Skip to content

fix(models): picker shows what runs — display names, one-level effort, clamped wire level - #10652

Closed
Skymore wants to merge 2 commits into
kirodotdev:mainfrom
Skymore:fix/reasoning-effort-default-sync
Closed

Skymore wants to merge 2 commits into
kirodotdev:mainfrom
Skymore:fix/reasoning-effort-default-sync

Conversation

@Skymore

@Skymore Skymore commented Sep 14, 2026 •

Copy link
Copy Markdown
Contributor

Problem / Motivation

The model picker says one thing and the session runs another, in four ways a user can see:

  • The effort control has two states for the same level: "Default · High" (a toggle labelled "Use configured default", slider disabled) and "High" (toggle off). Both run at high. The toggle also only knows the Settings → Chat tier, so a crew pinned to a different effort shows the wrong default, and the cached value goes stale after a Settings save.
  • A level the model does not accept is sent anyway. kiro-cli refuses it in a normal /effort result (success: false), the gateway reads only the reply text and reports success, and the session keeps its previous level while the picker shows the new one.
  • Every selector shows raw ids (claude-opus-4.8, gpt-5.6-sol) and each row spends a second line on the description; Auto carries a "Default" subtitle that reads as "the configured default model".
  • Picking a model waits for the gateway round trip before the chip and list move, so a slow gateway makes the picker feel stuck.

Why it matters

Effort and model are the two things a user changes mid-session most often. A picker that shows a level the model is not running, or that appears not to react to a click, teaches the user not to trust it; two visually different states for one level, plus three different meanings of "default" in one popover, are what every "why is my session slow / expensive" question comes back to.

What changed (motivation → approach → change)

One level, no default concept in the session picker. The slider shows the level in force: the session's own override if it has one, otherwise the level the crew pin → role → Settings → Chat chain resolves to, fetched from the backend resolver (/api/agents/resolved-model) so the picker and the Agents page agree. Every landing on the axis — a move, a click on the notch already shown, or Enter — writes that level as the session's own. There is no default marker, no "Use configured default" toggle and no reset link: defaults are configured in Settings → Chat and on the crew, and the session picker shows the result rather than explaining the chain. This is a deliberate reversal of the toggle #10244 introduced. The alternative kept ("explicit High" vs "inherited High" as two UI states) was dropped because users cannot tell them apart and never need to: when the chain resolves to nothing, the axis has no thumb and reads "Model decides" until the first pick, which is the only remaining non-level state.

Per-model levels, clamped on both ends. A table probed against kiro-cli 2.21.4 (Sonnet 4.6 has no xhigh; GPT-5.6 adds none, shown as "Off"; Opus 4.5 / Sonnet 4.x / Haiku / auto and the third-party models refuse effort) drives which notches the slider offers. The frontend and the backend share the table. On the kiro path the backend puts the nearest level the current model accepts on the wire — both the live /effort push and the spawn-time overlay — while keeping the stored pick as made. A /effort result carrying success: false is now a failed switch (rolled back, then the existing reset fallback), not a silently applied one. When the user switches to a model that lacks the session's level, the frontend rewrites the override to the clamped level and shows a 3-second notice ("Claude Sonnet 4.6 does not support Extra High; switched to High").

Settings ▸ Chat manages the selectable models in the page. The "Selectable Models" control was a searchable multi-select behind a summary trigger, so the state it edits (which models the picker offers) was hidden until opened. It now renders as a list standing in the page — same search, bulk actions, locked Auto row and keyboard travel, with the summary as a caption — via an inline mode on the shared MultiSelect; the dropdown shape is unchanged for any other consumer.

Display names everywhere a model is named. A static table maps kiro-cli ids to the vendor's spelling (Claude Opus 4.8, GPT-5.6 Sol, DeepSeek-V3.2, Qwen3-Coder-Next); ids the table does not know render verbatim (never an invented name). The name is used by the chat picker (single-chat and split-pane), the filter (matches name and id), the composer chip, the "set as default model for …" row, Settings → Chat (default / background / subagent / fallback model, the selectable-models list), the crew editor, scheduled jobs, Knowledge, Auto Research, Spec Builder and Issue Radar, and by the per-turn footer under each reply and its tooltip. The id stays the persisted value, in data-model-name / title, and in the Telemetry table where rows are grouped by id.

Compact rows. Each row is one line: the description sits behind a single-ring ⓘ, [Internal] stays beside the name, Auto's "Default" subtitle is gone (also a deliberate removal: it named a factory default while the same popover offers to set a default model). The picker is the shared shadcn Popover; keyboard focus on a row uses the hover background like the other menus rather than the outline ring, which the scrolling list clips.

