Fix: add litellm/ prefix for Agents SDK model routing - #7
Merged
Conversation
The OpenAI Agents SDK requires a litellm/ prefix to route non-OpenAI models through LiteLLM. Without it, models like anthropic/claude-sonnet-4-6 fail with "Unknown prefix". This adds the prefix at all Agent() call sites while keeping litellm.completion() calls unchanged. Also updates README quick start comments and model format docs.
Member
Author
Code reviewNo issues found. Checked for bugs and CLAUDE.md compliance. 🤖 Generated with Claude Code |
KylinMountain
added a commit
that referenced
this pull request
Jul 21, 2026
…es (#198) * feat(api): delete a knowledge base (POST /api/v1/kb/delete + openkb delete-kb) Physical, irreversible KB deletion with a type-the-name confirmation. - config.delete_kb: rmtree the KB directory + unregister it from the global registry (known_kbs / kb_aliases / default_kb, via the new unregister_kb). Guards against deleting a non-KB path; tolerates a ghost registry entry (directory already gone) by just unregistering. - POST /api/v1/kb/delete: confirm_name must equal kb, re-checked server-side. Lives in a new api_kbs_router (sibling of the config/graph/output routers) so api.py stays under the 800-line gate; delete_kb self-locks (config lock), no create_app closure needed. - openkb delete-kb NAME: type-the-name prompt (or --yes) then delete. Tests: endpoint (success + unregister, confirm mismatch, non-KB target) in test_api.py; config (physical delete + unregister, refuse non-KB, ghost tolerance) in test_config.py. Claude-Session: https://claude.ai/code/session_01XMxbhmAkxxVV8CFWCZDBaG * feat(api): delete a wiki page (POST /api/v1/page/delete) with backlink impact Delete a concept/entity page and degrade references cleanly instead of leaving them dangling: - page_ops.delete_wiki_page: dry_run reports the backlink pages (whose inbound [[wikilinks]] would be demoted) without touching anything; execute removes the page under the KB ingest lock, strips its index.md entry outright (compiler.remove_doc_from_index — the entry is a [[link]] line), and demotes the now-dangling inbound [[links]] to plain text in ONLY those backlink pages (lint.fix_broken_links restrict_to). The page ref is "section/stem", validated to concepts/ or entities/ only (path-traversal-safe). - New api_pages_router hosts POST /api/v1/page (read — relocated here from api.py to keep it under the 800-line gate) + POST /api/v1/page/delete. Tests: dry-run impact + execute (page gone, inbound link demoted to text, index entry removed, sibling kept), 404 on missing page, and rejection of unsafe / non-editable refs. Claude-Session: https://claude.ai/code/session_01XMxbhmAkxxVV8CFWCZDBaG * feat(api): edit a wiki page (PUT /api/v1/page) + link context (POST /api/v1/page/links) - page_ops.edit_wiki_page: replace a concept/entity page's BODY while preserving its code-managed OKF frontmatter (type/description/sources) verbatim; any frontmatter in the submitted content is dropped. Dead [[wikilinks]] the user typed are demoted to plain text on save (OpenKB keeps the wiki broken-link-free) and returned as ghosts_stripped so the UI can warn. Atomic write under the KB ingest lock. - page_ops.page_link_context (POST /api/v1/page/links): outbound + inbound links for the edit-impact panel. Editing the body does not break either (links are path-based), so this is context, not a blocker — the honest impact model. - PUT /api/v1/page added to api_pages_router. Tests: edit preserves frontmatter + keeps a resolvable link + demotes a dead one (reports ghosts_stripped) + replaces the body; 404 on a missing page; links reports out/backlinks. Claude-Session: https://claude.ai/code/session_01XMxbhmAkxxVV8CFWCZDBaG * feat(web): delete a knowledge base from the settings sheet (type-name confirm) Adds a Danger Zone section to KbSettingsSheet: a type-the-name confirmation gates