Reach the console's multi-model endpoints, and print what they cost

[its-janghoon]

Wires the console's multi-model endpoints into the CLI as two commands, and adds a goal-driven autopilot loop.

Needs console #127 deployed before paraphrase works against production — that PR fixes three bugs this client is what found.

variants.paraphrase / variants.compare

Both gated on the Redrob provider being connected: the route answers 404 when it is not, so a user on a local runtime or another vendor is told the feature is unavailable rather than watching a request fail for reasons that look like their own.

The cost is always shown. One request makes several charges, one per model. A user who believes they made one request will be surprised by the bill unless the surprise happens immediately, while they can still decide not to do it again.

Rewritten
  gpt-5.6-sol      $0.0005    4ms
  gpt-5.6-terra    $0.0006    9ms
  claude-opus-5    $0.0013   13ms
  total            $0.0024
  <the answer>

routedModel is what names the model that actually answered — under auto it is the only thing that does.

compare offers the answers in a selection list with previews, and only the pick enters the conversation. Session history cannot be pruned: Revert.State rewinds to a point and restores files with it, and fork makes a new session — neither can remove "these two of three messages". An answer appended just to try it would be permanent.

Three bugs that only a real terminal found

Result was invisible	It was a toast — absolutely positioned, capped at 60 columns, dismissed on a timer. It clipped the cost total and the entire answer, keeping only the first slot rows. The two things a reader most needs were the two it dropped. Now a dialog.
compare's chooser never opened	DialogAlert calls onConfirm and then clears the dialog stack itself, so a chooser pushed synchronously from that callback is opened and immediately wiped. Deferred by a tick.
Picking did nothing	The insert was a voided promise with no catch; the rejection was discarded. A user who has just been charged for several models must be told when the thing they paid for failed to land.

Verified the insert lands by reading the session store rather than the screen:

session_input: "로봇의 정의는\n\n(answered by gpt-5.6-sol)\n\nHere is a short answer…"

REDROB_CONSOLE_URL now applies to the whole CLI

A per-consumer Flag.REDROB_CONSOLE_URL ?? CONSOLE_URL made the override half apply: the variants service honoured it while the session's own inference path, the catalog fetch and the provider registration still went to production. Pointing the CLI at a local console produced one that answered /variants/paraphrase locally and 401'd every chat turn against the real one — which reads as a credential bug and is not.

Resolved once in the constants module. Measured after the change: local console inference calls went 0 → 2 on a session turn that previously 401'd.

An HttpClient requirement that reached serve

The HttpClient requirement propagated out of the variants functions, through the route handler, into the API's own requirement type, and out to every entry point building the API:

Type 'Effect<void, unknown, HttpClient | Service>' is not assignable to 'Effect<void, any, Service>'
  Type 'HttpClient' is not assignable to type 'Service'

That reads as a problem with serve and is not one. Provided inside the service instead. Correcting myself: I earlier reported these two typecheck errors as pre-existing on develop. They were not — clean develop typechecks at 0, and both were mine.

console-key.ts

Extracts credential resolution both consumers need. Three origins, nothing bridges them — REDROB_API_KEY, the Integration store, auth.json — with different shapes (type:"key" vs type:"api"). Reading only some already shipped once as a bug: desktop-app users saw six fallback models.

Runtime finding that typechecking could not give: Credential.Service is not in scope in HTTP handlers despite compiling. Resolved via Integration.Service.

loop.start / loop.stop

The first version re-sent a fixed instruction, which cannot notice circling because the input never changes. Each cycle now sends the goal, the ledger of approaches already rejected, and an instruction to assess against the goal before taking the next step.

Stops on: goal met, blocked, a repeated plan (three times — two is patience, since waiting on a build looks identical), a cycle backstop, or a session failure. Driven by session.idle rather than a timer, because a timer fires into a session that is still working and stacks turns.

Not yet verified: whether models emit the AUTOPILOT NEXT: marker reliably. If they do not, stall detection never fires. That is the most uncertain part of this design and wants a real run.

Gates

typecheck   0 errors across the monorepo
packages/tui    281 pass / 1 skip / 0 fail
packages/core   970 pass / 3 fail  — the same 3 fail on clean develop
                (ModelsDev static fallback: limit.output expects 32000, gets 64000)

Those three are unrelated to this branch and confirmed pre-existing by stashing everything and re-running.