Approved 2026-09-14. Coordinate through Linear and Git; no custom orchestration platform.
Provenance: Oliver's approval and SWE-2 correction in Codex task 01a09f3c-41a2-7e03-8103-00cadb4d71cb on 2026-09-14; Devin 3000.10.21 model inventory and Codex usage tool observed the same day. Project: https://linear.app/newth/project/lunch-money-mcp-5e3cbc5d645d
Durable storage: GBrain on m4mini, default source. Oliver explicitly requested saving this work there. Linear remains the execution tracker. Prior memory saves are historical copies; GBrain is the destination for subsequent durable records from this work.
| Task | Route |
|---|---|
| Dispatch, status, issue maintenance | Current coordinator |
| Bounded research, implementation, fixtures, tests, docs | Devin Local swe-2-high |
| Simple repetitive work | Devin Local swe-2-medium |
| Architecture, OAuth/tenant isolation, release review | Codex gpt-6-astra, high reasoning, through a delegated reviewer |
| Two unsuccessful fixes | Stop, re-diagnose, escalate with evidence |
Allowance ledger#
- Devin Local SWE-2: user confirms unlimited for the next month; exact expiry not supplied. CLI verified Medium/High/Max as Free on 2026-09-14. Recheck before dispatch. Do not infer cloud/Fusion/other models are free.
- Codex: 83% weekly allowance remaining at initial setup; live usage tool is authoritative at dispatch time. No paid credits or usage-reset credits reported.
- Other providers: unknown. No automatic billed fallback.
Execution contract#
- Start with at most two independent workers. Each gets one issue, explicit file ownership, source references, acceptance criteria and required validation.
- Use isolated feature-branch worktrees for repository edits. Discovery without an established repository uses separate scratch directories.
- Workers may not overwrite each other's work, deploy, merge, publish or change shared service configuration. Coordinator integrates reviewed changes within user authorization.
- Return findings/diff, exact validation commands/results, unresolved questions and continuation/session ID. Never include credentials or financial payloads.
- Record route and reason for each dispatch. Review evidence before dependent work; no issue is Done solely because an agent says so.
- Choose model based on risk, ambiguity, demonstrated results and verified allowance. The coordinator cannot change its own active model; stronger reasoning runs in a delegated task.
Initial batch#
- N-614 discovery: inspect public auth metadata and local auth references, identify real source/configuration gaps. SWE-2 High gathers evidence; stronger Codex review evaluates security implications.
- N-616 discovery: pin official Lunch Money v2 artifact, count operations, assess official client and production/preview distinction. SWE-2 High gathers evidence; implementation waits for repository/runtime decision.
This setup authorizes bounded initial discovery; the current batch does not create an external repository or deploy services.