feat: run each loop phase on its own strength (Thrifty, Balanced, Best)

AI Assisted
This commit is contained in:
jparkerweb committed 2026-11-04 19:23:52 -08:00
1 parent e963513a8a
commit 81d14c8e78
41 files changed
+3290 -81

No files matched your search

@@ -21,7 +21,7 @@ Run from the repo root (the folder holding `install.js` and `src/launcher/`). Ru
### Step 2 — Read the current files
Read `src/launcher/models.json`. It has four keys: `_note` (a string), `claude`, `devin` and `codex` (each an array of `{ "id", "label" }`; `codex` is empty today, because Codex's docs name no models to verify and the launcher skips the model question for it). Keep the `_note` text exactly as read — Step 9 writes it back verbatim and the Validation rules check it.
Read `src/launcher/models.json`. It has five keys: `_note` (a string), `claude`, `devin` and `codex` (each an array of `{ "id", "label" }`) and `strengths` (an object, see Step 4c). Keep the `_note` text exactly as read — Step 9 writes it back verbatim and the Validation rules check it.
Read `src/launcher/agents.json`. It has `schema` (currently `1`) and `agents`, a list of entries. Each entry carries `id`, `name`, `kind` (`builtin` or `template`), `command`, `prompt`, `args`, `bypassArgs`, `modelArgs`, and optional `note`. A `builtin` entry also carries `installUrl`, `install` (lists under `win`, `mac` and `linux`, each `{ label, run, note? }`) and `installDirs` (the same OS keys). Keep the file's shape and order exactly as read.
@@ -49,6 +49,10 @@ If the lookup fails or the page does not list aliases, say so, and switch to **g
Look up Codex's current model documentation (locate it on each run). Propose `codex` entries only for model ids the docs name explicitly. If they name none, say so and leave `codex` empty. If entries are added, the `_note` sentence about Codex goes stale; keep `_note` verbatim as everywhere else, but tell the maintainer so they can reword it by hand. If the lookup fails, use guided manual edits as in Step 3.
### Step 4c: Check the strengths mapping
`models.json` also has a top-level `strengths` object: for an agent id, the loop's `thrifty`, `balanced` and `best` words, each `{ "model", "effort"? }`. After Steps 3 to 4b settled the refreshed lists, check that every `strengths` row still names a `model` that is in that agent's refreshed list (an `effort` is kept only for an agent that takes one, and is `low`, `medium` or `high`). For a row whose model is **gone**, propose a replacement from the same family, one step lighter or stronger as the word implies (Thrifty lighter, Best stronger), and show it in Step 7's table as a `strengths.<agent>.<word>` change. Never add a mapping for an agent that has none: only agents already under `strengths` are checked.
### Step 5: Re-check the agent list against vendor docs
For each entry in `agents.json`, locate the vendor's current install documentation on each run (search for it; do not use a remembered URL, since docs move and `installUrl` itself may be stale). Compare the file against what the docs say today:
@@ -136,6 +140,7 @@ Check every rule before `npm test` (Step 10). They mirror the shipped-file tests
- Every entry in `claude`, `devin` and `codex` has a non-empty string `id` and a non-empty string `label`.
- Ids are unique within each list, and labels are unique within each list.
- No Devin id ends in `xhigh`, `max`, `fast` or `priority`.
- Every `strengths` row names a `model` that is in that agent's list in the same file, uses only the words `thrifty`, `balanced` and `best`, and no agent other than the ones already under `strengths` gains a mapping.
**`agents.json`:**