Instant model switch. The chip and the list update on click; the request still goes through the serialized switch protocol, which restores the last confirmed model if it fails and, for a rapid burst that fails entirely, the model in force before the burst. performSlotSwitch gains an optional optimistic argument for this; the adjudication rules are unchanged.

Smaller changes riding along. The footer's credits unit is localized (it was English in every locale); the API-cost branch uses the shared currency formatter. Scheduled jobs read the shared model list instead of a private query. A Settings → Chat effort save invalidates the picker's inherited-effort cache so an open tab does not keep the previous default. Slider gains onPick (every landing, including the shown notch) and drops its reference-marker props with their only consumer. Dead i18n keys for the removed toggle, marker and "Customize" page are deleted; effort_help no longer narrates the inheritance chain.

Tests

Frontend (Vitest): the effort control (level in force with no marker/toggle/reset; same-notch pick writes the level; a pick the session already holds writes nothing; optimistic thumb and rollback on a rejected write; per-model notches and clamping; "Model decides" state), Slider.onPick (every landing, Enter, inert without onPick), performSlotSwitch optimistic writes (immediate write, single-failure rollback to baseline, all-failure burst rollback to the first baseline, adoption of an older landed pick, baseline cleared after a confirmed pick), the per-turn footer (display name, localized unit, tooltip), fmtTurnModel, the resolver-backed resolveDefaultEffort, MultiSelect inline mode (no trigger, summary caption, search, disabled applying to the controls) and the Settings selectable-models list against it, and display-name assertions across the Settings, crew, Knowledge and picker suites. 204 files / 3373 tests in the model, effort, settings, footer and switch-protocol areas pass.

Backend (pytest): the level table (effort_levels_for_model), nearest_supported_effort, effort_for_model, model_supports_effort consulting the table (Opus 4.5 / Sonnet 4.x refused), change_effort clamping the wire level while keeping the pick, a success: false result rolling back and raising, clear_effort through the checked push, and the config-option ladder never stepping into none. 302 tests across the effort, provider, handler and codex-pair suites pass.

Gates: TypeScript build mode, ESLint on every changed file, all i18n:check lanes, phantom-class and theme-token gates, docs lint, comment-history gate against origin/main, production build.

Manual verification

kiro-cli 2.21.4 probed over ACP: /effort xhigh on claude-sonnet-4.6 answers {"success": false, "message": "invalid value 'xhigh' for 'output_config.effort', must be one of: low, medium, high, max"} as a normal result; /effort high answers {"success": true, "message": "Effort set to high"}; an effort-refusing model answers success: false with "Effort configuration is currently not available on …". The per-model table matches a 21-model probe.

A real Chromium run against the built SPA with stubbed gateway responses verified: display names in rows, no marker / toggle / reset, the inherited crew level shown as the level in force, a same-notch click writing that level, the switch to Sonnet 4.6 rewriting xhigh to high with the notice and a four-notch axis, and the footer naming the model by display name. The frontend half has been dogfooded on a private gateway for two days; the backend half (wire clamp, success: false handling) is covered by the unit tests and the kiro-cli probe above and has not yet run on a live gateway.

Screenshots / video

Before

Raw ids, two-line rows, Auto says "Default", effort is "Default · Extra High" with a toggle and a disabled slider; the footer shows the raw id.

Before: raw ids, two-line rows, Default toggle

After

One-line rows named like the vendor does; the effort axis shows the level in force (the crew's pin) with no marker, toggle or reset link; the footer uses the same name.

After: display names, one level, no default controls

Switching to a model without the session's level rewrites it and says so for three seconds; the chip and list update immediately.

After: switch to Sonnet 4.6 clamps xhigh to high with a notice

Settings ▸ Chat: the selectable-models list stands in the page (Auto locked, one hidden model unchecked, search and bulk actions on top).

Settings: selectable models as an in-page list

Before → After walkthrough with captions (same stubbed model data in both):

before-after.webm

Related Issues

Supersedes the effort-toggle interaction from #10244; replaces the model-ordering direction of #9969 with no ordering (Kiro's own catalog order is kept, hidden models remain the way to shorten the list).

Pattern harvest

Rule candidate: review-prompt
Pattern: "a slash-command reply is read as text and treated as success without checking its success field" — any command_result consumer should assert the structured outcome, not the presence of a message.

