Teknium review changes on the tiered policy:
1. threshold_pct default 10 -> 5 (listing budget = min(5% of context,
listing_max_tokens)); unknown-context fallback 20K -> 10K.
2. Tier 2 no longer leaves the model blind: when even names-only doesn't
fit, the bridge description now carries a one-line-per-server summary
('cloudflare (3320 tools)') plus an instruction to search FIRST rather
than substitute a generic tool or claim the capability is missing —
the measured tier-2 failure mode (core-tool substitution) at zero
meaningful token cost (~50 tokens/server).
3. Listing degradation is now PER SERVER, largest first: one oversized
server (Cloudflare) collapses to its summary line while small
co-attached servers (Linear) keep their full per-tool listings
('mixed' form). Previously global: attaching Cloudflare next to
Linear silently cost Linear its listing. Greedy fit is deterministic
(size then label) so the rendered block stays byte-stable per catalog
— prompt-prefix cache safe.
E2E on real captures (defaults, 200K ctx): linear alone -> tier 1 full;
unreal alone -> tier 2 groups (5% budget) / tier 1 names at 1M;
cloudflare alone -> tier 2 groups; linear+cloudflare -> tier 1 MIXED
(linear fully listed, cloudflare summarized). 48/48 tests.
Cloudflare's flat API MCP ships 61 property keys that violate Anthropic's
^[a-zA-Z0-9_.-]{1,64}$ pattern (query-filter params like 'issue_class~neq'
and 'meta.<field>[<operator>]'). One bad key anywhere in the tools array
400s the ENTIRE request — measured live: Anthropic, Bedrock, Google Vertex,
and Azure all rejected an eager 3,320-tool Cloudflare request at validation,
before token limits even applied.
- schema_sanitizer: rename non-conforming property keys deterministically
(bad chars -> '_', 64-char truncation, collision dedup with numeric
suffixes), nested schemas included; required[] remapped alongside
- unrename_tool_args(): reverse map applied in coerce_tool_args at dispatch,
so the MCP server receives the original wire names; recurses into object
values and array items
- deterministic on both sides: the rename map is recomputed from the
registry's original schema at dispatch time, no state carried
E2E on the real capture: 61 -> 0 violations across 3,320 tools; round-trip
verified on get_accounts_intel_attacksurfacereport_issues (5 '~neq' keys).
Tier 0: no MCP/plugin tools -> everything eager (pass-through).
Tier 1: deferred tools whose catalog listing fits min(threshold_pct%
of context, listing_max_tokens) -> bridge + skills-style
listing, degrading to names-only over budget.
Tier 2: listing over budget even names-only (Cloudflare's flat API
surface: 3,320 tools, names alone ~32K tokens) -> bare bridge,
discovery through tool_search only.
The old activation threshold (defer only when schemas > threshold_pct
of context) let mid-size catalogs ride eager and pay full schema cost;
with servers like Cloudflare (~597K tokens of schema, would not even
fit a 200K window) the binary gate is the wrong shape. Activation is
now driven purely by deferrable-tool presence; threshold_pct is
repurposed as the listing budget's context-relative leg.
- AssemblyResult gains tier + listing_form for observability
- listing_max_tokens default 4000 -> 20000 (cap 60000) so an 830-tool
catalog keeps a names-only listing while Cloudflare-scale drops to
bare bridge
- E2E verified against real captures: Linear 24 tools -> tier 1 full,
Epic UE 5.8 830 tools -> tier 1 names-only, Cloudflare 3,320 tools
-> tier 2 (both 200K and 1M context)
Bridge vs listing only (Opus 4.8, 830 real UE schemas, 3 reps/cell).
Excluding one both-modes mock artifact: listing 24/24 vs bridge 20/24,
searches/task 0.2 vs 4.0. Bridge failures: core-tool substitution at
frontier tier (ran the host test suite via terminal instead of
discovering RunTests, 2/3 reps), up to 8 searches to prove a negative,
and search-vocabulary misses on paraphrase. Listing asserts absence in
zero searches and answers a 5-way capability survey in 1 API call.
Scenarios target real confusion clusters in Epic's UE 5.8 catalog
(StaticMesh vs SkeletalMesh set_material, three tag systems, CurveTable
vs DataTable rows, Niagara Component vs System variables, four capture
variants, zero-keyword phrasing). Mocks return realistic editor errors
on wrong-type calls; scoring is strict (clean solve = correct tool with
zero distractor calls; first-call accuracy tracked separately).
Key result: first-call selection is unreliable in EVERY mode — eager
with all 199K of schemas in context managed 2/10 — but clean solves stay
75-95% because agents probe (get_components, get_material_slots) before
committing. The probe loop works through the 3-tool bridge at 1/4 the
cost of eager ($1.60-1.69 vs $6.49/task, Opus 4.8). On Haiku the
listing beats bare bridge 18/20 vs 15/20 (core-tool substitution again).
Zero distractor invocations across all 50 Opus runs.
Replays the actual tool schemas captured from Epic's UE 5.8
ModelContextProtocol + AllToolsets plugins (830 tools / 52 toolsets) as
live registry tools with mocked editor responses, then benchmarks
eager vs bare-bridge vs bridge+listing at two scales (62-tool editor
subset, full 830) on Claude Opus 4.8 (1M ctx; eager at 830 does not fit
any 200K model — first call requests ~266K tokens).
Headline (full 830, mean per task, rescored): eager 8/8 at 810,578
input tokens ($4.05); bare bridge 16/16 at 160,844 ($0.80); listing
16/16 at 257,264 ($1.29). Frontier model erases the accuracy gap in
every mode; cost is the differentiator. At 62 tools eager wins on cost
— consistent with the auto-threshold design.
Also parameterizes livetest harness model + listing_max_tokens via
env/args (TS_UE_MODEL, TS_UE_SCALE, TS_UE_MODES, TS_UE_LISTING_MAX).
Deferred MCP/plugin tools become invisible once the tool_search bridge
activates — live benchmarking (48 runs, Claude Haiku 4.5) showed models
substituting visible core tools (terminal/web_search/browser) for deferred
capabilities or declaring them nonexistent instead of searching: 16/24
task success vs 24/24 with eager loading.
Skills never had this failure mode because every skill keeps a ~21-token
name+description listing line in the system prompt. This ports that exact
pattern to the tool bridge: when tool_search activates, a grouped manifest
of every deferred tool (name + first sentence of description, clipped to
60 chars, grouped per MCP server / toolset) is embedded in the tool_search
bridge description.
- tools/tool_search.py: build_catalog_listing() with deterministic
ordering (byte-stable across assemblies -> prompt prefix stays
cacheable); token-budget fallbacks full -> names-only -> legacy bare
count; bridge_tool_schemas(listing=...) embeds it and instructs the
model to skip tool_search when the exact name is visible (one fewer
round-trip per use)
- config: tools.tool_search.listing auto|on|off (default auto),
listing_max_tokens (default 4000, clamped 200..20000); legacy bool
shapes keep working
- tests: 8 new tests (config parsing/clamps, short-desc clipping,
deterministic rendering, budget fallbacks, bridge embedding, assembly
on/off paths); full file green (47 passed)
- docs: tool-search.md config table + rationale
- scripts/tool_search_livetest2.py: benchmark harness v2 with real
per-call token accounting (normalize_usage spy) and a third 'listing'
mode for A/B/C comparison
The sidebar strip and the Channels page could contradict each other on
the same page load — "Gateway running" next to "The gateway is not
running." /api/status and /api/messaging/platforms each open-coded their
own liveness ladder: status probed GATEWAY_HEALTH_URL and scoped its
PID/state reads to the requested profile, messaging did neither and used
the uncached raw PID probe.
Three deployments hit the split: a cross-container gateway (no local PID,
only the health probe can see it), a profile-scoped dashboard (messaging
borrowed a DIFFERENT profile's runtime state, reporting a false
"connected" that hides a real outage — #71211), and a launch-service
managed gateway with no PID file.
Adds resolve_gateway_liveness() in gateway/status.py as the single ladder
(cached PID -> HTTP health probe -> runtime-status PID with
expected_home) and routes both endpoints, /api/messaging/platforms/{id}/test,
and the kanban dispatcher-presence probe through it. Probe callables are
injectable so the existing monkeypatch seams keep working, and
GatewayLiveness.probe_error distinguishes "down" from "couldn't tell" so
the kanban warning keeps failing OPEN instead of crying wolf.
Closes#71211.
Drives the same synthetic multi-tab streaming workload as `multitab`
(publishSessionState per session per flush, no backend, no credits) but
reports render attribution instead of frame pacing — sidebar_renders,
wasted_renders, and wasted_notifies.
Answers 'does the sidebar re-render while an agent is typing' directly.
Adds two dev-only counters that attribute re-renders and store
notifications during an interaction, so a perf claim can be answered with
a number instead of a hunch:
window.__RENDER_COUNTS__ what re-rendered, and why (props/state/parent)
window.__ATOM_CHURN__ which store published it, and whether it mattered
The `wasted` column in each is the fix list — components that re-rendered
with no changed input, and stores that published a value equal to the
last one. Both are inert until start(), so idle cost is one branch per
commit and per notify.
React 19.2 removed injectProfilingHooks from react-dom, so the mark*
profiling family is unavailable and onCommitFiberRoot is the only channel
left. <Profiler> can't answer the question either: React invokes onRender
for every Profiler in a committed tree including subtrees that bailed
out, and a bailed-out subtree still reports nonzero actualDuration. This
uses bippy's didFiberRender instead. bippy over react-scan because
react-scan/lite is a thin wrapper over it while the package pulls ~217
transitive deps and floats two on latest.
main.tsx imports the entry statically above react-dom because react-dom
captures the devtools hook at module init — a late install reports
renderers=0 and observes zero commits. Production exclusion is handled by
a build-time alias to a no-op module rather than tree-shaking, since a
static side-effect import can't be eliminated.
While the workspace pane shows a full page (Artifacts, Skills, Messaging, a
plugin route), a sidebar click on the ACTIVE session did nothing: onResumeSession
took focusOpenSession's `true` for the main-session branch as "already on
screen" and skipped the navigate, but fronting the workspace tab doesn't put the
chat back — the page is still routed. The user had to click some other session
and then the active one to get back.
focusOpenSession now reports WHICH surface it fronted ('main' | 'tile' | null),
and focusedSessionNeedsRoute decides: a tile never needs a route (its pane
renders the chat regardless), a main hit does while a page covers the workspace.
`/work` typed into a fresh Cmd+T tab loaded the skill in that tab and
printed "⚡ loading skill: work" there, then fired the skill's kickoff
prompt as a user message into whatever conversation was on screen.
The dispatcher resolves its target once, through resolveTargetSessionId,
and every other consumer of that answer already honors it: the output
writer binds to the target's stored id, and the busy gate reads the
target's own state. The send did not — `submitPromptText(message)` passed
no target at all, so submit fell back to `activeSessionIdRef`, which
names the foreground chat. #71805 fixed the two sibling leaks in this
same function; this is the third and the one that actually moved the
user's prompt.
Forward the resolved pair instead. Every target the dispatcher serves —
a tile, a background queue drain, a session this very call created —
was hitting the same fallback, so the fix covers the class rather than
the tab case that surfaced it.
Technical mode rendered the raw payload two different ways — a bare
block for most rows, a native `<details>` for file edits, whose
browser-drawn marker matches nothing else in the app. Both are now one
collapsed chevron disclosure at a smaller type size, with even padding
against the row body.
The strip's rule is an inset shadow painted in the container's last pixel
row, and full-height tabs covered it — so each tab read as overhanging
the bar by 1px. Inactive tabs compensated with their own border, stacking
a second translucent line that darkened the seam.
Inactive tabs now stop 1px short and draw no bottom border, leaving the
container as the sole owner of one continuous rule; the active tab keeps
full height so it alone cuts through. Hover also darkens rather than
lightens, since lightening moved a hovered tab toward the active
surface's look.
The arc reduced to a few faint dots on session rows. Two causes: the ring
was outset by 2px into a scroller that clips horizontally, losing its
left and right runs; and its tail color defaults to the chrome
background, which is invisible against the sidebar, leaving only the
bright stop of each gradient pass.
An `arc-row` variant sits flush and ties the tail back to the ring
color. The ring's radius is now derived from its standoff (r_host + gap)
rather than inherited, which keeps any outset host concentric instead of
pinched.
Content links read as a small primary-tinted chip instead of an
underline, dropping the trailing external-link arrow. The tint is
currentColor-relative, so one class carries text and fill in the same
hue across every theme, and `box-decoration-break: clone` gives a
wrapped link a chip per line fragment.
An agent told to work in a fresh git worktree does exactly that — creates
it, cds in, and runs every later command there — but the session stayed
pinned to the checkout it started in. The desktop kept labelling the chat
with the primary branch while all the work landed somewhere else.
The desktop half already existed: session.info carrying a moved cwd runs
followActiveSessionCwd, which refreshes the project tree and scopes the
sidebar into the new project. The backend just never reported the move.
Reconcile the session's cwd against terminal_tool's per-session record at
the end of a turn, when the agent has stopped moving and its recorded cwd
is a stable answer. A plain cd stays what it always was — not a workspace
move — so the reconcile only fires when the recorded cwd sits in a
different git working tree than the session's workspace.
The sidebar "+" now stacks a tab instead of replacing the surface, so the
prior session stays mounted and several chat surfaces can be on the page at
once. Helpers that waited for the old transcript to disappear from the page
timed out, and `.first()` locators / bare `document.querySelector` calls
started resolving against the wrong session (CI's "resolved to 2 elements"
strict-mode violation).
Target the most recently mounted `[data-composer-target]` surface instead,
and assert the NEW surface is empty rather than waiting for the old text to
vanish.
CuaDriverBackend caches a long-lived cua-driver subprocess for the life of
the Hermes process, and stop() was never called from anywhere — the driver
outlived the session that spawned it. #69903 stopped the orphan from pegging
a core by disabling the cursor overlay, but left the process behind; this is
item 3 of #28152 ("Hermes does not keep the driver alive after tool
completion").
Register an atexit hook, mirroring browser_tool's
atexit.register(_emergency_cleanup_all_sessions). atexit only, no signal
handlers, for the prompt_toolkit reason documented there. reset_backend_for_tests
now reuses the same teardown instead of repeating it.
The sidebar "+" ran startFreshSessionDraft, so it replaced whatever was on
screen — ⌘N behavior. With a conversation already open (possibly mid-turn)
that discards it to make room for a blank draft, which is not what a create
affordance should do. The tab-strip "+" and ⌘T already stack a new tab.
Route the sidebar button through the same path once a session is loaded,
falling back to the fresh-draft path when the surface is empty and there is
no tab worth preserving. openNewSessionTile now takes an optional cwd so the
new tab stays anchored to the clicked project/worktree lane.
Fold the null fallback into chatSurfaceRoot so both callers share one
target resolution, hoist the var names next to it, and drop the
duplicated rationale from the status-stack comment.
The existing `stream` scenario measures streaming into an empty thread, which
is exactly the case where per-delta transcript work does not show up. This adds
`--historyTurns`, which mounts a settled transcript and lets it drain BEFORE
the recorders start, so the measurement window contains streaming work only —
and `stream-history`, a report-only preset that turns it on.
Report-only (tier: manual) rather than gated: how much history a host can mount
varies, so the absolute number is not comparable across machines. It is meant
for same-machine before/after runs.
