LQ_AI backend — asks for Donna (relay index)

From: Donna frontend session · To: the LQ_AI backend session · Date: 2026-05-31 (status updated 2026-06-01) · Pin Donna is on: 945ad31

How to use this file. You (the LQ_AI session) work in the repo at /Users/kevinkeller/Code/lq-ai. These request docs live in the Donna repo at /Users/kevinkeller/Code/Donna/docs/upstream-requests/ — read them directly by absolute path. Each detailed doc lists the exact lq-ai files to change (also by absolute path). When you merge any of these, tell Kevin the merged SHA; Donna then bumps its vendor/lq-ai submodule pin, runs npm run gen:api, rebuilds, verifies live, and logs the bump in /Users/kevinkeller/Code/Donna/docs/decisions/lq-ai-pin.md.

Status (2026-06-01): P1.1–P1.4 all landed and Donna is pinned at 945ad31 — see Already landed below. No open asks remain. The P1.1–P1.4 sections are kept below for reference. (The bigger "autonomous workflows" item is not here — Donna tracks it in its own future roadmap: /Users/kevinkeller/Code/Donna/docs/roadmap/donna-future-roadmap.md. You're building that backend; the consumer-side requirements Donna will need are captured there for later.)

---

P1.1 — Make skill_inputs actually reach the model for non-templated skills

• Detailed request: /Users/kevinkeller/Code/Donna/docs/upstream-requests/lq-ai-skill-inputs-corpus.md
• lq-ai files:
  • /Users/kevinkeller/Code/lq-ai/gateway/app/skills/assembler.py (interpolate / _render_skill / assemble_skill_prompt)
  • test: /Users/kevinkeller/Code/lq-ai/gateway/tests/test_inference_skill_assembly.py
• Gist: MessageCreate.skill_inputs is accepted + forwarded as lq_ai_skill_inputs, but the assembler only substitutes {{placeholder}} tokens and silently drops unreferenced inputs. No built-in SKILL.md is templated, so collected inputs vanish for every built-in. Recommended fix (Option A): after interpolation, append any unreferenced bound inputs as a short labelled context block, so every skill benefits with no corpus edits.
• Unblocks (Donna): the deferred composer skill-input form.

P1.2 — Add MessageCreate.file_ids (per-message chat file attachment)

• Detailed request: /Users/kevinkeller/Code/Donna/docs/upstream-requests/lq-ai-chat-message-file-attach.md
• lq-ai files:
  • /Users/kevinkeller/Code/lq-ai/api/app/schemas/chats.py (add optional file_ids: string[] to MessageCreate)
  • the chat-message send handler + gateway forwarding (the path that already forwards lq_ai_skills)
• Gist: files can already be uploaded via POST /api/v1/files, but MessageCreate has no file field and there's no POST /chats/{id}/files, so a client holding file_ids can't say "this turn is about these files." Add optional file_ids?: string[], validate caller-owned ids (404/422, id-probing-safe), forward to the gateway as document context, and echo the applied ids on the message + SSE complete frame (like applied_skills). Please also document how this interacts with skill_inputs of type:"file".
• Unblocks (Donna): composer file picker / drop-zone for ad-hoc per-turn document attachments.

P1.3 — Add PATCH /api/v1/users/me (edit display_name / email)

• Detailed request: /Users/kevinkeller/Code/Donna/docs/upstream-requests/lq-ai-patch-users-me-profile.md
• lq-ai files:
  • /Users/kevinkeller/Code/lq-ai/api/app/api/users.py (add PATCH /users/me)
  • /Users/kevinkeller/Code/lq-ai/api/app/schemas/ (a UserProfileUpdate model)
• Gist: GET /users/me exposes display_name/email and PATCH /users/me/preferences exists, but there's no PATCH /users/me to edit the profile fields — /users/me has patch?: never. display_name-only is enough to unblock; email self-service is your call. This is the only thing blocking the editable-profile part of Donna's upcoming Settings page — everything else that page needs already exists.
• Unblocks (Donna): the profile-edit form in P7 Settings (the rest of P7 ships without it).

P1.4 — Expose deletion_scheduled_at on GET /users/me

• Detailed request: /Users/kevinkeller/Code/Donna/docs/upstream-requests/lq-ai-expose-deletion-status-on-users-me.md
• lq-ai files:
  • /Users/kevinkeller/Code/lq-ai/api/app/api/users.py (GET /users/me)
  • /Users/kevinkeller/Code/lq-ai/api/app/schemas/ (the User response model)
• Gist: deletion_scheduled_at lives only on AdminUserRow, not on the User schema from GET /users/me, so a normal session can't detect a pending deletion on load. Donna ships an always-visible "Cancel scheduled deletion" link as a workaround. Add deletion_scheduled_at (or a small status field/endpoint) to /users/me so Donna can show a proper conditional "Pending deletion — cancel by " banner instead.
• Unblocks (Donna): the conditional pending-deletion banner in P7-2 (the always-on cancel link ships without it).

---

After any of these merge

1. You: report the merged SHA to Kevin.
2. Donna: cd vendor/lq-ai && git fetch && git checkout <sha> → npm run gen:api → rebuild affected containers → verify live → log the bump in /Users/kevinkeller/Code/Donna/docs/decisions/lq-ai-pin.md → build the unblocked slice.

Already landed (no action — for your reference)

• P1.1 skill_inputs → non-templated skills (DE-328) — lq-ai #115 396e19f
• P1.2 MessageCreate.file_ids channel (Part A) — lq-ai #116 1592063
• P1.2 file content → model, verbatim (Part B) — lq-ai #117 6dae35d
• P1.3 PATCH /users/me display_name (UserProfileUpdate; email edit deferred → DE-329) — lq-ai #118 e9659da
• DE-329 filed + DE-328 marked resolved — lq-ai #119 badf83d
• P1.4 deletion_scheduled_at on GET /users/me — lq-ai 945ad31 (Donna pin PR #42); resolves lq-ai-expose-deletion-status-on-users-me.md. Consumed by the conditional pending-deletion banner on /settings/data.

All merged to lq-ai main; Donna pinned at 945ad31 as of 2026-06-01 (see docs/decisions/lq-ai-pin.md).

Earlier: anonymization-in-receipts (#102), streaming inference-routing-log (#103), /v1/models alias-field docs (#105) — all merged and previously pinned at 438198c.