Checklist

  • At most two commits (one is the norm), with a Conventional Commits title (feat|fix|docs|refactor|perf|test|chore|ci|build|revert: ...)
  • Existing tests pass and new tests added for new functionality
  • Self-review completed; code follows project style guidelines
  • Documentation updated (if applicable)
  • No secrets, credentials, or internal references in the diff

Contribution License Agreement

@github-actions github-actions Bot added fork Pull request from a fork (external contributor) readiness: action required A blocking check or review needs attention labels Sep 14, 2026
@dwu96

dwu96 commented Sep 14, 2026

Copy link
Copy Markdown
Contributor

👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review.

When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically.

1 similar comment
@dwu96

dwu96 commented Sep 14, 2026

Copy link
Copy Markdown
Contributor

👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review.

When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically.

@Skymore
Skymore marked this pull request as ready for review September 14, 2026 03:10
@Skymore
Skymore requested review from a team and cixuuz September 14, 2026 03:10
@github-actions

github-actions Bot commented Sep 14, 2026 •

Copy link
Copy Markdown
Contributor

UX Review (Fable 5, fork) — 🔴 BLOCK

UX-level review of 3ba732f3aeec3248a60de2a881e552faf949bc88 via the fork AI-review pipeline — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

UX-Verdict: BLOCK

Cannot evaluate: missing stills for the "Model decides" thumbless axis and the GPT "Off" notch — two new states no attachment shows.

Watch

  • Landing on the notch already shown silently pins the session: handlePick writes "an inherited level the user now owns outright" with zero visible change (thumb and header identical) and, with the reset toggle removed, no path back to inheriting — the session stops following crew/Settings defaults invisibly, for its whole life. Smallest fix: a brief inline acknowledgment ("High — set for this session") when a landing creates an override.
  • Term drift for one setting: the header now says "Thinking effort" while its own help text opens "Reasoning effort sets…", the aria-label is "Reasoning effort", and Settings calls it "Default Reasoning Effort" — the surfaces a user must connect use different names. Pick one term.

Evidence gaps

  • The "Model decides" state (axis with hidden thumb/fill, header "Model decides") — attach a still of a session with no override and no configured default.
  • The GPT six-level axis with the new "Off" notch (lib.effort.none) — attach a still of the effort control on a GPT-5.6 model.
  • Display names in the secondary pickers the PR lists (scheduled jobs, Knowledge, Auto Research, Spec Builder, Issue Radar crew editor) — one still of any of these selects open would close it.

Suggestions

  • effort_help should state the pin's permanence in the header's own words ("Picking a level here fixes it for this session — there is no reset") since the control gives no other cue.

[UX-REVIEWED] 3ba732f

@github-actions

github-actions Bot commented Sep 14, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Fable 5, fork) — 🟡 CONCERNS

Design-level review of 3ba732f3aeec3248a60de2a881e552faf949bc88 via the fork AI-review pipeline — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: CONCERNS

Two opposite clamp policies ship in one PR: the backend keeps the effort pick as made, the frontend destructively rewrites it on model switch.

Watch

  • Override rewrite contradicts this PR's own spec and backend, and can recreate the shown≠running bug. model-selection.md (this commit) and the backend test pin keep-the-pick ("the stored level is not rewritten, so it is xhigh again after a switch back"), but the ChatPage effect does dispatch(updateSlot(...reasoning_effort: clamped)) + api.chatSlotReasoningEffort(activeSlot, clamped). Pick xhigh on Opus → switch to Sonnet 4.6 (override rewritten to high, backend keys it under sonnet) → switch back to Opus: the backend's per-model store still resolves opus→xhigh while the slot shows High — the exact mismatch the PR exists to remove. The display clamp and the backend wire clamp already cover this case, so the rewrite buys nothing.
    Clears when: the ChatPage rewrite effect is dropped (or the spec/backend change to rewrite-everywhere), with a switch-away/switch-back test pinning one policy.
  • The kiro-cli 2.21.4 capability table is duplicated (Python + TS), unsynced by any test, and outranks the live source. enabled: … && staticLevels === undefined disables /api/effort-levels for known models, and _clamp_for_wire downshifts preemptively — so when a kiro-cli upgrade adds a level (e.g. xhigh on Sonnet 4.6), both ends silently cap it forever and the notice falsely says the model "does not support" it; the new success:false detection never gets to correct it.
    Clears when: the wire path tries the requested level and steps down on the structured rejection (the ladder shape _set_effort_config_option already has), or a sync check + version gate covers the two table copies.

Suggestions

  • Parse the rejection message's "must be one of: …" list to refresh per-model levels at runtime, shrinking the static table to a cold-start hint.

