feat: compare Maka and Pi agent in headless Harbor - #78
Closed
Astro-Han wants to merge 1695 commits into
Closed
Conversation
bundle rounds 13-15 of the 30-round polish loop. Closes the chrome deltas on the global overlay surfaces (command palette, help modal, toasts) so they pick up the dense scale used by the rest of the chat surface. Command palette: - .maka-palette-input 13 -> 12 - .maka-palette-item icon col 18 -> 16, 12 -> 11, transition 150ms ease-out - .maka-palette-hint 11 -> 10 - .maka-palette-footer padding 4/10 -> 3/8, 11 -> 10 Help modal: - .maka-help-body padding 10/14/14 -> 8/12/12, gap 14 -> 10 - .maka-help-section h3 margin 0/0/6 -> 0/0/4, 11 -> 10 - .maka-help-section dl gap 4/12 -> 3/12 - .maka-help-section dt 12 -> 11 - .maka-help-plus 11 -> 10 Toast: - .maka-toast grid 16 -> 14, gap 8 -> 7, radius 8 -> 7, padding 8/10 -> 7/9 - .maka-toast-copy strong 12 -> 11, small 11 -> 10 - .maka-toast-action radius 6 -> 5, padding 2/8 -> 2/7, 11 -> 10, transition 150ms ease-out - .maka-toast-close 20 -> 18, transition 150ms ease-out
bundle rounds 16-18 of the 30-round polish loop. Closes the chrome deltas on the inline tool-result surfaces (tool header, diff viewer, terminal output, load-tool preview) so they pick up the dense scale used elsewhere. Tool / overlay header: - .toolInline > header 11 -> 10, margin-bottom 4 -> 3 - .maka-overlay-preview margin 6/0/0 -> 5/0/0, 11 -> 10, max-height 200 -> 180 - .maka-overlay-close min-height 24 -> 22, padding-inline 8 -> 7 - .maka-load-tool-preview margin 6/0/0 -> 5/0/0, padding 6/10 -> 5/8, gap 3 -> 2, 11 -> 10 - .maka-load-tool-footer 12 -> 11 Diff viewer: - .maka-tool-diff border-radius 10 -> 8 - .maka-tool-diff-paths gap 8 -> 6, padding 6/10 -> 5/8, 11 -> 10 - .maka-tool-diff-body padding 6/0 -> 5/0, max-height 360 -> 320, 12.5 -> 11.5 - .maka-tool-diff-line padding 0/10 -> 0/8 Terminal output: - .maka-tool-terminal border-radius 10 -> 8 - .maka-tool-terminal-head gap 8 -> 6, padding 6/10 -> 5/8, 11 -> 10 - .maka-tool-terminal-exit padding 1/8 -> 1/6, 10 -> 9 - .maka-tool-terminal-empty padding 10 -> 8, 12 -> 11 - .maka-tool-terminal-stream padding 8/10 -> 6/8, max-height 200 -> 180, 12.5 -> 11.5 - .maka-tool-terminal-truncated-note gap 10 -> 8, padding 8/10 -> 6/8, 12 -> 11
Single tightening pass on the permission dialog chrome so it matches the dense scale used by the rest of the chat surface. Not part of a larger bundle — small fix dropped on its own pending WAWQAQ's pivot direction for the bigger visible structural work. - .maka-permission-header grid 28 -> 24, gap 10 -> 8 - .maka-permission-icon 28 -> 24, border-radius 8 -> 6 - .maka-permission-age 12 -> 11 - .maka-permission-body gap 12 -> 10 - .maka-permission-summary gap 6 -> 5 - .maka-permission-line 12 -> 11 - .permissionRemember + .credentialField gap 6 -> 5, margin-top 12 -> 10, 13 -> 12
VISIBLE polish — every diff / terminal / load-tool result now lands with a 350ms scale-in (opacity 0 -> 1, scale .97 -> 1, cubic-bezier(.16, 1, .3, 1)) instead of popping in flat. Ported from the upstream reference design's compact code-output mount keyframes. - new keyframe @Keyframes maka-tool-card-enter - applied to .maka-tool-diff, .maka-tool-terminal, .maka-load-tool-preview - prefers-reduced-motion drops the animation Counterweight to recent rounds of pure chrome tightening (per WAWQAQ msg 4e5e6450 — micro 1-2px polish reads as "no change"; need motion + structural changes to register visually).
VISIBLE polish — every chat message (user, assistant, system) now slides up 220ms ease-out (opacity 0 -> 1, translateY 6px -> 0) instead of slamming in. Streaming deltas, session-switch loads, and new turns all share the same land-feel. - new keyframe @Keyframes maka-message-row-enter (matches reference design's agents-slide-up cadence) - applied to .maka-message-row (the article wrapper used by every chat row in MessageList) - prefers-reduced-motion + visual-smoke fixture both drop the animation Ported from the upstream reference design's chat-mount motion vocabulary.
VISIBLE polish — composer send button now reacts physically on cursor: hover scales 1.06 with a soft accent-tinted shadow, active press snaps to scale 0.96 and drops the shadow so the button feels weighty. Cubic-bezier on transform matches the upstream reference design's button cadence. - transition: background 150ms / transform 180ms cubic-bezier(.16,1,.3,1) / shadow 220ms - :hover transform scale(1.06) + 0 4px 14px -6px accent-tinted shadow - :active transform scale(0.96), shadow cleared Pixel size / fill stay locked at 28x28 black per the existing contract test.
VISIBLE polish — composer toolbar buttons and chips now share the same hover physics that the send button got in f468b25: - .maka-composer-tool-button (mic, attach, context+) gets transition cadence + :hover scale(1.08) + :active scale(0.94) - .maka-composer-role-chip / .maka-composer-mode-chip / .maka-composer-model-chip pick up the 150ms background/color/border transition so the hover bg shift no longer feels instant; border darkens to 0.16 alpha on hover All four toolbar surfaces now read as one coherent control group with consistent motion vocabulary.
VISIBLE polish — row surfaces now nudge 2px right on hover, the same gesture the maka-chat-jump-bottom button uses to telegraph "click me." Keeps the motion vocabulary coherent across the chat surface. - .maka-skill-library-row :hover transform translateX(2px) + transition cadence on the row itself (180ms cubic-bezier(.16,1,.3,1)) - .maka-onboarding-card :hover transform translateX(2px) (bg alpha stays at 0.04 per the locked contract test in localized-main-shell) Whole composer toolbar + chat list now reads as one consistent motion system.
VISIBLE polish — the composer streaming-hint dot used to be a single 6px ping. Replace with a 3-dot bouncing typing indicator (center + ::before + ::after, 4px vertical hop, staggered phases) so the chat reads as actively "talking back" instead of a single static blip. - 3 dots at 5x5, 9px apart, staggered animation phases (-0.4s, -0.2s, 0) - 4px vertical bounce + opacity 0.55 -> 1, 1.05s cycle - composer-streaming-hint gap 4 -> 8 to give the dots room - prefers-reduced-motion drops all 3 animations Also adds 220ms cubic-bezier transition on .maka-composer-inner box-shadow so the focus ring fades in smoothly instead of snapping.
VISIBLE polish — when an empty-chat / onboarding hero appears (new session, app boot, blocked state, ready_empty), it now slides up gently instead of slamming in flat: 320ms cubic-bezier(.16,1,.3,1), opacity 0 -> 1, translateY 8px -> 0. - new keyframe @Keyframes maka-hero-enter - applied to .maka-hero, .emptyChat (legacy alias) - prefers-reduced-motion + visual-smoke fixture both drop the animation Ported from the upstream reference design's hero-mount cadence. Completes the chat surface motion vocabulary: hero lands -> messages slide up -> tool cards scale in -> composer chips/buttons hover physics -> streaming indicator types back.
VISIBLE polish — when the model starts extended thinking, the reasoning panel now slides up gently on mount: 280ms cubic-bezier(.16,1,.3,1), opacity 0 -> 1, translateY 4px -> 0. Closes the chat surface entrance vocabulary: hero, message rows, tool cards, and reasoning panel all share the same land-feel. - new keyframe @Keyframes maka-reasoning-panel-enter - applied to .maka-reasoning-panel - prefers-reduced-motion drops the animation Final round (30/30) of the WAWQAQ-scheduled polish loop anchored on msg 8e09de0f.
