_maybe_add_1m_suffix assumes [1m] model variants exist; gateway 400s when they don't

[fflory]

Summary

_maybe_add_1m_suffix infers the existence of a [1m] 1M-context model variant
from the version number alone. On a workspace whose AI Gateway does not serve
those variants, configure writes a model name the gateway rejects, and the
affected tier stops working with no actionable error.

Details

src/ucode/agents/claude.py:375-390:

should_suffix = (family == "opus" and (major, minor) >= (4, 6)) or (
    family == "sonnet" and (major, minor) >= (4, 6)
)
return f"{model}[1m]" if should_suffix else model

The decision is purely version-derived. Nothing checks whether the workspace
advertises the suffixed variant — even though /ai-gateway/anthropic/v1/models
returns an authoritative list, and on my workspace that list contains
databricks-claude-opus-4-8 but no [1m] form.

Result against that gateway (POST /ai-gateway/anthropic/v1/messages):

Model name	HTTP
databricks-claude-opus-4-8[1m]	400 Invalid name
databricks-claude-opus-4-8	200
databricks-claude-sonnet-4-6[1m]	400 Invalid name
databricks-claude-sonnet-4-6	200

There is also no escape hatch: the only UCODE_* env knobs are UCODE_DEBUG,
UCODE_HOME, and UCODE_WEB_SEARCH_MODEL, and no CLI flag disables the
suffixing. If discovery selects an affected model as a family model, that tier
is unusable until the settings file is hand-edited — which configure then
overwrites.

Why it is currently latent (and why that is fragile)

The regex at claude.py:88 requires a major and minor version:

r"^(?:system\.ai\.)?(?:databricks-)?claude-(opus|sonnet)-(\d+)-(\d+)(.*)$"

So claude-opus-5 and claude-sonnet-5 — no minor component — never match, and
never get suffixed. That is the only reason my configure produced a working
config: discovery happened to pick the -5 models. Selecting a 4.6/4.7/4.8
model as a family model, on this same workspace, would have produced a broken
tier. The safety here is incidental to a naming coincidence, not by design.

Suggested fix

Gate the suffix on the model actually being served, rather than on its version:
check the /v1/models listing (or the model-services listing) for the [1m]
form and only emit it when present. Falling back to the unsuffixed name is
strictly safer than emitting a name the gateway will reject — 1M context is an
enhancement, so losing it silently is a much better failure mode than a
non-functional tier.

A --no-1m / UCODE_DISABLE_1M escape hatch would also be a cheap
independent mitigation.

Environment

• ucode 0.0.0+174.g62c5536 (commit 62c5536, 2026-08-07)
• Databricks CLI 1.10.0, macOS 15 (Darwin 25.6.0)
• AI Gateway v2 (/ai-gateway/anthropic) on a Databricks-on-AWS workspace
• Agent: Claude Code (native install)