[DESIGN-REVIEWED] 3ba732f

@github-actions

github-actions Bot commented Sep 14, 2026 •

Copy link
Copy Markdown
Contributor

GPT 5.6 Review (fork) — 🔴 changes requested (blocking)

Reviewed 3ba732f3aeec3248a60de2a881e552faf949bc88 via the fork AI-review pipeline; updated in place on each push.

1 of 2 blocking finding(s) are security-class and were withheld from adjudication, so the blocking verdict stands.

BLOCKING -- website/src/pages/ChatPage.tsx:3793 -- Clamp write races failed model switch
api.chatSlotReasoningEffort(activeSlot, clamped).catch(() => {})
xhigh Opus → rejected optimistic Sonnet switch → queued clamp persists high on restored Opus.
Anchor: errors-use-error-notice
Fix: Persist the clamp only after model-switch confirmation and surface write failures.

BLOCKING -- website/src/lib/effort.ts:96 -- Static GPT ladder overrides adapter-reported levels (origin: validation)
if (m.includes('gpt')) return LEVELS_GPT
Codex openai.gpt-6-astra[max] → live high/xhigh/max ignored → unsupported pick persists while the session runs another level.
Anchor: residual/crash-data-loss-corruption
Fix: Apply the static table only to bare kiro-cli IDs; return undefined for provider-prefixed IDs so live levels win.

[BLOCK-MERGE] 3ba732f
[GPT-REVIEWED] 3ba732f

Adjudication (Opus 4.8) — is blocking on each finding proportionate?

I've traced the mechanism from both the diff (as data) and the on-disk base helpers. Let me record my rulings.

F1 — ChatPage.tsx clamp write races a failed model switch

The scenario: slot carries an explicit xhigh override that the current model (Opus) accepts. User picks a lower-capability model (Sonnet 4.6). The optimistic write (slotSwitch.ts performSlotSwitch → if (optimistic) write(optimistic.value), patch line 2787) flips shownModel to Sonnet synchronously. The clamp useEffect (deps [shownModel, activeSlot, _slotEffortRaw]) then fires, computes nearestSupportedEffort('xhigh', LEVELS_SONNET_46)='high' (effort.ts nearestSupportedEffort/effortLevelsForModel), dispatches updateSlot({reasoning_effort:'high'}) and persists it fire-and-forget (api.chatSlotReasoningEffort(...).catch(()=>{})). The server then rejects the switch and settleSlotSwitchFailure rolls the model back to the baseline Opus (slotSwitch.ts, restored-baseline branch), but the effort override stays clamped at high on an Opus that accepts xhigh.

  • Conditions confirmed: pre-existing higher override + switch to lower-capability model + server-side switch rejection; clamp is optimistic (fires before confirmation) and its persist is swallowed.
  • Recovery: none automatic — the reduced tier persists across reload; it is visible in the effort chip (shows "High"), and the user re-picks to restore. Server-side model-switch rejection is itself uncommon.
  • Harm rung: LOW — a recoverable, user-visible preference degradation on a rare condition combination; no data loss, no credential/governance exposure. The real fix (gate the clamp on switch confirmation, tracking optimistic-vs-confirmed model state, plus a surfaced-failure path replacing the fire-and-forget contract) is a new ordering contract between the clamp effect and performSlotSwitch carrying permanent maintenance load, which exceeds the harm prevented.

F2 — fenced. Static LEVELS_GPT for any id containing gpt. The condition requires a future GPT id whose real effort ladder differs from the static table and that reaches this path — the backend can serve new GPT model ids at any time, so this is plausible in real operation, not an input the system's writer cannot produce. I cannot complete a rarity record, so it holds as written.

[ADJUDICATION] 3ba732f total=1 uphold=0 downgrade=1
DOWNGRADE F1 website/src/pages/ChatPage.tsx:3793 reason=disproportionate-remedy
[GPT-ADJUDICATED] 3ba732f
[ADJUDICATION-FENCED] 3ba732f fenced=1 flagged=0
UPHOLD-FENCED F2 website/src/lib/effort.ts:96 -- A new provider-prefixed GPT id with a different real effort ladder is producible by the backend at any time, so the static table silently overriding live levels is plausible, not extreme.
[GPT-ADJUDICATED-FENCED] 3ba732f

@github-actions

github-actions Bot commented Sep 14, 2026 •

Copy link
Copy Markdown
Contributor

First Principles Review (Fable 5, fork) — 🔴 BLOCK

Premise-level review of 3ba732f3aeec3248a60de2a881e552faf949bc88 via the fork AI-review pipeline — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

First-Principles-Verdict: BLOCK

Product-shape change without accepted RFC: the effort picker's return-to-default capability is deleted — a declared reversal of #10244 and the base spec — with no recorded decision.

