bug(core): OpenCode Free pinned model list has rotted — 3 of 4 pinned models fail (429/400) and new free models never appear

[me2seeks]

What happened

OpenCode Free (the zero-credential default) is largely broken because its free-model set is a hardcoded pin that has rotted, and new free models never reach users.

The pin lives in packages/core/src/provider-registry.ts:604-609 (opencodeFreeModelIds) and the default enabled set at packages/core/src/provider-registry.ts:8-13:

• pinned: nemotron-3-ultra-free, mimo-v2.5-free, big-pickle, deepseek-v4-flash-free
• default enabled: nemotron-3-ultra-free, mimo-v2.5-free, deepseek-v4-flash-free

Live probes against https://opencode.ai/zen/v1/chat/completions (anonymous, 2026-08-21):

model	result
nemotron-3-ultra-free	200 OK
mimo-v2.5-free	429 FreeUsageLimitError — "Rate limit exceeded. Please try again later."
big-pickle	429 FreeUsageLimitError — same
deepseek-v4-flash-free	400 server_error — "Upstream request failed: Model is unavailable."

So 3 of the 4 pinned models are dead: users get stuck in Error: Rate limit exceeded loops, and the TUI even logs confusing notices like Model changed: mimo-v2.5-free → mimo-v2.5-free when re-electing within the stale set.

At the same time, the free catalog upstream has moved on. models.dev currently lists ~29 free opencode models; several work anonymously today but are absent from the pin, e.g.:

• x-preview-f-free — displayed as "Ox Alpha Free (Unlimited)", tool-capable, answers 200 anonymously
• laguna-s-2.1-free — 200
• nemotron-3.5-lightning-free — 200

Because the pin is as const, none of these ever appear in the model picker until someone edits the list by hand. The repo's own generated snapshot lags too: as of today's refresh (#3324) model-metadata.generated.ts still has no x-preview-f-free.

One more wrinkle for whoever fixes this: the anonymous-served set is strictly smaller than the models.dev free set. Many snapshot-listed free ids (glm-5-free, kimi-k2.5-free, grok-code, minimax-m3-free, …) answer 401 ModelError: Model … is not supported anonymously, so deriving the free list from cost.input === 0 alone would over-offer. Validation needs to hit the live endpoint (or at least treat 401 ModelError as "not anonymously served").

Suggested directions (maintainer's call):

• stop hardcoding the free set; derive candidates from the models.dev snapshot and health-check them (the testConnection fallback-probe machinery already does most of this)
• or keep a pin but add a scheduled check that fails CI/lint when a pinned id stops answering 200 anonymously
• treat 429 FreeUsageLimitError / 400 "Model is unavailable" as rotation signals so a dead enabled model doesn't trap the session

How to reproduce

1. Start Maka fresh with no credentials — the bootstrap seeds the opencode-free connection with enabled models [nemotron-3-ultra-free, mimo-v2.5-free, deepseek-v4-flash-free] (packages/runtime-host/src/server/bootstrap-runtime-policy.ts:76-83).
2. Switch the model to mimo-v2.5-free (or exhaust the per-IP quota on the default), then send any message.
3. Turn fails with Error: Rate limit exceeded; picking deepseek-v4-flash-free fails because the model is unavailable upstream; big-pickle is likewise 429.
4. Open the model picker on the OpenCode Free connection — only the stale four are offered; none of the newly added free models (e.g. Ox Alpha Free) exist.

Direct curl equivalent:

curl -s https://opencode.ai/zen/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"model":"mimo-v2.5-free","messages":[{"role":"user","content":"hi"}],"max_tokens":1}'
# {"type":"error","error":{"type":"FreeUsageLimitError","message":"… Rate limit exceeded …"}}

Environment

• Maka version or commit: 57e08d834 (main, 2026-08-21)
• OS and version: Linux
• Surface: all surfaces that use the opencode-free connection (Desktop / TUI / CLI)
• Node.js version, if running from source: 22.x

Logs, screenshots, or additional context

• This is the same failure class as fix(opencode-free): replace unhealthy big-pickle bootstrap default #2165, which swapped one unhealthy pinned default (big-pickle) for another; the pin has rotted again since. A one-off swap will keep recurring.
• Related: Fallback model lists still offer deprecated models for six providers #3355 (fallback lists offering retired models), Persisted connection catalogs keep model ids that the snapshot has retired #3354 (persisted catalogs keeping retired ids).
• Probe summary (anonymous, 2026-08-21): working — nemotron-3-ultra-free, x-preview-f-free, laguna-s-2.1-free, nemotron-3.5-lightning-free; rate-limited — mimo-v2.5-free, big-pickle; unavailable — deepseek-v4-flash-free; not anonymously served (401 ModelError) — most other -free ids in the snapshot.

