Architecture decision records
Every important or major change to ai-coach ships with an ADR in this folder; ADR-0001 says which changes count. Copy template.md, take the next number, and add a line below.
| ADR | Decision | Status |
|---|---|---|
| 0001 | Record every important change in an architecture decision record | accepted |
| 0002 | Coach in two tiers: offline Gemma, zero-knowledge Gemini Argon | accepted |
| 0003 | Reach the vault only through kb-template's MCP vault server | accepted |
| 0004 | Only the Tier 2 payload leaves the machine | superseded by ADR-0021 |
| 0005 | Run Tier 1 as a Socratic dialogue with banned buzzwords and bounded follow-ups | accepted |
| 0006 | Judge Tier 2 by blind claim extraction, and pass only flawless answers | accepted |
| 0007 | Schedule reviews on a fixed ladder read from the vault's session history | accepted |
| 0008 | Structure ai-coach as ports and adapters, enforced by Nx module boundaries | accepted |
| 0009 | Drive the coach from a terminal CLI that previews every write | accepted |
| 0010 | Test each layer on its own, and drop tests that need the kb-template server for now | accepted |
| 0011 | Source code carries no comments | accepted |
| 0012 | Spar book chapters, never the book index | superseded by ADR-0018 |
| 0013 | Set Gemma's context window explicitly and refuse prompts that would not fit | accepted |
| 0014 | Anchor Tier 1 on a focus, and keep the note out of the challenge and the dialogue | accepted |
| 0015 | Load secrets from a git-ignored .env file | accepted |
| 0016 | Check each Tier 1 gap against the transcript before it stands | accepted |
| 0017 | Study the gaps of a failed round with hints and the author's own words | accepted |
| 0018 | Spar concept notes against their core assertions; raw material is reference only | accepted |
| 0019 | Coach in one of four styles, chosen per note | accepted |
| 0020 | Asking for help in study never costs an attempt, and a bad model reply never ends a session | accepted |
| 0021 | Only the Tier 2 payload leaves the machine, and it may carry core assertions | accepted |
| 0022 | Ask which knowledge notes to keep before archiving a project | accepted |
| 0023 | Distill concept notes from the learner's own words, with the local model as helper | accepted |
| 0024 | Drop a Tier 1 gap only when the learner's own words cover it | proposed |
| 0025 | Judge each answer by the bar of the style it was sparred in | proposed |
| 0026 | Quote whole marked sentences, ask why an idea matters, and never repeat a study quote | proposed |
| 0027 | Decide a concept round by checking each core assertion on a larger local model | proposed |
| 0028 | Call making concept notes synthesis, and keep distillation for marking | proposed |
| 0029 | Plan the vault with agent skills that use only the vault tools | superseded by ADR-0030 |
| 0030 | Plan the vault with the full set of vault planning tools, and pass every preview back on execute | proposed |
| 0031 | Plan habits as definitions checked off per day, never carried over | proposed |
| 0032 | Publish the docs with Docusaurus, built from the Markdown where it lives | proposed |
| 0033 | Write commit messages as Conventional Commits, checked in pull requests | proposed |
| 0034 | Split the docs into a short README, a user guide and a contributing guide | proposed |