feat: /model autocomplete from a configurable :models list, plus /effort

[HelgeSverre]

/model argument autocomplete

The slash palette now completes /model's argument: once the input reads /model <partial>, it fuzzy-filters a picker list shown as Provider: Model Name — e.g. Anthropic: Claude Opus 4.6 claude-opus-4-6. Tab inserts the selected id, Enter runs it, and any id typed by hand still works, so the list never gates model choice. The active model is marked ●, and providers whose API-key env var is unset are annotated · no key. /model with no argument prints the same list.

The list is config data in init.sema under a new :models key, hot-reloaded like everything else:

:models
(list
  (provider "Anthropic"
    (list (model "claude-opus-4-6"  "Claude Opus 4.6")
          (model "claude-sonnet-5"  "Claude Sonnet 5")
          (model "claude-haiku-4-5" "Claude Haiku 4.5")))
  (provider "OpenAI"
    (list (model "gpt-5.6-sol"   "GPT-5.6 Sol")
          (model "gpt-5.6-terra" "GPT-5.6 Terra")
          (model "gpt-5.6-luna"  "GPT-5.6 Luna")
          (model "gpt-5.5"       "GPT-5.5"))))

All default ids are verified against the providers' published model lists (the GPT-5.6 family ids come from OpenAI's July 2026 release; the bare gpt-5.6 alias routes to Sol, so the explicit -sol/-terra/-luna ids are used).

The mechanism is generic: register-completions! attaches a completion source to any command. Sources receive the live state map and may mark an entry :active. The palette's Enter/Tab logic lives in pure palette-choice / palette-completion functions, covered headlessly by tests.

/effort

Sema's agent/run accepts a portable :reasoning-effort option (none/minimal/low/medium/high/xhigh) that maps to each provider's native control and no-ops on models without reasoning support.

• /effort <level> sets it for the session (with palette autocomplete, active level marked); /effort default resets; bare /effort shows the current level.
• (model id label {:effort "high"}) sets a per-model default. Precedence: /effort override → model default → config :effort → provider default.
• The TUI header shows effort <level> when a non-default level is active.

Sessions

• save-session! persists the session's effective effort alongside its model.
• Restoring a session — via the ⌃R modal or the new /resume <id> (which autocompletes from the saved-session list) — now brings back the model and effort it ran with and rebuilds the agent on them. Previously a resumed session silently continued on the current model.
• /config completes its edit argument.

Shadowing fix

coder.sema's top-level provider/model locals are renamed (llm-provider/cli-model) — in Sema's single namespace they would have clobbered the new config constructors, breaking init.sema on the first hot-reload. Same trap the file already documents for agent.

Notes for review

• Tests cover the constructors, model-entries/model-effort, config defaults, palette argument mode, the pure Enter/Tab selection logic, active-value marking, session effort round-trip, and the model/effort precedence helpers — but the suite was not executed (the sandbox couldn't install the sema interpreter). Please run jake coder.test before merging.

🤖 Generated with Claude Code

https://claude.ai/code/session_01U4ocGFxVK7FsgQaEhFmJCC