Coder presents model tiers, never vendor model names #40

Closed AtlantisPleb opened this 5d ago 2 comments

Encode the product invariant: the coder never shows a vendor model name to the user. The only names a reader sees are OpenAgents Coder tiers:

Coder Auto — the session's default: the server's default lane answers
Coder Flash — the fast tier
Coder Pro — the strong tier
Coder Local — a local model server answers; nothing leaves the machine

Scope:

The status line at the bottom of openagents coder names the tier, never the model id or a vendor label. A fresh session shows Coder Auto.
Shift+Tab cycles the tier: Auto → Flash → Pro → Local → Auto. Reasoning cycling moves to Tab.
Every other coder surface that named a model follows: fleet rows, resume picker, model-switch notices.
The tier map lives in one module, and the invariant is recorded in INVARIANTS.md with the test that holds it.

Acceptance: coder --dev starts a session whose bottom bar says Coder Auto; Shift+Tab flips through Flash, Pro, and Local; no vendor model name renders anywhere in the interface.

AtlantisPleb opened this issue 5d ago
AtlantisPleb closed this as completed in fb705a5 5d ago
A AtlantisPleb Author 5d ago

Shipped in fb705a506a: tier labels everywhere (status line, fleet, resume picker, notices, help), Coder Auto as the unpinned default with the proxy naming no model, shift+tab cycling Auto/Flash/Pro/Local with lazy source builds carrying the conversation, tab now cycling reasoning. INVARIANTS.md 'Coder Model Naming' records it; test/coder-tiers.test.ts holds it. 917/917 tests.
A AtlantisPleb Author 5d ago
Follow-up landed in 4773472fe2: the tier a reader reaches with shift+tab is now checked against what the deployment says it can answer.

The gap this closes is one the rename itself opened. chooseBackend already read availability from /api/v1/models for the model a session opens with, but buildTier pinned TIER_MODELS[tier] and opened a thread on it regardless — so the tier switch was the one path that never asked. Before #40 a reader who landed on a dead lane at least saw gemini-3.7-flash and could search for it; after #40 they saw Coder Flash fail and were told nothing. The friendlier name made the dead lane a better hiding place.

tierUnavailable in coder-tiers.ts reads the same catalog and the same available field, and applyPendingTier already did the right thing with a failed build, so a refused tier keeps the session on the tier that was answering and names the ones that can answer. An unreadable catalog allows the tier — "could not ask" is not "serves nothing" — and Coder Local is never refused here, since it answers from the reader's machine rather than the deployment.

Two things held deliberately:
- The refusal names tiers only. The invariant does not lapse when the news is bad, and a test asserts no vendor id reaches the message.
- Availability is the only claim. A test also asserts the refusal quotes no price, because the catalog declares a rate for some lanes and none for others — today gpt-5.6-luna behind Coder Pro has no pricing block at all — and a tier implying a cost the server never quoted would be a worse lie than the vendor name it replaced.
Eight tests in test/coder-tiers.test.ts, including a session-level one proving the session stays on the answering tier. Neutering the gate fails five of them. INVARIANTS.md "Coder Model Naming" records both rules.