13 commits

Author	SHA1	Message	Date
Kai	4c763f6763	feat: send prompt-attached videos directly with the prompt (#1999) ... * feat: send prompt-attached videos directly with the prompt Videos attached to a prompt (pasted in the TUI, uploaded in the web UI) previously reached the model only after it opened the file with ReadMediaFile — an extra tool round trip that could leave the video unseen when the model never made that call. They are now uploaded through the model provider's file channel and embedded directly in the user message as the provider-issued reference; ReadMediaFile stays as the fallback and as the model's own way to open video files. - agent-core: new uploadVideo agent RPC and Session.uploadVideo in the SDK; the TUI uploads pasted videos at submit time, falling back to the file-tag form on failure - kap-server: inline file-source video prompt parts at the REST edge and map provider file ids back to local uploads behind GET /files/llm/{llm_id} - kimi-web: play provider-referenced prompt videos after reload/resume * fix: fall back to inline video when the upload channel fails ReadMediaFile only used the provider's video upload channel when one was bound, and surfaced a hard tool error when the upload itself failed — on providers without a files endpoint (or a transient upload failure) the model was told the video could not be read at all. Now a missing or failing upload falls back to delivering the video inline (base64), the same shape providers without an upload channel already get. Both engines (agent-core and agent-core-v2) are fixed. * test: cover prompt video edge cases and failure paths Stress coverage for the prompt video pipeline: - agent-core uploadVideo RPC: extension/magic classification (.txt with video bytes accepted, extension trusted when magic is absent), exact 100MB boundary, directory and nonexistent paths - TUI: mixed per-video outcomes (one inlined, one tag fallback), submission order behind an in-flight upload, queued video messages carrying final uploaded parts - kap-server: per-video provider-id mappings, Range requests through the llm redirect, mapping persistence across a server restart * fix: surface auth rejections from the video upload channel The base64 fallback for a failing video upload must not mask auth rejections: a 401/403 (surfaced as provider.auth_error) drives the credential force-refresh and a clear auth error, while an inline payload would just be rejected again by the next request. Only a missing or broken upload channel (no files endpoint, network/server errors) falls back to inline delivery. * fix: keep the by-design no-hook video error from degrading to inline Main's contract for a provider with no video upload hook is an honest "does not support video upload" tool error — an inline payload would be dropped on that protocol's wire anyway. The base64 fallback now only applies to an upload channel that exists but failed at runtime. The no-hook throw gets a stable type (VideoUploadUnsupportedError) so the two cases are told apart without matching message text. * fix: constrain the llm video route param to a safe alphabet The provider file id is used as a blob-store key, and the node-fs backend joins scope and key into a storage path. An id containing an encoded path separator (%2F) could address a different storage path than the intended llm-video mapping; the route now only accepts the provider-id alphabet. * style(agent-core-v2): move added explanations into top-of-file comment blocks The package's comment convention keeps comments solely in the top-of-file block; relocate the new notes (url-source id pairing, video delivery fallback, VideoUploadUnsupportedError) from functions and schema fields into their module headers. * fix: reject prompt video uploads when the model lacks video input The uploadVideo RPC only checked the provider's upload channel, so an SDK caller on a text-only or unknown-capability model could obtain a valid-looking video_url part the current model is not supposed to accept. The TUI capability guard does not cover this public path, so the agent now gates on video_in itself. * fix: sniff uploaded video bytes before inlining them The inline branch trusted the upload-time content type; now the bytes are sniffed first, and anything magic-confirmed as a non-video kind (e.g. an image mislabeled as video/mp4) falls back to the file-tag form instead of being uploaded. Like the image gate, bytes are authoritative where the container format allows it — an MPEG-PS lookalike still rides the extension, matching ReadMediaFile. * test: keep the generation stub pending until the abort lands An immediately-answered 404 ends the turn (non-retryable) before the test's abort call, racing the cleanup into a 409; the stub now hangs generation until the abort cancels it. * fix: validate prompt file references before mutating session controls A stale file_id failed inside media resolution — after the model, thinking and permission overrides had already been