Repository navigation
feat(chat): mark the model chip with the session's harness when it is not the install's - #33404
Conversation
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks. The mark reuses the server's own Notes
[DESIGN-REVIEWED] dc1bfac Description read[DESCRIPTION-READ] f57d01e12206fa3357d49a8af2a7071e093f1bdd54c0a80f615663331b8e469f That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing blocks. Issue #33399 records the defect, the change reuses What this change shipsInventory (6 items) — 6 justifiedIntent: let a user see, without opening a menu, that a chat runs on a different harness than the install's own (for example Claude Code on a Kiro install). This is an ADDITION.
[FIRST-PRINCIPLES-REVIEWED] dc1bfac Description read[DESCRIPTION-READ] 134785ed7d130a7766c89640749e558e0cf948816e5f1a9cfb8e473a6ad6fb82 That sha256 names the PR description this verdict read. Recompute it with |
UX Review (Opus 5.5) — ✅ human override accepted@chenmingwei23 overrode this lane for |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
3000ff8 to
018213c
Compare
|
Hold lifted. The branch is rebased onto main with the bundle-size fix (#33665), and the owning worker session of conductor chat-3181-1793922654 is driving it to green. |
018213c to
bf4c118
Compare
bf4c118 to
7955bfc
Compare
7955bfc to
185a75c
Compare
185a75c to
fb36663
Compare
fb36663 to
aa01114
Compare
|
/ai-review override ux aa01114 not-applicable: owner ruled the bare mark is the intended design, no new footer text; the harness is named in the tooltip and aria-label. |
Human judgment recorded@chenmingwei23 recorded a not-applicable release (class
This decision applies only to this commit. A new push requires a new judgment. |
…hen it is not the install's A session created on another harness (session_create backend='claude') looked the same as a default one: nothing in the chat said Claude Code. The model chip now draws that harness's brand mark (HarnessLogo) before the model, named in its tooltip and accessible name (Runs on Claude Code). A session on the install's own harness is unchanged. Kiro CLI and KAS count as one harness.
aa01114 to
dc1bfac
Compare
|
/ai-review override ux dc1bfac not-applicable: owner ruled the bare mark is the intended design, no new footer text; the harness is named in the tooltip and aria-label. |
Human judgment recorded@chenmingwei23 recorded a not-applicable release (class
This decision applies only to this commit. A new push requires a new judgment. |
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
This review checked head dc1bfac. The mark reads the server's existing models_backend. own_models_backend fills it only when the session's harness is selectable and uses a different model namespace than the install's, on both the cold and the live capability paths. Both capability queries are keyed by slot, so one session's mark cannot leak onto another. Hosts that do not pass the new prop keep the old chip. Labels render as React text and attributes only. PR Readiness is green on this head.
Done well: the client does not work out a second "not the install's harness" check. So the chip, the model picker list and the harness error message cannot disagree. The new test also pins that the visible chip text stays the model name only.
Optional:
- The main chat wiring at
ComposerDockView.tsx:704(harnessBackend={modelsBackend}) has no test. One ChatPage case that mocksmodels_backend: 'claude'and asserts the chip title starts withRuns on Claude Code ·would pin it, the wayChatPane.sessionModelListpins the split pane. - The body says all 3 new tests fail with the chip change reverted. The "install harness unchanged" case passes on main, so only 2 fail. That case still guards the default session. Also, the
kascase checks only the span title and does not wait for the Kiro logo. - The inner mark span's
titlerepeats the start of the button's own tooltip. You could drop it and assert on the chip's aria-label instead.
Problem / Motivation
Goal: A chat whose session runs on another harness than the install's own (for example Claude Code on a Kiro install) shows that harness in the composer, without opening a menu.
A session can now be created on a chosen harness (
session_createwithbackend: "claude"). The chat never says which harness it runs on. A Claude Code session and a default session look the same: the header shows the title and status, and the composer footer shows the agent, the workspace and the model (opus). The harness name appears only in an error string inside the model picker.Why it matters
A user who opens a worker session cannot tell whether Claude Code or Kiro is answering. That decides which models, tools, sign-in and cost apply. Today they have to ask the agent or open Settings to find out.
