fix(acp): open Kimi sign-in in a terminal#24409

Problem

Adding Kimi as a provider stalls at “Starting sign-in…” instead of opening its login terminal. Reproduced with Kimi CLI 1.52.0 on Linux in both the Add provider wizard and saved Provider settings: Kimi advertises an untyped legacy terminal-auth method, and its ACP authenticate request never responds.

Change

Recognize valid legacy _meta["terminal-auth"] metadata as terminal authentication. Use its complete executable, arguments and environment through the existing PTY login and fresh-session verification flow. Typed terminal methods keep their existing launch behavior. A shared launcher supplies both the displayed command and execution.

Scope and approval

Fixes https://github.com/pingdotgg/t3code/issues/24302 GitCafe: https://git.cafe/pingdotgg/t3code/issues/24302

Triaged bug via the automated issue flow. This restores the existing Kimi sign-in flow; no new settings, contracts or persisted state.

Verification

• Red on 27e0c684c5: the regression test failed with expected ['login'] to deeply equal [], proving ACP authenticate was called instead of terminal login.
• Green after merging current main at 2c205d43bf: vp test run --config ../../vite.config.ts --dir . in packages/provider-acp-registry: 56 tests passed. Package vp exec tsc --noEmit, targeted vp lint on the three changed files, and git diff --check passed.
• Two isolated code reviews passed. Proof audit: all nine pairs sufficient. Independent head reconstruction matched layout and controls; differences above threshold16 were confined to the regenerated device code.

Evidence uses before 27e0c684c5 and after 14d89dd239. The final main merge changes none of the three fix files. Linux Chrome 154, DPR2, zoom1; CSS 1440×900 or 390×844, light/dark. Desktop still exports are1280×800; narrow exports 780×1688. Reduced motion, fonts loaded, matching scroll and frozen display clock for stills; each selected raw stability pair has zero changed pixels. Diff0 amplifies every changed pixel; DiffT marks maximum RGB-channel difference >16.

Fixture matrix · Wizard/narrow settings stills, stability pairs and provenance · Desktop settings replacements · Corrected recordings and indexed timestamps.

For every Kimi pair: start inactive, click Sign in, observe for at least 35 real seconds, Cancel to inactive Retry, retry a distinct attempt, then Cancel again. Before hangs without a terminal; after shows the real Kimi authorization prompt. All attempts were cancelled before publication. The linked indexes give per-action timestamps. Caption-only correction preserves every frame timestamp, track end and all app RGB pixels; original recordings and the postprocessing script remain public. K1 desktop-light after shows the final inactive state for 1.015s; no longer hold is claimed.

K1-1440-light — Add provider wizard, 1440 wide, light. Before recording · After recording

K1-1440-dark — Add provider wizard, 1440 wide, dark. Before recording · After recording

K1-390-light — Add provider wizard, 390 wide, light. Before recording · After recording

K1-390-dark — Add provider wizard, 390 wide, dark. Before recording · After recording

K3-1440-light — Saved Provider settings, 1440 wide, light. Before recording · After recording

K3-1440-dark — Saved Provider settings, 1440 wide, dark. Before recording · After recording

K3-390-light — Saved Provider settings, 390 wide, light. Before recording · After recording

K3-390-dark — Saved Provider settings, 390 wide, dark. Before recording · After recording

T1 — synthetic typed-terminal regression control, 1440×900 light. Before recording · After recording · Indexed actions, fixture, replay logs and provenance.

Both show SAFE+Enter and appended output, explicit authenticated resize RPC 101×31, cancellation/process exit/stale-response rejection, distinct retries, held verification failure, nonzero login exit, and held verification success ending Signed in. This is synthetic protocol proof, not real Kimi account completion. Supplemental stills below are decoded video frames; within-side pairs 500ms apart match exactly, while cross-side differences include encoding/render noise.

What else changed visibly: terminal insertion and output, readiness/account text and controls, modal height/recentering, downstream settings movement, and the background provider panel reflecting the same auth flow. Generated codes differ. K1 uses matching glassOpacity 100 and disabled backdrop blur to avoid export ghosting; default translucency is unverified. Captures make no “nothing else changed” claim.

Public reconstruction material: Kimi capture driver, setup driver, recipe, desktop retake drivers/provenance, T1 executable replay bundle. No evidence-only files are committed.

Unverified: real Kimi account completion and post-login success; Windows; native Electron/mobile; remote/relay/tunnel modes. T1 resize is RPC transport proof, not responsive-layout resize; Local ACP cannot distinguish its coincident base/resolved environment layers. Timing is approximate and videos were visually sampled, with exhaustive frame/pixel checks for caption corrections. Linux web prompt, cancellation and retry were exercised in the real app.

Implemented by Claude Opus 5.5 via the Claude harness; triaged, reviewed and verified by GPT-6.1-Sol via the Codex harness in T3 Code.