (llms.txt + llms-full.txt +
openapi.json behind a 15m TTL cache) with a static manifest generated from
the docs source, plus live per-page fetches. ~1,000 fewer lines of hand-
written code and one repo instead of two.

- scripts/sync-docs-manifest.ts walks apps/docs/content/docs/en and emits
  lib/copilot/generated/docs-manifest.ts. Each entry is simultaneously the
  docs/ VFS path and the docs.sim.ai URL path, so a read is a plain fetch.
  Section index pages fold onto their parent (fumadocs serves /workflows,
  not /workflows/index); academy/ and api-reference/ are excluded — they
  stay unmounted and unsearchable, reachable only via scrape_page.
- docs-manifest:generate / :check, with a CI step so a page added, renamed,
  or deleted without regenerating fails the build. Content edits don't.
- lib/copilot/docs/docs-corpus.ts + tools/handlers/vfs.ts: glob matches the
  manifest with no network, read fetches the page live, grep takes exactly
  ONE page (each is a fetch, so there is no corpus-wide grep). Opt-in like
  uploads/ — only an explicit docs/ prefix ever matches.
- search_docs now scopes to the docs/ tree instead of docs/documentation/,
  validates its path against the manifest (a bad path errors instead of
  silently returning nothing), and returns the docs/ path with every chunk
  so search chains into read. Unscoped searches drop rows the agent could
  not then read: unmounted sections, and pages gone since the last index
  rebuild.
- @docs tagging disabled: its query was the raw user message, a poor
  embedding query, and the mention UI it fed was already dead code.
- Reverts the apps/docs /openapi.json route, added only for the old
  api-reference VFS views.

Companion: simstudioai/mothership feat/enhance-search-agent

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Multi-agent review of the docs/ VFS change. Applied the behavior-preserving
fixes plus two agent-facing bugs that made real pages unreadable.

- docs read no longer hard-fails on oversized pages. Six+ live integration
  references exceed the inline cap (github.mdx is 354KB, sportmonks 513KB),
  so a plain read of them ALWAYS failed and cost a second fetch to recover.
  Truncate to the largest whole-line prefix that fits, keep the true
  totalLines, and tell the model how to page. An explicit offset/limit that
  still overflows is still an error — that one is a caller mistake.
- classify docs fetch failures. Everything collapsed to null, so a permanent
  404 was reported to the agent as "temporarily unavailable, retry shortly",
  inviting a retry loop on a page that will never exist. 4xx (except 429) is
  now permanent and says so; 5xx/429/network/timeout keep the retry wording.
- register search_documentation as a transitional alias for search_docs.
  sim and mothership deploy independently and the rename deleted the old id
  on both sides, so BOTH deploy orders broke docs lookup for the window
  between them. Old params are a subset of the new. Remove once both ship.
- extract the index-page fold (X/index.mdx <-> X.mdx) into docs-path.ts. It
  was re-derived in three places — the manifest generator, the
  source_document reverse mapping, and the search scope filter — which is
  the hand-synced-duplicate shape that has drifted in this repo before.
- grepDocsPage now goes through grepReadResult, the primitive files/ and
  uploads/ grep already use, instead of calling grep directly.
- couldMatchDocsScope delegates to isDocsPath; the bodies were identical.
- drop the dead 'docs' member from AgentContextType.

Tests: 404-vs-5xx-vs-429 classification, network failure, and a docs-path
round-trip asserting one source candidate reproduces every manifest entry.

Verified: tsc clean, 932 copilot tests, biome clean, docs-manifest:check,
check:utils and check:api-validation:strict both pass.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The SQL LIMIT is applied before the similarity-threshold and liveness
filters, so search_docs can return fewer hits than topK — or none, when
every candidate was filtered. An empty array is indistinguishable from "the
documentation does not cover this", which sends the agent off to guess
instead of rephrasing or falling back to glob.

searchDocs now returns the drop counts alongside the results, and the tool
attaches a note when anything was dropped: how many candidates the index
returned, why they went, and what to try next. Silent on the common path.