applied — so a rejected prompt still changed the session's controls. File references are now checked up front, keeping failed submits side-effect free. * fix: serialize prompt submissions per session A slow provider video upload let a later text-only request reach the queue ahead of an earlier video one, silently reordering the conversation for REST clients and multiple tabs. Submissions to the same session now chain, matching the ordering the TUI already guarantees locally. * fix: play reloaded provider videos through the authenticated fetch path A video recovered from an ms:// reference carried only the bare redirect URL, which 401s under daemon auth when loaded natively. The attachment now keeps the provider file id (llmFileId) end to end, and AuthMedia fetches the bytes with the Bearer credential through the daemon's llm redirect — the same blob-URL path uploaded files already use. * fix: fence pending video submits against session and model switches A slow paste-upload left the TUI idle, so /new, the session picker or /model could fire mid-upload; the continuation then dispatched the old session's provider reference into the newly selected session or a model that cannot resolve it. The dispatch now re-checks that the session and model are unchanged and asks for a resend instead. * fix: preserve the caller's video path verbatim in uploadVideo Trimming the path changed the filesystem target before validation and upload, so a name that legitimately starts or ends with whitespace resolved to the wrong file; the trim is now only the emptiness check. * fix: forward provider-issued ids on image URL prompt parts too The shared url-source schema accepts id for image and video parts, but only the video path forwarded it, dropping provider-keyed image ids between prompt acceptance and the model request. * fix(web): reconcile inlined video echoes into the optimistic user message The loose user-message matcher counted media parts and <video path> tags but not the [video:ms://…] text shape, so a racing server echo of an inlined upload slipped through as a duplicate user bubble. * fix: queue bash submits behind a pending video upload A bash-mode submit could start while a pasted video was still uploading, recording shell context before the earlier prompt was dispatched and reordering user actions; it now chains behind the upload like normal submits. * test: cast the driver through unknown for the session-switch fence test * fix: validate prompt file references before resolving the prompt agent A stale file_id posted to a fresh or cold session materialized the main agent (registering it in session metadata and igniting agent-scoped services) before the request was rejected. The file check now runs first, so failed submits create nothing and mutate nothing. * fix: key the playback mapping by the id embedded in the reference URL Projections and clients read the provider id from the ms:// URL, not the id field, so an upload that returns an id-less or mismatched reference would inline fine but 404 on playback. The mapping now derives its key from the URL, falling back to the explicit id. * fix: sanitize provider video ids on the write side of the playback map The read route got the safe-alphabet guard, but recordLlmVideoRef still used the provider-returned id verbatim as a blob-store key; a crafted provider response could write the mapping outside the llm-video namespace. Out-of-alphabet ids are now dropped on both put and get. * fix(web): keep recovered provider videos resendable through the edit path The composer reload path only honored fileId and fell back to a bare fetch(url) without the Bearer token, so editing a reloaded provider-video turn dropped the chip after a 401. llmFileId is now threaded through and the bytes are re-uploaded via the authenticated llm redirect. * fix: fall back to inline video for no-hook providers whose wire carries it The by-design no-hook error is only the honest answer when the wire would drop an inline payload anyway (the OpenAI family). Protocols that convert video_url (kimi, anthropic, google-genai, vertex) now take the base64 fallback instead of failing every video read; the registrar computes the flag from the model's protocol. * test: satisfy the full IBlobStore shape in the playback-map stub The tsgo typecheck job rejects a structural stub missing _serviceBrand and list even though plain tsc accepted it. * fix: queue prompt-producing slash commands behind pending uploads Skill activations and plugin commands started their turn immediately while a pasted video was still uploading, so the earlier video prompt queued behind them and user actions ran out of order. Both paths now chain behind inputSubmitChain (and re-check session/model at dispatch) like normal and bash submits. * fix: validate media kinds in the prompt file-reference preflight The preflight only proved a referenced file exists; a real upload used with the wrong kind (e.g. a PDF submitted as video) still passed it and mutated session controls before assertMediaFile rejected the request. The kind assertion now runs up front with the existence check. * fix(web): play sent and recovered videos in the file preview The media preview returned early for every kind except image, so a user-turn video chip's play action was a no-op even with llmFileId threaded through. The preview now handles video: bytes come from the authenticated file/llm fetch into a blob URL and render in a native player. * fix(web): preview recovered videos with the authenticated blob URL The llm re-upload branch fetched bytes with auth but kept the protected redirect URL as the chip preview, which 401s as a native video src; the fetched blob now becomes the preview URL, mirroring the fileId branch. * style(agent-core-v2): move the inline-fallback note into the module header Same package comment convention as before: rationale lives in the top-of-file block, not beside the registration call. * fix: fence delayed bash submits to the originating session The chained bash callback ran runShellCommandFromInput against whatever session was active at dispatch time, so a command submitted in session A could execute in session B's workspace and be recorded there after a mid-upload switch. The originating session is now captured at submit and re-verified at dispatch, like the prompt and skill paths. * fix: emit prompt video telemetry from the agent scope video_upload is an agent-level event requiring ambient agent identity, but the uploader was built with the Core-scoped session view, leaving prompt-upload events unattributable. The route now resolves telemetry from the target agent for the uploader while image compression keeps the session-scoped view. * fix: preserve provider image ids through legacy projections and web mappers The url-source id accepted by the prompt schema was dropped again by the legacy message projection and the web wire mapper, so provider- keyed image references lost their id across messages, snapshots, and undo responses. It now flows through both directions. * fix(web): revoke recovered video blob URLs before dropping attachments A failed llm re-upload removed the attachment without revoking the freshly created preview blob URL, pinning the whole video in the browser blob store until page unload on every failed edit attempt. * fix: serialize foreground slash commands behind pending uploads /compact and /init started their turn while a pasted video was still uploading, so the earlier message landed after them — and compaction summarized the context without it. Prompt-producing builtin commands now share the same queueBehindPendingUploads chain (with the session/model dispatch fence) as skill, plugin, and bash submits. * fix: defer session controls until media preparation succeeds Media resolution now runs before any profile/model/thinking/permission/ denylist mutation, with the uploader resolved transiently from the requested (or currently bound) model — a failed submission leaves the session's controls untouched. A concurrent model switch during preparation is rejected with session.busy instead of enqueueing a reference uploaded for the previous model. * chore: bump the new SDK video upload API as a minor release Session.uploadVideo is new public API surface, not a patch-level tweak. * fix: re-check busy state before draining upload-queued commands A slash command deferred behind a video upload ran the moment the video prompt dispatched, landing on an already-running turn: beginSessionRequest wiped the active turn's live pane, and /init or /compact started on top of it. The deferred callbacks for skills, plugin commands, /compact and /init now re-run the resolver's busy check at dispatch and show the same blocked message the user would get when typing while streaming. * fix: resolve profile-bound models before choosing the uploader A first prompt carrying "profile" without "model" resolved the upload model from the still-unbound alias, so no uploader was installed and every attached video fell back to a tool-read tag even when the configured default model supports provider upload. The transient resolution now mirrors AgentProfileService.bind: an explicit body model wins, a profile bind falls back to the configured default model, and only otherwise does the currently bound alias apply. * refactor: resolve prompt videos at request time inside the engine Move prompt-video delivery out of the submission edge: the TUI submits synchronously with a local file:// part the v1 turn resolves before the message enters history, and kap-server carries an internal kimi-file:// reference the v2 requester resolves against the effective model with an app-scoped upload cache. History keeps the durable local file id, the /messages projection emits structured video parts, and the web plays videos back through the authenticated /files channel - deleting the submit fences, per-session serialization, provider-id reverse mapping, redirect endpoint, and the unreleased SDK upload API. * fix: propagate abort through video upload delivery instead of degrading A turn cancelled mid-upload used to be treated as an ordinary upload failure: v1 fell back to an inline base64 part and appended the degraded message to history, and the v2 resolver memoized the tag fallback for the rest of the agent's lifetime. Both catch sites now check the delivery signal itself - abort rejections vary in shape by provider - and re-throw so cancellation ends the turn (v1, classified as cancelled via the abort reason) or the request (v2, not memoized, so the next turn uploads). * fix: keep the tag form for no-upload providers whose wire drops inline video An OpenAI-family model configured with video_in but no provider upload channel used to receive prompt videos as an inline base64 part - which chat completions rejects and the Responses adapter degrades to an omitted-video placeholder, persisting ~4/3x the file size in history for bytes the model never sees. The prompt path now mirrors the v2 resolver's protocol gate and degrades to the <video path> tag instead; ReadMediaFile's own delivery is unchanged. Also merges the prompt-video changesets into a single user-facing entry. * fix: escape the NUL separator in the video upload cache key The cache-key template literal contained a literal NUL byte instead of the \0 escape, which made Git classify the whole source file as binary - no inline diffs, unreliable text tooling. The escape produces the byte-identical runtime string, so hashed cache keys are unchanged. * fix: check the abort signal before the inline video fallback The no-uploader inline path (and the post-upload-failure fall-through) never consulted the delivery signal, so cancelling a turn while the video bytes were being read still base64-encoded the file and appended the degraded message to history. The inline branch now re-throws the abort reason first, matching the upload catch. * fix: retry transient prompt video upload failures on later steps A generic upload failure used to memoize its tag fallback for the rest of the agent's lifetime, freezing a transient files-endpoint error into a permanently degraded video. The resolver now marks failure-born fallbacks as non-memoizable: the current request keeps the lightweight tag form and the next step retries the upload. Structural outcomes (successful uploads, capability and sniff fallbacks, no-hook inline) stay memoized for step-retry stability.	2026-07-22 13:46:00 +08:00
liruifengv	83e175399f	feat: auto-background timed-out foreground bash commands (#1591) ... * feat: auto-background timed-out foreground bash commands * fix: discourage blocking TaskOutput waits on background tasks * test: avoid unsafe string conversion in bash timeout test * fix: align bash timeout description with auto-background opt-out * feat(agent-core-v2): auto-background timed-out bash commands - detach timed-out foreground Bash tasks and re-arm the background deadline - align TaskOutput guidance with non-blocking background task handling - add klient SEA end-to-end coverage for the v2 server * fix(klient): clean up lint errors in auto-background e2e example --------- Co-authored-by: haozhe.yang <yanghaozhe@moonshot.ai>	2026-07-13 21:08:21 +08:00
Kai	65d30177ad	feat(agent-core): record llm request trace in wire.jsonl (#1448) ... * 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	2026-07-07 14:09:19 +08:00
liruifengv	e9db9cafcf	feat: record model response id in wire logs (#1349) ... * feat: record model response id in wire logs * chore: include sdk in response id changeset * chore: include agent-core in response id changeset	2026-07-03 17:15:04 +08:00
Kai	78a058acd2	chore(agent-core): remove experimental micro compaction (#1317) ... * chore(agent-core): remove experimental micro compaction * fix(docs): drop micro compaction row from env-vars table	2026-07-02 19:50:51 +08:00
Kai	86e0c9201e	feat(agent-core): rework compaction to keep only user prompts and summary (#1214) ... * feat(agent-core): rework compaction to keep only user prompts and summary * refactor(agent-core): rewrite compaction summary as first-person handoff Rework the full-compaction summary to read as the agent's own continuing notes instead of a third-party report: - compaction-instruction.md: free-form first-person continuation that preserves exact commands, paths and outcomes, states the precise next action, and flags claimed-but-unverified work rather than trusting it. - compaction-summary-prefix.md: skeptical "your own working notes" framing; drop the collaborative third-party prefix. - system.md: add compaction-awareness guidance so the model continues naturally from a summary and re-checks any reported "done". - Rename the compaction helpers module to handoff.ts. Update tests and regenerate snapshots for the new prompt text, and fill in contextSummary in the restored-compaction replay expectations. * fix(agent-core): count image/audio/video parts in token estimation estimateTokensForContentPart returned 