* feat(kosong): classify HTTP 413 request-body-too-large as a dedicated error type
* feat(agent-core): lower default image downscale cap to 2000px and make it configurable
* feat(agent-core): strip media to text markers and retry when the compaction request is too large
* feat(agent-core): cap model-initiated image reads with a configurable byte budget
* feat(agent-core): resend with degraded media when the provider rejects the request body as too large
* test(agent-core): add explicit timeouts to encode-heavy image budget tests
* feat: add WebP decoding support with wasm integration
- Introduced a new WebP decoding module using @jsquash/webp's wasm decoder.
- Implemented functions to decode WebP images and check for animated WebP formats.
- Updated image compression tests to include scenarios for WebP handling, including encoding and decoding.
- Enhanced error handling for API request size limits to accommodate various error messages.
- Updated pnpm lockfile to include new dependencies for WebP encoding and decoding.
* chore(changeset): consolidate this PR's entries into one
* fix(nix): update pnpmDeps hash for merged lockfile
* feat(agent-core): refuse HEIC/HEIF reads with platform-matched conversion guidance
The r1/r2/r3 reminders injected into repeated tool results led with
prohibition verdicts and, in r2, echoed the repeated tool name and full
arguments back into the context, reinforcing the very pattern they were
meant to break. Rewrite them to state the situation factually and hand
the model a concrete next action: an expectation-setting sentence for
the next call (r1), a forced decision menu of falsify / ask-user /
conclude (r2), and a final hand-off summary without further tool calls
(r3). Detection, thresholds (3/5/8/12), force-stop, and telemetry are
unchanged.
* fix: surface provider auth error for unavailable models
When an OAuth-managed model returns 401 after a forced token refresh, the token is valid but the provider rejected it for that model (the account lacks access). Emit provider.auth_error carrying the provider's message instead of auth.login_required with a misleading "OAuth login expired. Send /login" prompt.
* fix(agent-core): preserve provider auth errors through compaction
Treat provider.auth_error like auth.login_required in the compaction path so an auth rejection during compaction surfaces the provider's message instead of being wrapped as a generic compaction failure.
Align progressive tool disclosure with the discard-on-compaction model:
compaction no longer rebuilds loaded dynamic tool schemas. The boundary
announcement re-lists every loadable name, the model re-selects what it
still needs, and a from-memory call to a no-longer-loaded tool is
rejected by preflight with select guidance.
This removes the keep-all rebuild and its half-trigger budget heuristics
entirely: the post-compaction floor is back to users + summary, which is
structurally outside the auto-compaction trigger band, and the guard
baseline degenerates to summary + reinjected reminders. Every downstream
mechanism already treated the empty loaded set as its consistent base
state (ledger scan, pending clear at the compaction boundary, deferred
extras, preflight wording), so this is a strict simplification.
Co-authored-by: fengchenchen <fengchenchen@moonshot.ai>
* fix(agent-core): report EXIF-rotated image dimensions and raise edge cap to 3000px
Image compression now reports original dimensions in the decoded
(EXIF-rotated) space, matching the coordinate system of the sent image
and of ReadMediaFile region readback; previously portrait JPEGs
(orientation 5-8) got swapped width/height in captions. The longest-edge
downscale cap rises from 2000px to 3000px, and the default jimp resize
path is documented as the anti-aliased area-average one so it is not
accidentally switched to a point-sampled interpolation mode.
* test: shrink oversized image fixtures to fit CI timeouts
The 3600x3600 fixtures introduced for the 3000px edge cap nearly doubled
the pixel area jimp has to decode and deflate, pushing the slowest
compression tests past the 5s vitest timeout on CI runners. 3600x1800
keeps every fixture over the cap while restoring roughly the workload of
the old 2600x2600 fixtures that CI handled comfortably.
* test: pin anti-aliased downscale quality with executable guards
A 1px checkerboard probe pins the compressor to full-coverage averaging
at integer and fractional ratios, with jimp's point-sampled BILINEAR
mode kept as the executable aliasing counter-example (it collapses the
50%-gray pattern to solid black at 4:1). Also guards the other classic
downscale bugs: transparent-pixel color bleed, mean-brightness drift,
iterative recompression degradation, and zero-size collapse on extreme
aspect ratios.
* fix(agent-core): report decoded EXIF-rotated dimensions in ReadMediaFile notes
The media note derived its original-dimensions line from the header
sniff, which reports pre-rotation values for EXIF orientation 5-8
JPEGs. The sent image and region readback both live in the decoded
(rotated) space, so portrait photos got axis-swapped coordinate
guidance. Once a decode has happened — compression or crop — its
dimensions now overwrite the sniffed ones.
* fix(agent-core): improve handling of EXIF orientation in image dimensions and metadata
* fix(agent-core): sniff EXIF orientation and step budget fallback through 2000px
Two follow-ups to the EXIF and 3000px-cap changes:
sniffImageDimensions now reads the JPEG EXIF Orientation tag (pure
header parse, both byte orders) and reports display-space dimensions
for orientations 5-8. Passthrough images — never decoded — previously
kept the pre-rotation header size in compression results and media
read notes, disagreeing with the decoded space that region readback
uses.
encodeWithinBudget steps the over-budget fallback through 2000px
before the 1000px last resort. Raising the cap to 3000px had left a
regression window: an image whose 2000px encode fits the byte budget
was sent at 1000px where the old 2000px cap used to send it at
2000px.
* fix(kimi-code): record pasted image dimensions in display space
The TUI paste path recorded attachment and original dimensions from its
raw header parser, which ignores EXIF orientation. For a portrait JPEG
the submit-time caption then contradicted the sent image's aspect and
region readback coordinates were axis-swapped. Dimensions now come from
the compression result, which reports display space on both the
compressed and passthrough paths; parseImageMeta remains only the
format/mime gate.
* feat(agent-core): add image compression and crop telemetry
Every image ingestion path now reports an image_compress event —
outcome (compressed / passthrough fast, guard, unsupported, unhelpful,
error), input/output formats, byte and pixel sizes, EXIF transposition,
and duration — and region readback reports an image_crop event with a
failure classification and the region's share of the original area.
Wiring is per call site via a new CompressImageOptions.telemetry
option, so the outcome split and timing are measured inside the
compressor while each caller only names its source: ReadMediaFile
(tool construction, like GrepTool), MCP tool results (McpOutputOptions),
server prompt ingestion (ICoreProcessService now exposes the host
telemetry client), ACP prompts (session track adapter), and TUI paste
(host.track adapter). Properties are numeric/enum only — never paths
or content — and a throwing client can never affect the compression
result.
* fix(agent-core): run the full JPEG quality ladder at fallback sizes
The fallback rescales encoded only at quality 20, so a JPEG whose
ladder failed at the fitted size collapsed straight to the lowest
quality even when the smaller size left budget headroom for a higher
rung (the realistic window is the 1000px step, where the 4x pixel
drop pays for q80/q60). Each fallback edge now walks the same
q80-to-q20 ladder as the fitted size.
* test: shrink heavy JPEG fixtures and add explicit timeouts
The fallback-ladder test runs ~11 pure-JS JPEG encodes and the EXIF
paste test decodes, rotates, and re-encodes a 6.5MP frame; both sat at
the edge of the 5s vitest timeout on CI runners. Narrower fixtures cut
the pixel area (the ladder test keeps its width above 2000px so the
full fallback chain still runs) and explicit 15s timeouts absorb runner
variance.
