feat: add --model flag to ucode copilot

[larsmoan]

Copilot has no way to pick a model per launch: there's no --model flag (unlike claude), Copilot's own /models picker can't be trusted against the Databricks gateway, and ucode overwrites COPILOT_MODEL on every 30-minute token refresh anyway. The only override today is an admin-only managed config.

Adds --model to ucode copilot, mirroring claude. The explicit model is threaded past the refresh loop (_refresh_forever/_refresh_token_once) so it stays pinned for the whole session instead of being silently reclaimed by default_model() at the next refresh.

Closes #384.

[Copilot]

Pull request overview

Adds per-launch --model selection to ucode copilot and preserves it across token refreshes.

Changes:

• Adds the --model CLI option.
• Threads model overrides through Copilot dispatch and refresh logic.
• Adds CLI, dispatch, and refresh-loop tests.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Summary	Findings
tests/test_cli.py	Tests CLI model handling.	None
tests/test_agents_init.py	Tests dispatch behavior.	None
tests/test_agent_copilot.py	Tests model selection and refresh pinning.	None
src/ucode/cli.py	Defines and forwards --model.	Critical (2 votes): Explicit models can fail when no fallback model exists because resolution occurs before the explicit override.
src/ucode/agents/copilot.py	Pins the model across refreshes.	None
src/ucode/agents/__init__.py	Dispatches Copilot model overrides.	None

Suppressed comments (1)

src/ucode/cli.py:1961

• resolved_model can differ from the raw CLI model: for example, a managed budget recommendation is selected by managed_launch_model, written by configure_tool, but not stored in state. Passing None for a Copilot launch without an explicit --model makes copilot.launch call default_model(state) again and immediately rewrite that selection, and refreshes keep doing so. Pass the already resolved model here so managed/recommended choices survive the initial write and refresh loop.

        launch_agent(tool, state, ctx.args, model=model if tool == "copilot" else None)

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[larsmoan]

Fixed in 0ff75e7 — launch_agent now gets resolved_model (which already absorbs --model, a managed default, and any budget recommendation) instead of the raw --model flag. Manually tested end-to-end against a real workspace: ucode copilot --model system.ai.claude-opus-4-8 launches and pins correctly.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated no new comments.

Suppressed comments (2)

src/ucode/cli.py:2284

• This option is applied only after _launch_tool has completed its setup and model resolution. On a first ucode copilot --model X (or when Copilot is not yet in available_tools), _auto_configure_tool("copilot") runs first, and configure_single_tool/validate_env select and validate default_model(state) without seeing X; if discovery has no default or the automatic model is unsupported, the command exits before the requested model is ever written or launched. Thread the explicit model through the auto-config/validation path, or skip default-model setup when --model is supplied.

            help="Launch on a specific Databricks model id (e.g. a UC "
            "`<catalog>.<schema>.<name>`). Outranks the automatic sonnet/opus/haiku/codex "
            "pick and stays pinned across ucode's automatic token refreshes.",

tests/test_agents_init.py:182

• This class is inserted before the existing test_pi_available_* methods, so those availability tests are now collected under TestLaunch instead of TestCheckGatewayEndpoint. They still run, but the grouping is misleading and makes future fixture/setup changes error-prone; place the new TestLaunch class after the complete gateway-endpoint test class.

class TestLaunch:

[larsmoan]

Addressed both review findings in 321e693:

1. Critical: a first-time ucode copilot --model X (copilot not yet in available_tools) ran _auto_configure_tool before the --model handling in _launch_tool, so it configured and smoke-tested the automatic sonnet/opus/haiku/codex default instead of X. Threaded the explicit model through configure_single_tool/_configure_one (config write) and validate_tool/copilot.validate_env (smoke test) so X is what's actually configured and validated.
2. Test ordering: moved TestLaunch out of the middle of TestCheckGatewayEndpoint, which had split off its test_pi_available_* tests into a misleadingly-grouped class.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated no new comments.

Suppressed comments (1)

src/ucode/cli.py:1742

• This forwards the explicit model into first-time setup, but the normal launch path still calls resolve_launch_model(tool, state, managed_model) before applying model at line 1900. If discovery/state has no fallback model (for example with --skip-preflight or a failed model listing), that resolver raises before the requested Copilot model is ever considered, so ucode copilot --model X cannot launch. Pass the Copilot override into model resolution or bypass the fallback resolution when it is present.

            _auto_configure_tool(tool, model=model if tool == "copilot" else None)

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated no new comments.