yolo/arr
3
0
Fork 0
Code Issues 12 Pull Requests Actions Packages Projects Releases Wiki Activity

docs: pair two models per difficulty tier

Browse Source
This commit is contained in:
Claude
2026-08-22 18:52:35 +01:00
parent dfd8440e10
commit 18fc1d1d44
1 changed files with 22 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CLAUDE.md
+22 -15
View File
@@ -131,35 +131,42 @@ alone. The loop:
### Model selection
Sessions are spawned with `aoe`. **Four models, and only these four** — Opus 5
and Fable 5 through the `claude` tool, Sol and Terra through `codex`.
Sessions are spawned with `aoe`. Each difficulty tier has **two models**, and
issues at that tier are distributed between them — not run twice. Splitting
across two providers keeps a single provider's rate limits or an outage from
stalling the whole queue, and keeps one model's blind spots from shaping the
entire codebase.
| Difficulty | Model | Spawn |
| Difficulty | Models | Spawn |
|---|---|---|
| `difficulty/hard` | Fable 5 | `aoe add <path> --tool claude --extra-args "--model claude-fable-5" -l` |
| | Sol | `aoe add <path> --tool codex --extra-args "-m gpt-5.6-sol" -l` |
| `difficulty/moderate` | Opus 5 | `aoe add <path> --tool claude --extra-args "--model claude-opus-5" -l` |
| `difficulty/easy` | Sol | `aoe add <path> --tool codex --extra-args "-m gpt-5.6-sol" -l` |
| | Sol, low effort | `aoe add <path> --tool codex --extra-args "-m gpt-5.6-sol -c model_reasoning_effort=low" -l` |
| `difficulty/easy` | Sonnet 5 | `aoe add <path> --tool claude --extra-args "--model claude-sonnet-5" -l` |
| | Terra | `aoe add <path> --tool codex --extra-args "-m gpt-5.6-terra" -l` |
| `difficulty/trivial` | Terra | `aoe add <path> --tool codex --extra-args "-m gpt-5.6-terra" -l` |
Alternate within a tier rather than draining one model first, so a bad run is
visible early instead of after ten issues.
Use `-w <branch> -b` to put each session in its own git worktree, which is what
makes parallel issues safe.
Two caveats on the table above, both cheap to fix and worth fixing before a long
unattended run:
Caveats worth resolving before a long unattended run:
- `gpt-5.6-terra` is confirmed — it is the default in `~/.codex/config.toml`.
`gpt-5.6-sol` is inferred from the sibling naming and unverified.
- Opus 5 and Fable 5 are both frontier models; the split between them at hard
and moderate is about reasoning depth, not capability tier. Sol above Terra is
an assumption — swap the rows if it is backwards.
- `area/web` issues run on Fable 5 regardless of their difficulty label, since
they go through the `impeccable` skill.
- `-c model_reasoning_effort=low` is the documented codex config-override form
but has not been exercised here.
Escalate one tier if a session fails CI twice on the same issue. Never
de-escalate mid-issue.
**`area/web` issues run on Fable 5 regardless of their difficulty label**, since
they go through the `impeccable` skill and the UI is the reason this project
has a frontend at all (see the design section above).
`area/web` issues ignore this table's lower tiers — see the design section
above. They go through the `impeccable` skill regardless of difficulty label.
Escalate one tier if a session fails CI twice on the same issue — and escalate
to the *other* model at that tier first, before going up. Never de-escalate
mid-issue.
### What a driver must not do