Choose a model and a reasoning effort before a Thread starts - #34
Closed
QuintinBotes wants to merge 1 commit into
Closed
Choose a model and a reasoning effort before a Thread starts#34QuintinBotes wants to merge 1 commit into
QuintinBotes wants to merge 1 commit into
Conversation
`LaunchConfig.model` existed and already emitted `--model`. Nothing ever set it, so the plumbing was there and unreachable: every Thread ran whatever the CLI happened to default to, and the one thing that most changes what a Thread costs was the one thing the composer could not express. `--effort` is now confirmed present on the shipped 2.1.220 binary, taking `low`, `medium`, `high`, `xhigh` and `max`. It had been assumed to be a slash command only. `LaunchConfig` carries it alongside the model and the adapter emits it. The options are declared by the adapter, not listed in the UI. This is the rule the mode picker already follows: a control offering a choice the runtime would reject is worse than one offering none. `AgentRuntime::launch_options()` returns empty by default, so a runtime that takes neither draws no controls, and Codex does not get a picker full of Claude aliases. Models are aliases rather than pinned identifiers, because `claude --help` documents them as tracking whatever is current. A pinned `claude-opus-4-1` would rot in the worst possible way: the old name still resolves, so it would fail by quietly running last year's model rather than by erroring. The effort list, by contrast, has to be exact, and a test pins it. An unrecognised `--effort` value is only a warning: the CLI prints one line, falls back to the default and runs anyway. A typo there produces a session that appears to run at the requested effort and does not, which is the exact class of silent mismatch this project exists to make visible. Both are launch flags, so the pickers appear only while nothing is running. The alias and the model it resolved to are shown together once the session reports back, since they differ and the difference is the cost. Two smaller things follow from the same reasoning. An empty selection means "whatever the profile already chooses" and is never sent as an empty flag value. And switching profile clears both, because a profile can switch runtime and an alias one runtime resolves is one another may reject or, worse, misread. `launch_options` rides on `agents_overview` rather than `agents_discovery`: the composer draws these before anything is probed, and every adapter declares them statically, so it costs a lock and a map. rust 688 to 691, vitest 324 to 328. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This was referenced Aug 3, 2026
QuintinBotes
added a commit
that referenced
this pull request
Aug 4, 2026
A handoff for whoever picks this up. Two jobs: say what is true right now, and stop the next person paying again for things that cost hours here. The traps are the point. Editing Rust — or switching branches — while `pnpm app` runs restarts the user's Threads. Merging with `--delete-branch` auto-closes any PR based on that branch and a closed PR cannot be reopened, which is how #34 was lost. This repository squash-merges only, so the PR body is the commit message. The established facts are there so nobody re-derives them: the gate must print nothing when it does not object, because `defer` ends the turn and reports success; `--effort` exists and an unknown value warns rather than errors; a plan only exists if the Thread *started* in plan mode; `TERM=dumb` disables ZLE; a pty read blocks uninterruptibly. Also the two patterns behind nearly every bug found: capabilities built in Rust and unreachable from the interface, and the app doing something without saying so. Six of the first in one day. The second is what the user meant by "unclear what to do where and when", and it is fixed instance by instance and never as a pattern. Next steps are ordered, with the first being an instruction not to write code: open a real zsh and check `bindkey ^T` by hand before touching the completion driver again. The last session lost several rounds patching that blind.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #33, which is why the base is
tervin-self-reportingrather thanmain. Both touchAgentsOverview. Retarget tomainonce #33 lands.LaunchConfig.modelalready existed and already emitted--model. Nothing ever set it, so the plumbing was there and unreachable: every Thread ran whatever the CLI defaulted to, and the single thing that most changes what a Thread costs was the one thing the composer could not express.What is new
--effortis confirmed present on the shipped 2.1.220 binary, takinglow,medium,high,xhigh,max. It had been assumed to be a slash command only, so this was verified against the binary rather than the docs.The options are declared by the adapter, not listed in the UI
The rule the mode picker already follows: a control offering a choice the runtime would reject is worse than one offering none.
AgentRuntime::launch_options()returns empty by default, so a runtime that takes neither draws no controls, and Codex does not get a picker full of Claude aliases.Models are aliases, not pinned identifiers, because
claude --helpdocuments them as tracking whatever is current. A pinnedclaude-opus-4-1rots in the worst way: the old name still resolves, so it fails by quietly running last year's model rather than by erroring.The effort list has to be exact, and a test pins it. An unrecognised
--effortvalue is only a warning — the CLI prints one line, falls back to the default, and runs anyway. A typo there yields a session that appears to run at the requested effort and does not, which is exactly the silent mismatch this project exists to surface.Details that follow from the same reasoning
launch_optionsrides onagents_overview, notagents_discovery: the composer draws before anything is probed, and the declarations are static.rust 688 to 691, vitest 324 to 328. clippy and fmt clean.
Not verified in the running app
The flags and levels were checked against the real binary, but the pickers themselves have not been exercised in a running Tervin.