Not justified as shipped

  • Item 1 — product-shape without recorded decision: "This is a deliberate reversal of the toggle feat(model-picker): streamline model picker controls #10244 introduced" — the description is the proposal, not the decision (evidence in Blockers).
  • Item 4 — framing contradicted by the diff: the spec hunk promises "the stored level is not rewritten" while ChatPage persists the clamped level (evidence in Blockers).
  • Item 3 — duplicate of website/src/model_registry.json: the levels table is hand-mirrored in effort.py/effort.ts ("the two tables must agree") beside the JSON both sides already import, whose supports_effort now contradicts it on 2 entries (sonnet-4.5, opus-4.8/claude-opus-4.5).
    Subtraction: fold the probed levels into model_registry.json; delete both mirrors.
    Clears when: one table holds the per-model levels.
  • Item 5 — duplicate of website/src/model_registry.json: 10 entries already carry display (grep display_list / displayModels); src/lib/modelDisplayName.ts gives 9 of the same acp ids a second name.
    Subtraction: put vendor spellings in the registry's display; delete modelDisplayName.ts and its eslint.i18n.config.js exemption.
    Clears when: one naming table.
  • Item 9 — unjustified move: "was hidden until opened" names no person who failed to find or operate the dropdown.
    Subtraction: drop MultiSelect's inline prop (one consumer: ChatPanel) and keep the dropdown.
    Clears when: a report names who failed to see or manage the set.
  • Item 10 — undeclared: the fix's own comment cites "the effort 'use configured default' link does exactly that" — a control this same diff deletes.
    Subtraction: delete the composedPath branch and useFilteredDropdown.detachedTarget.test.tsx, or name a surviving self-unmounting control.
    Clears when: a surviving in-menu control that unmounts on click is named.
  • Item 7 — rides along: shell refactor inside a fix PR.

What this change ships

Inventory (10 items) — 3 justified

Intent: make the model/effort picker show exactly what the session runs, instantly — framed as a FIX; several items are redesign. More than 10 differences exist (localized credits unit, JobForm shared list, resolver-backed default effort, cache invalidations, deleted i18n keys); the 10 most noticeable are kept.

  1. Effort control loses its default toggle, marker and reset; any landing writes a session override — product-shape without recorded decision
  2. A level kiro-cli refuses now fails the switch (rollback + reset) instead of silently passing — justified
  3. Slider offers only per-model levels; backend clamps the wire level — duplicate of website/src/model_registry.json
  4. Switching models rewrites the stored effort override to the clamped level, with a toast — framing contradicted by the diff
  5. Every picker shows vendor model names from a new static table — duplicate of website/src/model_registry.json
  6. Rows are one line; descriptions behind ⓘ; Auto's "Default" subtitle gone — justified
  7. Picker shell moved onto the shared Radix Popover — rides along
  8. Model chip and list update on click, with failure rollback — justified
  9. Selectable Models list stands in the Settings page instead of a dropdown — unjustified move
  10. Menus survive a click on a self-unmounting button — undeclared

Blockers

product-shape change without accepted RFC — the one-level redesign removes the "Use configured default" toggle and reset ("no 'Use configured default' toggle and no reset link … a deliberate reversal of the toggle #10244 introduced"); the base spec pins the removed shape (docs/system-specs/common/model-selection.md:79: "Enabling the configured effort default moves the slider thumb to that level"), and the base RFC list names nothing covering the picker.
Clears when: an RFC covering the one-level session effort picker is merged non-draft on the base branch and this branch is pushed or rebased, or a maintainer pushes the branch to this repository and records /ai-review override first-principles 3ba732f3aeec3248a60de2a881e552faf949bc88: <decision>.

Spec contradicted by its own diff — model-selection.md adds "the stored level is not rewritten, so it is xhigh again after a switch back to a model that offers it", while ChatPage's model-switch effect persists the clamped level (api.chatSlotReasoningEffort(activeSlot, clamped)), duplicating _clamp_for_wire, which already puts the accepted level on the wire while preserving the pick. Subtraction: delete the rewrite (the updateSlot dispatch + API write), keeping the toast.
Clears when: the ChatPage rewrite is deleted, or the spec sentence matches what ships.

[FIRST-PRINCIPLES-REVIEWED] 3ba732f

@github-actions

github-actions Bot commented Sep 14, 2026 •

Copy link
Copy Markdown
Contributor

Opus 4.8 Review (fork) — 🔴 changes requested (blocking)

Reviewed 3ba732f3aeec3248a60de2a881e552faf949bc88 via the fork AI-review pipeline; updated in place on each push.