This does not change which rows are returned or how many — the ordering
issue behind the shortfall is a pre-existing bug the deleted
search-documentation.ts had too, and pushing the threshold into SQL is its
own change.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@j15z
j15z requested a review from a team as a code owner July 25, 2026 00:37
@vercel

vercel Bot commented Jul 25, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment
Project Deployment Actions Updated (UTC)
docs Skipped Skipped Jul 25, 2026 2:08am

Request Review

@cursor

cursor Bot commented Jul 25, 2026

Copy link
Copy Markdown

PR Summary

Medium Risk
Copilot doc access now depends on live docs.sim.ai fetches and manifest/index alignment; search behavior changes for agents and during the search_documentation rename shim.

Overview
Adds a read-only docs/ copilot VFS backed by a CI-checked generated manifest (docs-manifest:check). Glob lists manifest paths (opt-in via docs/ patterns only); read and grep lazy-fetch page markdown from docs.sim.ai, with truncation, single-page grep limits, and user-facing DocsCorpusError handling wired into existing VFS tools.

search_docs replaces inline search_documentation: vector search over docs_embeddings with optional docs/ path scoping, similarity/stale filtering, topK clamping, and results that include the VFS path to read next. search_documentation stays registered as a hidden transitional alias for mixed deploys.

@docs chat context is disabled (no more server-side doc search injection). Smaller fixes: workspace resource mention IDs prefer non-empty id/path, and Zod validation failures on chat POST log at warn level.

Reviewed by Cursor Bugbot for commit 01c4d70. Bugbot is set up for automated code reviews on this repo. Configure here.

@github-actions github-actions Bot added the requires-mothership-merge Has a companion PR on the mothership/copilot side — merge in lockstep label Jul 25, 2026
@github-actions

Copy link
Copy Markdown

⚠️ Cross-repo companion check

One or more companion PRs aren't merged into staging yet. Merging this without them will leave copilot and sim out of sync — merge them in lockstep.

  • simstudioai/mothership#371OPEN, not merged (targets staging) — feat(search): replace search_documentation with path-scoped search_docs

@greptile-apps

greptile-apps Bot commented Jul 25, 2026

Copy link
Copy Markdown
Contributor

Greptile Summary

The PR completes the documentation-tool migration and repairs the previously reported mixed-version dispatch path.

  • Retains search_documentation as a hidden, Sim-routed catalog alias backed by the search_docs server handler.
  • Adds coverage for catalog classification, server registration, legacy parameters, and hidden-tool behavior.
  • Introduces a manifest-backed read-only docs/ VFS with live page reads, single-page grep, and scoped semantic search.
  • Adds CI validation to keep the generated documentation manifest synchronized.

Confidence Score: 5/5

The PR appears safe to merge.

The previously reported transitional-alias dispatch failure is fully addressed, and no blocking failure remains.

Important Files Changed

Filename Overview
apps/sim/lib/copilot/tools/server/router.ts Registers both tool identifiers against the new documentation-search handler, completing the previously missing dispatch path.
apps/sim/lib/copilot/tools/server/docs/search-docs-alias.test.ts Verifies that the legacy identifier remains known, Sim-routed, registered, hidden, and compatible with legacy parameters.
apps/sim/lib/copilot/generated/tool-catalog-v1.ts Adds the new tool and preserves the old identifier as a hidden transitional catalog entry.
apps/sim/lib/copilot/docs/docs-corpus.ts Implements the manifest-backed documentation VFS namespace, live reads, globbing, and single-page grep.
apps/sim/lib/copilot/docs/docs-search.ts Implements manifest-aware semantic documentation search with path scoping and stale-result filtering.

Reviews (4): Last reviewed commit: "fix(copilot): make the search_docs topK ..." | Re-trigger Greptile

Comment thread apps/sim/lib/copilot/tools/server/router.ts
The registry alias added for the rename never fired: executeTool gates on
isKnownTool(toolId) — membership in the generated TOOL_CATALOG — before it
consults baseServerToolRegistry, so an id absent from the catalog is
rejected as unknown and routed to the app-tool path instead. The alias was
dead code, and worse, it advertised a mixed-version safety net that did not
exist.

The companion PR restores search_documentation to the contract as a hidden,
deprecated entry, so the id now passes the catalog gate and reaches this
alias. Marks it hidden in the UI set too — a call only ever arrives from an
older Mothership build and renders as the search_docs it maps onto.

Adds a test pinning every link in the dispatch chain (catalog membership,
sim route, registered handler, hidden, param shape) — the assertion that
would have caught this.