icon.png shipped as a 1254² full-bleed image (artwork edge-to-edge, zero transparent margin), so macOS rendered it noticeably larger than sibling dock icons. Regenerate on the standard 1024² canvas with the squircle scaled to ~80% (824px) and centered, leaving the ~100px transparent margin macOS expects. Affects the dev `npm start` dock icon (app.dock.setIcon) and the BrowserWindow title-bar icon, both of which load assets/icon.png. Claude-Session: https://claude.ai/code/session_01KtNufPddivogjCYaQijf5t
BIG VISIBLE redesign per WAWQAQ msg 5a461e0c — Skills page and 定时任务 (Plan/ Reminder) page now render as lifted white cards on the gray plate, matching the upstream reference design's aux-panel signature instead of sitting flat. Each page shell gets: - 1px solid var(--border) outline - 10px border-radius - soft lift shadow: 0 1px 2px + 0 8px 24px -16px (foreground-tinted) - generous inner padding (16/18px Skills, 18px Plan) - 320ms cubic-bezier(.16,1,.3,1) entrance (opacity 0->1, translateY 6->0) Surfaces touched: - .maka-skill-library outer container — new card chrome + padding 16/18, gap 12 -> 14, page-shell entrance animation - .maka-plan-shell outer container — new card chrome + padding 18, gap 10 -> 12, top padding tightened from clamp(16,3vh,32) to (16,3vh,24) - new shared @Keyframes maka-page-shell-enter - prefers-reduced-motion + visual-smoke fixture both drop the animation This is the structural pivot WAWQAQ asked for: pages no longer read as flat lists on a flat canvas — they read as composed cards on a plate. Inner sub-cards (.maka-skill-library-list, individual reminder cards) keep their existing 1px borders, giving the reference layout's nested card hierarchy. NOTE: icon test (dist/main/__tests__/localized-main-shell-contract.test.js:341) is pre-existing red on origin/main after the icon PR #75 merged, unrelated to this change. Verified by stashing this change and re-running — same red.
) Reshape the desktop titlebar/cursor UI: - align the titlebar icon clusters (sidebar header, collapsed strip, workspace top-actions) to one vertical baseline via a single source-of-truth var - adopt the native cursor convention: arrow on controls, hand only on links, guarded by a static contract test with an exact `.maka-markdown-link-internal` allowlist - trim the baseline's explanatory comments to one line each Claude-Session: https://claude.ai/code/session_01VWXzHoyncpGYuwbaDEDizD
…g modal (#80) The search modal's migration to the Base UI Dialog primitives drifted the real-window smoke diagnostic's selectors, failing two programmatic checks on a clean tree. Bind the diagnostic to a stable `maka-dialog-backdrop` hook and detect the focus trap structurally; drop the dead `.maka-search-modal-backdrop` rule. All 6 programmatic smoke checks pass. Claude-Session: https://claude.ai/code/session_01VWXzHoyncpGYuwbaDEDizD
Astro-Han
force-pushed
the
codex/pi-agent-headless-smoke
branch
from
June 21, 2026 15:53
0fc419f to
5fefa1e
Compare
jackwener
force-pushed
the
main
branch
5 times, most recently
from
June 21, 2026 18:59
85f1f75 to
77a75f0
Compare
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
4 participants
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.
Summary
pi-agentheadless backend bridge that maps Pi JSON-mode output into Maka runtime events and token usageMAKA_PI_COMMANDand secret-file API key passthroughbackends.tsstays fake-only internal plumbingpi-agentandfakeslugsReview fixes
toolcall_endand tool results only frommessage_end, so runtime traces keep final args/results instead of partial start payloads.MAKA_BACKEND=pi-agentno longer parsesMAKA_MODELthrough Maka's AI SDK provider whitelist.npm install -gfor Pi; benchmark images/specs must providepior setMAKA_PI_COMMAND.Verification
npm run -w @maka/runtime test— 590 pass, 0 failnpm run -w @maka/headless test— 211 pass, 0 failReal Harbor smoke
Using the same
deepseek-v4-promodel and DeepSeek API key for both Maka and Pi:Across these five scored samples, both harnesses passed 5/5. Pi used about 9x the estimated cost under DeepSeek official
deepseek-v4-propricing.filter-js-from-htmlwas excluded from the scored comparison because both agent cells completed but Harbor's Selenium verifier did not finish before the batch run was interrupted.Note: these real Pi smoke runs used a container setup that made the Pi CLI available. After review hardening, the adapter verifies an existing Pi binary instead of installing it at root during adapter setup.