2026-04-03: Organizational Model + Qwen Strategy

journal-2026-04-03-org-model-qwen Doc

private journal

Project

Context

Session started with Marcus unable to see phone numbers on the admin player profile. Led to three connected threads: fixing the bug, reviving Nemo, and rethinking how we organize work for small model execution.

The Phone Bug (DONE)

API contract mismatch: /players/{id} returns nested parent: { phone }, frontend expects flat parent_phone
/admin/players list endpoint missing parent_phone entirely
Two PRs merged: basketball-api#277, westside-landing#204
Validated live via Playwright. Sandra Apaisa's phone (385) 232-9501 visible on admin profile.
Marcus notified via GroupMe DM with the phone number.

API Contract Audit (IN PROGRESS)

Spike #278 found 9 mismatches, 4 HIGH severity
Root cause: API returns nested objects, frontend expects flat fields
Decision: fix at API layer (flatten responses)
7 tickets created (#280-#286), 5 reviewed READY, 2 re-scoped and approved
8 dev agents dispatched in parallel — PRs arriving
All touch different files so no conflicts until merge — then sequential rebase needed

Nemo Revival (SCOPED)

Pod running but dead — Anthropic credits at $0 since March 31
Model: switching to qwen3.5:4b (3.4 GB, downloaded locally, being pulled to cluster Ollama)
Architecture decision: purpose-built tools with typed I/O, not generic api_get
Architecture decision: tools inside Nemo for V1, extract to MCP server for V2 when second consumer exists
Architecture decision: use basketball-api REST endpoints (not direct SQL) — single source of truth
Two tickets: #30 infra (APPROVED, agent running), #29 app (NEEDS_REFINEMENT — oversized, decided to keep as one)
Superseded: #16 (provider flag) and #27 (SQL tool) — both closed

Nemo Tool Schemas (defined)

search_player(name: str) → str — player + parent contact info
list_roster(team_name: str) → str — all players on a team
payment_status(name?: str) → str — payment/subscription info
program_stats() → str — dashboard numbers

Tool handlers call basketball-api and return pre-formatted strings. qwen3.5:4b never sees raw JSON.

Organizational Model Brainstorm

Current system feels like bloat. 20 templates, 20+ conventions, 7 label categories. It works but the mental model is unclear.

Proposed hierarchy (Lucas's words)

Projects = collections of user stories. "Projects are actually users and what I do to make a user happy."
User Stories = collections of repos and kanban
Kanban = organized tickets
Tickets can have child tickets or child kanban if scope increases

Schema thesis (Lucas's key insight)

"If there is a schema of defined inputs and outputs then we can have good measurements of what is happening." The ticket IS the input schema. The PR IS the output schema. The review gate IS the validation layer. When every boundary has a typed interface, you can swap the executor (Claude → qwen) without changing the pipeline. All pressure is on ticket quality, not model quality.

DORA connection

The pipeline is a schema-driven execution chain: User Story → Ticket → Forgejo Issue → Agent → PR → QA → Validation. Every arrow is a typed interface. DORA metrics fall out naturally from measuring time through the chain.

Session Stats

2 bug fix PRs merged and validated
1 spike completed (9 mismatches found)
9 tickets created (7 API contract + 2 Nemo)
All 9 reviewed through gate
8+ dev agents dispatched
1 GroupMe DM sent to Marcus
2 superseded tickets closed

Next Session

Merge API contract PRs (sequential rebase needed)
QA + validate each
Close spike #278
Fix Nemo app ticket #29 scope, review, dispatch
Validate Nemo end-to-end: "What's the phone number for Apaisa?" in GroupMe
Organizational model redesign — continue from brainstorm
DORA reasoning for qwen — Lucas to articulate fully