Model-switch effect silently and permanently overwrites the user's stored effort pick with the clamped value, defeating this PR's own keep-the-pick design.

[BLOCK-MERGE] 3ba732f

BLOCKING — website/src/pages/ChatPage.tsx:2899

dispatch(updateSlot({ key: activeSlot, reasoning_effort: clamped }))
api.chatSlotReasoningEffort(activeSlot, clamped).catch(() => {})
Opus slot at xhigh → switch model to Sonnet 4.6 (no xhigh) → this new effect fires, clamped = "high", and persists it: api_chat_slot_reasoning_effort writes slot.reasoning_effort = "high" and change_effort("high") stores the pick as high. The xhigh pick is destroyed — switching back to Opus now runs high, directly contradicting the spec text this same PR adds ("the stored level is not rewritten, so it is xhigh again after a switch back") and the backend test test_kiro_change_effort_clamps_wire_level_to_model_but_keeps_pick, which pins _effort_per_model["claude-sonnet-4.6"] == "xhigh". The toast tells the user it's a display substitution while the write makes it permanent.
Fix: Clamp for display only — drop the updateSlot dispatch and the api.chatSlotReasoningEffort(activeSlot, clamped) write, keep only the toast; the backend already clamps on the wire and preserves the pick.

FINDING — website/src/lib/effort.ts:2497 — unanchored m.includes('gpt') / /(sonnet|opus)-(\d+)/ matchers catch the provider-prefixed/suffixed ids config-option harnesses expose (claude-opus-4-8[1m], openai.gpt-6-astra[max], anthropic.claude-sonnet-4-5…) that the backend's anchored _KIRO_CLAUDE_ID/_KIRO_GPT_ID deliberately return None for, so effortLevelsForModel returns a static ladder that sets staticLevels, disabling the authoritative /api/effort-levels live query (or hiding the control) for exactly the backends the PR comment claims it "mirrors" → Fix: anchor the frontend patterns to the bare-id spelling and return undefined for prefixed/suffixed ids, mirroring the backend so the live query stays authoritative.

[OPUS-REVIEWED] 3ba732f

@Skymore Skymore closed this Sep 14, 2026
@github-actions github-actions Bot removed the readiness: action required A blocking check or review needs attention label Sep 14, 2026
@Skymore Skymore reopened this Sep 14, 2026
@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running and removed readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running labels Sep 14, 2026
@Skymore
Skymore force-pushed the fix/reasoning-effort-default-sync branch from bd97f9b to 6672bb5 Compare September 14, 2026 06:44
@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention and removed readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running labels Sep 14, 2026
@Skymore
Skymore force-pushed the fix/reasoning-effort-default-sync branch from 6672bb5 to 7601c62 Compare September 14, 2026 07:03
@Skymore
Skymore force-pushed the fix/reasoning-effort-default-sync branch from 4a148ff to 8b7da2b Compare September 15, 2026 20:12
@dwu96

dwu96 commented Sep 15, 2026

Copy link
Copy Markdown
Contributor

👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review.

When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically.

@github-actions github-actions Bot added the merge conflict Branch has merge conflicts with its base — author must resolve before merge label Sep 16, 2026
@Skymore
Skymore force-pushed the fix/reasoning-effort-default-sync branch from 8b7da2b to aa1a13c Compare September 16, 2026 08:32
@dwu96

dwu96 commented Sep 16, 2026

Copy link
Copy Markdown
Contributor

👋 Hi! This PR is currently in draft status. Workflow runs won't be auto-approved until it's marked as ready for review.

When you're ready, click "Ready for review" and the workflows will be approved on the next cycle automatically.

@Skymore
Skymore marked this pull request as ready for review September 16, 2026 08:38
@github-actions github-actions Bot removed the merge conflict Branch has merge conflicts with its base — author must resolve before merge label Sep 16, 2026
@Skymore
Skymore force-pushed the fix/reasoning-effort-default-sync branch from aa1a13c to fb06424 Compare September 16, 2026 08:46
@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention and removed readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running labels Sep 16, 2026
Rui Tao added 2 commits September 16, 2026 10:28
…, instant switch

Model picker
- Rows are one line: the description moves behind an info tip, `[Internal]`
  stays beside the name, Auto loses its "Default" subtitle, and every row
  names the model by the vendor's spelling (`Claude Opus 4.8`, `GPT-5.6 Sol`,
  `DeepSeek-V3.2`); the id stays the persisted value and is what the filter
  also matches. The same names are used by every other model selector
  (Settings ▸ Chat, crew editor, scheduled jobs, Knowledge, Auto Research,
  Spec Builder, Issue Radar) and by the per-turn footer and its tooltip.
- The picker is a shadcn Popover; keyboard focus on a row uses the hover
  background like the other menus, not the outline ring.
- Picking a model updates the chip and the list at once. The request still
  goes through the serialized switch protocol; if it fails the chip returns
  to the last confirmed model, and a burst that fails entirely returns to
  the model in force before the burst.

