Skip to content

Maple Work: catalog-only models with a server-suggested default - #1360

Merged
AnthonyRonning merged 1 commit into
masterfrom
agent/work-model-defaults
Oct 29, 2026
Merged

AnthonyRonning merged 1 commit into
masterfrom
agent/work-model-defaults

Conversation

@AnthonyRonning

@AnthonyRonning AnthonyRonning commented Oct 29, 2026 •

Copy link
Copy Markdown
Contributor

Summary

Maple Work picks its models more carefully and stops flashing the model chip at launch.

  • One endpoint. Maple Work reads its model list, labels and voice endpoints from /v1/models/catalog only, cached on disk; it no longer calls /v1/models. API-only models such as gpt-oss-safeguard-120b drop out of the menu, and the chip never shows a raw model id before its label.
  • Server-suggested default. The catalog gains an optional defaults.agent: DeepSeek V4.1 Flash on paid plans, gpt-oss-120b on Free. Older clients ignore it. Changing the default later is a backend change only.
  • The user's pick comes first. A new task starts on the user's last manual pick while the catalog still offers it, then on defaults.agent, then on the built-in default (now deepseek-v4-1-flash), then on the catalog's first model. A removed model is never loaded by default. Tasks still lock to the model of their first message.
  • No launch jitter. List, labels and pick land in one update. Until the model resolves, the composer controls are laid out but hidden, then fade in once (motion::fade_in, 140 ms) in their final positions.

Compatibility

  • Backend: additive JSON field. Older Maple Work builds and Research ignore it.
  • Saved pick: stored under a new account setting, agent.model_choice. agent.defaults is rewritten on every runtime start and uses deny_unknown_fields, so a new field there would break settings loading in older builds. The pick is still mirrored into defaultModel for older builds.
  • Older saved defaults: a defaultModel saved by an older build counts as a pick, once, unless it equals the old built-in glm-5-3. Someone who deliberately chose glm-5-3 moves to the new default.
  • Older remote computers: if the execution computer runs an older build, its list (from /v1/models) is filtered to models the controller's catalog describes. Its default still follows the old logic.
  • No catalog at all (offline, first launch failure): the pick, else the model the runtime started on, as before.
  • Disk cache: the cache.model_catalog entry from older builds (bare ids) is treated as a miss and refetched.

Free-plan gating (locks on Pro models in Free's menu) is out of scope; Maple Work is not offered to Free users.

Validation

Local mde stack (OpenSecret + billing + Continuum), real app driven with cua-driver, launches recorded and inspected frame by frame (30 fps):

Master This PR
Default (Pro, no pick) GLM 5.3 DeepSeek V4.1 Flash
Launch frames, composer row "Model" + "Starting Maple…", then "GLM 5.3": chip widens, row shifts, mic pops in "Starting Maple…" alone, then the full row fades in at its final position; nothing moves
Menu includes raw gpt-oss-safeguard-120b labelled catalog models only
  • Picked Kimi K3, relaunched, and signed out and back in: Kimi K3 kept, ahead of the server suggestion.
  • Free fixture: defaults to GPT-OSS.
  • Real run: a new task on DeepSeek V4.1 Flash answered a no-tools prompt.
  • apps/maple-agent: just clippy (all feature sets) and just test (workspace + headless) pass. New tests cover the pick, suggestion and fallback order, cache persistence across restart, filtering of disabled, deprecated, alias and non-chat entries, and adoption of an older build's pick.
  • services/opensecret: fmt, clippy --all-features -D warnings, and the model_config tests pass, including a new test that defaults.agent names a listed, enabled, tool-capable model the plan can use.
  • sdk/rust: fmt, clippy, lib tests pass. TS SDK change is type-only; its test suite was not run locally (needs a backend on :3000).

🤖 Generated with Claude Code


View with [code]smith Autofix with [code]smith
Need help on this PR? Tag @codesmith-bot with what you need.

Maple Work now reads its model list, labels and voice endpoints from
/v1/models/catalog alone, cached on disk, instead of joining /v1/models ids
with catalog labels in two requests. API-only models such as
gpt-oss-safeguard-120b no longer appear, and the chip never shows a raw id
before its label.

The catalog gains an optional defaults.agent: the concrete model a new
Maple Work task starts on (DeepSeek V4.1 Flash on paid plans, the free
quick model otherwise). Older clients ignore it. A new task starts on the
user's last pick while the catalog offers it, then the suggestion, then the
built-in default (now deepseek-v4-1-flash), then the catalog's first model,
so a removed model is never loaded by default. Picks are stored under their
own account setting, because agent.defaults is rewritten on every runtime
start and rejects unknown fields in older builds; a pick saved by an older
build is adopted once.

At launch the composer controls stay laid out but hidden until the model
resolves, then fade in once in their final positions.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@AnthonyRonning
AnthonyRonning enabled auto-merge (squash) October 29, 2026 02:44
@AnthonyRonning
AnthonyRonning merged commit b4cfef9 into master Oct 29, 2026
27 checks passed
@AnthonyRonning
AnthonyRonning deleted the agent/work-model-defaults branch October 29, 2026 02:44
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant