Make the Brain the Chat's living background; float components on legibility-floored Glass

[kateebonner]

Closes#56

Integration branch for the Living Chat redesign, built slice-by-slice from the sub-issue DAG (#59#60#61#62#63; design-of-record: ADR 0002, merged via #58).

Landed:

• Mount the Brain chat-wide and absorb the inline brain strip #59 — Brain mounted chat-wide (session + landing draft), inline brain strip absorbed (one render loop; TimelineRow.Brain retired; shimmer kept); live event feed extracted to pure, tested deriveBrainEvents.

• Derive two AA-floored, chat-theme-keyed Glass tint tiers with a contrast test #60 — two-tint AA-floored Glass token system (data-glass="standard|dense", one 18px constant blur, chat-theme-keyed per-mode tints derived against the per-theme reference frame — dark #fff676 / light #8f8000 — with a certified contrast test; @supports fallback).

• Float chat components on the two Glass tiers (standard prose/bubbles/composer, dense code/diffs/plots/tools) #61 — everything floats: standard tier on bubbles/prose/composer, dense on code/diffs/run-plots/tool cards; opaque fills removed so the Brain lives in the gaps; muted ink dense-backed by measurement; browser-verified over the live Brain, both themes.

• Drive the Brain by session activity: adaptive heartbeat, sparse-seed rest, bounded in-place growth #62 — adaptive heartbeat (setActive + shared draw-gate: full tempo working, ~8fps rest), reduced-motion hard-pause (burst-then-still), sparse-seed boot (core alone; grows only from session activity), total-population cap with recency eviction, fixed camera (densify in place). Orphaned skeleton dataset removed at integration.

• Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63 — perf governor: ease ladder under sustained over-budget p95 (window-scoped, hysteresis dead-band, no oscillation) → terminal hard-pause to a static blurred field; blur/tint pinned untouched by test; DEV ?brainForceActive hook; headless perf-trace proxy (e2e/perf/brain-governor-trace.spec.ts). Issue left open for its hitl acceptance criterion — see the sign-off checklist below.

Final gate on this branch: ui 402 pass / 0 fail · app 464 pass / 0 fail · typecheck 23/23 · lint 0 errors.

Honest pre-gate evidence (headless, dev build — NOT the gate): forced full tempo + 11s continuous scroll of a 144-message thread → dark p95 25.4ms / light 25.2ms, mean ≈16.3ms, 11–15 frames >33ms. The tail misses the 16.7ms budget in that proxy environment (dev-mode Solid, headless vsync, sampler riding the measured loop) — which is exactly what the pre-agreed governor valve is for. The real call happens on the reference laptop (checklist below).

Known follow-up candidates (from #61's browser pass, non-blocking): composer placeholder hint and reasoning-part still composite on standard/bare over the Brain.

🤖 Generated with Claude Code

[kateebonner]

HITL sign-off checklist — reference-laptop 60fps gate (#63)

The one open acceptance criterion on this PR. Run on the reference laptop, battery-powered:

• Pin the machine: model, OS, battery %, VS Code + webview (Chromium) version — write them here.
• Launch the dev lane: opencode serve --port 4099 + bun --cwd packages/app dev --port 5990 with VITE_OPENCODE_SERVER_PORT=4099 (the force hook is DEV-gated by design).
• Open the Chat in the VS Code webview on a real thread of ≥40 cards with ?brainForceActive appended; verify in devtools: window.__amicoBrainStats() → active: true, motion: "full".
• Record a ≥10s continuous scroll (devtools Performance panel, or the rAF sampler in Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63's notes ~600 frames).
• Pass = p95 ≤ 16.7ms AND zero frames > 33ms. Repeat in both themes (flip the chat scheme mid-session — Brain/Glass re-key losslessly).
• Write the numbers here and sign off.

If the gate fails: remove ?brainForceActive and confirm the governor engages under load — watch __amicoBrainStats().motion step eased-1 → eased-2 → full-stop → hard-paused with the frosted look intact — then decide accept-with-valve vs iterate. (Headless dev-proxy already suggests the tail will need the valve; see PR body.)

🤖 Generated with Claude Code

[kateebonner]

Live design-tuning batch (Kate, 2026-07-25) — commit 69ebf7e

The design evolved significantly beyond the original slices through a live review session. Landed:

• Glass — now purely blur + tint + radius: dark derives to fully transparent (brightness 0.8), no border, no top sheen, no box-shadow. AA kept for light, waived+recorded for dark standard (the landing constellation is yellow-free).
• Composer — one uniform glass surface (two-tone fixed); full-bleed bottom dock, ~35% viewport height, edge-to-edge; wordmark + project/folder picker removed.
• Landing — starter-chip wall removed; discovery moved into the composer as a rotating typewriter placeholder (types the real starters letter-by-letter, blinking caret, Tab-to-accept, reduced-motion static). Constellation contrast bumped.
• Session parity — the 3D constellation is now the background in-session too (not the sparse 2D graph); launch + session are one surface.

Queued (next): live-thought node pulsing over the constellation (data-true layer), card width containment, and folding the effort/variant selector into the model dropdown.

Gates: ui 438 / app 473 / typecheck 23/23 / lint 0 errors.

🤖 Generated with Claude Code

[kateebonner]

Overnight batch 2 — 05fa001d0

All three queued items from the live-tuning session, plus the new ask (session components on the launch-composer feel):

1. Live thought over the constellation ✨

The "+ live thought" half of constellation + live thought is in. As amico works, each real touch flares its constellation node in the thought color (#fff676 dark / #8f8000 light — the brand law holds per scheme): fast rise, ~1.6s ease-out decay, an expanding ring, glints down the node's incident edges, and the touched label ("session.tsx", "amico-vault") named beside the flare.

• Labels hash deterministically into their category lobe (file work lights the code lobe, skills the skills lobe…), then walk to a visible node — a flare you can't see isn't thought.
• History replays silently on mount; re-touches refresh their flare instead of stacking; capped at 24 so a torrent stays a shimmer.
• Reduced motion: the tableau holds one statically lit node — where amico is now.
• Dev preview without a real turn: window.__amicoBrainTouch({label:"x.ts", type:"resource"}) in the console (the ?-param pattern doesn't survive the router redirect, so this rides the __amicoBrainStats hook pattern instead).
• +7 tests: flare paint, light-scheme ink, silent replay, refresh-not-stack, decay-to-rest, reduced-motion tableau, and byte-determinism under touches. The latent-web color law tests still pin zero yellow with zero touches.

2. Thinking effort lives in the model menu

The standalone Default/None/High Select is gone (both composer paths). A ModelVariantRow renders inside the paid popover and the unpaid dialog (the dev/free path — it would otherwise have lost the control entirely): chips on --accent-fill-soft/--accent-edge, picking one keeps the menu open. The collapsed trigger now reads "DeepSeek V4 Fl… · High" so effort stays visible at rest.

3. Session glass sweep — the launch-composer feel everywhere

Glass = blur + tint + radius, nothing else, now enforced in-session too:

• dock-prompt (permission/question card): 1px ring removed, footer border-top removed (one surface, same law as the composer footer), radius 10px → --radius-lg.
• In-session composer: focus-within ring removed (the caret carries focus; drag keeps its dashed dropzone).
• Comment chips: borderless dense tint.
• Width containment: tool collapsibles ("Explored · 2 searches"), the diffs group, reasoning, and error cards all share the prose measure — min(88%, 72ch) — so the thread reads as floating cards over the Brain, never full-width bands.

Verified live in both themes on real sessions (silent replay ✓, live flare mid-session ✓, permission card ✓).

Gate: ui 445 / app 473 / typecheck / lint 0 errors.

[kateebonner]

Parked 2026-07-26 — converting to draft. Aaron's call: not pursuing this UI direction right now. Everything through 83d070ae4 is on the branch (living-chat glass redesign, thought-tracking camera, bottom-stacked edge-justified flow, effort-in-model-menu, component audit + width law). The branch is the complete archive of the direction if we return to it; the component library artifact documents every surface.

[jack-champagne]

@kateebonner — question before this ages further: is anything here still worth keeping?

Asking because ADR 0002, which this implements, is marked superseded on mainline:

Status: Superseded by ADR 0003 (direction parked 2026-07-26; the Brain moved to the context-tree top panel instead of becoming the chat background)

and ADR 0003 (Accepted, 2026-07-27, on Aaron's 2026-07-26 feedback) is explicit:

Supersedes: ADR 0002 (Brain as chat background — direction parked)
… Aaron's feedback reframed what the graph is for: not ambience, but orientation … The graph moves out of the timeline into a pinned top panel … the ambient background direction is parked.

The ADR-0003 replacement has fully landed

All of this is on local/amicode and shipped in v1.17.3-amicode.11:

• c8468a416 context tree goes organic — Obsidian-like force-settled graph (Context tree: organic Obsidian-like graph layout (stacked on #68) #70)
• ecb05faaa / 7610867bc side panel owns its width; vault fills it by default (Side panel sizing + vault-by-default + store sync; pending-chip hints #71)
• 5a1b6d219 context tree never surfaces user prompt text (fix(amicode): context tree never surfaces user prompt text #83)
• e64255834 context tree marks proprietary vault context locked (feat(amicode): context tree marks proprietary vault context locked #84/fix(amicode): land #84 on mainline (locked vault context) — stacked-merge miss #86)
• d5f2d60b8 header backdrop-blur off while the context-tree band is open (fix(session): header backdrop-blur off while the context-tree band is open #97)

So the top-panel + first-class vault surface direction is done and released.

What's still unique to this branch

brain-engine.ts and context-tree-engine.ts are already on main — the graph engine itself made it across. What has no counterpart is the presentation layer for the parked direction:

only in #64
glass.css (778 lines), glass-tokens.ts, script/generate-glass.ts
brain-atmosphere.tsx / .css
brain-constellation.ts
brain-perf-governor.test.ts, brain-events.ts, e2e/perf/brain-governor-trace.spec.ts

16 added files, 54 modified, 1 deleted.

Why I'm asking rather than just closing

Two things in there look potentially valuable independent of the parked background direction:

1. The legibility machinery — glass-tokens.ts plus the WCAG-AA-floored tint derivation. ADR 0002 records that the earlier attempt failed exactly one AA check (muted grey on the most translucent tier), so this encodes a real lesson. Worth salvaging if any surface still layers text on translucency — the session header's backdrop-blur comes to mind, given fix(session): header backdrop-blur off while the context-tree band is open #97 just had to switch it off over the context-tree band.
2. brain-perf-governor + the e2e perf trace — frame-budget governance for the graph. The organic context tree that did land has the same cost profile, so a perf governor may be wanted regardless of where the graph renders.

If neither is worth extracting, this is a clean close — the direction is parked and the replacement is released. If either is, they'd be small standalone PRs rather than resurrecting 71 files.

For context on the state: it is still a draft, CONFLICTING/DIRTY, 72 commits behind mainline, and its checks are all CANCELLED from before #2 fixed the runner labels — so CI has never actually evaluated this branch.