Co-authored-by: Jakub Wolniewicz <frizikk@users.noreply.github.com>
Once the coalescing floor had already elapsed, the next delta was scheduled
with requestAnimationFrame. Chromium pauses rAF for a renderer it considers
hidden, and that is not something this code can verify: backgroundThrottling:
false and the process-level switches in electron/main.ts cover the blurred and
occluded cases, but not a minimized window, a fully off-screen one, or a
renderer the compositor has otherwise parked. In those states the callback is
accepted and never runs, so a finished answer sits in the queue until some
later focus or input event happens to wake a frame — the reply looks stalled,
then lands all at once on refocus.
Always use a timer. The coalescing cadence is unchanged (the floor above is
what enforces it), timers are clamped rather than suspended in background
renderers, and disable-background-timer-throttling already opts out of that
clamp. The teardown path loses its cancelAnimationFrame branch with it.
The regression test parks rAF the way an occluded renderer does and asserts
the delta still arrives. It has to send a delta, let it flush, then idle past
the floor before the delta under test, because the frame-gated branch was only
reachable on that second scheduling — a single-delta version of this test
passes on main and proves nothing.
Co-authored-by: NetRunner2037 <rerdi92@users.noreply.github.com>
The composer and the out-of-flow status stack published their measured
heights onto document.documentElement, but both components mount once per
chat surface. Session tiles render a full ChatView beside the workspace
pane, so N surfaces raced for one value: a background tab with a tall
status stack inflated the foreground thread's bottom clearance and pushed
its jump-to-bottom button into mid-screen. Whichever surface unmounted
last also cleared the var for everyone still showing.
Publish onto the surface's own root instead, and re-declare the clearance
calc there — `:root` substitutes the root measurements once, so scoping
only the inputs would leave every thread reading the same value.
The workspace pane read the global $messages/$busy atoms — a mirror of
whichever session was active — while every ⌘T tile rendered from its own
$sessionStates slice. With two turns in flight, navigating away from a
still-streaming session left it painting into the surface now showing a
different conversation: the wrong transcript under the right route.
Point the primary view at the active session's own slice, keeping the
global atoms as the draft surface for a chat that has no runtime id yet.
Previews correct themselves on read, but a title already written to the DB
stays wrong. This regenerates those titles from what the user actually typed,
dry-run by default.
Two guards, both hit on a real store: a candidate that isn't title-shaped is
rejected rather than replacing a serviceable title with command output, and a
unique-title collision dedupes through the lineage the way the live auto-titler
does instead of leaving the leaked title in place.
`preview` is the head of the first user message and the title fallback on
every surface — sidebar rows, pickers, exports, the desktop's sessionTitle().
An untitled /skill session therefore read `[IMPORTANT: The user has invoked
the "work" skill, indicatin...` wherever it appeared.
A scaffolded row now selects a wide enough excerpt to reach the typed
instruction (head + tail spliced for a long body) and shapes it through
describe_skill_invocation(). Because previews are computed on read, existing
sessions are corrected without a migration.
The six copies of the preview subquery and four copies of its shaping collapse
into one expression and one helper along the way, and the /rewind picker gets
the same treatment.
generate_title() sent the first 500 characters of the user turn to the
auxiliary model. On a /skill invocation those characters are the skill's own
opening prose, so the session got named after the skill instead of the request
— /work sessions came back as "Isolated Git Worktree Setup".
Route the turn through describe_skill_invocation() first, so the titler sees
what the user typed. Also keep only the first line of the response: a model
that ignores "return ONLY the title" and answers the prompt would otherwise
have a shell transcript stored as the title, truncated mid-command.
A /skill invocation expands into a message that embeds the whole skill body.
Anything that summarizes a user turn from its raw content reads the skill's
prose as if the user had written it.
describe_skill_invocation() sits next to the existing extractor and reuses its
markers, returning `/work — fix the title leak` for an invocation with an
instruction and `/work` for a bare one. It also exports the SQL LIKE pattern
and excerpt-joint sentinel that listing queries need to recognize scaffolding
before a row reaches Python.
Teknium