Reasoning effort
- The slider shows one level: the session's override, else the level the
  crew pin → role → Settings ▸ Chat chain resolves to (from the backend
  resolver, so it agrees with the Agents page), clamped to what the shown
  model accepts. No default marker, no toggle, no reset link.
- Every landing on the axis — a move, a click on the notch already shown, or
  Enter — makes that level the session's own. When the chain resolves to
  nothing the axis has no thumb and reads "Model decides" until the first
  pick.
- Notches come from a per-model table probed against kiro-cli (Sonnet 4.6
  has no xhigh; GPT-5.6 adds "Off"). Switching to a model that lacks the
  session's level rewrites the override to the nearest lower level and says
  so in a 3-second notice.
- The footer's credits unit is localized; a Settings save invalidates the
  inherited-effort cache so an open picker does not keep a stale default.

Slider grows `onPick` (fires on every landing, including the shown notch);
its reference-marker props go with their only consumer. Scheduled jobs read
the shared model list instead of a private query.
… failure

- A per-model level table probed against kiro-cli 2.21.4 (mirrors the
  frontend's) backs `effort_levels_for_model` / `nearest_supported_effort`.
  On the kiro path the level put on the wire — the live `/effort` push and
  the spawn overlay — is the nearest level the current model accepts, while
  the stored pick is kept as made. `model_supports_effort` consults the table
  first, so Opus 4.5 and Sonnet 4.x are not offered a control kiro refuses.
- kiro-cli reports a refused level as a normal result with `success: false`;
  that is now a failed switch (rolled back, then the handler's reset path),
  not a silently applied one. `EFFORT_LEVELS` gains `none` (GPT-5.6); the
  config-option ladder never steps down into it.
@Skymore
Skymore force-pushed the fix/reasoning-effort-default-sync branch from fb06424 to 3ba732f Compare September 16, 2026 10:28
@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention merge conflict Branch has merge conflicts with its base — author must resolve before merge and removed readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running labels Sep 16, 2026
@dwu96 dwu96 added the needs-pr-triage PR scanner: awaiting automated triage label Sep 20, 2026
@NicholasRBowers NicholasRBowers added needs-author-decision PR blocked on author input and removed needs-pr-triage PR scanner: awaiting automated triage labels Sep 20, 2026
@NicholasRBowers

Copy link
Copy Markdown
Contributor

🤖 Kiro Crew [operator: NicholasRBowers#a942f9ca]: This PR has been inactive for 7+ days. I reviewed the blockers but they require your input:

  • The gating decision is a product one: you are deleting a capability that feat(model-picker): streamline model picker controls #10244 deliberately added. First Principles Review is 🔴 BLOCK on 3ba732f3a for removing the "Use configured default" toggle and its reset link — which the description itself calls "a deliberate reversal of the toggle feat(model-picker): streamline model picker controls #10244 introduced" — while the base spec still pins the removed shape (docs/system-specs/common/model-selection.md:79: "Enabling the configured effort default moves the slider thumb to that level"). I checked the lane's evidence myself: no RFC on current main covers the one-level session effort picker. Its exit conditions are an RFC merged non-draft on the base, or a maintainer recording /ai-review override first-principles. Which are you pursuing? A description is a proposal, not the decision, and I will not re-add the toggle you removed on purpose.

  • Separately, three reviewers independently say the same hunk contradicts your own spec — and that one is a fix, but it is a policy choice you should make explicitly. First Principles' second blocker, Opus's blocking finding at ChatPage.tsx:2899, GPT's F1 at ChatPage.tsx:3793 and Design's Watch item all land on the model-switch effect:

    dispatch(updateSlot({ key: activeSlot, reasoning_effort: clamped }))
    api.chatSlotReasoningEffort(activeSlot, clamped).catch(() => {})
    

    Pick xhigh on Opus, switch to Sonnet 4.6, switch back: the backend's per-model store resolves Opus to xhigh while the slot shows High — the exact shown≠running mismatch this PR exists to remove. It also contradicts the spec sentence this same PR adds ("the stored level is not rewritten, so it is xhigh again after a switch back") and your own backend test test_kiro_change_effort_clamps_wire_level_to_model_but_keeps_pick. All four reviewers converge on deleting the updateSlot dispatch and the API write and keeping only the toast, since _clamp_for_wire already puts the accepted level on the wire while preserving the pick. Confirm that keep-the-pick is the policy you want (rather than rewrite-everywhere, which would mean changing the spec and the backend test instead) and this becomes mechanical.