[kateebonner]

Closes#56

Integration branch for the Living Chat redesign, built slice-by-slice from the sub-issue DAG (#59#60#61#62#63; design-of-record: ADR 0002, merged via #58).

Landed:

• Mount the Brain chat-wide and absorb the inline brain strip #59 — Brain mounted chat-wide (session + landing draft), inline brain strip absorbed (one render loop; TimelineRow.Brain retired; shimmer kept); live event feed extracted to pure, tested deriveBrainEvents.

• Derive two AA-floored, chat-theme-keyed Glass tint tiers with a contrast test #60 — two-tint AA-floored Glass token system (data-glass="standard|dense", one 18px constant blur, chat-theme-keyed per-mode tints derived against the per-theme reference frame — dark #fff676 / light #8f8000 — with a certified contrast test; @supports fallback).

• Float chat components on the two Glass tiers (standard prose/bubbles/composer, dense code/diffs/plots/tools) #61 — everything floats: standard tier on bubbles/prose/composer, dense on code/diffs/run-plots/tool cards; opaque fills removed so the Brain lives in the gaps; muted ink dense-backed by measurement; browser-verified over the live Brain, both themes.

• Drive the Brain by session activity: adaptive heartbeat, sparse-seed rest, bounded in-place growth #62 — adaptive heartbeat (setActive + shared draw-gate: full tempo working, ~8fps rest), reduced-motion hard-pause (burst-then-still), sparse-seed boot (core alone; grows only from session activity), total-population cap with recency eviction, fixed camera (densify in place). Orphaned skeleton dataset removed at integration.

• Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63 — perf governor: ease ladder under sustained over-budget p95 (window-scoped, hysteresis dead-band, no oscillation) → terminal hard-pause to a static blurred field; blur/tint pinned untouched by test; DEV ?brainForceActive hook; headless perf-trace proxy (e2e/perf/brain-governor-trace.spec.ts). Issue left open for its hitl acceptance criterion — see the sign-off checklist below.

Final gate on this branch: ui 402 pass / 0 fail · app 464 pass / 0 fail · typecheck 23/23 · lint 0 errors.

Honest pre-gate evidence (headless, dev build — NOT the gate): forced full tempo + 11s continuous scroll of a 144-message thread → dark p95 25.4ms / light 25.2ms, mean ≈16.3ms, 11–15 frames >33ms. The tail misses the 16.7ms budget in that proxy environment (dev-mode Solid, headless vsync, sampler riding the measured loop) — which is exactly what the pre-agreed governor valve is for. The real call happens on the reference laptop (checklist below).

Known follow-up candidates (from #61's browser pass, non-blocking): composer placeholder hint and reasoning-part still composite on standard/bare over the Brain.

🤖 Generated with Claude Code

[kateebonner]

HITL sign-off checklist — reference-laptop 60fps gate (#63)

The one open acceptance criterion on this PR. Run on the reference laptop, battery-powered:

• Pin the machine: model, OS, battery %, VS Code + webview (Chromium) version — write them here.
• Launch the dev lane: opencode serve --port 4099 + bun --cwd packages/app dev --port 5990 with VITE_OPENCODE_SERVER_PORT=4099 (the force hook is DEV-gated by design).
• Open the Chat in the VS Code webview on a real thread of ≥40 cards with ?brainForceActive appended; verify in devtools: window.__amicoBrainStats() → active: true, motion: "full".
• Record a ≥10s continuous scroll (devtools Performance panel, or the rAF sampler in Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63's notes ~600 frames).
• Pass = p95 ≤ 16.7ms AND zero frames > 33ms. Repeat in both themes (flip the chat scheme mid-session — Brain/Glass re-key losslessly).
• Write the numbers here and sign off.

If the gate fails: remove ?brainForceActive and confirm the governor engages under load — watch __amicoBrainStats().motion step eased-1 → eased-2 → full-stop → hard-paused with the frosted look intact — then decide accept-with-valve vs iterate. (Headless dev-proxy already suggests the tail will need the valve; see PR body.)

🤖 Generated with Claude Code

[kateebonner]

Live design-tuning batch (Kate, 2026-07-25) — commit 69ebf7e

The design evolved significantly beyond the original slices through a live review session. Landed:

• Glass — now purely blur + tint + radius: dark derives to fully transparent (brightness 0.8), no border, no top sheen, no box-shadow. AA kept for light, waived+recorded for dark standard (the landing constellation is yellow-free).
• Composer — one uniform glass surface (two-tone fixed); full-bleed bottom dock, ~35% viewport height, edge-to-edge; wordmark + project/folder picker removed.
• Landing — starter-chip wall removed; discovery moved into the composer as a rotating typewriter placeholder (types the real starters letter-by-letter, blinking caret, Tab-to-accept, reduced-motion static). Constellation contrast bumped.
• Session parity — the 3D constellation is now the background in-session too (not the sparse 2D graph); launch + session are one surface.

Queued (next): live-thought node pulsing over the constellation (data-true layer), card width containment, and folding the effort/variant selector into the model dropdown.

Gates: ui 438 / app 473 / typecheck 23/23 / lint 0 errors.

🤖 Generated with Claude Code

[kateebonner]

Overnight batch 2 — 05fa001d0

All three queued items from the live-tuning session, plus the new ask (session components on the launch-composer feel):

1. Live thought over the constellation ✨

The "+ live thought" half of constellation + live thought is in. As amico works, each real touch flares its constellation node in the thought color (#fff676 dark / #8f8000 light — the brand law holds per scheme): fast rise, ~1.6s ease-out decay, an expanding ring, glints down the node's incident edges, and the touched label ("session.tsx", "amico-vault") named beside the flare.

• Labels hash deterministically into their category lobe (file work lights the code lobe, skills the skills lobe…), then walk to a visible node — a flare you can't see isn't thought.
• History replays silently on mount; re-touches refresh their flare instead of stacking; capped at 24 so a torrent stays a shimmer.
• Reduced motion: the tableau holds one statically lit node — where amico is now.
• Dev preview without a real turn: window.__amicoBrainTouch({label:"x.ts", type:"resource"}) in the console (the ?-param pattern doesn't survive the router redirect, so this rides the __amicoBrainStats hook pattern instead).
• +7 tests: flare paint, light-scheme ink, silent replay, refresh-not-stack, decay-to-rest, reduced-motion tableau, and byte-determinism under touches. The latent-web color law tests still pin zero yellow with zero touches.

2. Thinking effort lives in the model menu

The standalone Default/None/High Select is gone (both composer paths). A ModelVariantRow renders inside the paid popover and the unpaid dialog (the dev/free path — it would otherwise have lost the control entirely): chips on --accent-fill-soft/--accent-edge, picking one keeps the menu open. The collapsed trigger now reads "DeepSeek V4 Fl… · High" so effort stays visible at rest.

3. Session glass sweep — the launch-composer feel everywhere

Glass = blur + tint + radius, nothing else, now enforced in-session too:

• dock-prompt (permission/question card): 1px ring removed, footer border-top removed (one surface, same law as the composer footer), radius 10px → --radius-lg.
• In-session composer: focus-within ring removed (the caret carries focus; drag keeps its dashed dropzone).
• Comment chips: borderless dense tint.
• Width containment: tool collapsibles ("Explored · 2 searches"), the diffs group, reasoning, and error cards all share the prose measure — min(88%, 72ch) — so the thread reads as floating cards over the Brain, never full-width bands.

Verified live in both themes on real sessions (silent replay ✓, live flare mid-session ✓, permission card ✓).

Gate: ui 445 / app 473 / typecheck / lint 0 errors.

[kateebonner]

Parked 2026-07-26 — converting to draft. Aaron's call: not pursuing this UI direction right now. Everything through 83d070ae4 is on the branch (living-chat glass redesign, thought-tracking camera, bottom-stacked edge-justified flow, effort-in-model-menu, component audit + width law). The branch is the complete archive of the direction if we return to it; the component library artifact documents every surface.

[jack-champagne]

@kateebonner — question before this ages further: is anything here still worth keeping?

Asking because ADR 0002, which this implements, is marked superseded on mainline:

Status: Superseded by ADR 0003 (direction parked 2026-07-26; the Brain moved to the context-tree top panel instead of becoming the chat background)

and ADR 0003 (Accepted, 2026-07-27, on Aaron's 2026-07-26 feedback) is explicit:

Supersedes: ADR 0002 (Brain as chat background — direction parked)
… Aaron's feedback reframed what the graph is for: not ambience, but orientation … The graph moves out of the timeline into a pinned top panel … the ambient background direction is parked.

The ADR-0003 replacement has fully landed

All of this is on local/amicode and shipped in v1.17.3-amicode.11:

• c8468a416 context tree goes organic — Obsidian-like force-settled graph (Context tree: organic Obsidian-like graph layout (stacked on #68) #70)
• ecb05faaa / 7610867bc side panel owns its width; vault fills it by default (Side panel sizing + vault-by-default + store sync; pending-chip hints #71)
• 5a1b6d219 context tree never surfaces user prompt text (fix(amicode): context tree never surfaces user prompt text #83)
• e64255834 context tree marks proprietary vault context locked (feat(amicode): context tree marks proprietary vault context locked #84/fix(amicode): land #84 on mainline (locked vault context) — stacked-merge miss #86)
• d5f2d60b8 header backdrop-blur off while the context-tree band is open (fix(session): header backdrop-blur off while the context-tree band is open #97)

So the top-panel + first-class vault surface direction is done and released.

What's still unique to this branch

brain-engine.ts and context-tree-engine.ts are already on main — the graph engine itself made it across. What has no counterpart is the presentation layer for the parked direction:

only in #64
glass.css (778 lines), glass-tokens.ts, script/generate-glass.ts
brain-atmosphere.tsx / .css
brain-constellation.ts
brain-perf-governor.test.ts, brain-events.ts, e2e/perf/brain-governor-trace.spec.ts

16 added files, 54 modified, 1 deleted.

Why I'm asking rather than just closing

Two things in there look potentially valuable independent of the parked background direction:

1. The legibility machinery — glass-tokens.ts plus the WCAG-AA-floored tint derivation. ADR 0002 records that the earlier attempt failed exactly one AA check (muted grey on the most translucent tier), so this encodes a real lesson. Worth salvaging if any surface still layers text on translucency — the session header's backdrop-blur comes to mind, given fix(session): header backdrop-blur off while the context-tree band is open #97 just had to switch it off over the context-tree band.
2. brain-perf-governor + the e2e perf trace — frame-budget governance for the graph. The organic context tree that did land has the same cost profile, so a perf governor may be wanted regardless of where the graph renders.

If neither is worth extracting, this is a clean close — the direction is parked and the replacement is released. If either is, they'd be small standalone PRs rather than resurrecting 71 files.

For context on the state: it is still a draft, CONFLICTING/DIRTY, 72 commits behind mainline, and its checks are all CANCELLED from before #2 fixed the runner labels — so CI has never actually evaluated this branch.

[kateebonner]

Closes#56

Integration branch for the Living Chat redesign, built slice-by-slice from the sub-issue DAG (#59#60#61#62#63; design-of-record: ADR 0002, merged via #58).

Landed:

• Mount the Brain chat-wide and absorb the inline brain strip #59 — Brain mounted chat-wide (session + landing draft), inline brain strip absorbed (one render loop; TimelineRow.Brain retired; shimmer kept); live event feed extracted to pure, tested deriveBrainEvents.

• Derive two AA-floored, chat-theme-keyed Glass tint tiers with a contrast test #60 — two-tint AA-floored Glass token system (data-glass="standard|dense", one 18px constant blur, chat-theme-keyed per-mode tints derived against the per-theme reference frame — dark #fff676 / light #8f8000 — with a certified contrast test; @supports fallback).

• Float chat components on the two Glass tiers (standard prose/bubbles/composer, dense code/diffs/plots/tools) #61 — everything floats: standard tier on bubbles/prose/composer, dense on code/diffs/run-plots/tool cards; opaque fills removed so the Brain lives in the gaps; muted ink dense-backed by measurement; browser-verified over the live Brain, both themes.

• Drive the Brain by session activity: adaptive heartbeat, sparse-seed rest, bounded in-place growth #62 — adaptive heartbeat (setActive + shared draw-gate: full tempo working, ~8fps rest), reduced-motion hard-pause (burst-then-still), sparse-seed boot (core alone; grows only from session activity), total-population cap with recency eviction, fixed camera (densify in place). Orphaned skeleton dataset removed at integration.

• Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63 — perf governor: ease ladder under sustained over-budget p95 (window-scoped, hysteresis dead-band, no oscillation) → terminal hard-pause to a static blurred field; blur/tint pinned untouched by test; DEV ?brainForceActive hook; headless perf-trace proxy (e2e/perf/brain-governor-trace.spec.ts). Issue left open for its hitl acceptance criterion — see the sign-off checklist below.

Final gate on this branch: ui 402 pass / 0 fail · app 464 pass / 0 fail · typecheck 23/23 · lint 0 errors.

Honest pre-gate evidence (headless, dev build — NOT the gate): forced full tempo + 11s continuous scroll of a 144-message thread → dark p95 25.4ms / light 25.2ms, mean ≈16.3ms, 11–15 frames >33ms. The tail misses the 16.7ms budget in that proxy environment (dev-mode Solid, headless vsync, sampler riding the measured loop) — which is exactly what the pre-agreed governor valve is for. The real call happens on the reference laptop (checklist below).

Known follow-up candidates (from #61's browser pass, non-blocking): composer placeholder hint and reasoning-part still composite on standard/bare over the Brain.

🤖 Generated with Claude Code

[kateebonner]

HITL sign-off checklist — reference-laptop 60fps gate (#63)

The one open acceptance criterion on this PR. Run on the reference laptop, battery-powered:

• Pin the machine: model, OS, battery %, VS Code + webview (Chromium) version — write them here.
• Launch the dev lane: opencode serve --port 4099 + bun --cwd packages/app dev --port 5990 with VITE_OPENCODE_SERVER_PORT=4099 (the force hook is DEV-gated by design).
• Open the Chat in the VS Code webview on a real thread of ≥40 cards with ?brainForceActive appended; verify in devtools: window.__amicoBrainStats() → active: true, motion: "full".
• Record a ≥10s continuous scroll (devtools Performance panel, or the rAF sampler in Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63's notes ~600 frames).
• Pass = p95 ≤ 16.7ms AND zero frames > 33ms. Repeat in both themes (flip the chat scheme mid-session — Brain/Glass re-key losslessly).
• Write the numbers here and sign off.

If the gate fails: remove ?brainForceActive and confirm the governor engages under load — watch __amicoBrainStats().motion step eased-1 → eased-2 → full-stop → hard-paused with the frosted look intact — then decide accept-with-valve vs iterate. (Headless dev-proxy already suggests the tail will need the valve; see PR body.)

🤖 Generated with Claude Code

[kateebonner]

Live design-tuning batch (Kate, 2026-07-25) — commit 69ebf7e

The design evolved significantly beyond the original slices through a live review session. Landed:

• Glass — now purely blur + tint + radius: dark derives to fully transparent (brightness 0.8), no border, no top sheen, no box-shadow. AA kept for light, waived+recorded for dark standard (the landing constellation is yellow-free).
• Composer — one uniform glass surface (two-tone fixed); full-bleed bottom dock, ~35% viewport height, edge-to-edge; wordmark + project/folder picker removed.
• Landing — starter-chip wall removed; discovery moved into the composer as a rotating typewriter placeholder (types the real starters letter-by-letter, blinking caret, Tab-to-accept, reduced-motion static). Constellation contrast bumped.
• Session parity — the 3D constellation is now the background in-session too (not the sparse 2D graph); launch + session are one surface.

Queued (next): live-thought node pulsing over the constellation (data-true layer), card width containment, and folding the effort/variant selector into the model dropdown.

Gates: ui 438 / app 473 / typecheck 23/23 / lint 0 errors.

🤖 Generated with Claude Code

[kateebonner]

Overnight batch 2 — 05fa001d0

All three queued items from the live-tuning session, plus the new ask (session components on the launch-composer feel):

1. Live thought over the constellation ✨

The "+ live thought" half of constellation + live thought is in. As amico works, each real touch flares its constellation node in the thought color (#fff676 dark / #8f8000 light — the brand law holds per scheme): fast rise, ~1.6s ease-out decay, an expanding ring, glints down the node's incident edges, and the touched label ("session.tsx", "amico-vault") named beside the flare.

• Labels hash deterministically into their category lobe (file work lights the code lobe, skills the skills lobe…), then walk to a visible node — a flare you can't see isn't thought.
• History replays silently on mount; re-touches refresh their flare instead of stacking; capped at 24 so a torrent stays a shimmer.
• Reduced motion: the tableau holds one statically lit node — where amico is now.
• Dev preview without a real turn: window.__amicoBrainTouch({label:"x.ts", type:"resource"}) in the console (the ?-param pattern doesn't survive the router redirect, so this rides the __amicoBrainStats hook pattern instead).
• +7 tests: flare paint, light-scheme ink, silent replay, refresh-not-stack, decay-to-rest, reduced-motion tableau, and byte-determinism under touches. The latent-web color law tests still pin zero yellow with zero touches.

2. Thinking effort lives in the model menu

The standalone Default/None/High Select is gone (both composer paths). A ModelVariantRow renders inside the paid popover and the unpaid dialog (the dev/free path — it would otherwise have lost the control entirely): chips on --accent-fill-soft/--accent-edge, picking one keeps the menu open. The collapsed trigger now reads "DeepSeek V4 Fl… · High" so effort stays visible at rest.

3. Session glass sweep — the launch-composer feel everywhere

Glass = blur + tint + radius, nothing else, now enforced in-session too:

• dock-prompt (permission/question card): 1px ring removed, footer border-top removed (one surface, same law as the composer footer), radius 10px → --radius-lg.
• In-session composer: focus-within ring removed (the caret carries focus; drag keeps its dashed dropzone).
• Comment chips: borderless dense tint.
• Width containment: tool collapsibles ("Explored · 2 searches"), the diffs group, reasoning, and error cards all share the prose measure — min(88%, 72ch) — so the thread reads as floating cards over the Brain, never full-width bands.

Verified live in both themes on real sessions (silent replay ✓, live flare mid-session ✓, permission card ✓).

Gate: ui 445 / app 473 / typecheck / lint 0 errors.

[kateebonner]

Parked 2026-07-26 — converting to draft. Aaron's call: not pursuing this UI direction right now. Everything through 83d070ae4 is on the branch (living-chat glass redesign, thought-tracking camera, bottom-stacked edge-justified flow, effort-in-model-menu, component audit + width law). The branch is the complete archive of the direction if we return to it; the component library artifact documents every surface.

[jack-champagne]

@kateebonner — question before this ages further: is anything here still worth keeping?

Asking because ADR 0002, which this implements, is marked superseded on mainline:

Status: Superseded by ADR 0003 (direction parked 2026-07-26; the Brain moved to the context-tree top panel instead of becoming the chat background)

and ADR 0003 (Accepted, 2026-07-27, on Aaron's 2026-07-26 feedback) is explicit:

Supersedes: ADR 0002 (Brain as chat background — direction parked)
… Aaron's feedback reframed what the graph is for: not ambience, but orientation … The graph moves out of the timeline into a pinned top panel … the ambient background direction is parked.

The ADR-0003 replacement has fully landed

All of this is on local/amicode and shipped in v1.17.3-amicode.11:

• c8468a416 context tree goes organic — Obsidian-like force-settled graph (Context tree: organic Obsidian-like graph layout (stacked on #68) #70)
• ecb05faaa / 7610867bc side panel owns its width; vault fills it by default (Side panel sizing + vault-by-default + store sync; pending-chip hints #71)
• 5a1b6d219 context tree never surfaces user prompt text (fix(amicode): context tree never surfaces user prompt text #83)
• e64255834 context tree marks proprietary vault context locked (feat(amicode): context tree marks proprietary vault context locked #84/fix(amicode): land #84 on mainline (locked vault context) — stacked-merge miss #86)
• d5f2d60b8 header backdrop-blur off while the context-tree band is open (fix(session): header backdrop-blur off while the context-tree band is open #97)

So the top-panel + first-class vault surface direction is done and released.

What's still unique to this branch

brain-engine.ts and context-tree-engine.ts are already on main — the graph engine itself made it across. What has no counterpart is the presentation layer for the parked direction:

only in #64
glass.css (778 lines), glass-tokens.ts, script/generate-glass.ts
brain-atmosphere.tsx / .css
brain-constellation.ts
brain-perf-governor.test.ts, brain-events.ts, e2e/perf/brain-governor-trace.spec.ts

16 added files, 54 modified, 1 deleted.

Why I'm asking rather than just closing

Two things in there look potentially valuable independent of the parked background direction:

1. The legibility machinery — glass-tokens.ts plus the WCAG-AA-floored tint derivation. ADR 0002 records that the earlier attempt failed exactly one AA check (muted grey on the most translucent tier), so this encodes a real lesson. Worth salvaging if any surface still layers text on translucency — the session header's backdrop-blur comes to mind, given fix(session): header backdrop-blur off while the context-tree band is open #97 just had to switch it off over the context-tree band.
2. brain-perf-governor + the e2e perf trace — frame-budget governance for the graph. The organic context tree that did land has the same cost profile, so a perf governor may be wanted regardless of where the graph renders.

If neither is worth extracting, this is a clean close — the direction is parked and the replacement is released. If either is, they'd be small standalone PRs rather than resurrecting 71 files.

For context on the state: it is still a draft, CONFLICTING/DIRTY, 72 commits behind mainline, and its checks are all CANCELLED from before #2 fixed the runner labels — so CI has never actually evaluated this branch.

[kateebonner]

Closes#56

Integration branch for the Living Chat redesign, built slice-by-slice from the sub-issue DAG (#59#60#61#62#63; design-of-record: ADR 0002, merged via #58).

Landed:

• Mount the Brain chat-wide and absorb the inline brain strip #59 — Brain mounted chat-wide (session + landing draft), inline brain strip absorbed (one render loop; TimelineRow.Brain retired; shimmer kept); live event feed extracted to pure, tested deriveBrainEvents.

• Derive two AA-floored, chat-theme-keyed Glass tint tiers with a contrast test #60 — two-tint AA-floored Glass token system (data-glass="standard|dense", one 18px constant blur, chat-theme-keyed per-mode tints derived against the per-theme reference frame — dark #fff676 / light #8f8000 — with a certified contrast test; @supports fallback).

• Float chat components on the two Glass tiers (standard prose/bubbles/composer, dense code/diffs/plots/tools) #61 — everything floats: standard tier on bubbles/prose/composer, dense on code/diffs/run-plots/tool cards; opaque fills removed so the Brain lives in the gaps; muted ink dense-backed by measurement; browser-verified over the live Brain, both themes.

• Drive the Brain by session activity: adaptive heartbeat, sparse-seed rest, bounded in-place growth #62 — adaptive heartbeat (setActive + shared draw-gate: full tempo working, ~8fps rest), reduced-motion hard-pause (burst-then-still), sparse-seed boot (core alone; grows only from session activity), total-population cap with recency eviction, fixed camera (densify in place). Orphaned skeleton dataset removed at integration.

• Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63 — perf governor: ease ladder under sustained over-budget p95 (window-scoped, hysteresis dead-band, no oscillation) → terminal hard-pause to a static blurred field; blur/tint pinned untouched by test; DEV ?brainForceActive hook; headless perf-trace proxy (e2e/perf/brain-governor-trace.spec.ts). Issue left open for its hitl acceptance criterion — see the sign-off checklist below.

Final gate on this branch: ui 402 pass / 0 fail · app 464 pass / 0 fail · typecheck 23/23 · lint 0 errors.

Honest pre-gate evidence (headless, dev build — NOT the gate): forced full tempo + 11s continuous scroll of a 144-message thread → dark p95 25.4ms / light 25.2ms, mean ≈16.3ms, 11–15 frames >33ms. The tail misses the 16.7ms budget in that proxy environment (dev-mode Solid, headless vsync, sampler riding the measured loop) — which is exactly what the pre-agreed governor valve is for. The real call happens on the reference laptop (checklist below).

Known follow-up candidates (from #61's browser pass, non-blocking): composer placeholder hint and reasoning-part still composite on standard/bare over the Brain.

🤖 Generated with Claude Code

[kateebonner]

HITL sign-off checklist — reference-laptop 60fps gate (#63)

The one open acceptance criterion on this PR. Run on the reference laptop, battery-powered:

• Pin the machine: model, OS, battery %, VS Code + webview (Chromium) version — write them here.
• Launch the dev lane: opencode serve --port 4099 + bun --cwd packages/app dev --port 5990 with VITE_OPENCODE_SERVER_PORT=4099 (the force hook is DEV-gated by design).
• Open the Chat in the VS Code webview on a real thread of ≥40 cards with ?brainForceActive appended; verify in devtools: window.__amicoBrainStats() → active: true, motion: "full".
• Record a ≥10s continuous scroll (devtools Performance panel, or the rAF sampler in Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63's notes ~600 frames).
• Pass = p95 ≤ 16.7ms AND zero frames > 33ms. Repeat in both themes (flip the chat scheme mid-session — Brain/Glass re-key losslessly).
• Write the numbers here and sign off.

If the gate fails: remove ?brainForceActive and confirm the governor engages under load — watch __amicoBrainStats().motion step eased-1 → eased-2 → full-stop → hard-paused with the frosted look intact — then decide accept-with-valve vs iterate. (Headless dev-proxy already suggests the tail will need the valve; see PR body.)

🤖 Generated with Claude Code

[kateebonner]

Live design-tuning batch (Kate, 2026-07-25) — commit 69ebf7e

The design evolved significantly beyond the original slices through a live review session. Landed:

• Glass — now purely blur + tint + radius: dark derives to fully transparent (brightness 0.8), no border, no top sheen, no box-shadow. AA kept for light, waived+recorded for dark standard (the landing constellation is yellow-free).
• Composer — one uniform glass surface (two-tone fixed); full-bleed bottom dock, ~35% viewport height, edge-to-edge; wordmark + project/folder picker removed.
• Landing — starter-chip wall removed; discovery moved into the composer as a rotating typewriter placeholder (types the real starters letter-by-letter, blinking caret, Tab-to-accept, reduced-motion static). Constellation contrast bumped.
• Session parity — the 3D constellation is now the background in-session too (not the sparse 2D graph); launch + session are one surface.

Queued (next): live-thought node pulsing over the constellation (data-true layer), card width containment, and folding the effort/variant selector into the model dropdown.

Gates: ui 438 / app 473 / typecheck 23/23 / lint 0 errors.

🤖 Generated with Claude Code

[kateebonner]

Overnight batch 2 — 05fa001d0

All three queued items from the live-tuning session, plus the new ask (session components on the launch-composer feel):

1. Live thought over the constellation ✨

The "+ live thought" half of constellation + live thought is in. As amico works, each real touch flares its constellation node in the thought color (#fff676 dark / #8f8000 light — the brand law holds per scheme): fast rise, ~1.6s ease-out decay, an expanding ring, glints down the node's incident edges, and the touched label ("session.tsx", "amico-vault") named beside the flare.

• Labels hash deterministically into their category lobe (file work lights the code lobe, skills the skills lobe…), then walk to a visible node — a flare you can't see isn't thought.
• History replays silently on mount; re-touches refresh their flare instead of stacking; capped at 24 so a torrent stays a shimmer.
• Reduced motion: the tableau holds one statically lit node — where amico is now.
• Dev preview without a real turn: window.__amicoBrainTouch({label:"x.ts", type:"resource"}) in the console (the ?-param pattern doesn't survive the router redirect, so this rides the __amicoBrainStats hook pattern instead).
• +7 tests: flare paint, light-scheme ink, silent replay, refresh-not-stack, decay-to-rest, reduced-motion tableau, and byte-determinism under touches. The latent-web color law tests still pin zero yellow with zero touches.

2. Thinking effort lives in the model menu

The standalone Default/None/High Select is gone (both composer paths). A ModelVariantRow renders inside the paid popover and the unpaid dialog (the dev/free path — it would otherwise have lost the control entirely): chips on --accent-fill-soft/--accent-edge, picking one keeps the menu open. The collapsed trigger now reads "DeepSeek V4 Fl… · High" so effort stays visible at rest.

3. Session glass sweep — the launch-composer feel everywhere

Glass = blur + tint + radius, nothing else, now enforced in-session too:

• dock-prompt (permission/question card): 1px ring removed, footer border-top removed (one surface, same law as the composer footer), radius 10px → --radius-lg.
• In-session composer: focus-within ring removed (the caret carries focus; drag keeps its dashed dropzone).
• Comment chips: borderless dense tint.
• Width containment: tool collapsibles ("Explored · 2 searches"), the diffs group, reasoning, and error cards all share the prose measure — min(88%, 72ch) — so the thread reads as floating cards over the Brain, never full-width bands.

Verified live in both themes on real sessions (silent replay ✓, live flare mid-session ✓, permission card ✓).

Gate: ui 445 / app 473 / typecheck / lint 0 errors.

[kateebonner]

Parked 2026-07-26 — converting to draft. Aaron's call: not pursuing this UI direction right now. Everything through 83d070ae4 is on the branch (living-chat glass redesign, thought-tracking camera, bottom-stacked edge-justified flow, effort-in-model-menu, component audit + width law). The branch is the complete archive of the direction if we return to it; the component library artifact documents every surface.

[jack-champagne]

@kateebonner — question before this ages further: is anything here still worth keeping?

Asking because ADR 0002, which this implements, is marked superseded on mainline:

Status: Superseded by ADR 0003 (direction parked 2026-07-26; the Brain moved to the context-tree top panel instead of becoming the chat background)

and ADR 0003 (Accepted, 2026-07-27, on Aaron's 2026-07-26 feedback) is explicit:

Supersedes: ADR 0002 (Brain as chat background — direction parked)
… Aaron's feedback reframed what the graph is for: not ambience, but orientation … The graph moves out of the timeline into a pinned top panel … the ambient background direction is parked.

The ADR-0003 replacement has fully landed

All of this is on local/amicode and shipped in v1.17.3-amicode.11:

• c8468a416 context tree goes organic — Obsidian-like force-settled graph (Context tree: organic Obsidian-like graph layout (stacked on #68) #70)
• ecb05faaa / 7610867bc side panel owns its width; vault fills it by default (Side panel sizing + vault-by-default + store sync; pending-chip hints #71)
• 5a1b6d219 context tree never surfaces user prompt text (fix(amicode): context tree never surfaces user prompt text #83)
• e64255834 context tree marks proprietary vault context locked (feat(amicode): context tree marks proprietary vault context locked #84/fix(amicode): land #84 on mainline (locked vault context) — stacked-merge miss #86)
• d5f2d60b8 header backdrop-blur off while the context-tree band is open (fix(session): header backdrop-blur off while the context-tree band is open #97)

So the top-panel + first-class vault surface direction is done and released.

What's still unique to this branch

brain-engine.ts and context-tree-engine.ts are already on main — the graph engine itself made it across. What has no counterpart is the presentation layer for the parked direction:

only in #64
glass.css (778 lines), glass-tokens.ts, script/generate-glass.ts
brain-atmosphere.tsx / .css
brain-constellation.ts
brain-perf-governor.test.ts, brain-events.ts, e2e/perf/brain-governor-trace.spec.ts

16 added files, 54 modified, 1 deleted.

Why I'm asking rather than just closing

Two things in there look potentially valuable independent of the parked background direction:

1. The legibility machinery — glass-tokens.ts plus the WCAG-AA-floored tint derivation. ADR 0002 records that the earlier attempt failed exactly one AA check (muted grey on the most translucent tier), so this encodes a real lesson. Worth salvaging if any surface still layers text on translucency — the session header's backdrop-blur comes to mind, given fix(session): header backdrop-blur off while the context-tree band is open #97 just had to switch it off over the context-tree band.
2. brain-perf-governor + the e2e perf trace — frame-budget governance for the graph. The organic context tree that did land has the same cost profile, so a perf governor may be wanted regardless of where the graph renders.

If neither is worth extracting, this is a clean close — the direction is parked and the replacement is released. If either is, they'd be small standalone PRs rather than resurrecting 71 files.

For context on the state: it is still a draft, CONFLICTING/DIRTY, 72 commits behind mainline, and its checks are all CANCELLED from before #2 fixed the runner labels — so CI has never actually evaluated this branch.

[kateebonner]

Closes#56

Integration branch for the Living Chat redesign, built slice-by-slice from the sub-issue DAG (#59#60#61#62#63; design-of-record: ADR 0002, merged via #58).

Landed:

• Mount the Brain chat-wide and absorb the inline brain strip #59 — Brain mounted chat-wide (session + landing draft), inline brain strip absorbed (one render loop; TimelineRow.Brain retired; shimmer kept); live event feed extracted to pure, tested deriveBrainEvents.

• Derive two AA-floored, chat-theme-keyed Glass tint tiers with a contrast test #60 — two-tint AA-floored Glass token system (data-glass="standard|dense", one 18px constant blur, chat-theme-keyed per-mode tints derived against the per-theme reference frame — dark #fff676 / light #8f8000 — with a certified contrast test; @supports fallback).

• Float chat components on the two Glass tiers (standard prose/bubbles/composer, dense code/diffs/plots/tools) #61 — everything floats: standard tier on bubbles/prose/composer, dense on code/diffs/run-plots/tool cards; opaque fills removed so the Brain lives in the gaps; muted ink dense-backed by measurement; browser-verified over the live Brain, both themes.

• Drive the Brain by session activity: adaptive heartbeat, sparse-seed rest, bounded in-place growth #62 — adaptive heartbeat (setActive + shared draw-gate: full tempo working, ~8fps rest), reduced-motion hard-pause (burst-then-still), sparse-seed boot (core alone; grows only from session activity), total-population cap with recency eviction, fixed camera (densify in place). Orphaned skeleton dataset removed at integration.

• Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63 — perf governor: ease ladder under sustained over-budget p95 (window-scoped, hysteresis dead-band, no oscillation) → terminal hard-pause to a static blurred field; blur/tint pinned untouched by test; DEV ?brainForceActive hook; headless perf-trace proxy (e2e/perf/brain-governor-trace.spec.ts). Issue left open for its hitl acceptance criterion — see the sign-off checklist below.

Final gate on this branch: ui 402 pass / 0 fail · app 464 pass / 0 fail · typecheck 23/23 · lint 0 errors.

Honest pre-gate evidence (headless, dev build — NOT the gate): forced full tempo + 11s continuous scroll of a 144-message thread → dark p95 25.4ms / light 25.2ms, mean ≈16.3ms, 11–15 frames >33ms. The tail misses the 16.7ms budget in that proxy environment (dev-mode Solid, headless vsync, sampler riding the measured loop) — which is exactly what the pre-agreed governor valve is for. The real call happens on the reference laptop (checklist below).

Known follow-up candidates (from #61's browser pass, non-blocking): composer placeholder hint and reasoning-part still composite on standard/bare over the Brain.

🤖 Generated with Claude Code

[kateebonner]

HITL sign-off checklist — reference-laptop 60fps gate (#63)

The one open acceptance criterion on this PR. Run on the reference laptop, battery-powered:

• Pin the machine: model, OS, battery %, VS Code + webview (Chromium) version — write them here.
• Launch the dev lane: opencode serve --port 4099 + bun --cwd packages/app dev --port 5990 with VITE_OPENCODE_SERVER_PORT=4099 (the force hook is DEV-gated by design).
• Open the Chat in the VS Code webview on a real thread of ≥40 cards with ?brainForceActive appended; verify in devtools: window.__amicoBrainStats() → active: true, motion: "full".
• Record a ≥10s continuous scroll (devtools Performance panel, or the rAF sampler in Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63's notes ~600 frames).
• Pass = p95 ≤ 16.7ms AND zero frames > 33ms. Repeat in both themes (flip the chat scheme mid-session — Brain/Glass re-key losslessly).
• Write the numbers here and sign off.

If the gate fails: remove ?brainForceActive and confirm the governor engages under load — watch __amicoBrainStats().motion step eased-1 → eased-2 → full-stop → hard-paused with the frosted look intact — then decide accept-with-valve vs iterate. (Headless dev-proxy already suggests the tail will need the valve; see PR body.)

🤖 Generated with Claude Code

[kateebonner]

Live design-tuning batch (Kate, 2026-07-25) — commit 69ebf7e

The design evolved significantly beyond the original slices through a live review session. Landed:

• Glass — now purely blur + tint + radius: dark derives to fully transparent (brightness 0.8), no border, no top sheen, no box-shadow. AA kept for light, waived+recorded for dark standard (the landing constellation is yellow-free).
• Composer — one uniform glass surface (two-tone fixed); full-bleed bottom dock, ~35% viewport height, edge-to-edge; wordmark + project/folder picker removed.
• Landing — starter-chip wall removed; discovery moved into the composer as a rotating typewriter placeholder (types the real starters letter-by-letter, blinking caret, Tab-to-accept, reduced-motion static). Constellation contrast bumped.
• Session parity — the 3D constellation is now the background in-session too (not the sparse 2D graph); launch + session are one surface.

Queued (next): live-thought node pulsing over the constellation (data-true layer), card width containment, and folding the effort/variant selector into the model dropdown.

Gates: ui 438 / app 473 / typecheck 23/23 / lint 0 errors.

🤖 Generated with Claude Code

[kateebonner]

Overnight batch 2 — 05fa001d0

All three queued items from the live-tuning session, plus the new ask (session components on the launch-composer feel):

1. Live thought over the constellation ✨

The "+ live thought" half of constellation + live thought is in. As amico works, each real touch flares its constellation node in the thought color (#fff676 dark / #8f8000 light — the brand law holds per scheme): fast rise, ~1.6s ease-out decay, an expanding ring, glints down the node's incident edges, and the touched label ("session.tsx", "amico-vault") named beside the flare.

• Labels hash deterministically into their category lobe (file work lights the code lobe, skills the skills lobe…), then walk to a visible node — a flare you can't see isn't thought.
• History replays silently on mount; re-touches refresh their flare instead of stacking; capped at 24 so a torrent stays a shimmer.
• Reduced motion: the tableau holds one statically lit node — where amico is now.
• Dev preview without a real turn: window.__amicoBrainTouch({label:"x.ts", type:"resource"}) in the console (the ?-param pattern doesn't survive the router redirect, so this rides the __amicoBrainStats hook pattern instead).
• +7 tests: flare paint, light-scheme ink, silent replay, refresh-not-stack, decay-to-rest, reduced-motion tableau, and byte-determinism under touches. The latent-web color law tests still pin zero yellow with zero touches.

2. Thinking effort lives in the model menu

The standalone Default/None/High Select is gone (both composer paths). A ModelVariantRow renders inside the paid popover and the unpaid dialog (the dev/free path — it would otherwise have lost the control entirely): chips on --accent-fill-soft/--accent-edge, picking one keeps the menu open. The collapsed trigger now reads "DeepSeek V4 Fl… · High" so effort stays visible at rest.

3. Session glass sweep — the launch-composer feel everywhere

Glass = blur + tint + radius, nothing else, now enforced in-session too:

• dock-prompt (permission/question card): 1px ring removed, footer border-top removed (one surface, same law as the composer footer), radius 10px → --radius-lg.
• In-session composer: focus-within ring removed (the caret carries focus; drag keeps its dashed dropzone).
• Comment chips: borderless dense tint.
• Width containment: tool collapsibles ("Explored · 2 searches"), the diffs group, reasoning, and error cards all share the prose measure — min(88%, 72ch) — so the thread reads as floating cards over the Brain, never full-width bands.

Verified live in both themes on real sessions (silent replay ✓, live flare mid-session ✓, permission card ✓).

Gate: ui 445 / app 473 / typecheck / lint 0 errors.

[kateebonner]

Parked 2026-07-26 — converting to draft. Aaron's call: not pursuing this UI direction right now. Everything through 83d070ae4 is on the branch (living-chat glass redesign, thought-tracking camera, bottom-stacked edge-justified flow, effort-in-model-menu, component audit + width law). The branch is the complete archive of the direction if we return to it; the component library artifact documents every surface.

[jack-champagne]

@kateebonner — question before this ages further: is anything here still worth keeping?

Asking because ADR 0002, which this implements, is marked superseded on mainline:

Status: Superseded by ADR 0003 (direction parked 2026-07-26; the Brain moved to the context-tree top panel instead of becoming the chat background)

and ADR 0003 (Accepted, 2026-07-27, on Aaron's 2026-07-26 feedback) is explicit:

Supersedes: ADR 0002 (Brain as chat background — direction parked)
… Aaron's feedback reframed what the graph is for: not ambience, but orientation … The graph moves out of the timeline into a pinned top panel … the ambient background direction is parked.

The ADR-0003 replacement has fully landed

All of this is on local/amicode and shipped in v1.17.3-amicode.11:

• c8468a416 context tree goes organic — Obsidian-like force-settled graph (Context tree: organic Obsidian-like graph layout (stacked on #68) #70)
• ecb05faaa / 7610867bc side panel owns its width; vault fills it by default (Side panel sizing + vault-by-default + store sync; pending-chip hints #71)
• 5a1b6d219 context tree never surfaces user prompt text (fix(amicode): context tree never surfaces user prompt text #83)
• e64255834 context tree marks proprietary vault context locked (feat(amicode): context tree marks proprietary vault context locked #84/fix(amicode): land #84 on mainline (locked vault context) — stacked-merge miss #86)
• d5f2d60b8 header backdrop-blur off while the context-tree band is open (fix(session): header backdrop-blur off while the context-tree band is open #97)

So the top-panel + first-class vault surface direction is done and released.

What's still unique to this branch

brain-engine.ts and context-tree-engine.ts are already on main — the graph engine itself made it across. What has no counterpart is the presentation layer for the parked direction:

only in #64
glass.css (778 lines), glass-tokens.ts, script/generate-glass.ts
brain-atmosphere.tsx / .css
brain-constellation.ts
brain-perf-governor.test.ts, brain-events.ts, e2e/perf/brain-governor-trace.spec.ts

16 added files, 54 modified, 1 deleted.

Why I'm asking rather than just closing

Two things in there look potentially valuable independent of the parked background direction:

1. The legibility machinery — glass-tokens.ts plus the WCAG-AA-floored tint derivation. ADR 0002 records that the earlier attempt failed exactly one AA check (muted grey on the most translucent tier), so this encodes a real lesson. Worth salvaging if any surface still layers text on translucency — the session header's backdrop-blur comes to mind, given fix(session): header backdrop-blur off while the context-tree band is open #97 just had to switch it off over the context-tree band.
2. brain-perf-governor + the e2e perf trace — frame-budget governance for the graph. The organic context tree that did land has the same cost profile, so a perf governor may be wanted regardless of where the graph renders.

If neither is worth extracting, this is a clean close — the direction is parked and the replacement is released. If either is, they'd be small standalone PRs rather than resurrecting 71 files.

For context on the state: it is still a draft, CONFLICTING/DIRTY, 72 commits behind mainline, and its checks are all CANCELLED from before #2 fixed the runner labels — so CI has never actually evaluated this branch.

[kateebonner]

Closes#56

Integration branch for the Living Chat redesign, built slice-by-slice from the sub-issue DAG (#59#60#61#62#63; design-of-record: ADR 0002, merged via #58).

Landed:

• Mount the Brain chat-wide and absorb the inline brain strip #59 — Brain mounted chat-wide (session + landing draft), inline brain strip absorbed (one render loop; TimelineRow.Brain retired; shimmer kept); live event feed extracted to pure, tested deriveBrainEvents.

• Derive two AA-floored, chat-theme-keyed Glass tint tiers with a contrast test #60 — two-tint AA-floored Glass token system (data-glass="standard|dense", one 18px constant blur, chat-theme-keyed per-mode tints derived against the per-theme reference frame — dark #fff676 / light #8f8000 — with a certified contrast test; @supports fallback).

• Float chat components on the two Glass tiers (standard prose/bubbles/composer, dense code/diffs/plots/tools) #61 — everything floats: standard tier on bubbles/prose/composer, dense on code/diffs/run-plots/tool cards; opaque fills removed so the Brain lives in the gaps; muted ink dense-backed by measurement; browser-verified over the live Brain, both themes.

• Drive the Brain by session activity: adaptive heartbeat, sparse-seed rest, bounded in-place growth #62 — adaptive heartbeat (setActive + shared draw-gate: full tempo working, ~8fps rest), reduced-motion hard-pause (burst-then-still), sparse-seed boot (core alone; grows only from session activity), total-population cap with recency eviction, fixed camera (densify in place). Orphaned skeleton dataset removed at integration.

• Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63 — perf governor: ease ladder under sustained over-budget p95 (window-scoped, hysteresis dead-band, no oscillation) → terminal hard-pause to a static blurred field; blur/tint pinned untouched by test; DEV ?brainForceActive hook; headless perf-trace proxy (e2e/perf/brain-governor-trace.spec.ts). Issue left open for its hitl acceptance criterion — see the sign-off checklist below.

Final gate on this branch: ui 402 pass / 0 fail · app 464 pass / 0 fail · typecheck 23/23 · lint 0 errors.

Honest pre-gate evidence (headless, dev build — NOT the gate): forced full tempo + 11s continuous scroll of a 144-message thread → dark p95 25.4ms / light 25.2ms, mean ≈16.3ms, 11–15 frames >33ms. The tail misses the 16.7ms budget in that proxy environment (dev-mode Solid, headless vsync, sampler riding the measured loop) — which is exactly what the pre-agreed governor valve is for. The real call happens on the reference laptop (checklist below).

Known follow-up candidates (from #61's browser pass, non-blocking): composer placeholder hint and reasoning-part still composite on standard/bare over the Brain.

🤖 Generated with Claude Code

[kateebonner]

HITL sign-off checklist — reference-laptop 60fps gate (#63)

The one open acceptance criterion on this PR. Run on the reference laptop, battery-powered:

• Pin the machine: model, OS, battery %, VS Code + webview (Chromium) version — write them here.
• Launch the dev lane: opencode serve --port 4099 + bun --cwd packages/app dev --port 5990 with VITE_OPENCODE_SERVER_PORT=4099 (the force hook is DEV-gated by design).
• Open the Chat in the VS Code webview on a real thread of ≥40 cards with ?brainForceActive appended; verify in devtools: window.__amicoBrainStats() → active: true, motion: "full".
• Record a ≥10s continuous scroll (devtools Performance panel, or the rAF sampler in Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63's notes ~600 frames).
• Pass = p95 ≤ 16.7ms AND zero frames > 33ms. Repeat in both themes (flip the chat scheme mid-session — Brain/Glass re-key losslessly).
• Write the numbers here and sign off.

If the gate fails: remove ?brainForceActive and confirm the governor engages under load — watch __amicoBrainStats().motion step eased-1 → eased-2 → full-stop → hard-paused with the frosted look intact — then decide accept-with-valve vs iterate. (Headless dev-proxy already suggests the tail will need the valve; see PR body.)

🤖 Generated with Claude Code

[kateebonner]

Live design-tuning batch (Kate, 2026-07-25) — commit 69ebf7e

The design evolved significantly beyond the original slices through a live review session. Landed:

• Glass — now purely blur + tint + radius: dark derives to fully transparent (brightness 0.8), no border, no top sheen, no box-shadow. AA kept for light, waived+recorded for dark standard (the landing constellation is yellow-free).
• Composer — one uniform glass surface (two-tone fixed); full-bleed bottom dock, ~35% viewport height, edge-to-edge; wordmark + project/folder picker removed.
• Landing — starter-chip wall removed; discovery moved into the composer as a rotating typewriter placeholder (types the real starters letter-by-letter, blinking caret, Tab-to-accept, reduced-motion static). Constellation contrast bumped.
• Session parity — the 3D constellation is now the background in-session too (not the sparse 2D graph); launch + session are one surface.

Queued (next): live-thought node pulsing over the constellation (data-true layer), card width containment, and folding the effort/variant selector into the model dropdown.

Gates: ui 438 / app 473 / typecheck 23/23 / lint 0 errors.

🤖 Generated with Claude Code

[kateebonner]

Overnight batch 2 — 05fa001d0

All three queued items from the live-tuning session, plus the new ask (session components on the launch-composer feel):

1. Live thought over the constellation ✨

The "+ live thought" half of constellation + live thought is in. As amico works, each real touch flares its constellation node in the thought color (#fff676 dark / #8f8000 light — the brand law holds per scheme): fast rise, ~1.6s ease-out decay, an expanding ring, glints down the node's incident edges, and the touched label ("session.tsx", "amico-vault") named beside the flare.

• Labels hash deterministically into their category lobe (file work lights the code lobe, skills the skills lobe…), then walk to a visible node — a flare you can't see isn't thought.
• History replays silently on mount; re-touches refresh their flare instead of stacking; capped at 24 so a torrent stays a shimmer.
• Reduced motion: the tableau holds one statically lit node — where amico is now.
• Dev preview without a real turn: window.__amicoBrainTouch({label:"x.ts", type:"resource"}) in the console (the ?-param pattern doesn't survive the router redirect, so this rides the __amicoBrainStats hook pattern instead).
• +7 tests: flare paint, light-scheme ink, silent replay, refresh-not-stack, decay-to-rest, reduced-motion tableau, and byte-determinism under touches. The latent-web color law tests still pin zero yellow with zero touches.

2. Thinking effort lives in the model menu

The standalone Default/None/High Select is gone (both composer paths). A ModelVariantRow renders inside the paid popover and the unpaid dialog (the dev/free path — it would otherwise have lost the control entirely): chips on --accent-fill-soft/--accent-edge, picking one keeps the menu open. The collapsed trigger now reads "DeepSeek V4 Fl… · High" so effort stays visible at rest.

3. Session glass sweep — the launch-composer feel everywhere

Glass = blur + tint + radius, nothing else, now enforced in-session too:

• dock-prompt (permission/question card): 1px ring removed, footer border-top removed (one surface, same law as the composer footer), radius 10px → --radius-lg.
• In-session composer: focus-within ring removed (the caret carries focus; drag keeps its dashed dropzone).
• Comment chips: borderless dense tint.
• Width containment: tool collapsibles ("Explored · 2 searches"), the diffs group, reasoning, and error cards all share the prose measure — min(88%, 72ch) — so the thread reads as floating cards over the Brain, never full-width bands.

Verified live in both themes on real sessions (silent replay ✓, live flare mid-session ✓, permission card ✓).

Gate: ui 445 / app 473 / typecheck / lint 0 errors.

[kateebonner]

Parked 2026-07-26 — converting to draft. Aaron's call: not pursuing this UI direction right now. Everything through 83d070ae4 is on the branch (living-chat glass redesign, thought-tracking camera, bottom-stacked edge-justified flow, effort-in-model-menu, component audit + width law). The branch is the complete archive of the direction if we return to it; the component library artifact documents every surface.

[jack-champagne]

@kateebonner — question before this ages further: is anything here still worth keeping?

Asking because ADR 0002, which this implements, is marked superseded on mainline:

Status: Superseded by ADR 0003 (direction parked 2026-07-26; the Brain moved to the context-tree top panel instead of becoming the chat background)

and ADR 0003 (Accepted, 2026-07-27, on Aaron's 2026-07-26 feedback) is explicit:

Supersedes: ADR 0002 (Brain as chat background — direction parked)
… Aaron's feedback reframed what the graph is for: not ambience, but orientation … The graph moves out of the timeline into a pinned top panel … the ambient background direction is parked.

The ADR-0003 replacement has fully landed

All of this is on local/amicode and shipped in v1.17.3-amicode.11:

• c8468a416 context tree goes organic — Obsidian-like force-settled graph (Context tree: organic Obsidian-like graph layout (stacked on #68) #70)
• ecb05faaa / 7610867bc side panel owns its width; vault fills it by default (Side panel sizing + vault-by-default + store sync; pending-chip hints #71)
• 5a1b6d219 context tree never surfaces user prompt text (fix(amicode): context tree never surfaces user prompt text #83)
• e64255834 context tree marks proprietary vault context locked (feat(amicode): context tree marks proprietary vault context locked #84/fix(amicode): land #84 on mainline (locked vault context) — stacked-merge miss #86)
• d5f2d60b8 header backdrop-blur off while the context-tree band is open (fix(session): header backdrop-blur off while the context-tree band is open #97)

So the top-panel + first-class vault surface direction is done and released.

What's still unique to this branch

brain-engine.ts and context-tree-engine.ts are already on main — the graph engine itself made it across. What has no counterpart is the presentation layer for the parked direction:

only in #64
glass.css (778 lines), glass-tokens.ts, script/generate-glass.ts
brain-atmosphere.tsx / .css
brain-constellation.ts
brain-perf-governor.test.ts, brain-events.ts, e2e/perf/brain-governor-trace.spec.ts

16 added files, 54 modified, 1 deleted.

Why I'm asking rather than just closing

Two things in there look potentially valuable independent of the parked background direction:

1. The legibility machinery — glass-tokens.ts plus the WCAG-AA-floored tint derivation. ADR 0002 records that the earlier attempt failed exactly one AA check (muted grey on the most translucent tier), so this encodes a real lesson. Worth salvaging if any surface still layers text on translucency — the session header's backdrop-blur comes to mind, given fix(session): header backdrop-blur off while the context-tree band is open #97 just had to switch it off over the context-tree band.
2. brain-perf-governor + the e2e perf trace — frame-budget governance for the graph. The organic context tree that did land has the same cost profile, so a perf governor may be wanted regardless of where the graph renders.

If neither is worth extracting, this is a clean close — the direction is parked and the replacement is released. If either is, they'd be small standalone PRs rather than resurrecting 71 files.

For context on the state: it is still a draft, CONFLICTING/DIRTY, 72 commits behind mainline, and its checks are all CANCELLED from before #2 fixed the runner labels — so CI has never actually evaluated this branch.