Remove all of it once search_docs is live in prod on both sides.
@j15z

j15z commented Jul 25, 2026

Copy link
Copy Markdown
Collaborator Author

Confirmed — you're right, and this was a real hole rather than a nit.

executeTool gates on isKnownTool(toolId) (i.e. toolId in TOOL_CATALOG) before it consults baseServerToolRegistry:

const canUseRegisteredHandler =
  isKnownTool(toolId) && (isSimExecuted(toolId) || ...)
if (!canUseRegisteredHandler) return executeAppTool(toolId, appParams)  // → tool not found

So dropping search_documentation from the generated catalog made the registry alias unreachable, and the mixed-version window it was meant to cover was still open — worse than having no alias, because the PR claimed a safety net that didn't exist.

Fixed across both repos:

  • mothership#371 (aa46b879): registers search_documentation as a hidden, deprecated catalog entry routed to sim. It's in no agent's AllowedTools, so nothing is ever offered it — it exists only so an inbound call from an older build resolves.
  • this PR (06f3a3756f): regenerated catalog now contains the id, so it passes the gate and reaches the alias. Also added to the UI hidden set, since such a call renders as the search_docs it maps onto.
  • Added search-docs-alias.test.ts pinning every link in the chain — catalog membership, sim route, registered handler, hidden flag, param shape. That's the assertion that would have caught this.

One thing your flowchart implies that I want to state explicitly: this only fixes the sim-first direction. New Mothership emitting search_docs at an old Sim can't be fixed from either side, so deploy order remains sim first, then mothership. Noted on both PRs.

940 copilot tests, tsc, and all 11 CI audits pass.

@j15z

j15z commented Jul 25, 2026

Copy link
Copy Markdown
Collaborator Author

@greptile

@j15z

j15z commented Jul 25, 2026

Copy link
Copy Markdown
Collaborator Author

@cursor review

Comment thread apps/sim/lib/copilot/docs/docs-search.ts
…g search

A directory scope matched only `<section>/%`, which covers an overview stored
as `<section>/index.mdx` but not one stored as a sibling `<section>.mdx`.
Fumadocs accepts both layouts and page scope already handles both via
docsSourceCandidates, so a scoped section search could silently omit the
overview chunks — and the doc comment claimed it did not.

Every section in the tree currently uses the index.mdx layout, so nothing is
broken today; this closes the gap before someone adds a sibling overview and
gets quietly incomplete results.
@j15z

j15z commented Jul 25, 2026

Copy link
Copy Markdown
Collaborator Author

@greptile

@j15z

j15z commented Jul 25, 2026

Copy link
Copy Markdown
Collaborator Author

@cursor review

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

Reviewed by Cursor Bugbot for commit e52205d. Configure here.

The clamp guarded magnitude but not type: Math.min/Math.max propagate NaN,
so a non-numeric topK reached the query as `.limit(NaN)`. The `?? DEFAULT`
only caught undefined. Nothing enforced this but the generated Ajv schema,
and searchDocs is also called directly, so it should not depend on that.

Extract clampTopK, which falls back to the default for anything non-finite
(NaN, Infinity, a string that slipped through) and clamps the rest to
[1, 25].

The clamp was completely untested because the db mock's .limit() stub
discarded its argument — the mock now records it. Covers default, cap,
floor, truncation, and the non-finite fallback. Worth pinning: staging's
search_documentation documented "max 10" and enforced nothing, so this
bound is new behavior, not just a bigger number.
@j15z

j15z commented Jul 25, 2026

Copy link
Copy Markdown
Collaborator Author

@greptile

Chips read "Searched docs" with no indication of what was searched. The
query-aware title existed earlier on this branch and this commit's own
predecessor dropped it: removing search_docs from the catalog deleted the
display case and its test, and putting the tool back only restored the
static map entry. The generic "every visible catalog tool has a title"
assertion still passed, because it checks that a title exists, not that it
is the useful one.

Chips now read: Searching docs for "how to read workflow logs and view
executions" -> Searched docs for "...". The gerund flip already preserves
the suffix, so the completed state needs no extra handling — the test now
pins that too, since it was the part most likely to regress silently.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

requires-mothership-merge Has a companion PR on the mothership/copilot side — merge in lockstep

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant