Repository navigation
Releases
Kite v1.0.10
Kite v1.0.10
Date: 2026-11-01
Highlights
A reliability and continuity release. Provider prompt-cache payloads now stay on compatible routes; failed LiteLLM prewarm imports can retry; sessions can be forked; and bundled defaults no longer block ordinary git push before its approval gate.
Provider and cache reliability
- Anthropic
cache_controlcontent blocks are sent only to Anthropic-compatible routes. OpenRouter, Azure, Bedrock, and Vertex requests keep their ordinary text content. - Cache-hit ratios use normalized input totals without double-counting cache-creation tokens.
- A failed background LiteLLM import clears its prewarm handle so the next call can retry.
Sessions and subagents
kite sessions --fork <id>creates a new session containing the source conversation and current provider/model/reasoning settings. It does not copy source audit events, checkpoints, or sidecars.- Unknown subagent profiles now suggest close matches; profile details show model and tool settings.
- The bundled system prompt documents scoped-worker write boundaries and uses Kite's bounded file-inspection tools consistently across platforms.
Defaults and UI
- Ordinary
git pushnow reaches the serious-action approval flow; destructive force-push remains blocked.--force-with-leaseretains its supported behavior. - Bundled tool progress interval matches the 2-second runtime default.
Performance and maintenance
- Retained the user-recorded controlled-run performance comparisons in the 1.0.10 changelog; these are local measurements, not portable timing guarantees.
- Offline behavioral coverage exercises cache request shape, prewarm retry, session fork isolation, and the packaged push defaults.
Upgrade
git pull
./scripts/install.sh --no-clone
pytest -q
kite --version # 1.0.10Full changelog
See CHANGELOG.md for the [1.0.10] entry.
Kite v1.0.9
Kite v1.0.9
Date: 2026-10-23
Highlights
A performance and cleanup release. Session tool events no longer rewrite the
metadata sidecar on every append, slash/plugin/skill machinery loads only when
needed, KITE_HOME and config reads avoid redundant filesystem work, and roughly
1,000 lines of unwired modules, unused helpers, and misnamed tests are gone.
CLI and REPL behavior is unchanged.
Session and config performance
- Durable events (
tool_start,tool_end, approvals, checkpoints, …) now append
to the transcript only. An event changes neither provider/model/reasoning nor
the message count carried by the sidecar, so the per-event atomic write +
fsync bought no durability. Recency continues to use the transcript/sidecar
mtime ranking key; message appends andnote_runtimestill refresh the
sidecar. record_eventresolves the persistence policy once per event and reuses it for
redaction instead of re-readingconfig.tomlfor each row.- Windows appends skip
Path.chmod(600): Windows only maps that to the read-only
attribute, while the inherited user-profile ACL already restricts the file. - A
KITE_HOMEoverride caches its resolved path, andUserConfig.loaduses a
single stat rather than checkingis_file()and then stat-ing again.
Illustrative local measurement for the append path (same machine, not a portable
timing guarantee): ~7.3 ms/event before this release, ~0.95 ms/event after.
Startup and system performance
- Plugin, skill, and markdown-command loaders are imported on demand instead of
at module import. The slash module no longer pulls in Rich/theme setup when a
plain task is about to run. - Plain REPL input is rejected before building the full slash index; only input
that starts with/pays for plugin/skill/command discovery. - Plain
kite runtasks skip the slash module entirely; only prompts that start
with/import it. - Illustrative local measurements: cold
import kite.ui.repl~655 ms →
~590-630 ms; isolatedimport kite.cli.slash~333 ms → ~110 ms. Direct
verbatim commands on a quiet machine are the only reliable comparison.
Cleanup
- Removed unwired application-layer modules (
application/context.py,
persistence.py,ui.py,event_payloads.py), the token-report/routing/tool-tier
experiments, and dead UI wrappers. None were reachable from the CLI or REPL. - Removed unused
HarnessDependenciesslots and miscellaneous dead helpers. - Renamed test bundles that said "token efficiency" while covering live
reasoning/cache, registry, read, spill, and error behavior; tests for removed
modules were deleted. - Consolidated duplicate interrupt/persistence assertions and merged the
question-tool, stream-stall, and diff-render tests into cohesive behavioral
cases. Collected cases: 465 → 460 (457 passed, 3 platform skips on Windows).
Compatibility
- LiteLLM 1.106 imports its MCP bridge for every non-empty tools list, even when
the tools are ordinary function tools. Kite now opts out of that bridge when
it is available, so the latest LiteLLM no longer triggers the optionalmcp
dependency on tool calls.
Upgrade
git pull
./scripts/install.sh --no-clone
pytest -q
kite --version # 1.0.9Full changelog
See CHANGELOG.md for the [1.0.9] entry.
Kite v1.0.8
Kite v1.0.8
Date: 2026-10-12
Highlights
Safer worker boundaries and more useful provider failures: scoped subagents cannot write outside their declared paths, Antigravity retries only transient failures, and auth/model errors now explain the next step. Restricted-mode pushes remain approval-controlled rather than being rejected outright.
Model and provider reliability
- Antigravity network, service, rate-limit, and 5xx failures use Kite's bounded retry policy. Authentication, quota, and malformed-response errors are not retried.
- Provider and model errors distinguish common connection, service, rate-limit, and authentication failures, redact sensitive values, and give a next action.
- Antigravity is shown as linked only after
agy modelsverifies the session.
Security and subagents
- A subagent's optional
scopebounds workspace file mutations before execution. An empty scope allows no writes; scoped workers cannot use mutating bash commands because their targets cannot be reliably constrained. - User memory and gathered repository metadata are marked as untrusted prompt data, with forged closing markers escaped.
- SSRF-protected HTTP fetches connect directly instead of honoring proxy environment variables, so connected-peer checks validate the destination. Configure any required proxy outside these guarded requests.
- In restricted mode, ordinary
git pushfollows the existing network approval policy. Force-push remains blocked.
Upgrade
git pull
./scripts/install.sh --no-clone
pytest -q
kite --version # 1.0.8Full changelog
See CHANGELOG.md for the [1.0.8] entry.
Kite v1.0.7
Kite v1.0.7
Date: 2026-10-09
Highlights
Runtime hot-path optimizations and correctness fixes reduce startup, streaming,
and test-suite costs while keeping behavior covered by focused, isolated tests.
Performance and correctness
- Reduced
LitellmModelconstruction overhead and headless startup-to-first-query
latency. - Improved streamed-answer and reasoning accumulation performance, and reduced
path-completion and picker costs for large directories and option lists. - Fixed correctness issues found while optimizing runtime paths; the test suite
now runs as independent tests rather than relying on shared batching.
Developer workflow
- Added offline end-to-end smoke coverage, isolated worktrees, a faster CI loop,
and CLI profiling support. - Documented performance and behavior changes, runtime traces, test policy, and
the development loop.
Upgrade
git pull
./scripts/install.sh --no-clone
pytest -q
kite --version # 1.0.7Full changelog
See CHANGELOG.md for the [1.0.7] entry.
Kite v1.0.6
Kite v1.0.6
Date: 2026-10-05
Highlights
Quiet release: the terminal stopped shouting. Tool output collapses to a
glanceable line, every truncated body is retrievable again with /last, and
the session store stops leaking files.
The scrollback
A single edit used to paint 16 preview lines at start and 40 diff rows at
end. A todo_write painted the checklist, a start card, a done card and the
raw item JSON — every task item two or three times.
- The task list is now one line:
● Tasks 1/3 █████░░░░░░ · edit render.py,
with the itemized checklist behind/tasks. todo_write/todo_readno longer echo their payload under the checklist
that already shows it. Status, timing and exit code still print.- Scroll-printed bodies get a 5-line budget (
PREVIEW_LINES): tool output,
write/edit previews, bash blocks and diffs. Truncation always names the
command that reveals the rest. - Approval previews deliberately keep a 40-line window — that is a pause where
you judge a patch, not scrollback.
/last
The 5-line cap initially shipped a lie: its markers named a /diff command
that does not exist, and /expand only widens later tool calls — so a body
truncated three turns ago had no way back.
/last re-prints the previous tool call in full: the record is captured at
tool-end (bounded at 200k chars per payload) and rendered through the existing
formatters, uncollapsed and uncapped. A diff comes back whole.
A regression test cross-checks every overflow marker against the registered
command list, so a hint naming a command that does not exist fails the suite.
Session storage
- Deleted sessions leaked a file.
delete_sessionreaped the transcript,
meta and trajectory but never the.stats.jsonsidecar — one orphan per
deletion, forever.delete_all_sessionsnow also sweeps sidecars orphaned by
a crash between unlinks. list_sessionsparsed metadata for every session before applyinglimit.
It now pre-ranks candidates by transcript/sidecar mtime and parses a bounded
head. The sidecar matters:note_runtimebumpsupdated_atwithout touching
the transcript, so ranking on transcript mtime alone reorders wrongly._trajectory_pathwas duplicated byte-for-byte across two modules; one copy
is gone.- Crash tolerance. Both readers raised
AttributeErroron valid-JSON-but-
not-an-object rows, andUnicodeDecodeErrorescaped uncaught — it is not an
OSError.list_session_eventswrapped its whole loop in one
JSONDecodeErrorhandler, so a single torn line discarded every event. A
truncated transcript is the normal post-crash state; all three now degrade to
the rows that did decode. - One saved checkpoint counted twice (the loop writes a
context_checkpoint
row and acheckpointevent for the same id). Deduplicated by id.
Footer
- A turn ending within 125 ms of its last paint could leave the footer frozen
on the finalworkingline until you typed.clear_running()is a turn
boundary and now forces its repaint. - A flash notification set at an idle prompt outlived its 8 s TTL, because
expiry was only checked insidetouch().active_flashexpires on read, and
the composer toolbar uses it. context_pctdivided by a negative context window, renderingctx -800%.
Tests
141 collected (141 passed, 1 platform skip) — up from 119, still under the 150
CI budget. New: scrollback density, tool-output retrieval, session hygiene,
analytics hygiene, footer throttle.
Upgrade
git pull
./scripts/install.sh --no-clone
pytest -q
kite --version # 1.0.6Full changelog
See CHANGELOG.md for the [1.0.6] entry.
Kite v1.0.5
Kite v1.0.5
Date: 2026-10-03
Reliability release: kills the infinite-thinking hang, makes long tasks
visible and resumable, tightens the submit gate against hallucinated
"done", and cleans up answer streaming.
Hang fix
- Provider establishment is now bounded and cancel-responsive
(_bounded_completion_call): a held-open connection raisesTimeoutError
past the model timeout (default 180s) and surfaces as a retryable
ProviderFaultinstead of spinning forever. Esc interrupts promptly. - Loop-level watchdog: per-attempt hard bound (2× model timeout + 60s)
with interrupt polling, so no model can hang the worker thread. agent_endalways fires (hardenedfinally), and the REPL turn loop
resets busy/running/spinner state intry/finally— no more stuck
"working" footer after Ctrl-C or save failures.- Long-reasoning streams emit
stream_thinking_long("thinking N chars")
instead of looking frozen; coalescer flushes onturn_end;
interrupt/errorclear stale running labels.
Answer text
- No split-style flicker on streamed
##/-lines; fenced code blocks
cap streamed output at 4k chars with an explicit truncation pointer. - Tool output capped centrally (12k, head+tail) with
...[truncated]
markers and spill-file pointers; tables truncate with…, never silent
slices. - Prompts require
path:line/command citations and propersubmit-tool
finishes; casual chat stays literal (no checklist forhi).
Long tasks
- Headless display renders progress heartbeats, checkpoints, compaction,
context ratio, retries/faults, approval gates, and job elapsed times —
long runs never look hung. - Every non-success exit logs a mission/done/next/todos continuity brief
pluskite resume <sid> "continue from checkpoint";tasks runprints
row errors with resume hints. - Phantom
"running"jobs eliminated (failed drain marksjob_end);
two new bench cases cover the exit path.
Hallucination gate
- Submit requires
## Done+## Changed(+## Verificationwith✓
when checks exist) for edited sessions; unrelated passing tests no
longer satisfy evidence;py_compile/node --checkcount as checks. - Relaxed path (
verify_before_submit=False) and no-edit turns keep prior
behavior — casual finishes still submit.
Connection health (verified, no change)
- Boot stays off-network (fast probe, lazy resolve); misconfigured
providers fail fast withkite keys/kite loginhints; transient
faults retry with countdown, auth fails hard, sessions survive for
resume --retry. Known gaps logged for next pass: Codex probe without
timeout, generic 429Retry-Afterunparsed, one fault rendering 3×.
Tests
114 collected (113 passed, 1 platform skip) — unchanged count, still
under the 150 CI budget.
Upgrade
git pull
./scripts/install.sh --no-clone
pytest -q
kite --version # 1.0.5Full changelog
See CHANGELOG.md for the [1.0.5] entry.
Kite v1.0.4
Kite v1.0.4
Date: 2026-10-03
Security-hardening release. Fixes all 9 findings from the adversarial audit
(trust boundaries, session durability, usage accounting, context freshness)
and consolidates the test suite below the 150-test CI budget.
Security fixes (High)
- Project extensions gated on trust (F-01). Project-local
.kite/extensionsno longer executes on load. Only user-global
~/.kite/extensionsloads until the workspace is explicitly trusted;
headless runs (no trust prompt) stay untrusted.no_extensionsstill
disables everything.SECURITY.mdnow states the exact behavior. - GitHub token scoped to lone
gh(F-02). AmbientGH_TOKEN/
GITHUB_TOKENare injected only into a singlegh …subprocess, never
into a chained shell (gh …; <other>, pipes,&&, substitutions).
Compound commands can no longer exfiltrate the token to sibling processes.
Correctness fixes (Medium)
- Session events survive rewrites (F-03).
_write_metare-appends
durableevent/context_checkpointrows, soset_exit,save, and
compaction snapshots no longer wipe the audit trail. - Atomic session snapshots (F-04). Session files rewrite via
temp-file +os.replacewith owner-only perms; an interrupted write
leaves the prior transcript intact.
Edge-case fixes (Low)
- Non-object JSONL skipped (F-05).
null, lists, strings, and numbers
in session files are skipped instead of abortingload_session/ resume. - Cache accounting (F-06). OpenAI-style
cached_tokens(a subset of
prompt_tokens) no longer double-counts: 1000 prompt / 500 cached now
reports 1000 input and 50% hit ratio instead of 1500 and 33%. - Stale context cache (F-07). Instruction-file and commit fingerprint
invalidates the 120s project-context cache;git statusrefreshes on
cache hits so edits surface immediately. - Authorized
git push(F-08). Plainpush/cloneflow to the
approval path (SERIOUS remote-write) instead of a permanent hard-block;
only--force/-f(except--force-with-lease) stays hard-denied. - Agy prompt ceiling (F-09).
flatten_promptenforces the 24k-char
hard bound after composing system + retained turns, and theagy
subprocess inherits a filtered environment (no credential keys).
Tests
Suite consolidated 305 → 114 collected (113 passed, 1 platform skip)
with identical assertions — batched per domain module, headroom for
regression tests under the 150 cap. New regression tests cover every
finding above.
Upgrade
git pull
./scripts/install.sh --no-clone
pytest -q
kite --version # 1.0.4Full changelog
See CHANGELOG.md for the [1.0.4] entry.
Kite v1.0.3
Kite v1.0.3
Date: 2026-09-29
Highlights
Thinking variants go first-class (kite variants / /variants, strictly
limited to what your model supports, shown as provider/model#variant),
the model can now ask you clarifying questions mid-turn (opencode-style
question tool), Antigravity subscriptions answer through your signed-in
agy CLI without an API key, and the terminal UI compacts itself inside
Orca relay sessions. Theme pickers are typed-only now — no more frozen
option lists or stray escape bytes.
Upgrade
git pull
./scripts/install.sh --no-clone
pytest -q
kite --version # 1.0.3Full changelog
See CHANGELOG.md for the [1.0.3] entry.
Kite v1.0.2
Kite v1.0.2
Date: 2026-09-25
Highlights
- Token Efficiency & Harness Pass — Comprehensive Eric Zakariasson token efficiency pass covering telemetry, prompt split, spill management, tiering, error handling, history compaction, and routing.
- Antigravity & Auth Caching — Enhanced verification via the official
agyflow and cached OAuth status probes with hermetic test isolation. - Startup Performance — Background prewarming of LiteLLM imports and inlined CLI parser structure for lightning-fast startup.
- Robust Self-Update & Windows Handoff — Detached background handoffs, shim-aware helpers, install lock retries, and clean uninstall flows for Windows.
- Crew & Multi-Agent Orchestration — Crew coalescing, whole-tree cost aggregation, and richer worker context for orchestrated agents.
- NVIDIA & Provider Resiliency — Gated parallel tools on explicit support for NVIDIA, stream fallback, gateway key loading, instruction preservation, and cross-platform CRLF normalization.
Upgrade
git pull
./scripts/install.sh --no-clone
pytest -q
kite --version # 1.0.2Windows: .\scripts\install.ps1. Global CLI: uv tool upgrade kite.
See CHANGELOG.md for the [1.0.2] entry.
Kite v1.0.1
Kite v1.0.1
Date: 2026-09-22
Highlights
- Antigravity OAuth —
kite login antigravitylinks a Google
subscription through the officialagyCLI (local browser flow or the
SSH manual URL loop). Status/logout, catalog entry, and fallback Gemini
models included; model calls useGEMINI_API_KEY, same split as the
Claude Code subscription. - Composer fold — long pastes collapse to first lines +
+N lines;
any key expands,Enteralways submits the full text,F9toggles. - Steering that continues — mid-turn corrections interrupt the agent
but keep the pinned composer alive instead of stopping the session. - Metering that meters — cost falls back to usage-based pricing when
providers omitresponse_cost; the footer leaves $0.000 behind. - Plan mode acts —
/plan <text>and/build <text>start work
immediately; bare/planreports a resumed checklist. - Leaner suite, same coverage — tests consolidated 425 → ~217 with
every assertion preserved;examples/clampuntracked and purged.
Upgrade
git pull
./scripts/install.sh --no-clone
pytest -q
kite --version # 1.0.1Windows: .\scripts\install.ps1. Global CLI: uv tool upgrade kite.
See CHANGELOG.md for the [1.0.1] entry.