Not a goal
What changed (motivation → approach → change)
The owner asked for as little new text as possible. So the composer's model chip now draws the session's harness brand mark (the existing
HarnessLogo, 13 px, in the chip's muted colour) right before the model name, so it reads[mark] opus. It shows only when the session's harness is not the install's own (configuredAcpBackend). A session on the install's harness looks exactly as before, with zero new pixels. Kiro CLI and KAS count as one harness here: a stored Kiro CLI choice runs on KAS once kiro-cli offers it, and both draw the same Kiro mark, so marking one against the other would only add noise.The mark's tooltip and the chip's accessible name say
Runs on Claude Code, using the sameacpBackendNamelabel as the rest of the app. One new string,components.chatInput.harness_mark, is translated in every catalog.No backend change was needed. The slot's selection-capabilities read already returns
models_backend, which the server'sown_models_backendsets only when the session's harness is outside the install's model namespace (Kiro CLI and KAS share one), for a live session and before it starts. That is exactly the "not the install's own" rule, so the client does not re-derive it. The main chat (ComposerDockView.tsx, which already receives it asmodelsBackend) and the split pane (ChatPane.tsx) pass it toChatInputasharnessBackend(chat-input/props.ts), and the chip (ModelChipinchat-input/ContextShelf.tsx) draws the mark when it is set. While the read is still loading, no mark is drawn rather than a guess. The chip loadsHarnessLogolazily, asChatInputalready does forSessionEnvironmentChip, so the brand marks stay out of the chat bundle (the App chunk otherwise went 918 B over its Bundle Size Gate budget).HarnessLogo.tsxonly gets its header comment updated. The new string lives inwebsite/src/i18n/locales/(every catalog,en-XAregenerated). The model-selection spec records the rule.Backwards compatibility
Compatible: one optional prop and one new string. A session on the install's harness renders the same chip as before.
Removes nothing: the change only adds a mark next to the model name.
Tests
website/src/test/ChatInput.harnessChip.test.tsx(new): a Claude Code session (models_backend: "claude") draws the Claude mark, with tooltipRuns on Claude Codeand an accessible name that starts with it, and the chip text is still justopus. A Kiro session on another harness's install ("kas") draws the Kiro mark namedRuns on Kiro Agent. A session with nomodels_backend(the install's own harness) draws no mark and keeps its old accessible name. With the chip change reverted, all 3 fail. The lazy logo's wait uses a named 5 s ceiling (namedCeiling), like the other lazy-mount tests; the file passed 3 more runs with--sequence.shuffle.website/src/test/ChatPane.sessionModelList.test.tsx: two tests find the chip by its title on a pane that runs on Claude Code or Codex, so their selectors now expect the chip's title to start withRuns on <harness>.ChatPage.reasoningEffort,ChatInput.shelfChipsFloor,ChatInput,CrewChatWindow(one test inCrewChatWindowtimed out at 5 s while several suites ran together, and passed when run alone).tsc -p tsconfig.app.json, eslint on the changed files andnpm run i18n:checkare clean.Proof: not provable -- frontend test
Manual verification
On an isolated gateway (its own data home and port, not the real install, which runs KAS), a conductor session created two workers with
session_create: one withbackend: "claude"and one with no backend. Their selection-capabilities reads returnedbackend: "claude"andbackend: "kas". In the dashboard, the Claude Code session's chip had one mark, and its accessible name began withRuns on Claude Code, then the modelopusand the effort. The default session's chip had no mark, and its accessible name began with the modelclaude-opus-5.5, as on main.Screenshots / video
Claude Code session (install runs KAS): the Claude mark before
opus.Default session on the same install: unchanged.
Full window, Claude Code session
Full window, default session
Related Issues
Closes #33399
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)