* fix(server): scope prompt image compression telemetry to the session
The prompt-ingestion image_compress events were emitted with the bare
host telemetry client, while every agent-side source inherits a
session-scoped client — so prompt_inline/prompt_file events could not
be correlated with their session. The route now wraps the client with
withTelemetryContext({ sessionId }) like rpc/core-impl does for
session telemetry.
* chore(changeset): consolidate image compression changesets
One entry covering the cap raise and the EXIF dimension fix, listed
for both the CLI and the SDK so the SDK changelog's compression
description (previously pinned at 2000px) stays accurate.
Pass windowsHide:true when spawning the hook process so a visible console
no longer flashes and steals focus on Windows. The Bash-tool path was
already hardened (KAOS buildLocalSpawnOptions); the hook runner missed the
flag even though its own taskkill helper already set it.
Extract the spawn options into a pure builder and add a regression test
asserting windowsHide, mirroring the existing KAOS spawn-options test.
Relates to #1298.
* feat(agent-core): record llm request trace in wire.jsonl
Add three observability record types so every request sent to the model
can be reconstructed from the wire log at the logical-request level:
- llm.tools_snapshot: content-addressed snapshot of the top-level tools
table as sent (post deferred-strip), written once per unique table
- llm.request: one record per outbound request (retries, strict resends,
and compaction rounds included) carrying the effective request params
and hash links to the system prompt and tools snapshot
- mcp.tools_discovered: the server's verbatim tools/list result plus the
agent's gating (allow-list, collisions), deduplicated by content hash
Observability records never feed state rebuild; replay only restores the
write-dedup cursors. The records/types.ts contract now documents the two
record classes explicitly (persisted is not the same as replayed).
Recording happens at the single Agent.generate choke point. The
LLMRequestLogFields side channel gains kind/projection/maxTokens/
droppedCount, chatWithRetry preserves caller-set fields, and compaction
tags its requests. The vis wire view renders the new record kinds.
* fix(agent-core): record the provider-clamped completion cap in the request trace
The llm.request trace recorded the client-requested budget cap, but
chat-completions providers tighten the actual wire value inside
withMaxCompletionTokens (remaining-context sizing, transport ceilings,
model-default resolution) — with the default budget the clamp is active
on nearly every non-empty-context request, so the recorded value did not
match what was sent.
Providers now expose the effective cap they computed as a readonly
maxCompletionTokens field on the clone, and the recorder reads it from
the effective provider at the Agent.generate choke point. This replaces
the side-channel recomputation, which is removed along with the
appliedCompletionBudgetCap helper.
* fix(agent-core): park pre-replay MCP discovery records and hash the collision outcome
Two wire-hygiene fixes for the mcp.tools_discovered trace:
Parking: the real Session ordering connects MCP servers concurrently with
agent construction, so ToolManager can observe a connected server before
agent.resume() has replayed the wire. Recording at that point bypassed
the restored dedup cursor (duplicating a 1-50KB record on every resume)
and appended a stray metadata record ahead of replay. AgentRecords now
exposes a one-shot opened latch — set when replay completes (after the
migration rewrite flushes) or when the first live record is logged — and
ToolManager parks discoveries until then, re-running the dedup check at
drain time. A frozen range-limited replay never opens; those agents are
transient previews.
Collision hashing: the dedup hash now covers the collision outcome, not
just the raw list and allow-list. Collisions depend on which other
servers hold a sanitized qualified name at registration time, so a
server can re-register with identical tools but a flipped outcome; that
gating change must produce a new record instead of being suppressed.
* fix(agent-core): skip the request trace for pre-flight-aborted calls
Mirror kosong generate()'s pre-flight abort check at the Agent.generate
choke point: a call whose signal is already aborted never reaches the
wire (generate throws before dispatching), so it must not leave an
llm.request/llm.tools_snapshot trace or a diagnostic log line claiming a
request was sent. Recording stays before dispatch for every call that
passes the gate, preserving the crash-safety of the trace.
* chore(agent-core): remove a leftover adaptive-thinking override hook
The adaptiveThinkingOverride option was a temporary local hook explicitly
marked for removal before commit. Nothing passes it, so resolution falls
back to the alias-level adaptiveThinking value in all cases; drop the
option and the dead indirection.
* fix(kosong): derive the exposed completion cap from generation kwargs
maxCompletionTokens was a field stored only by withMaxCompletionTokens,
so caps that reach the wire through other paths were invisible to the
request trace: with completion budgeting disabled via env, Anthropic
still sends the constructor-resolved max_tokens (required by the
Messages API), and constructor-level kwargs like OpenAILegacyOptions
maxTokens were likewise unreported.
Replace the stored field with a getter derived from each provider's
generation kwargs — the single source the request body reads — covering
constructor defaults, direct withGenerationKwargs configuration, and
budget application in one place. Kimi mirrors its request-time legacy
max_tokens alias normalization; openai-legacy reuses the same
normalizeGenerationKwargs the request path uses.
* feat(agent-core): add thinkingKeep passthrough for Kimi providers and update tests
- drop the session-wide absolute drain deadline that gated print-mode turn holds
- hold the turn until background subagents reach a terminal state, bounded by each subagent's own timeout
- fixes late or long-running subagents being abandoned (results suppressed) in long `kimi -p` runs
* fix: stop rendering <system> notes from tool results in the terminal and web UIs
Tool results carry <system> blocks as side-channel notes for the model (ReadMediaFile summaries, Read status, MCP image captions, error/empty sentinels). Keep them in history for the model, but strip them at every core-to-UI boundary so they no longer render as plain text. vis is intentionally left untouched to preserve the model's-eye view for debugging.
* fix: keep error/empty status text visible when stripping tool-result <system> tags
Unwrap the tool error/empty sentinels (<system>ERROR: ...</system>, <system>Tool output is empty.</system>) instead of deleting them: keep the human-readable text and drop only the tags. Otherwise a failed or empty tool result rendered as a blank output, indistinguishable from a rendering bug. The model still reads the wrapped form in history.
* refactor: move tool-result metadata into a structured note side channel
Tool-produced model-facing metadata (ReadMediaFile summaries, Read status
lines, MCP image-compression captions) was baked into tool output as
<system> text, so every UI had to strip it back out and three copies of
the model-view normalization had silently drifted apart.
- ExecutableToolResult gains `note`: content rendered to the model but
never to UIs; records and history now store the raw output plus the
structured isError/note fields
- the model view is rendered exactly once at the LLM projection boundary
by renderToolResultForModel; the transcript and vis hand-copies are
deleted (vis now calls the same function for its model view, fixing
their drifted empty-output checks)
- ReadMediaFile / Read / MCP captions write `note`; tool outputs stay
pure data, and text-only results keep a single text part (note joined
with a newline) so provider tool content stays a plain string
- all UI-side <system> stripping is removed; failed tools show their own
error text with the structured isError flag
- wire protocol 1.4 -> 1.5 migrates existing records' tool-produced
<system> blocks into `note` on resume
* fix: provider-neutral wording, no wire migration, direct optional fields
- "The attached image was downsampled" replaces directional wording that
depended on provider serialization order (inline media vs flatten-and-
re-attach)
- drop the 1.4 -> 1.5 wire migration: legacy records replay verbatim, so
the model view of old sessions stays byte-identical to what the model
originally saw and UIs show the legacy <system> text as-is; this also
removes the risk of the migration misclassifying user data that quotes
tool metadata, and the additive note field needs no version bump
- pass optional result fields as undefined instead of conditional spreads
(repo convention)
* fix: enforce the note contract at the trust boundary; narrow the TUI system-tag guard
- normalizeToolResult now keeps a note only when it is a non-empty string:
tools and finalize hooks are arbitrary JS, and a malformed note (null,
number, object) would previously persist into the record and crash every
subsequent LLM projection of the session. Everything downstream now
trusts note to be string | undefined.
- the TUI tool body suppression matches the full <system-reminder> tag
instead of any <system prefix: reminder piggy-backing stays hidden,
while real output that merely starts with a literal <system> tag (file
contents, MCP text) stays visible, covered through the real
ToolCallComponent path.
* fix: return MCP compression captions as data instead of extracting them from text
compressImageContentParts now returns { parts, captions } — captions come
back from the compressor as structured data and are never inserted into
the parts, so the MCP pipeline no longer pattern-matches text to move
them into the note side channel. Tool output that merely quotes a
caption (a doc, a log, a test fixture) stays verbatim in the output.
Also corrects the stale claim that prompt ingestion uses this helper
(it compresses per image while constructing the part).
* docs: correct the image-compression re-export comment; export CompressedContentParts
The package-root comment still described compressImageContentParts as the
input-stage helper every ingestion site calls; prompt ingestion compresses
per image with compressBase64ForModel / compressImageForModel, and the MCP
pipeline is the walker's only caller. Also export the walker's
CompressedContentParts return type so public-API consumers can name it.
* feat: wrap tool status sentinels in <system> so the model can tell harness verdicts from tool output
The error/empty status text is model-only after the note refactor (UIs
render the raw output and style failures via the structured isError
flag), so the earlier plain-text wording served no remaining audience.
Wrapping the statuses in <system> gives every piece of system-generated
text inside a tool result the same marker:
- failed calls get '<system>ERROR: Tool execution failed.</system>'
unconditionally — the ERROR:-prefix guard is removed, so the harness
verdict can no longer be confused with tool output that happens to
start with error-like text
- empty outputs render as '<system>Tool output is empty.</system>'; the
plain placeholder the loop layer bakes into records is still
recognized and upgraded at projection time
* style: collapse an internal helper docstring per the services subtree convention
* feat(agent-core): enable Preserved Thinking by default on the Anthropic provider
Default thinking.keep to "all" for the Anthropic provider (Claude and Kimi in Anthropic-compatible mode) while Thinking is on, via a context_management clear_thinking_20251015 edit, mirroring the Kimi default. Reuses [thinking] keep and KIMI_MODEL_THINKING_KEEP (env > config > default "all"); off-values disable it.
* feat(kosong): route Anthropic Preserved Thinking through the beta Messages API
Force the beta endpoint (client.beta.messages.create) when thinking.keep is enabled, since clear_thinking_20251015 is only honored there. Also prepend clear_thinking to any existing context-management edits (for example clear_tool_uses) instead of replacing them, keeping it first as Anthropic requires when combining edits.
* docs: clarify Anthropic beta endpoint and compaction keep behavior
Note in code comments and bilingual docs that enabling Anthropic Preserved Thinking routes requests to the beta Messages API (client.beta.messages.create), with keep=off as the escape hatch back to the standard endpoint. Correct the resolveThinkingKeep comment to reflect that compaction shares ConfigState.provider and intentionally carries the same keep.
* test(kosong): cover Anthropic beta endpoint (streaming and forced betaApi)
Add a streaming beta-endpoint capture and a test that withThinkingKeep forces the beta endpoint even when constructed with betaApi: false, pinning down the documented behavior.
Add an `interrupt_reason` field to the `turn_interrupted` telemetry event so the data can tell a deliberate user cancel (`user_cancelled`) apart from a programmatic abort (`aborted`), max-steps exhaustion (`max_steps`), an error (`error`), or a hook-filtered turn (`filtered`).
The user-cancel signal comes from the existing UserCancellationError carried as the abort signal's reason, reused here without changing any loop control or external protocol semantics.
* feat(agent-core): feed AskUserQuestion answers back as question text and option labels
The flattened answers record the model receives was keyed by synthesized
ids (q_0 / opt_0_1), forcing a cross-message positional lookup against the
original tool call to understand what the user picked — both unreadable in
transcripts and a real model-misreads-the-choice badcase.
- toAgentCoreResponse now takes the original broker request and translates
wire ids back to question text (keys) and option labels (values);
unknown ids are kept verbatim, missing request falls back to raw ids
- wire protocol unchanged: clients still answer with option ids; the
resolve route reads the pending request before settling it
- question texts must be unique per call and option labels unique per
question, enforced in the tool execution path (AJV cannot express the
zod refine) and mirrored on the exported schemas
- web transcript card resolves both the new label form and legacy id
transcripts; TUI and ACP paths already produced the text form
* fix(agent-core): align multi-select answer join across clients and harden question schema
- Join multi-select labels with ', ' in the server translator, matching
what the TUI reverse-RPC path already emits, so the model sees one
format regardless of which client answered
- Trim segments in the web transcript resolver before label matching:
TUI-answered multi-select transcripts (', '-joined) previously lost
their highlight to a spurious leading-space Other row
- Move the question-text/legacy-q_<i> answer lookup out of the SFC into
askUserToolParse as answerFor(), per that module's testability intent
- Require non-empty question text and option labels (.min(1)) so empty
strings are rejected by AJV at the tool boundary instead of failing
deeper in the protocol layer
* fix(agent-core): resolve option ids only within the answered question
The translator's option-id lookup was a single flat map across all
questions, so a stale or malformed response pairing one question with
another question's option id (q_1 + opt_0_0) was silently translated
into a label that was never offered for that question. Scope the lookup
to the answered question's own options; cross-question and unknown ids
now both pass through verbatim, staying diagnosable.
* feat(agent-core): progressive tool disclosure via select_tools
Keep MCP tool schemas out of the immutable top-level tools[] and let the
model load them on demand, preserving the provider prompt cache:
- kosong: Message.tools (append-only load primitive, serialized as Kimi
messages[].tools with type:function wrapping and no content),
Tool.deferred (stripped once in generate() so loaded tools stay
executable without re-entering the top level), select_tools capability
bit (UNKNOWN/catalog default false).
- select_tools builtin: load-by-exact-name, three-branch semantics
settled per name (Loaded / Already available / Unknown), schemas read
from the live registry, injection-origin schema messages survive undo.
- ToolsDiffInjector: <tools_added>/<tools_removed> announcements at turn
boundaries and post-compaction, folded from history (undo/compaction/
resume self-heal), appended only when the loadable set changes.
- Loaded-tools ledger = history scan + defer-window pending set (cleared
on /clear); loop re-reads the executable table per step so a selected
tool dispatches on the next step of the same turn; preflight
distinguishes not-loaded from loaded-but-disconnected.
- Cross-cuts: projection strips protocol context for non-select_tools
models (lossless mid-session model switch both ways), compaction
filters it from the summarizer input and rebuilds loaded schemas
keep-all after folding, token estimation counts message.tools, request
logging reflects the post-strip wire tools.
- Three-condition gate: capability.select_tools x capability.tool_use x
tool-select experimental flag (KIMI_CODE_EXPERIMENTAL_TOOL_SELECT).
Any gate closed reproduces the inline request byte-for-byte; all
current models keep the capability off, so behavior is unchanged
until a supporting model is catalogued. The SDK catalog-to-alias
mapping forwards the capability so catalog-driven setups can enable it.
* feat(kosong): skip tool-declaration-only messages in non-Kimi providers
Message-level tool declarations (messages[].tools) are a Kimi wire
feature. The other providers' explicit field construction already keeps
the tools field off the wire, but the content-free leftover message
would be rejected (OpenAI: system message without content) or serialize
as a garbage <system></system> turn (Anthropic/Google system-to-user
wrapping). Skip such messages entirely via a shared predicate; a message
that also carries content only loses the tools field, as before.
Unreachable in kimi-code (the projection gate strips dynamic-tool
context for models without the select_tools capability before any
provider sees it) — defense-in-depth for direct kosong consumers.
* fix(agent-core): survive runtime flag flips and align tool table with post-compaction state
Two fixes from PR review:
- Register select_tools unconditionally and gate only its exposure in
loopTools. The tool-select flag can flip at runtime (config reload
calls setConfigOverrides on the live resolver) without
initializeBuiltinTools re-running; previously the disclosure shape
activated while the tool itself was unregistered, cutting the session
off from MCP entirely until a model/cwd change rebuilt the builtins.
A profile listing the name explicitly still never surfaces it in
inline mode, and execution guards the flip race defensively.
- Resolve the per-step tool table AFTER beforeStep, next to
buildMessages. beforeStep can run full compaction, which trims loaded
schemas and rewrites the ledger; a table captured before it could
still dispatch a tool whose schema the model no longer has. The
executable table and the request messages now always reflect the same
state, so a trimmed tool is rejected with select guidance instead of
executed.
* fix(agent-core): drop unused Tool import in dynamic-tools
* fix(agent-core): baseline compaction guard after post-compaction reinjection
The reinjected reminders (loadable-tools manifest, goal) are re-appended
after every compaction, but the nothing-new-since-compaction baseline was
captured before injectAfterCompaction. With a large manifest the guard
could re-trigger auto-compaction against a floor that cannot shrink.
Raise the baseline to the true post-compaction floor once reinjection
completes; the earlier capture stays as a fallback when reinjection
throws.
---------
Co-authored-by: fengchenchen <fengchenchen@moonshot.ai>
Default `thinking.keep` to "all" when Thinking is on so prior `reasoning_content` is kept across turns. Add `[thinking] keep` to config.toml and keep `KIMI_MODEL_THINKING_KEEP` as an override (env > config > default); off-values disable it.
- add GET /api/v1/workspaces/{workspace_id}/skills backed by a listWorkspaceSkills core RPC and ISkillService.listForWorkDir, reusing the session skill-loading path so results match a new session
- web: populate the composer slash menu from workspace skills before a session exists, then fall back to session skills once one is active
* feat(session): rebuild session index on boot and self-describe workDir
- persist workDir into state.json so session dirs are self-describing and
summaries do not depend on the index's one-way-hashed workDir
- relax readSessionIndex so a stale or non-absolute index workDir no longer
drops an otherwise valid entry
- serialize in-process index appends to avoid torn jsonl lines
- add SessionStore.reindex() and run it once at server boot so the
scan-free request path can find sessions whose index line is missing
or stale
* chore: add changeset for session index rebuild
* fix(session): repair index entries with a stale workDir during reindex
- send tool declarations, system prompt, and sampling/thinking settings in the camelCase shape the Google SDK forwards so tool calls reach the model
- thread tool-call extras (thought signatures) through the loop tool-call event into context so Gemini 3 can resume a tool turn
- update and add tests for the corrected request shape and signature round-trip
Background (detached) shell commands were exempt from the 16 MiB output
ceiling, so a runaway background command could fill the disk or crash
the process. Apply the same cap to background shell commands and stop
feeding the disk write chain once it trips. Scope the ceiling to
process tasks so subagent and user-question results, which are appended
once and must be persisted, are left untouched.
* feat: hold print-mode turn until background subagents drain
In `kimi -p` (print mode), when the main agent ends a turn while background
subagents (`kind === 'agent'`) are still running, hold the turn open and
idle-wait until they finish, flushing their completions into the turn so the
model can react before the run exits.
Previously, the main agent could end its turn after launching background
subagents; the print flow then drained them with their completion
notifications suppressed, so the main agent never saw the results and the run
exited with the work abandoned (e.g. no nomination). This was the root cause
of the swarm-alpha-mining eval failures.
The hold is gated on a new `drainAgentTasksOnStop` session option (set by the
print flow), only affects `kind === 'agent'` background tasks, and is bounded
by `background.printWaitCeilingS`. Backfill / fan-out is handled by
re-enumerating active tasks. Other background task kinds and non-print modes
are unaffected.
* feat(agent-core): guide the model away from repeating denied or failed tool calls
- system.md: add a diagnose-before-retrying paragraph next to the existing
permission-denial guidance, covering failed tool calls
- permission: when the user rejects an approval on the main agent, tell the
model not to re-attempt the exact same call (sub agents already had an
equivalent hint)
* fix(agent-core): close abandoned tool exchanges and dedupe duplicate tool_use ids
A turn that dies between a recorded tool.call and its paired tool.result
(e.g. a transcript write failure mid-batch) used to leave
pendingToolResultIds open forever: every later message was stranded in
deferredMessages and user input was silently swallowed.
- runOneTurn now defensively closes any dangling tool calls when a turn
ends (completed, cancelled, or failed), synthesizing an error result
that names the cause, with a warn log and a tool_exchange_abandoned
telemetry event
- the projector drops assistant tool calls whose id already appeared
earlier (first occurrence wins): a duplicate id is wire-invalid on
strict providers and not repairable by the strict resend; reported via
the existing projection-repair log and telemetry
- resume-side closePendingToolResults now logs what it closes (warn for
a mid-history gap, info for the routine trailing interruption)
* chore: add changesets for tool exchange fixes
* fix(agent-core): scope duplicate tool_use id dedup to the strict resend
Unconditional dedup regressed providers that emit per-response counter
ids (e.g. call_0 in every step) and accept their own duplicates: later
tool exchanges silently vanished from the projected history, and a
duplicate call's own recorded result was left dangling.
- the dedupe pass is now opt-in via dedupeDuplicateToolCalls and enabled
only in strictMessages, so the normal projection keeps the history the
provider produced
- the pass also drops every tool result after the first for an id, so no
dangling tool message survives; when the kept call has no result of
its own, the surviving one is reattached by the adjacency repair
- kosong now classifies the Anthropic "tool_use ids must be unique" 400
as a recoverable request-structure error so it triggers the strict
resend
* feat(cli): wait for background subagents before exiting kimi -p
When `background.keep_alive_on_exit` is enabled, `kimi -p` now waits for
all background subagents to reach a terminal state before exiting, bounded
by `background.print_wait_ceiling_s` (default 3600s). This lets concurrent
background subagents run to completion in single-turn runs instead of being
torn down when the main agent's turn ends.
* fix(agent-core): route image-compression captions through hidden system reminders
Prompt ingestion (server upload/base64 route, TUI paste, ACP) annotates a
compressed image with an inline <system> caption inside the user's own
message. That raw markup rendered verbatim in every user-visible history
projection (TUI session replay, web UI) and leaked into session titles.
Split the caption out at the appendUserMessage chokepoint and deliver it
through the built-in system-reminder injection (origin
{kind: 'injection', variant: 'image_compression'}), which every UI already
hides. The model still receives the full note; ingestion sites and the wire
protocol are unchanged. Session titles/lastPrompt strip the caption the same
way. Tool-result captions (MCP) keep the established <system> convention.
Covered by unit tests plus an end-to-end smoke suite that drives
rpc.prompt/steer through the real turn pipeline and asserts the provider
wire request, stored history, replay records, and resume parity.
* chore: tighten changeset wording per gen-changesets conventions
* feat(agent-core): guide the model away from repeating denied or failed tool calls
- system.md: add a diagnose-before-retrying paragraph next to the existing
permission-denial guidance, covering failed tool calls
- permission: when the user rejects an approval on the main agent, tell the
model not to re-attempt the exact same call (sub agents already had an
equivalent hint)
* fix(agent-core): close abandoned tool exchanges and dedupe duplicate tool_use ids
A turn that dies between a recorded tool.call and its paired tool.result
(e.g. a transcript write failure mid-batch) used to leave
pendingToolResultIds open forever: every later message was stranded in
deferredMessages and user input was silently swallowed.
- runOneTurn now defensively closes any dangling tool calls when a turn
ends (completed, cancelled, or failed), synthesizing an error result
that names the cause, with a warn log and a tool_exchange_abandoned
telemetry event
- the projector drops assistant tool calls whose id already appeared
earlier (first occurrence wins): a duplicate id is wire-invalid on
strict providers and not repairable by the strict resend; reported via
the existing projection-repair log and telemetry
- resume-side closePendingToolResults now logs what it closes (warn for
a mid-history gap, info for the routine trailing interruption)
* chore: add changesets for tool exchange fixes
* fix(agent-core): scope duplicate tool_use id dedup to the strict resend
Unconditional dedup regressed providers that emit per-response counter
ids (e.g. call_0 in every step) and accept their own duplicates: later
tool exchanges silently vanished from the projected history, and a
duplicate call's own recorded result was left dangling.
- the dedupe pass is now opt-in via dedupeDuplicateToolCalls and enabled
only in strictMessages, so the normal projection keeps the history the
provider produced
- the pass also drops every tool result after the first for an id, so no
dangling tool message survives; when the kept call has no result of
its own, the surviving one is reattached by the adjacency repair
- kosong now classifies the Anthropic "tool_use ids must be unique" 400
as a recoverable request-structure error so it triggers the strict
resend
Compaction runs at the point of maximum context for the task, and the next
turn resumes with less. So the handoff note now records the plan for the
remaining work — upcoming steps, settled decisions, and foreseeable obstacles,
plus any work that can be pre-committed — instead of only the immediate next
command. Update the affected compaction snapshots and one hardcoded
input-token assertion (the instruction is ~163 tokens longer).
* fix(agent-core): recover sessions bricked by orphan tool results
A stray `tool` message with no preceding assistant `tool_calls` permanently
bricked a session on OpenAI-compatible providers: every turn re-sent the same
malformed history and got a 400, and switching model/provider did not help.
Two independent gaps caused this:
- kosong did not recognize the OpenAI / DeepSeek / vLLM / Qwen phrasings of the
tool-exchange structural 400 (`role 'tool' must be a response to a preceding
message with 'tool_calls'` and the mirror `assistant message with 'tool_calls'
must be followed by tool messages`), so the post-400 strict-resend fallback
that drops the orphan never fired.
- The legacy-restore compaction path kept a verbatim tail
`history.slice(compactedCount)`; when the cut landed inside a tool exchange the
tail began with an orphan tool result whose assistant was summarized away. The
normal projection does not repair a leading orphan, so the malformed history
was baked in and re-sent every turn.
Recognize the additional phrasings so the strict resend un-bricks any session,
and trim leading tool results from the legacy-restore tail so the orphan is
never persisted in the first place.
* fix(agent-core): drop orphan tool results at the projection boundary
Rework the legacy-restore half of the fix based on review feedback: mutating
`_history` at restore time desyncs every consumer that models the history from
the wire records — the transcript reducer's fold length would overcount and
make MessageService skip unflushed live-tail messages.
Keep the restored history faithful to the wire records instead, and drop a
`tool` result whose call is nowhere in the history at the projection boundary,
on every request-building projection: the normal wire (`messages`), the
post-400 strict resend (`strictMessages`), and the compaction summarizer. An
orphan is wire-invalid on strict providers and useless to the model either
way, so it never reaches a provider — no longer relying on recognizing the
provider's 400 phrasing to recover. Fragment projections (e.g. token-estimating
a history slice) leave results untouched, since a matching call may
legitimately sit outside the slice.
* feat(agent-core): keep head and tail of user messages during compaction
Compaction used to keep only the most recent 20k tokens of real user
input, so the original task statement was the first thing to vanish in
long sessions. Now, when the user-message pool fits the 20k budget it is
still kept whole; when it overflows, the oldest 2k tokens and the most
recent 18k are kept instead, with an elision marker between the two
segments telling the model what was omitted and that the summary covers
it. The summary prefix and the default system prompt describe the new
shape as well.
The new `keptHeadUserMessageCount` record field keeps restore and the
wire-transcript folded length consistent: records without it (written by
older versions) restore with the original tail-only selection that
produced them, and the vis model-mode projection mirrors the same
head/marker/tail rebuild.
* style(agent-core): drop redundant spread over slice in head selection
* feat(agent-core): announce image compression and keep originals readable
Every image ingestion point (ReadMediaFile, MCP tool results, clipboard
paste, REST upload/inline base64, ACP) now places a <system> caption next
to a compressed image stating the original vs. delivered dimensions, byte
size, and format, so downsampling is never silent to the model.
Originals stay readable: file uploads point at the stored blob, and
in-memory images are persisted into the session's media-originals dir
(content-addressed, size-capped, removed with the session; temp-dir
fallback when no session is known).
ReadMediaFile gains region (crop in original-image pixel coordinates,
delivered at full fidelity) and full_resolution (skip downscaling, with
an explicit error over the per-image byte limit), so the model can zoom
into fine detail instead of silently degrading on large images.
* fix(agent-core): exempt compression captions from the MCP text budget
The caption announcing an image's compression was inserted before the
shared 100K text budget was applied, so a chatty MCP result (page text +
screenshot) consumed the budget first and the caption was evicted — or
sliced mid-string into an unclosed <system> fragment — while the
downsampled image survived, silently reintroducing the exact degradation
the caption exists to report, and orphaning the persisted original.
Split the size-limit pass in two and reorder the pipeline: the text
budget now runs on the tool's own text BEFORE compression inserts
captions (exempt by construction), and the per-part binary cap still
runs after compression so compressible screenshots are kept.
* fix(agent-core): harden crop error reporting and document readback semantics
- cropImageForModel rejects non-finite region coordinates with a clean
message instead of surfacing the codec's internal validation dump
- the full_resolution and cropped-region over-budget errors now include
exact byte counts alongside the rounded sizes, so a file a hair over
budget no longer reads "is 3.8 MB, over the 3.8 MB limit"
- read-media.md notes that re-reading a file without region or
full_resolution reproduces the same downsampled image
* feat(agent-core): align model-facing prompts with actual tool behavior
A hunk-by-hunk accuracy pass over every model-visible prompt surface
(system.md, tool .md descriptions, zod describes, profile role prompts,
and injected reminder strings), with each claim verified against the
implementation and, where possible, empirically (ripgrep semantics).
Fix descriptions that drifted from the code:
- Grep `glob` matches against each file's absolute path, so
`src/**/*.ts` silently matches nothing — document the working forms
- Glob `path` accepts relative paths; results are files-only
- FetchURL no longer promises a content-type-to-mode mapping the
default provider does not honor
- cron: a pinned-date 5-field expression repeats yearly unless
`recurring: false`; drop a bench-only env knob from cron-list
- skill `args` expansion covers $NAME/$1/$ARGUMENTS and the trailing
ARGUMENTS: line; goal reminder no longer cites a nonexistent
developer-message channel
Disclose enforced-but-silent behavior:
- cron fires deliver only while the session is idle; expressions with
no fire within 5 years are rejected at create time
- VCS metadata directories are always excluded from Glob/Grep, even
with include_ignored; sensitive-file guard exemptions
(.env.example/.env.sample/.env.template, public SSH keys)
- large images may be downsampled while the <system> block reports
original dimensions; subagent summaries under the length floor are
sent back for expansion; background-disabled Agent calls are
rejected before launch; AGENTS.md beyond ~32 KB triggers a
performance warning (surfaced in the /init prompt)
Resolve cross-surface contradictions:
- AskUserQuestion background describe/envelope no longer teach polling
- AgentSwarm subagent_type documents that resume keeps original types
- bash.md scopes &&-chaining to dependent commands and steers
independent read-only commands to parallel calls
- the shared system prompt no longer names tools that read-only
subagent profiles lack
Add missing guidance:
- denied/rejected tool calls mean the user declined that action —
adjust, don't retry or route around (root agent)
- plan subagent now knows it is read-only; coder subagent knows its
final message is the entire handoff; explore subagent knows web
tools are in scope
- gh CLI routing for GitHub-hosted work; FetchURL login-wall note;
a dual-use content-safety boundary; scope discipline,
surrounding-idiom, and dependency-verification norms; file:line
citation convention; progress notes on long multi-phase tasks
* fix(agent-core): let the model fetch a background answer after the completion notice
In sessions with background persistence (any agent with a homedir), a
background question's answer is flushed to output.log and the completion
notification carries an <output-file> pointer, not the answer text. The
previous envelope wording ("use TaskOutput only to re-read the answer if
you missed the notification") gated the normal post-completion fetch
behind a missed-notification condition, so a model could acknowledge the
notice and continue without ever reading the user's answer.
Reword the envelope to state that the completion notice may carry a
pointer and to direct the model to read that file (or call TaskOutput
once) for the answer, while still forbidding polling before the user
responds. Align the background param describe the same way ("notified
automatically" rather than "the answer arrives", polling scoped to the
pending window).
* fix
Moonshot / Kimi (OpenAI-compatible) rejects a history whose tool message
references a tool_call_id with no matching tool_calls entry in the preceding
assistant message as `400 tool_call_id is not found`. The
TOOL_EXCHANGE_ADJACENCY_MESSAGE_PATTERNS only covered Anthropic's
tool_use/tool_result phrasing, so isRecoverableRequestStructureError returned
false, the strict-resend fallback in executeLoopStep never fired, and the
session stayed permanently stuck re-sending the same rejected history every
turn (observed in the field after a manual compaction busted the prompt cache
and forced full revalidation of a latently misordered prefix).
Add the tool_call_id-anchored pattern so the whole recovery chain — strict
projection (adjacency repair, orphan-result drop, synthetic results) plus the
one-shot resend — now also covers the default provider. Covered by classifier
unit tests and an e2e resend-and-recover case.
* feat(agent-core): sharpen the compaction handoff prompt
Refine the first-person handoff note the model writes at compaction so it
preserves what actually gets dropped instead of what already survives:
- Lead with the intent of the latest request, not a verbatim re-transcription
(the recent user messages are already kept verbatim beside the summary);
name which request governs when several are in play.
- Carry forward tool results — the concrete values, key lines, schemas — not
just the commands that produced them.
- Keep settled decisions separate from still-open questions, and name the
context the next turn must go and re-check.
- Write in the conversation's language, keep the note proportional to the task,
and don't re-transcribe the auto-attached TODO list.
Also correct the system prompt's description of the post-compaction shape: the
recent user messages come first, followed by a first-person summary (not a
rigidly sectioned report), and a newer kept message supersedes the summary.
Update the affected inline snapshots and the compaction request token count.
* fix(agent-core): preserve an oversized latest request in the handoff note
selectRecentUserMessages truncates a kept user message to the size cap,
keeping only its prefix, so when the latest request itself exceeds the cap
only its head survives verbatim beside the summary. Telling the summary
"don't re-transcribe, it survives verbatim" then permanently dropped the
tail — often the actual ask. Keep the intent-not-transcription guidance,
but require preserving the at-risk parts of a long current request.
* fix(agent-core): cap foreground shell output to prevent OOM crash
A foreground command that streams a very large or unbounded amount of output (e.g. `b3sum --length 18446744073709551615`) grew the live-output buffer until Node aborted with a JavaScript heap out-of-memory error (exit 134). The per-command output is now capped at 16 MiB: on breach the command is gracefully terminated (SIGTERM -> grace -> SIGKILL) and the result carries a message pointing at redirecting large output to a file. The per-task output ring buffer is also made O(1) per chunk (was O(n^2)), which previously starved the event loop and the foreground timeout. Background/detached tasks are exempt.
* fix(agent-core): stop buffering output after the foreground cap trips
After the 16 MiB foreground ceiling tripped, appendOutput still enqueued every subsequent chunk into the per-task disk write chain during the SIGTERM grace window. A producer that ignores SIGTERM could keep that chain — and the chunk strings each pending write retains — growing until SIGKILL, re-introducing the same out-of-memory risk the cap prevents. Once the cap has tripped, keep only the bounded in-memory ring buffer and stop feeding the disk write chain. Interrupt/timeout capture behaviour is unchanged (they do not set outputLimitTripped).
* feat: add KIMI_MODEL_THINKING_EFFORT to force a thinking effort
Send thinking effort only when the model declares it in support_efforts, and add the KIMI_MODEL_THINKING_EFFORT environment variable as an escape hatch to force a specific effort regardless of declared support.
* test: align thinking effort expectations with support_efforts gating
Update the kimi adapter e2e and compaction tests that asserted the previous pass-through behavior on models without support_efforts.
* feat: redesign web ui & add design system
* feat(web): add motion to redesigned UI and add changesets
Animate toast enter/leave, dialog open, and workspace-list and tool-row expand/collapse instead of snapping, and add the changesets covering the web redesign.
* fix(web): remove undo message exit animation
* fix(web): route agent tools to AgentTool and focus dialog on open
- toolRegistry matched the raw 'agent' name, but normalizeToolName folds
agent/subagent into 'task', so agent calls fell through to GenericTool
and lost the inline Open button for the subagent detail panel.
- Dialog's focus watcher only fired on change; callers that mount with
open already true (Login, Settings, ...) never moved focus into the
modal. Run it immediately so initial focus and restore-on-close work.
* feat(web): add logo long-press design-system easter egg
Hold the sidebar logo for 3 seconds to open a dialog showing the design system page. Also trim and rebalance the redesign changesets.
* fix(web): silence Sidebar v-show warning by making it single-root
Nest the design-system Teleport inside the sidebar <aside> so the component has a single element root. App applies v-show to Sidebar, which needs an element root to attach to; the fragment root logged a "non-element root node" warning on every reactive update and the collapse did not take effect.
* fix(web): thinking toggle, tool-group i18n, agent detail button
- Show the default-thinking switch as on whenever thinking is effectively enabled (enabled !== false), matching the core resolver.
- Route the grouped tool-call header and status through vue-i18n.
- Hide the subagent "Open detail" button when no matching task exists (e.g. a completed foreground subagent after a refresh).
* fix(web): use strict equality in agent detail button guard
oxlint eqeqeq flagged the loose != null check; resolveAgentTaskId returns string | undefined, so compare with !== undefined.
* chore(web): remove stray design mockups and screenshots
Remove the design exploration mockups, screenshots, prompts, and notes that were accidentally committed under apps/kimi-web/design. Keep design-system.html, which the sidebar logo easter egg still references.
* fix(web): keep sessions on continuation failure, treat absent thinking as on
- Keep sessions already loaded from earlier pages when a continuation page fetch fails, instead of replacing the workspace with an empty page.
- Treat an absent thinking config as enabled in the settings toggle, matching the core resolver (thinking is on unless explicitly disabled).
* feat(web): open design-system easter egg on 10 logo clicks
Replace the 3-second long-press trigger with 10 consecutive clicks on the Kimi mark; the count resets after a short idle. The long-press was unreliable because pointerleave cancelled the timer on any drift.
* fix(web): treat cancelled swarm members as finished
phaseForTask now lets a terminal status (completed/failed/cancelled) override a stale subagentPhase, so a cancelled swarm member no longer stays live and suppresses the finished AgentSwarm card.
* feat(web): use 1s long-press for design-system easter egg
Switch the logo easter egg back to a long-press, shortened to 1 second, and make it reliable this time: use pointer capture plus touch-action:none so a slight drift no longer cancels the hold.
* fix(web): use plain Spinner for activity notices
ActivityNotice renders for non-chat loading states (e.g. compaction), so it must use Spinner per the design-system rule that reserves MoonSpinner for the chat first-response state.
* docs(web): add a11y guidance to design system
* docs(web): drop stale design README link
* fix(web): restore model search focus and define panel header weight
- Bind the model picker's search Input to searchRef so useDialogFocus moves focus into it on open instead of the dialog's close button.
- Use the defined --weight-semibold token for panel header titles (--weight-bold is not declared, so the shorthand was invalid).
* fix(web): let any open dialog own Escape over the side panel
Track open design-system Dialog instances in a shared count and include it in App.vue's anyOverlayOpen, so a dialog whose open state lives outside App.vue (such as the sidebar session search) captures Escape before the background side panel closes.
* fix(web): base dock-work flag on filtered dock task lists
Foreground subagents are excluded from the dock task lists, so a session whose only task is a foreground subagent no longer renders an empty workbar above the composer.
* fix(web): create subagent task before forwarding text deltas
A client that subscribes from a snapshot after subagent.spawned already fired never received the lifecycle taskCreated; the reducer only applies taskProgress to existing tasks, so assistant text deltas were dropped and the live subagent detail stayed blank. Emit taskCreated (via patchSubagent) before the text progress, mirroring the tool-progress path.
* fix(web): keep plan, swarm, and goal mode toggles per session
Plan, swarm, and goal modes were stored as global scalars on the web client and a single localStorage key each, so they leaked across sessions. Bind them to the active session via per-session maps, persist per session, and apply server status/events to the originating session so background sessions keep independent state.
* fix(web): keep subagent detail reachable for synthesized tasks
When the web client subscribes after a subagent already spawned, the synthesized subagent task has no parentToolCallId, so the Agent tool's Open-detail button was hidden and the panel would not open. Fall back to the single unmapped subagent task when resolving the detail target in both the button visibility and the panel open paths.
* fix(web): keep session kebab menu from being clipped
Teleport the SessionRow kebab menu to body and anchor it with fixed positioning so the collapsing group-sessions list's overflow:hidden no longer clips the dropdown.
* fix(web): apply staged modes to the created session by id
When starting the first prompt, apply the staged plan/swarm/goal modes to the just-created session's per-session maps by id instead of via the activeSessionId-based setters, so a session switch during the selectSession await can't drop the modes for this session or pollute another.
* refactor(web): unify confirmation dialogs into a single modal
Replace the inconsistent confirmation patterns (native confirm(),
two-step menu arming, hand-rolled inline strips, bare buttons) with one
modal ConfirmDialog driven by a global useConfirmDialog() composable,
and consolidate the duplicated confirm/cancel i18n keys.
* feat(web): inline message queue with separate stop button
- Send button always sends/enqueues; interrupt moves to a separate red Stop
button shown only while running, so the two can no longer be confused.
- Queued prompts now render inline at the tail of the transcript (after the
running turn) instead of behind the dock panel: click to edit, remove, drag
the grip to reorder, with image thumbnails and a "next up" marker.
- Remove the dock queue panel and the QueuePane component; Steer stays on
Ctrl/Cmd+S.
* chore: add changeset for web queue UX
* feat(web): prompt reliability, sidebar menu, and composer/markdown polish
- Fix spurious errors when question/approval/task actions were already complete
- Add loading feedback to question and approval prompts; block double-clicks
- Make the question "Other" option selectable by row click and let Enter advance/submit
- Consolidate workspace section actions into an overflow menu
- Tighten markdown prose line-height and block spacing
- Recall input history only when the caret is at text start
* chore(telemetry): track conversation undo, shell mode, and effort changes
- conversation_undo: fires when an undo reverts history (double-Esc selector, /undo, /undo N)
- shell_command: fires when a ! shell command runs in the TUI
- thinking_toggle: now fires on any effort change (not just on/off flips) and carries { enabled, effort, from }; align the no-session TUI path to the same payload
* chore(telemetry): track conversation_undo in core undoHistory
Tracking it in the TUI only covered the TUI; web/acp/REST undos reach
core.rpc.undoHistory directly and were missed. Emit the event from the
agent undoHistory RPC after the undo succeeds so every host is covered,
and drop the now-redundant TUI track.
Addresses review feedback on #1271.
* feat: add model alias overrides
Preserve user model overrides across provider catalog refreshes and resolve effective model metadata for runtime, TUI, protocol, and ACP consumers.
* fix: apply model display name overrides
Show overridden model display names in the footer, welcome panel, status output, and model switch confirmations.
* fix: pass through kimi effort when undeclared
Keep support_efforts authoritative when declared, but pass requested Kimi thinking effort through when the model does not declare support_efforts.
* fix: honor model overrides in effort commands
Use effective model metadata for /effort choices and for always_thinking clamping when resolving thinking effort.
* feat(agent-core): compress oversized images before sending to the model
Downsample images to a 2000px longest-edge and per-image byte budget at the
single prompt-ingestion chokepoint (the prompt/steer RPC) and on tool results
(ReadMediaFile, MCP), so every client transport — CLI, web, desktop, ACP, SDK —
is covered uniformly inside the core. PNG screenshots stay lossless and only
degrade to JPEG when the byte budget cannot otherwise be met. Best-effort: the
original image is sent unchanged if compression fails.
* fix(agent-core): serialize prompt/steer RPCs to avoid a turn-claim race
The prompt/steer RPC handlers await image compression before turn.launch()
synchronously claims the active turn, so two overlapping calls could both
compress first — letting the faster-to-compress one win the turn and strand the
other on agent_busy. Run these two RPCs through a per-agent serialization chain
so they claim in submit order; cancel and the other RPCs stay immediate.
* fix: update flake.nix pnpmDeps hash for the jimp dependency
Adding jimp to the workspace changed pnpm-lock.yaml, so the pnpmDeps
fixed-output hash was stale and the nix build failed. Update it to the value
the CI nix build reported.
* fix(agent-core): guard image compression against decompression bombs
A tiny-byte, huge-dimension image (e.g. a solid 30000x30000 PNG) would be fully
decoded into a multi-gigabyte bitmap by Jimp before any resize — an OOM vector
the byte budget never catches. Skip compression when the sniffed pixel count
exceeds MAX_DECODE_PIXELS (~100 MP), before the decode; oversized images pass
through uncompressed as they did before compression existed.
* fix(agent-core): cap decode byte size before compressing images
Compression runs before downstream size caps (e.g. the 10MB MCP per-part
limit), so a huge or invalid base64 image from an MCP tool was Buffer.from-
decoded — and handed to Jimp — just to be dropped afterward. Add a
MAX_DECODE_BYTES ceiling (64MB, overridable) checked before the base64 decode
and before Jimp, the byte-side complement to the pixel-count guard; oversized
payloads pass through uncompressed.
* refactor(agent-core): compress images at ingestion, not on the turn RPC
Move image compression off the prompt/steer RPC path and back to each ingestion
site (CLI paste, server upload resolution, ACP conversion; ReadMediaFile and MCP
already compressed at their producers). Compressing on the RPC control path put
an async step before the synchronous turn-claim, which spawned a series of
races: prompt/steer interleaving, and — with a cancel arriving mid-compression —
an ineffective abort that let a cancelled prompt launch anyway.
Treating compression as a pure input-stage transform (done while the content
part is built, before it ever enters the agent loop) removes those races
structurally: rpc.prompt/steer are plain synchronous handlers again, and the
serialization/cancel-window machinery is gone. Records stay compressed, resume
stays consistent, and coverage degrades gracefully (a new client that skips
compression just sends a larger image, as before this feature).
* fix: compress inline base64 prompts and honor ACP cancels mid-compression
Two contained ingestion-site follow-ups:
- server: resolvePromptMediaFiles now also compresses images submitted as an
inline `{ kind: 'base64' }` source, not just uploaded files, so the REST
inline-base64 path gets the same downsampling.
- acp-adapter: AcpSession tracks a pending-abort flag while prompt() awaits
image compression (before any turn exists). A session/cancel in that window
flips it, so the prompt returns `cancelled` instead of launching a turn the
client already stopped.
* fix(acp-adapter): cover all concurrent pre-turn prompts on cancel
The pending-abort marker was a single session field, so with two
`session/prompt` requests compressing large inline images at once the later
one overwrote it and a `session/cancel` could mark only one — the other
launched after the client had cancelled. Track a token per in-flight prompt in
a set and flip them all on cancel so every pre-turn prompt is covered.
* chore(node-sdk): declare jimp as a devDependency
The SDK re-exports the image compressor, whose lazy `import('jimp')` (inside
the bundled agent-core code) is inlined into the published dist. jimp was
resolved only transitively via agent-core, so declare it as an explicit build
input here — matching the CLI — to make the bundling reliable rather than
phantom. It stays a devDependency: jimp is bundled, not a runtime dependency.
* fix(provider): honor base_url for google-genai and vertexai providers
The google-genai and vertexai provider types silently ignored a configured
base_url and always hit generativelanguage.googleapis.com (e.g. a Gemini-
compatible proxy URL + key could not be used). Plumb the endpoint through to
the @google/genai SDK via httpOptions.baseUrl:
- kosong: add baseUrl to GoogleGenAIOptions and inject it into the client's
httpOptions alongside the existing headers (the SDK merges headers and
overrides the base host).
- agent-core: forward provider.baseUrl in the google-genai and vertexai
branches, with GOOGLE_GEMINI_BASE_URL / GOOGLE_VERTEX_BASE_URL env
fallback. vertexai keeps deriving location from an aiplatform host.
- docs: document base_url for both providers, noting the host root only
must be given because the SDK appends the API version itself.
Covered by unit tests asserting the URL reaches the kosong config and the
SDK client's httpOptions.
* fix(provider): use the effective base_url for vertex location detection
The vertexai branch forwarded the endpoint from config `base_url` OR the
GOOGLE_VERTEX_BASE_URL env fallback, but service-account detection
(`hasVertexAIServiceEnv` / `vertexAILocation`) still derived the region from
`provider.baseUrl` only. Supplying the regional endpoint via the env fallback
(with a project but no explicit GOOGLE_CLOUD_LOCATION) therefore left location
undefined and silently downgraded Vertex ADC to API-key Gemini routing.
Resolve the effective base URL once and use it for both forwarding and location
derivation, so the env fallback behaves exactly like `base_url`. Add a
changeset for the kosong + agent-core patch release.
* feat(agent-core): slim WebSearch to query-only; fetch page content via FetchURL
The coding search backend no longer honors `limit`/`enable_page_crawling` and
always returns full page content, which was token-heavy and frequently
truncated. Realign the web tools around a search+fetch split:
- WebSearch request sends only `text_query`; the tool exposes only `query`.
- Search results drop inline page content and add the source site; full page
content is now fetched on demand via FetchURL.
- Add a citation reminder to both WebSearch and FetchURL results (in the
FetchURL front note so it survives body truncation).
- Update tool descriptions and reference docs accordingly.
* fix(agent-core): satisfy no-base-to-string lint and drop redundant | undefined
- Assert the exact serialized WebSearch request body instead of String()-coercing
a BodyInit, fixing the type-aware no-base-to-string lint error.
- Drop redundant `| undefined` from WebSearchResult optional fields per AGENTS.md.
- Add changeset.
* chore(changeset): bump agent-core to minor for WebSearch input-contract change
Removing the `limit`/`include_content` tool inputs tightens a closed
(`additionalProperties: false`) schema, so previously-valid args are now
rejected — an incompatible change for a released package. Bump minor rather
than patch.