deleteKb() (POST /api/v1/kb/delete); on success KbDetail navigates back to the KB list. New deleteKb client + kbSettings danger* keys + common cancel (zh + en, identical key sets). Claude-Session: https://claude.ai/code/session_01XMxbhmAkxxVV8CFWCZDBaG * feat(web): in-reader page edit + delete with impact preview (F2/F3) For an open concept/entity page, the reader header now shows Edit and Delete: - Delete: a dry-run first (deletePage dryRun) surfaces the backlink pages whose [[links]] will demote to plain text; a red confirm card lists them, then the real delete closes the page and refreshes the inventory. - Edit: a body textarea (frontmatter is preserved server-side), Save via PUT /api/v1/page, a links panel (getPageLinks: out/backlinks), a "recompile may overwrite" note, and a toast listing any dead links demoted to text. Adds the deletePage/editPage/getPageLinks wiki clients and kb:pageOps.* locale keys (zh + en, identical key sets). Build green (i18n guard + tsc + vite). Claude-Session: https://claude.ai/code/session_01XMxbhmAkxxVV8CFWCZDBaG * fix(workbench): address code-review findings (content-mgmt correctness + concurrency) Backend (page_ops / kb_admin / lint): - delete/edit now do their read-modify-write INSIDE the KB ingest lock and re-check the page exists under it: no stale backlink snapshot, no resurrecting a concurrently-deleted page, no 500 on a racing unlink (missing_ok=True). [#4,#5,#6] - delete_kb serializes rmtree under the ingest lock (was fully unlocked). Moved delete_kb/unregister_kb into a new openkb/kb_admin.py so config.py drops back under the per-file line gate (751 lines). [#1,#15] - page deletion no longer scans or rewrites lint's _EXCLUDED_FILES (AGENTS.md/SCHEMA.md/log.md); fix_broken_links(restrict_to) also refuses them, which protects `openkb remove` too. [#2] - edit_wiki_page treats `content` as the body verbatim — no naive frontmatter split that silently dropped a body opening with a '---' block. [#3] - delete uses the real on-disk stem for the exact-match index-entry removal (case-insensitive FS), and adds index.md to the demotion set so a [[target]] embedded in another entry's brief no longer dangles. [#7,#9] API / CLI: - delete-KB endpoint & CLI can now clean a ghost registry entry (registered name whose directory was removed by hand), catch delete_kb's ValueError -> 400, and the endpoint is typed `-> KbDeleteResponse`. [#8,#13,E5] Frontend (KbDetail): - Deleting a KB dispatches `openkb:reload-kbs` so the sidebar refreshes immediately (was stale until a manual reload) — live-test fix. - Removed the dead `pageReloadSeq` path (editPage always returns content). [#14] Tests: excluded-doc-untouched, '---'-body preservation, ghost-KB unregister; non-KB-target test updated to the new resolution model. Backend 1234 passed, mypy/ruff clean, frontend build green. Not changed (intentional): pages_linking_to is NOT folded into build_graph (that excludes explorations/ backlinks) [#10]; the preview+confirm delete inherently scans per request [#11]; EDITABLE_SECTIONS stays duplicated in the frontend, which can't import the Python constant [#12]. Claude-Session: https://claude.ai/code/session_01XMxbhmAkxxVV8CFWCZDBaG * feat(workbench): allow editing summary pages (edit ⊇ delete section sets) Summaries are compiled markdown like concepts/entities (a recompile regenerates them), so they are now editable via PUT /api/v1/page. They remain NOT independently deletable — a summary is removed by deleting its source document. Split page_ops into EDITABLE_SECTIONS (concepts/entities/summaries) vs DELETABLE_SECTIONS (concepts/entities); validate_page_ref takes the allowlist. Frontend ReaderBody gates Edit on canEdit (incl. summaries), Delete on canDelete. Test: summary editable (frontmatter preserved) + summary delete rejected (400). Claude-Session: https://claude.ai/code/session_01XMxbhmAkxxVV8CFWCZDBaG * fix(api): cross-platform delete_kb + endpoint OSError handling (xhigh review #1,#2) - delete_kb no longer holds the ingest lock DURING rmtree (the lock file lives inside the KB dir; Windows cannot delete an open file — the prior review-fix regressed this). It now takes the lock as a BARRIER (drain + wait out any in-flight mutation), releases it, re-checks existence, then rmtrees. [#1] - delete-KB endpoint maps FileNotFoundError to an idempotent success (a concurrent delete already removed the tree) and other OSError to a clean 500 with a message, instead of an uncaught 500 stack trace. [#2] Tests: endpoint OSError -> clean 500, FileNotFoundError -> 200 deleted. Claude-Session: https://claude.ai/code/session_01XMxbhmAkxxVV8CFWCZDBaG
KylinMountain
added a commit
that referenced
this pull request
Jul 22, 2026
…ane (#197) (#199) * feat(web): read a document's converted source text in the Documents pane (#197) The Documents pane listed each ingested doc (name/type/hash) with no way to read the converted full text ingestion produced under wiki/sources/. Backend: new POST /api/v1/document/source resolves a doc hash to its source text — short docs read <doc_name>.md; long docs concatenate the per-page <doc_name>.json (page text joined by a thematic break). Hash is the identifier (unique, avoids doc_name/stem collisions); resolution prefers the registry's stored source_path then falls back to the wiki/sources/<doc_name>.{md,json} convention, with a path-traversal guard. Read-only (sources are do-not-edit). Frontend: document rows are now clickable and open a wide read-only slide-out reader (MarkdownView) — ESC/overlay/close to dismiss, independent scroll, content cached + memoized per hash, native find-in-page preserved (no virtualization). Delete stays inline (stopPropagation). Closed drawer is inert. Known limitation: images embedded in long-doc pages are not rendered inline yet. * fix(web): address xhigh code-review findings for the document reader (#197) Correctness / a11y: - Rebuild the reader drawer on Radix Dialog (like KbSettingsSheet) instead of a hand-rolled overlay: proper modal focus trap, initial + return focus, Escape, and background inert (was: aria-modal with none of it) [#4]. This also removes the hand-rolled window keydown listener that re-subscribed every render [#7]. - Restructure each document row so the open-reader target is a real <button> and the delete control is a SIBLING, not nested. Keyboard-activating delete no longer bubbles into opening the reader, and the invalid nested-interactive markup is gone [#1, #5]. - Resolve a source file by the doc's own type (long → .json first, else .md), so two docs sharing a doc_name each resolve to their own file rather than whichever extension is tried first [#2]. - Guard source reads: skip non-dict page entries, reject non-list JSON, and return a controlled 500 on corrupt/unreadable sources instead of an unhandled exception [#3]. - Invalidate the per-hash content cache when the inventory changes, so a reopen after recompile refetches instead of serving stale text [#6]. Fetch/cache/memoized body moved to DocumentsPane so they survive the drawer's unmount-on-close. #8 (frontmatter stripping) intentionally not applied: source docs render verbatim (a user's own frontmatter is content, unlike wiki-page OKF metadata). Adds tests for the collision and malformed-JSON paths.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
litellm/prefix to allAgent(model=...)calls so non-OpenAI models (e.g.,anthropic/claude-sonnet-4-6,gemini/gemini-2.5-pro) are correctly routed through LiteLLM by the OpenAI Agents SDKlitellm.completion()andPageIndexClientcalls remain unchanged (they use LiteLLM format directly).env.exampleTest plan
gpt-5.4,openai/gpt-5.4,anthropic/claude-sonnet-4-6,gemini/gemini-2.5-pro) work with bothlitellm.completion()andAgent(model=litellm/...)usingLLM_API_KEY