[kateebonner]

Closes#56

Integration branch for the Living Chat redesign, built slice-by-slice from the sub-issue DAG (#59#60#61#62#63; design-of-record: ADR 0002, merged via #58).

Landed:

• Mount the Brain chat-wide and absorb the inline brain strip #59 — Brain mounted chat-wide (session + landing draft), inline brain strip absorbed (one render loop; TimelineRow.Brain retired; shimmer kept); live event feed extracted to pure, tested deriveBrainEvents.

• Derive two AA-floored, chat-theme-keyed Glass tint tiers with a contrast test #60 — two-tint AA-floored Glass token system (data-glass="standard|dense", one 18px constant blur, chat-theme-keyed per-mode tints derived against the per-theme reference frame — dark #fff676 / light #8f8000 — with a certified contrast test; @supports fallback).

• Float chat components on the two Glass tiers (standard prose/bubbles/composer, dense code/diffs/plots/tools) #61 — everything floats: standard tier on bubbles/prose/composer, dense on code/diffs/run-plots/tool cards; opaque fills removed so the Brain lives in the gaps; muted ink dense-backed by measurement; browser-verified over the live Brain, both themes.

• Drive the Brain by session activity: adaptive heartbeat, sparse-seed rest, bounded in-place growth #62 — adaptive heartbeat (setActive + shared draw-gate: full tempo working, ~8fps rest), reduced-motion hard-pause (burst-then-still), sparse-seed boot (core alone; grows only from session activity), total-population cap with recency eviction, fixed camera (densify in place). Orphaned skeleton dataset removed at integration.

• Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63 — perf governor: ease ladder under sustained over-budget p95 (window-scoped, hysteresis dead-band, no oscillation) → terminal hard-pause to a static blurred field; blur/tint pinned untouched by test; DEV ?brainForceActive hook; headless perf-trace proxy (e2e/perf/brain-governor-trace.spec.ts). Issue left open for its hitl acceptance criterion — see the sign-off checklist below.

Final gate on this branch: ui 402 pass / 0 fail · app 464 pass / 0 fail · typecheck 23/23 · lint 0 errors.

Honest pre-gate evidence (headless, dev build — NOT the gate): forced full tempo + 11s continuous scroll of a 144-message thread → dark p95 25.4ms / light 25.2ms, mean ≈16.3ms, 11–15 frames >33ms. The tail misses the 16.7ms budget in that proxy environment (dev-mode Solid, headless vsync, sampler riding the measured loop) — which is exactly what the pre-agreed governor valve is for. The real call happens on the reference laptop (checklist below).

Known follow-up candidates (from #61's browser pass, non-blocking): composer placeholder hint and reasoning-part still composite on standard/bare over the Brain.

🤖 Generated with Claude Code

[kateebonner]

HITL sign-off checklist — reference-laptop 60fps gate (#63)

The one open acceptance criterion on this PR. Run on the reference laptop, battery-powered:

• Pin the machine: model, OS, battery %, VS Code + webview (Chromium) version — write them here.
• Launch the dev lane: opencode serve --port 4099 + bun --cwd packages/app dev --port 5990 with VITE_OPENCODE_SERVER_PORT=4099 (the force hook is DEV-gated by design).
• Open the Chat in the VS Code webview on a real thread of ≥40 cards with ?brainForceActive appended; verify in devtools: window.__amicoBrainStats() → active: true, motion: "full".
• Record a ≥10s continuous scroll (devtools Performance panel, or the rAF sampler in Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63's notes ~600 frames).
• Pass = p95 ≤ 16.7ms AND zero frames > 33ms. Repeat in both themes (flip the chat scheme mid-session — Brain/Glass re-key losslessly).
• Write the numbers here and sign off.

If the gate fails: remove ?brainForceActive and confirm the governor engages under load — watch __amicoBrainStats().motion step eased-1 → eased-2 → full-stop → hard-paused with the frosted look intact — then decide accept-with-valve vs iterate. (Headless dev-proxy already suggests the tail will need the valve; see PR body.)

🤖 Generated with Claude Code

[kateebonner]

Live design-tuning batch (Kate, 2026-07-25) — commit 69ebf7e

The design evolved significantly beyond the original slices through a live review session. Landed:

• Glass — now purely blur + tint + radius: dark derives to fully transparent (brightness 0.8), no border, no top sheen, no box-shadow. AA kept for light, waived+recorded for dark standard (the landing constellation is yellow-free).
• Composer — one uniform glass surface (two-tone fixed); full-bleed bottom dock, ~35% viewport height, edge-to-edge; wordmark + project/folder picker removed.
• Landing — starter-chip wall removed; discovery moved into the composer as a rotating typewriter placeholder (types the real starters letter-by-letter, blinking caret, Tab-to-accept, reduced-motion static). Constellation contrast bumped.
• Session parity — the 3D constellation is now the background in-session too (not the sparse 2D graph); launch + session are one surface.

Queued (next): live-thought node pulsing over the constellation (data-true layer), card width containment, and folding the effort/variant selector into the model dropdown.

Gates: ui 438 / app 473 / typecheck 23/23 / lint 0 errors.

🤖 Generated with Claude Code

[kateebonner]

Overnight batch 2 — 05fa001d0

All three queued items from the live-tuning session, plus the new ask (session components on the launch-composer feel):

1. Live thought over the constellation ✨

The "+ live thought" half of constellation + live thought is in. As amico works, each real touch flares its constellation node in the thought color (#fff676 dark / #8f8000 light — the brand law holds per scheme): fast rise, ~1.6s ease-out decay, an expanding ring, glints down the node's incident edges, and the touched label ("session.tsx", "amico-vault") named beside the flare.

• Labels hash deterministically into their category lobe (file work lights the code lobe, skills the skills lobe…), then walk to a visible node — a flare you can't see isn't thought.
• History replays silently on mount; re-touches refresh their flare instead of stacking; capped at 24 so a torrent stays a shimmer.
• Reduced motion: the tableau holds one statically lit node — where amico is now.
• Dev preview without a real turn: window.__amicoBrainTouch({label:"x.ts", type:"resource"}) in the console (the ?-param pattern doesn't survive the router redirect, so this rides the __amicoBrainStats hook pattern instead).
• +7 tests: flare paint, light-scheme ink, silent replay, refresh-not-stack, decay-to-rest, reduced-motion tableau, and byte-determinism under touches. The latent-web color law tests still pin zero yellow with zero touches.

2. Thinking effort lives in the model menu

The standalone Default/None/High Select is gone (both composer paths). A ModelVariantRow renders inside the paid popover and the unpaid dialog (the dev/free path — it would otherwise have lost the control entirely): chips on --accent-fill-soft/--accent-edge, picking one keeps the menu open. The collapsed trigger now reads "DeepSeek V4 Fl… · High" so effort stays visible at rest.

3. Session glass sweep — the launch-composer feel everywhere

Glass = blur + tint + radius, nothing else, now enforced in-session too:

• dock-prompt (permission/question card): 1px ring removed, footer border-top removed (one surface, same law as the composer footer), radius 10px → --radius-lg.
• In-session composer: focus-within ring removed (the caret carries focus; drag keeps its dashed dropzone).
• Comment chips: borderless dense tint.
• Width containment: tool collapsibles ("Explored · 2 searches"), the diffs group, reasoning, and error cards all share the prose measure — min(88%, 72ch) — so the thread reads as floating cards over the Brain, never full-width bands.

Verified live in both themes on real sessions (silent replay ✓, live flare mid-session ✓, permission card ✓).

Gate: ui 445 / app 473 / typecheck / lint 0 errors.

[kateebonner]

Parked 2026-07-26 — converting to draft. Aaron's call: not pursuing this UI direction right now. Everything through 83d070ae4 is on the branch (living-chat glass redesign, thought-tracking camera, bottom-stacked edge-justified flow, effort-in-model-menu, component audit + width law). The branch is the complete archive of the direction if we return to it; the component library artifact documents every surface.

[jack-champagne]

@kateebonner — question before this ages further: is anything here still worth keeping?

Asking because ADR 0002, which this implements, is marked superseded on mainline:

Status: Superseded by ADR 0003 (direction parked 2026-07-26; the Brain moved to the context-tree top panel instead of becoming the chat background)

and ADR 0003 (Accepted, 2026-07-27, on Aaron's 2026-07-26 feedback) is explicit:

Supersedes: ADR 0002 (Brain as chat background — direction parked)
… Aaron's feedback reframed what the graph is for: not ambience, but orientation … The graph moves out of the timeline into a pinned top panel … the ambient background direction is parked.

The ADR-0003 replacement has fully landed

All of this is on local/amicode and shipped in v1.17.3-amicode.11:

• c8468a416 context tree goes organic — Obsidian-like force-settled graph (Context tree: organic Obsidian-like graph layout (stacked on #68) #70)
• ecb05faaa / 7610867bc side panel owns its width; vault fills it by default (Side panel sizing + vault-by-default + store sync; pending-chip hints #71)
• 5a1b6d219 context tree never surfaces user prompt text (fix(amicode): context tree never surfaces user prompt text #83)
• e64255834 context tree marks proprietary vault context locked (feat(amicode): context tree marks proprietary vault context locked #84/fix(amicode): land #84 on mainline (locked vault context) — stacked-merge miss #86)
• d5f2d60b8 header backdrop-blur off while the context-tree band is open (fix(session): header backdrop-blur off while the context-tree band is open #97)

So the top-panel + first-class vault surface direction is done and released.

What's still unique to this branch

brain-engine.ts and context-tree-engine.ts are already on main — the graph engine itself made it across. What has no counterpart is the presentation layer for the parked direction:

only in #64
glass.css (778 lines), glass-tokens.ts, script/generate-glass.ts
brain-atmosphere.tsx / .css
brain-constellation.ts
brain-perf-governor.test.ts, brain-events.ts, e2e/perf/brain-governor-trace.spec.ts

16 added files, 54 modified, 1 deleted.

Why I'm asking rather than just closing

Two things in there look potentially valuable independent of the parked background direction:

1. The legibility machinery — glass-tokens.ts plus the WCAG-AA-floored tint derivation. ADR 0002 records that the earlier attempt failed exactly one AA check (muted grey on the most translucent tier), so this encodes a real lesson. Worth salvaging if any surface still layers text on translucency — the session header's backdrop-blur comes to mind, given fix(session): header backdrop-blur off while the context-tree band is open #97 just had to switch it off over the context-tree band.
2. brain-perf-governor + the e2e perf trace — frame-budget governance for the graph. The organic context tree that did land has the same cost profile, so a perf governor may be wanted regardless of where the graph renders.

If neither is worth extracting, this is a clean close — the direction is parked and the replacement is released. If either is, they'd be small standalone PRs rather than resurrecting 71 files.

For context on the state: it is still a draft, CONFLICTING/DIRTY, 72 commits behind mainline, and its checks are all CANCELLED from before #2 fixed the runner labels — so CI has never actually evaluated this branch.