_{Diagnosis and endpoint probes above were prepared with assistance from an AI coding agent (opencode); the reporter verified the results.}

[cunninghamcard-bit]

I would like to help

[me2seeks]

Update: the free tier is shrinking in real time, which makes the pin rot faster than any manual swap can track.

Same-day snapshot diff (models.dev, 2026-08-21):

• deepseek-v4-flash-free — now lifecycle: 'deprecated' (matches the 400 "Model is unavailable" probe above)
• laguna-s-2.1-free — now lifecycle: 'deprecated' too (it answered 200 this morning)
• active tool-capable free set is down to 7 ids, including x-preview-f-free ("Ox Alpha Free (Unlimited)")

A minimal un-pinned design exists and is prototyped locally:

1. scripts/sync-model-metadata.mjs emits an isFree: true marker when models.dev prices a model at zero input cost (one-line change; the generator already parses cost).
2. opencodeFreeModelIds / OPENCODE_FREE_DEFAULT_ENABLED_MODELS become a derivation over that marker (isFree && functionCalling && !deprecated, default model first) instead of a hardcoded list. A module-load guard fails loudly if the snapshot ever loses the default.
3. No runtime discovery or rotation machinery needed for this fix — the existing testConnection fallback probing already skips unhealthy candidates, and new/retired free models now flow through the routine snapshot refresh (chore: refresh the models.dev catalog snapshot #3324-style chores).

With this, today's prune lands automatically on the next refresh: deprecated ids drop out, x-preview-f-free appears, and fresh installs no longer seed dead defaults.

_{Follow-up analysis prepared with the same AI agent assistance disclosed in the issue body.}

[shaokeyibb]

Maybe consider read https://opencode.ai/zen/go/v1/models to get a full list of opencode models, tag the free model when the suffix end with Free?

[me2seeks]

Thanks for the suggestion — the live /models idea is directionally right and worth recording, but two corrections from probing the endpoints today:

1. Wrong endpoint for the free tier. https://opencode.ai/zen/go/v1/models is the OpenCode Go (paid plan) catalog — it returns minimax-m3, kimi-k3, etc., essentially no free models. The anonymous free tier lives on https://opencode.ai/zen/v1/models, which does work without credentials (64 ids today, free subset matches the derived set exactly).

2. A "Free" suffix rule is not a reliable tag. Current counter-examples from the live catalog:

id	displayName	ends with "Free"?
big-pickle	Big Pickle	no "Free" at all
x-preview-f-free	Ox Alpha Free (Unlimited)	no — "(Unlimited)" suffix

Display names are not a contract. models.dev's cost.input === 0 is structured, authoritative, and already flows through our sync pipeline (scripts/sync-model-metadata.mjs parses cost for pricing), so the prototyped fix derives the free set from that marker plus functionCalling and lifecycle !== 'deprecated'.

One more caveat if you pursue discovery: the served /zen/v1/models list is not a health filter either — today it still lists deepseek-v4-flash-free and laguna-s-2.1-free, both deprecated and returning errors. Lifecycle filtering stays necessary regardless of source.

A good future enhancement would be intersecting the snapshot-derived free set with the anonymous zen/v1/models listing to also drop "declared free but not anonymously served" ids. For the minimal fix the derivation alone already resolves this issue's rot.

[cunninghamcard-bit]

Opened #3429 implementing the derivation approach discussed above — isFree marker from models.dev's cost.input === 0 in the metadata sync, free set derived as isFree && functionCalling && !deprecated with a loud module-load guard on the default model, plus a snapshot refresh so today's prune lands immediately (deprecated ids out, x-preview-f-free in).

Thanks @me2seeks for the detailed probes and the derivation design — the endpoint corrections and the "display names are not a contract" counter-examples shaped the fix directly — and @shaokeyibb for raising the live /models discovery idea; the snapshot-vs-anonymous-listing intersection is recorded in the PR as a possible follow-up.