  • GPT's second blocking finding is a scope question about where the capability table lives. effort.ts:96 — if (m.includes('gpt')) return LEVELS_GPT — makes the static kiro-cli 2.21.4 ladder outrank adapter-reported levels, so openai.gpt-6-astra[max] ignores live high/xhigh/max; the same unanchored matching catches claude-opus-4-8[1m] and anthropic.claude-sonnet-4-5…. The narrow fix is to apply the static table only to bare kiro-cli ids and return undefined for provider-prefixed ones. But First Principles wants the table gone entirely (items 3 and 5): the per-model levels are hand-mirrored in effort.py/effort.ts beside website/src/model_registry.json, which both sides already import and whose supports_effort now contradicts the mirrors on sonnet-4.5 and opus-4.8/claude-opus-4.5; likewise src/lib/modelDisplayName.ts gives 9 acp ids a second name where 10 registry entries already carry display. Do you want the point fix, or the subtraction into one table? Those ship different PRs and I should not pick.

  • UX Review is 🔴 BLOCK on missing evidence plus one invisible-state concern. Needed stills: the "Model decides" thumbless axis, the GPT six-level axis with the new "Off" notch (lib.effort.none), and any one of the secondary pickers you list. I can capture and attach those — but the concern underneath is yours: landing on the notch already shown silently pins the session for its whole life with zero visible change, and with the reset toggle removed there is no path back to inheriting. UX also flags term drift across surfaces a user must connect — header "Thinking effort", help text "Reasoning effort sets…", aria-label "Reasoning effort", Settings "Default Reasoning Effort". Pick one term.

Not waiting on you: all CI is green, and the merge conflict is four files (ChatPane.tsx, ChatPage.tsx, MultiSelect.tsx, ChatPage.switchLabelOptimistic.test.tsx) which I expect to be tractable once the decisions above fix what the diff should contain — resolving them before then would just be work thrown away.

When you've addressed these, the pipeline will re-assess on its next cycle.

Copy link
Copy Markdown
Contributor

Triage (automated, 2026-10-05) — DISTINCT, needs rebase.

  • Nothing on origin/main (a97659d) implements this. There is no per-model effort table, no wire clamp, no success:false handling for /effort, no display-name map, no inline MultiSelect and no optimistic performSlotSwitch. Nothing here is superseded.
  • The branch is 2608 commits behind main and git merge-tree shows 38 conflicted paths, including effort.py and providers/acp.py. Main now has resolve_effort_for_model(levels=…) driven by ACP_BACKENDS_EFFORT_FROM_ADVERTISED_OPTION, so the static kiro-cli table must be reconciled with that path when you rebase.
  • Breaking/take-away change: removing the "Use configured default" toggle reverses merged feat(model-picker): streamline model picker controls #10244. That concern is already raised (First Principles BLOCK, plus the 2026-09-20 operator comment), so this is not a separate request-changes. It still needs a recorded decision or a maintainer override before merge.
  • Overlap to coordinate: feat(dashboard): show what each session actually runs #16906 has the same "show what the session runs" goal, implemented by reading back served_* from the live session. It touches the same spec, acp.py and composer files, and is complementary. fix(dashboard): reset model state on agent switch #9115 also rewrites website/src/lib/slotSwitch.ts, so sequence the two performSlotSwitch changes.

Recommendation: rebase, settle the open product decisions, then continue.

@iamwhatever

Copy link
Copy Markdown
Contributor

Thank you for the careful write-up of where the picker and the session disagree. We are closing this because main has moved a long way since it was opened: model display names and handling of a refused effort level have landed through other changes, effort levels now come from the model's advertised options rather than a static table, and the branch conflicts in 38 paths. Removing the 'Use configured default' toggle reverses #10244 and would need its own decision first. If you still want any piece of this, such as the optimistic model switch or the inline multi-select, a small PR per piece on current main is very welcome. The 'show what the session actually runs' goal is also being worked on in #16906.

@iamwhatever iamwhatever closed this Nov 1, 2026
@github-actions github-actions Bot removed the readiness: action required A blocking check or review needs attention label Nov 1, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

fork Pull request from a fork (external contributor) merge conflict Branch has merge conflicts with its base — author must resolve before merge needs-author-decision PR blocked on author input

Projects

None yet

Development

Successfully merging this pull request may close these issues.

5 participants