A provider can accept a request, return response headers, and then never
send a byte of body data. The connection-phase request timeout was cleared
as soon as headers arrived, so nothing bounded that wait and the agent turn
hung indefinitely after a tool call completed: step-finish:tool-calls was
recorded and the next step-start never arrived, with the HTTP server still
responsive.
Extend the same configured timeout deadline to the wait for the response
body's first byte. The connection-phase timer covers the fetch up to
headers; once headers arrive, the remaining deadline is handed to a
first-byte guard that aborts the request if no data arrives. After the
first byte the guard becomes a passthrough, so idle gaps inside an already
streaming response (reasoning, buffering, slow token generation) are never
touched and remain opt-in via chunkTimeout.
This is a transport-level signal (bytes on the wire, before any content)
rather than the absence of normalized AI SDK events, so it cannot fire on
long prompt processing or reasoning the way the reverted stream watchdog
did. timeout: false still disables the bound entirely.
Adds a hermetic regression test that injects a simulated stalled socket
through the provider's own fetch option via the plugin config hook, so the
SDK, Kilo's fetch wrapper, SSE parsing, the processor and the agent loop
all stay production code. The stalled request is transient, so the test
asserts the turn recovers by retrying and completing instead of freezing.
The test goes red without the fix (no retry, frozen at step-finish) and
green with it.
Refs #8656
* fix(cli): enforce permissions on shell commands the parser fails to scan
* fix(cli): fail closed on error chunks without command names, move pwsh execution test to kilo file
* test(cli): cover TUI startup outside package
* fix(cli): use native preload path in TUI test
---------
Co-authored-by: Johnny Eric Amancio <johnnyeric@gmail.com>
* fix(cli): drain session ingest queue on shutdown and flush terminal batches promptly
* fix(cli): drain the session ingest queue on process shutdown
* fix(cli): pin drain bound expiry, add changeset, conform to naming rule
* fix(cli): keep kilo-sessions out of the CLI startup import graph
* fix(cli): never let the ingest drain task reject the shutdown sequence
* test(cli): pin drain-before-dispose ordering on the KiloCli shutdown path
* fix(cli): make the guarded ingest drain non-rejecting and correct the lazy-import rationale
* test(cli): cover the retryable-status drain path under shutdown
* test(cli): decouple cli-shutdown drain assertions from declaration order
* fix(cli): advertise the instance from enableRemote so /remote registers as a spawn target
Enabling the remote relay from the TUI `/remote` slash command connected the
socket and mirrored sessions, but never advertised the instance, so the CLI
never appeared as a spawn target in the mobile "Run on" picker. Only the
explicit `kilo remote` command called setInstanceAdvertisement.
The advertisement now runs on every successful enableRemote() entry, before the
already-connected and coalescing early returns. That ordering matters: bootstrap
auto-enable frequently connects first, so `/remote` usually hits
`if (remote) return` and an advertisement placed in the connection-setup body
would leave the defect unfixed in the common case. `ingestDisabled` returns
before the advertisement and stays unadvertised.
The ensure helper is a no-op when an advertisement is already set, so it fires no
extra heartbeat, while explicit setInstanceAdvertisement keeps its existing
replace semantics. buildInstanceAdvertisement moves to a shared module so the
command path and the enable path derive it identically.
* fix(cli): report pending question and permission on the session heartbeat
The heartbeat built each session's status from SessionStatus.Service, whose
union is idle/retry/busy/offline and which never consults Question.Service or
Permission.Service. deriveStatus() already did consult both, but only fed the
ingest session_status sync. So a session genuinely blocked on a question was
advertised as busy on the heartbeat, and the mobile app — which takes live row
status from the heartbeat — showed no needs-input badge.
Extract the precedence (permission, then question, then SessionStatus) into a
shared helper used by both deriveStatus and the heartbeat, so the two channels
cannot drift.
The heartbeat runs on a ~10s timer across every session, and deriveStatus makes
service calls per session, so the permission and question lists are fetched once
per tick and indexed by session id rather than queried per session. A test pins
the call count.
Behaviour note beyond the strict fix: sharing the derivation also means a
SessionStatus of offline now reports as retry on the wire, matching what
deriveStatus has always sent to ingest. Nothing consumes offline from the
heartbeat — the transport forwards only idle and busy, and the mobile row treats
both as non-attention — so the effect is that the two channels now agree. The
detach fence test is parameterised accordingly; its assertion that the status
clears on detach is unchanged.
* chore(cli): widen the promise-facade allowlist for the heartbeat attention tests
The DEF-3 heartbeat tests raise and reply to real Question and Permission
requests through the global AppRuntime, which took kilo-sessions.test.ts from 4
classified references to 29 and failed the allowlist check.
Bumping the count rather than restructuring the tests is deliberate: the
heartbeat resolves attention status from the global Question.Service and
Permission.Service, so asserting it requires driving those same services.
Scoped layers cannot express that — the global-runtime coupling is the thing
under test — and it is the same integration pattern this entry already
sanctioned for the detach fence. The reason string records that.
* refactor(cli): run remote sessions in one process with safe per-session exit
Consolidate remote session handling into a single CLI process instead of
spawning one process per remote-created session (addresses the PR review):
- restore in-process create_session (accepts an absent sessionId and targets
the connection directory); remove the session spawner, the
KILO_REMOTE_ATTACH_SESSION attach-on-boot path, the child-advertisement gate,
and their tests
- retain instance advertisement and fire one immediate out-of-band heartbeat on
(re)connect when advertising, so a headless `kilo remote` host is discoverable
without delay
Make /exit (wire command exit_cli, unchanged for compatibility) detach only the
target session instead of terminating the CLI:
- AttachedState.detach with a presence-suppression tombstone; detach also clears
the target's SessionStatus so the negative-containment heartbeat fence resolves
deterministically for busy/retry/offline sessions
- exit_cli handler verifies ownership, cancels the active prompt, detaches and
awaits the detach heartbeat, then ACKs; the interactive RemoteExit callback is
invoked only after the ACK when the last owned session exits; a headless
`kilo remote` host stays alive and advertising at zero sessions
- add an optional canExitSession boolean to the list_commands v1 catalog
(always true, independent of exitAvailable) so clients can detect safe
session-exit semantics
History and stored sessions are preserved on exit.
* fix(cli): break module-load cycle in remote session prompt-cancel
The K1 in-process exit_cli seam added a static `import { SessionPrompt }`
to kilo-sessions.ts. @/session/prompt evaluates KiloSessionPrompt at module
load, so the new static edge raced that init and left the namespace in TDZ,
crashing unrelated test files with 'undefined is not an object (evaluating
KiloSessionPrompt.shouldAskPlanFollowup)'. Defer to a dynamic import at the
single call site, mirroring remote-command.ts.
* fix(cli): correct AttachedState announce/detach concurrency and rollback
Address review findings on the shared-process session lifecycle:
- announce/detach no longer join the OPPOSITE in-flight operation. Joining
detach's negative-containment fence made announce resolve success for a
detached id (and vice versa: detach joined announce and resolved success
while still attached, which exit_cli treats as license to ACK/close). Each
path now joins only a same-kind in-flight op and, when the opposite op is
in flight, awaits it to settle and then performs the real work.
- Failed-detach rollback now releases the suppression tombstone, so a
still-attached session is not dropped by the next setPresence (the tombstone
loop would otherwise remove the still-present id and never clear).
- Both catch/rollback branches now honor the lifecycle generation guard
(mirroring the success path); a stale in-flight op that rejects after
reset() no longer mutates the new lifecycle's presence/pending/suppressed
sets (reset clears the same Set instances).
Adds regression tests for each fix, plus AC6f covering the remote-ws
detachSessionId negative-containment waiter.
* fix: avoid write-on-read mutation of config file
* fix: update test description for schema injection case
* Apply suggestion from @kilo-code-bot[bot]
Co-authored-by: kilo-code-bot[bot] <240665456+kilo-code-bot[bot]@users.noreply.github.com>
* chore: add changeset
* fix(cli): preserve as first config property via getInsertionIndex
* fix(cli): close missing brackets in config test preventing syntax error
* Removed test for preserving environment variables when adding $schema to config.
Removed test for preserving environment variables when adding $schema to config.
* fix(cli): add ordering assertions to existing injection tests
* docs(changeset): clarify per-load churn scope for comment-first JSONC
* test(cli): use FSUtil in config schema tests after upstream refactor
* chore: combine changeset lines
---------
Co-authored-by: kilo-code-bot[bot] <240665456+kilo-code-bot[bot]@users.noreply.github.com>
* feat(opencode): record per-step token throughput metrics
Capture prompt-processing and text-generation tokens/sec on every
StepFinishPart. The metrics helper prefers provider-reported rates
from llama.cpp / vLLM timings and falls back to wall-clock computation.
A new kilocode tui usage route renders PP/TG inline.
chore(sdk): regenerate types for StepFinishPart.metrics
feat(tui): render PP/TG in sidebar usage panel
feat(vscode): per-message and aggregated token throughput display
Surface throughput on each AssistantMessage badge (behind the
showTokenThroughput toggle) and as a compact PP/TG row in the
expanded TaskHeader. Adds session helpers, i18n entries in 20 locales,
and StepFinishPart.metrics to extension/webview messages.
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* fix(vscode): wire token throughput toggle and drop unreachable provider branch
- Replace the dead sendThroughputSetting() private with the shared
buildThroughputSettingMessage() helper and add validThroughputSetting
to handleUpdateSetting so the showTokenThroughput setting has the same
guard as the chat/indexing twins (fixes Knip regression).
- Bind the DisplayTab Switch to the local settings draft so the toggle
flips on click instead of waiting for a Save round-trip (the user-facing
kill switch for #6579).
- Narrow StepThroughputMetrics.source to "computed"; backend hard-codes
computed metrics today because the upstream AI SDK drops provider
timings. Drop the unused provider branches from AssistantMessage and
TaskHeader so the rendering code has no dead paths.
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* fix(opencode): centralize token throughput labels and tighten type guard
- Replace inline PP/TG labels in the CLI sidebar with a throughputLabel
constant in model-usage so a future i18n sweep is one file instead
of every rendering site.
- Tighten isStepMetrics in sidebar-usage back to a real discriminator
check after dropping the unreachable "provider" union member.
- Drop formatPP/formatTG exports from model-usage since callers already
use the shared formatRateValue; mirror the swap in the TUI usage test.
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* feat(token-throughput-v2): wire aggregation through DisplayProvider and test computed-only
- aggregateMetrics adopts the first non-empty computed sample per field
across every step-finish in the session, replacing the dead provider-
ranked last-wins strategy that shipped with the unreachable branch.
- Share the throughputVisible signal through DisplayProvider so every
AssistantMessage and the TaskHeader row react to a single onMount
requestThroughputSetting round-trip, instead of each message posting
its own handshake.
- Drop the unused routes/session/usage.tsx TUI route (no remaining
imports) and add the chat-layout badge/header pill styles it was
gating on.
- Refresh session-utils tests to exercise only source: "computed"
samples and follow the new first-wins rule.
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* fix(token-throughput-v2): seed settings on hydration, drop dead provider branch
Seed the DisplayTab "Show Token Throughput" Switch on initial load by
mirroring the throughputSettingLoaded message into settings() (same
pattern as chat.shiftTabCyclesVariant). Without this, a persisted-true
setting renders unchecked on first open because the Switch was bound to
settings()["showTokenThroughput"] but no handler ever populated it.
Drop the dead data-source attributes on the per-message and task-header
throughput surfaces now that StepThroughputMetrics.source is narrowed
to "computed" only — the attribute was always the literal string.
Drop the unreachable chat.throughput.badge.provider and
.chat.throughput.badge.tooltip.provider i18n keys across all 20 locales.
The badges no longer branch on source === "provider" since the
provider-source branch is removed (the AI SDK adapter upstream strips
llama.cpp timings before they reach providerMetadata).
Co-Authored-By: Claude <noreply@anthropic.com>
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* fix(token-throughput-v2): drop PP display until llama.cpp wiring lands
PP (prompt-processing rate) has no signal in this build: the AI SDK
adapter upstream strips llama.cpp's `prompt_per_second` before it reaches
providerMetadata, and computeMetrics has nothing else to derive it from.
Ship the TG (text-generation) rate only — the UI no longer renders the
"PP –" placeholder that made the feature look broken.
CLI sidebar drops the PP row; per-message badge and aggregated header
pill both lose the "PP – ·" prefix. The wire shape keeps the optional
prompt field so the follow-up that wires the upstream metadataExtractor
can populate it without another schema bump.
The `throughputLabel` constant on the opencode side and the `formatPP`
helper on the webview side are removed; tests that fabricated prompt
values are pruned to match.
Co-Authored-By: Claude <noreply@anthropic.com>
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* fix(token-throughput-v2): integrate TG into Tokens row, plain-text style
The standalone blue pills read as loud for what is secondary session
info. Move the aggregated TG into the existing Tokens row as another
spanned value (alongside ↑ input, ↑ cache, ↓ output) and restyle the
per-message badge as plain text in descriptionForeground so both surfaces
match the tokens family.
TaskUsage now accepts a `throughput` prop and renders `TG <rate> t/s`
inline in the Summary component when the toggle is on. TaskHeader no
longer emits a standalone [data-slot="task-header-throughput"] element;
its [data-slot="task-header-throughput"] CSS rule is removed. The
throughputText / throughputTooltip memos and the unused formatTG
import are dropped — the values flow straight into TaskUsage.
Co-Authored-By: Claude <noreply@anthropic.com>
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* fix(token-throughput-v2): finish removing memory badge from AssistantMessage
The "Merge origin/main into feat/token-throughput-v2" resolution kept the
throughput branch's memory-badge code (already removed from main by
28d015f8fe), which broke the kilo-ui-contract test and the i18n-keys test.
Drop the dead code: `useMemory`/`MemoryMarkerMeta` imports, `mem`, the
`meta`/`recall`/`fmt`/`count`/`items`/`verbose` createMemos, the `tip`
function, and the `<Show when={mem.enabled() && recall()}>` block. The
file lands at 349 lines (down from 391), matching main + throughput only.
Verified locally:
- i18n-keys + kilo-ui-contract: 53 pass, 0 fail
- Full kilo-vscode suite: failures 138 → 136 (+2 from the two fixes)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* fix(token-throughput-v2): finish removing memory badge from AssistantMessage
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* formatting fixes
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* fix(token-throughput-v2): address Marius review comments
- Rename 'TG' to 'Generation speed' in en.ts and add a 'gauge' icon to
packages/kilo-ui so the per-message badge and the Tokens row show
'<icon> Generation speed <rate> t/s' instead of the cryptic 'TG <rate>'.
Centralize the opencode sidebar label in throughputLabel.generation.
- Switch aggregateMetrics (both webview and CLI) to the latest non-empty
step-finish snapshot so only the most recent assistant turn's generation
rate is shown rather than a session-wide aggregate. Update tests and
comments to match.
- Translate the throughput strings in no.ts to Norwegian; mirror the new
key shape across the other locales (English fallback for untranslated
strings).
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* fix(token-throughput-v2): close unclosed CSS block and apply prettier formatting
The throughput rebases left a duplicated .vscode-session-turn-diffs
selector and let three files drift from prettier's expectations. Fix
the CSS unclosed-block (which broke the Storybook preview build) and
re-run prettier --write on the touched files.
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* ci: re-run after fixing CSS unclosed-block + prettier drift
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
* fix(token-throughput-v2): address kilo-code-bot review
Kilo:
- CRITICAL: Remove the section === '' guard in KiloProvider.ts that
blocked persistence of every top-level setting key. The throughput
validator is now redundant, so drop it from throughput-settings.ts.
- WARNING: Reset samples in sidebar-usage.tsx when props.session_id
changes via a keyed createEffect, so a session switch no longer
blends step-finish metrics from the previous session and the array
no longer grows without bound across long-lived plugin instances.
- WARNING: Pass {speed} to language.t for the TaskUsage throughput
tooltip and reuse the shared formatTG helper instead of reformatting
the value inline. Drop the dead [data-component='assistant-memory-badge']
rule whose target component no longer exists in the tree.
- SUGGESTION: Drop redundant guards in model-usage.ts (undefined check
after Number.isFinite, and the ?? 0 on an already-required number
field). Use typeof === 'number' for the type narrowing.
* feat(token-throughput-v2): weighted per-turn rate, plain text footer
Address Marius's review of the throughput UI:
Calculation
- Persist per-step timing (start/end/elapsed) on step-start and
step-finish parts in the session processor.
- Add wire schemas in core/src/v1/session.ts and packages/sdk/openapi.json
so the new time field round-trips end-to-end.
- Replace the last-wins 'latest step rate' snapshot with a weighted
aggregate: sum(output + reasoning tokens) / sum(active generation
duration) across the turn's step-finish parts. Tool execution and
idle waiting are excluded.
- The CLI sidebar (model-usage.ts) gains the same weighted semantics
when timing is available, falling back to last-wins otherwise so
older callers keep working.
Presentation
- Strip the per-message badge to plain muted text (no icon, no label,
no border). The chip in the upstream action row reads as metadata.
- Move throughput out of the task header Tokens row so each turn owns
its own value (no flicker across turns, single source of truth).
- Read the throughput memo from the full message parts in the data
store rather than the chunked row slice, so step-finish in any
chunk produces the badge.
i18n
- Replace chat.throughput.speed.{label,row,tooltip,tooltip.missing}
with chat.throughput.tooltip and chat.throughput.tooltip.missing
across all 19 locale files.
Tests
- Add messageThroughput and sessionThroughput describe blocks
exercising the weighted aggregate across multiple steps.
- Cover weighted + fallback paths in the CLI aggregateMetrics tests.
* fix(token-throughput-v2): render t/s inline beside copy/feedback buttons
Move the throughput badge from a footer line below the assistant message
into the copy/feedback action row of the text part that carries the copy
button. This avoids the extra vertical space the footer consumed.
Also apply prettier formatting to drifted PR files (i18n line wraps,
TaskHeader/session-utils/test reflows).
* fix(token-throughput-v2): correct changeset package name to @kilocode/cli
* fix(token-throughput-v2): annotate step-start time field with kilocode_change
---------
Co-authored-by: Thomas Brugman <thomas@kilocode.ai>
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
Co-authored-by: marius-kilocode <marius@kilocode.ai>
Compaction writes the output budget into agent.options.maxOutputTokens.
LLMRequestPrep.prepare merges agent.options into providerOptions, and the
AI SDK serializes it verbatim into the request body; strict
OpenAI-compatible upstreams reject it with
'Unsupported parameter(s): maxOutputTokens'.
Strip the injection site. The cap is still enforced through two
independent paths: the constrained model passed downstream and the
llm.ts re-cap on prepared.params.maxOutputTokens.
Marius