[kateebonner]

Closes#56

Integration branch for the Living Chat redesign, built slice-by-slice from the sub-issue DAG (#59#60#61#62#63; design-of-record: ADR 0002, merged via #58).

Landed:

• Mount the Brain chat-wide and absorb the inline brain strip #59 — Brain mounted chat-wide (session + landing draft), inline brain strip absorbed (one render loop; TimelineRow.Brain retired; shimmer kept); live event feed extracted to pure, tested deriveBrainEvents.

• Derive two AA-floored, chat-theme-keyed Glass tint tiers with a contrast test #60 — two-tint AA-floored Glass token system (data-glass="standard|dense", one 18px constant blur, chat-theme-keyed per-mode tints derived against the per-theme reference frame — dark #fff676 / light #8f8000 — with a certified contrast test; @supports fallback).

• Float chat components on the two Glass tiers (standard prose/bubbles/composer, dense code/diffs/plots/tools) #61 — everything floats: standard tier on bubbles/prose/composer, dense on code/diffs/run-plots/tool cards; opaque fills removed so the Brain lives in the gaps; muted ink dense-backed by measurement; browser-verified over the live Brain, both themes.

• Drive the Brain by session activity: adaptive heartbeat, sparse-seed rest, bounded in-place growth #62 — adaptive heartbeat (setActive + shared draw-gate: full tempo working, ~8fps rest), reduced-motion hard-pause (burst-then-still), sparse-seed boot (core alone; grows only from session activity), total-population cap with recency eviction, fixed camera (densify in place). Orphaned skeleton dataset removed at integration.

• Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63 — perf governor: ease ladder under sustained over-budget p95 (window-scoped, hysteresis dead-band, no oscillation) → terminal hard-pause to a static blurred field; blur/tint pinned untouched by test; DEV ?brainForceActive hook; headless perf-trace proxy (e2e/perf/brain-governor-trace.spec.ts). Issue left open for its hitl acceptance criterion — see the sign-off checklist below.

Final gate on this branch: ui 402 pass / 0 fail · app 464 pass / 0 fail · typecheck 23/23 · lint 0 errors.

Honest pre-gate evidence (headless, dev build — NOT the gate): forced full tempo + 11s continuous scroll of a 144-message thread → dark p95 25.4ms / light 25.2ms, mean ≈16.3ms, 11–15 frames >33ms. The tail misses the 16.7ms budget in that proxy environment (dev-mode Solid, headless vsync, sampler riding the measured loop) — which is exactly what the pre-agreed governor valve is for. The real call happens on the reference laptop (checklist below).

Known follow-up candidates (from #61's browser pass, non-blocking): composer placeholder hint and reasoning-part still composite on standard/bare over the Brain.

🤖 Generated with Claude Code

[kateebonner]

HITL sign-off checklist — reference-laptop 60fps gate (#63)

The one open acceptance criterion on this PR. Run on the reference laptop, battery-powered:

• Pin the machine: model, OS, battery %, VS Code + webview (Chromium) version — write them here.
• Launch the dev lane: opencode serve --port 4099 + bun --cwd packages/app dev --port 5990 with VITE_OPENCODE_SERVER_PORT=4099 (the force hook is DEV-gated by design).
• Open the Chat in the VS Code webview on a real thread of ≥40 cards with ?brainForceActive appended; verify in devtools: window.__amicoBrainStats() → active: true, motion: "full".
• Record a ≥10s continuous scroll (devtools Performance panel, or the rAF sampler in Add the perf governor (ease motion, never blur) and gate 60fps on the reference laptop #63's notes ~600 frames).
• Pass = p95 ≤ 16.7ms AND zero frames > 33ms. Repeat in both themes (flip the chat scheme mid-session — Brain/Glass re-key losslessly).
• Write the numbers here and sign off.

If the gate fails: remove ?brainForceActive and confirm the governor engages under load — watch __amicoBrainStats().motion step eased-1 → eased-2 → full-stop → hard-paused with the frosted look intact — then decide accept-with-valve vs iterate. (Headless dev-proxy already suggests the tail will need the valve; see PR body.)

🤖 Generated with Claude Code

[kateebonner]

Live design-tuning batch (Kate, 2026-07-25) — commit 69ebf7e

The design evolved significantly beyond the original slices through a live review session. Landed:

• Glass — now purely blur + tint + radius: dark derives to fully transparent (brightness 0.8), no border, no top sheen, no box-shadow. AA kept for light, waived+recorded for dark standard (the landing constellation is yellow-free).
• Composer — one uniform glass surface (two-tone fixed); full-bleed bottom dock, ~35% viewport height, edge-to-edge; wordmark + project/folder picker removed.
• Landing — starter-chip wall removed; discovery moved into the composer as a rotating typewriter placeholder (types the real starters letter-by-letter, blinking caret, Tab-to-accept, reduced-motion static). Constellation contrast bumped.
• Session parity — the 3D constellation is now the background in-session too (not the sparse 2D graph); launch + session are one surface.

Queued (next): live-thought node pulsing over the constellation (data-true layer), card width containment, and folding the effort/variant selector into the model dropdown.

Gates: ui 438 / app 473 / typecheck 23/23 / lint 0 errors.

🤖 Generated with Claude Code

[kateebonner]

Overnight batch 2 — 05fa001d0

All three queued items from the live-tuning session, plus the new ask (session components on the launch-composer feel):

1. Live thought over the constellation ✨

The "+ live thought" half of constellation + live thought is in. As amico works, each real touch flares its constellation node in the thought color (#fff676 dark / #8f8000 light — the brand law holds per scheme): fast rise, ~1.6s ease-out decay, an expanding ring, glints down the node's incident edges, and the touched label ("session.tsx", "amico-vault") named beside the flare.

• Labels hash deterministically into their category lobe (file work lights the code lobe, skills the skills lobe…), then walk to a visible node — a flare you can't see isn't thought.
• History replays silently on mount; re-touches refresh their flare instead of stacking; capped at 24 so a torrent stays a shimmer.
• Reduced motion: the tableau holds one statically lit node — where amico is now.
• Dev preview without a real turn: window.__amicoBrainTouch({label:"x.ts", type:"resource"}) in the console (the ?-param pattern doesn't survive the router redirect, so this rides the __amicoBrainStats hook pattern instead).
• +7 tests: flare paint, light-scheme ink, silent replay, refresh-not-stack, decay-to-rest, reduced-motion tableau, and byte-determinism under touches. The latent-web color law tests still pin zero yellow with zero touches.

2. Thinking effort lives in the model menu

The standalone Default/None/High Select is gone (both composer paths). A ModelVariantRow renders inside the paid popover and the unpaid dialog (the dev/free path — it would otherwise have lost the control entirely): chips on --accent-fill-soft/--accent-edge, picking one keeps the menu open. The collapsed trigger now reads "DeepSeek V4 Fl… · High" so effort stays visible at rest.

3. Session glass sweep — the launch-composer feel everywhere

Glass = blur + tint + radius, nothing else, now enforced in-session too:

• dock-prompt (permission/question card): 1px ring removed, footer border-top removed (one surface, same law as the composer footer), radius 10px → --radius-lg.
• In-session composer: focus-within ring removed (the caret carries focus; drag keeps its dashed dropzone).
• Comment chips: borderless dense tint.
• Width containment: tool collapsibles ("Explored · 2 searches"), the diffs group, reasoning, and error cards all share the prose measure — min(88%, 72ch) — so the thread reads as floating cards over the Brain, never full-width bands.

Verified live in both themes on real sessions (silent replay ✓, live flare mid-session ✓, permission card ✓).

Gate: ui 445 / app 473 / typecheck / lint 0 errors.

[kateebonner]

Parked 2026-07-26 — converting to draft. Aaron's call: not pursuing this UI direction right now. Everything through 83d070ae4 is on the branch (living-chat glass redesign, thought-tracking camera, bottom-stacked edge-justified flow, effort-in-model-menu, component audit + width law). The branch is the complete archive of the direction if we return to it; the component library artifact documents every surface.

[jack-champagne]

@kateebonner — question before this ages further: is anything here still worth keeping?

Asking because ADR 0002, which this implements, is marked superseded on mainline:

Status: Superseded by ADR 0003 (direction parked 2026-07-26; the Brain moved to the context-tree top panel instead of becoming the chat background)

and ADR 0003 (Accepted, 2026-07-27, on Aaron's 2026-07-26 feedback) is explicit:

Supersedes: ADR 0002 (Brain as chat background — direction parked)
… Aaron's feedback reframed what the graph is for: not ambience, but orientation … The graph moves out of the timeline into a pinned top panel … the ambient background direction is parked.

The ADR-0003 replacement has fully landed

All of this is on local/amicode and shipped in v1.17.3-amicode.11:

• c8468a416 context tree goes organic — Obsidian-like force-settled graph (Context tree: organic Obsidian-like graph layout (stacked on #68) #70)
• ecb05faaa / 7610867bc side panel owns its width; vault fills it by default (Side panel sizing + vault-by-default + store sync; pending-chip hints #71)
• 5a1b6d219 context tree never surfaces user prompt text (fix(amicode): context tree never surfaces user prompt text #83)
• e64255834 context tree marks proprietary vault context locked (feat(amicode): context tree marks proprietary vault context locked #84/fix(amicode): land #84 on mainline (locked vault context) — stacked-merge miss #86)
• d5f2d60b8 header backdrop-blur off while the context-tree band is open (fix(session): header backdrop-blur off while the context-tree band is open #97)

So the top-panel + first-class vault surface direction is done and released.

What's still unique to this branch

brain-engine.ts and context-tree-engine.ts are already on main — the graph engine itself made it across. What has no counterpart is the presentation layer for the parked direction:

only in #64
glass.css (778 lines), glass-tokens.ts, script/generate-glass.ts
brain-atmosphere.tsx / .css
brain-constellation.ts
brain-perf-governor.test.ts, brain-events.ts, e2e/perf/brain-governor-trace.spec.ts

16 added files, 54 modified, 1 deleted.

Why I'm asking rather than just closing

Two things in there look potentially valuable independent of the parked background direction:

1. The legibility machinery — glass-tokens.ts plus the WCAG-AA-floored tint derivation. ADR 0002 records that the earlier attempt failed exactly one AA check (muted grey on the most translucent tier), so this encodes a real lesson. Worth salvaging if any surface still layers text on translucency — the session header's backdrop-blur comes to mind, given fix(session): header backdrop-blur off while the context-tree band is open #97 just had to switch it off over the context-tree band.
2. brain-perf-governor + the e2e perf trace — frame-budget governance for the graph. The organic context tree that did land has the same cost profile, so a perf governor may be wanted regardless of where the graph renders.

If neither is worth extracting, this is a clean close — the direction is parked and the replacement is released. If either is, they'd be small standalone PRs rather than resurrecting 71 files.

For context on the state: it is still a draft, CONFLICTING/DIRTY, 72 commits behind mainline, and its checks are all CANCELLED from before #2 fixed the runner labels — so CI has never actually evaluated this branch.