Staging batch: Library AI by book count, Opus 5.5 default, Copilot on the relay

[manager]

Staging batch. Everything on this branch goes to staging; OffSec stays behind its dev-only gate there as it does now. Authorized by Wolf on 2026-09-25.

New on staging with this merge:

• Library: more than 15 books opens the AI shelf and the magic books without the flag (Library: more than 15 books opens the AI without the flag #244).
• Library: AI shelf and magic books default to Claude Opus 5.5; cover hotspots find their library by owner id.
• Copilot: Claude Sonnet 5 through the subscription relay, the paid OpenAI fallback removed; replies arrive whole instead of streamed.
• AI Atlas push work already released to production through feat(ai-atlas): Terminal pushes the Atlas guide, no rebuild needed #242 and feat(ai-atlas): the push sets the map, the cards and every page word #243 (merged from main, no new change).

Checks run on the merged tree: node scripts/release/library-batch-check.cjs PASS, tsc --noEmit clean.

🤖 Generated with Claude Code

[github-actions]

Claude finished @manager's task in 8m 1s —— View job

---

Review of PR #245

• Gather context (PR description, changed files list)
• Read diff (git diff origin/dev...HEAD — origin/dev's local copy is a shallow, unrelated-history ref, so I diffed each changed file directly against origin/dev instead)
• Review Library AI flag / cover hotspot changes
• Review AI Atlas adapter/store/stripGuide changes
• Review Copilot relay (llmClient, concierge, concierge-landing)
• Post inline comments for issues found
• Post final summary

One inline comment posted on src/lib/aiAtlas/stripGuide.ts (lines 116-126): the entries/steps/system.nodes validation only checks that id is a string — title/text/detail/children pass through unchecked, and adaptGuide()'s try/catch (the stated safety net) doesn't actually throw for most bad values in these fields, so a malformed Terminal push can be accepted and stored, then crash the page client-side later for a real visitor instead of being refused at push time. Details and a fix link are in the inline comment.

Everything else looked solid

• opensLibraryAi/countBooks (src/lib/library/flags.ts) are correct: > LIBRARY_AI_BOOKS_OVER (15) matches "more than 15 books," countBooks safely no-ops on a null library via optional chaining, and the server-side gate in ownerOfLibrary (src/lib/library/owner.ts) fetches the library with the caller's own token, so the private-shelves book count is always accurate for the real owner regardless of what the client UI has hydrated.
• /api/ai-atlas/guide.ts: bearer-token check hashes both sides with SHA-256 before timingSafeEqual, correctly avoiding the classic length-mismatch throw/timing leak; refuses cleanly when the key isn't configured; journals refused/failed/stored/live outcomes without leaking to unauthenticated callers.
• InteractiveCover/coverHotspots.ts: switching hotspot matching from username to a stable ownerId is a real correctness fix (usernames can change and silently break the binding); no hooks or SSR issues introduced.
• llmClient.ts / concierge.ts / concierge-landing.ts: removing the direct Anthropic/OpenAI calls in favor of the relay is a clean refactor, askClaudeJson's parameter order is consistent at both call sites, and the streaming→single-shot change matches what the PR description says was authorized. One thing worth keeping an eye on (not blocking, and explicitly called out as intended): the relay's job-polling path can in theory take much longer than the old direct/streaming call for a live chat widget — worth confirming the widget's own timeout/UX tolerates that.
• No App Router patterns, no Tailwind/CSS-in-JS, no new state libraries, no <img> for SVGs, no global CSS imports outside _app.tsx, no empty/named-export barrels touched. tsc --noEmit and library-batch-check.cjs are reported passing per the PR body.

I was not able to run eslint/tsc myself in this sandbox (network/tool restrictions on this run), so import-order/lint conformance is based on manual inspection, not a tool run — flagging that gap per instructions.

[github-actions]

Correctness: title/text/detail/children on entries, steps and system.nodes aren't type-checked, so a malformed push can slip past this "guide it cannot draw is refused" gate.

isRecordList (line 21-26) only checks that id is a string; everything else is copied through pick() unchanged. The comment at line 132 says "the page must be able to draw what is stored; a guide it cannot draw is refused," relying on adaptGuide(guide) throwing to catch bad data — but adaptGuide mostly assigns these fields raw (e.g. title: entry.title at adapter.ts:84) and only throws for a handful of hardcoded ids that happen to call a string method (node(id).title.toLowerCase() for layerIds, adapter.ts:217). A title that's an object on an arbitrary entry sails through stripGuide → gets stored → gets served, and only breaks when a visitor opens that specific dossier, since ai-atlas.tsx renders {data.title} directly as a JSX child (e.g. line 745) with no error boundary around the page — a full client-side crash for that visitor rather than a push-time refusal.

Worth adding an explicit string+length check for title/text/detail (and that children is an array of strings) on entries/steps/system.nodes, mirroring the rigor already applied to isCards/isCopy/stagesError just below.

Fix this →

[manager]

Fixed in 20428ab: title, location, text, role and basis must be text, detail and children lists of texts, each up to 4000 characters, on steps, entries and system nodes. A guide that fails is refused at push time and the page keeps the last stored one. Checked against the bundled guide (accepted), an object title and a number in a detail list (both refused).

[github-actions]

Claude finished @manager's task in 1m 15s —— View job

---

Review of PR #245 (updated push)

• Gather context (diff against origin/dev on the new HEAD — same shallow/unrelated-history caveat as before, diffed the changed commit directly)
• Check whether the previously flagged stripGuide.ts validation gap was addressed
• Review any other changes introduced by this push
• Post inline comments for issues found (none needed this round)
• Post final summary

This push (commit 20428ab) fixes the issue flagged in the previous review

src/lib/aiAtlas/stripGuide.ts now type-checks every kept field before a Terminal-pushed guide is stored:

• TEXT_FIELDS = ['title', 'location', 'text', 'role', 'basis'] must each be a string ≤ 4000 chars.
• LIST_FIELDS = ['detail', 'children'] must each be an array of such strings.
• fieldsError() runs over steps, entries, and system.nodes and is checked before the guide is assembled/stored.

I checked this against how each field is actually consumed in adapter.ts to make sure the shape matches reality, not just the strip logic:

• detail is read as an array in both places it's used — entry.detail || [entry.text] (line 85) and describe(systemRef(id), node(id).detail) (line 198) — so treating it as a list-of-strings is correct for both entries and system nodes.
• role/basis are read directly as scalars (badge: node(id).role, why: node(id).basis) — correctly TEXT_FIELDS.
• children is consumed as an array of entry ids (entries.get(child)) — correctly a LIST_FIELDS of strings.
• Every field in the pick() lists for steps/entries/system.nodes is now covered by either TEXT_FIELDS or LIST_FIELDS; nothing passes through unchecked anymore. null values are also caught (null !== undefined but typeof null !== 'string', so isText correctly rejects it).

This closes the gap cleanly: a malformed push with e.g. an object title is now refused at push time with a clear error instead of getting stored and crashing the page for a real visitor later.

One non-blocking note, not introduced by this PR: scripts/ai-atlas/strip-guide.mjs (the script that refreshes the bundled fallback, called out in the stripGuide.ts file header as keeping "the same field lists") doesn't have the equivalent type check — it still just does the raw pick(). That script isn't touched by this PR and its input is presumably admin-controlled rather than an arbitrary Terminal push, so it's lower risk, but worth a follow-up if that source ever becomes less trusted.

No other files changed in this push — everything else from the earlier full review (Library AI-flag gate, cover hotspot owner-id fix, Copilot relay refactor, AI Atlas guide endpoint auth) stands as previously assessed.