0 for image_url/audio_url/video_url, so auto-compaction triggers, the overflow-shrink budget, the kept-user budget, and the reported context size all went blind to media — a media-heavy session could overflow the model window while the estimate reported a near-empty context. Media parts now carry a fixed estimate (MEDIA_TOKEN_ESTIMATE), and the content-part switch is exhaustive so a new ContentPart kind must declare its estimate rather than silently count as zero. * feat(agent-core): re-surface active background tasks after compaction Folding the live context to [recent user prompts, summary] drops the messages that started background tasks and their status updates, so the model could forget a task is still running and spawn a duplicate. injectAfterCompaction now appends a system-reminder listing active background tasks (with guidance to use TaskOutput/TaskList/TaskStop instead of re-spawning). It runs only post-compaction and carries an injection origin, so the next compaction drops and rebuilds it rather than stacking copies; the all-user-role post-compaction shape is preserved (no tool-pairing reintroduced). * test(agent-core): add compaction scenario guards and risk probes Adds compaction-scenarios.test.ts driving the real Agent/ContextMemory/ FullCompaction machinery: - A guard test locking in that repeated compaction folds the prior summary into the new one instead of stacking two summaries. - Seven `it.fails` probes that executably reproduce known, currently-accepted edge-case defects so the suite stays green while documenting each one precisely; any of them will flip red (forcing removal of `.fails`) the day the behavior is fixed. They cover: assistant/tool appended during an in-flight summarizer call being dropped; unbounded shrink on empty summaries; the fixed 20k kept-user budget overflowing a small model window; a tool result orphaned when compaction starts mid-exchange; legacy compaction records dropping their verbatim tail on replay; micro-compaction clearing recent tool results in an overflow-shrunk suffix; and media being discarded when the oldest kept user message is truncated. * fix(agent-core): repair tool_use/tool_result adjacency in projected context A tool call and its result can end up non-adjacent in history — a background-task notification or flushed steer lands between them, or an interrupted/nested step delays the result — which strict providers reject with HTTP 400. The projector now moves each tool_use's result up to immediately follow it (projection-time only; the stored history is untouched), and full compaction projects its summarizer input with a synthetic result for any still-open call so the summary request stays well-formed. Micro-compaction only surfaced this latent ordering by busting the prompt cache, so it now defaults off. Includes projector adjacency regression tests, a context-level integration test, and a compaction synthesize-missing guard; the prior "keeps an unresolved tool exchange out of the compaction prompt" test is updated to the now-well-formed (synthetic-result) behavior. * fix(agent-core): preserve the verbatim tail when restoring legacy compactions A pre-rework `context.apply_compaction` record used `[summary, ...history.slice(compactedCount)]` semantics and kept a verbatim recent tail, but it has no `keptUserMessageCount`. The reworked applyCompaction re-folded such records into the all-user shape, dropping the recent assistant/tool tail — so resuming a session compacted by an older version silently lost its most recent context. On restore of such a record (gated on records.restoring, no keptUserMessageCount, and compactedCount < history length) reproduce the old shape instead. The forward/live path is unchanged; the projector's tool-adjacency repair keeps the restored tail well-formed, and compaction only runs at clean step boundaries so the tail has no open exchange. The legacy-tail probe now passes as a regression guard via the real restore path. * fix(agent-core): align legacy compaction foldedLength with live restore The transcript reducer re-derived foldedLength for pre-rework context.apply_compaction records (no keptUserMessageCount) using the new kept-user+summary rule, but ContextMemory's restore now reproduces the legacy [summary, ...history.slice(compactedCount)] shape for those records. The two diverged for legacy sessions, so MessageService's foldedLength-vs-live-history comparison could mis-handle GET /messages (miss or misorder recent output). The reducer now mirrors the live legacy fold: when compactedCount is below the pre-compaction length it computes 1 + (length - compactedCount); otherwise it falls back to the kept-user derivation. The MessageService transcript test's fixture is corrected to a new-format record, matching its all-user live mock. * fix(kosong): merge a follow-up user turn into the preceding tool_results The Anthropic message merge keyed on isToolResultOnly(last) === isToolResultOnly(converted), which left a tool_result-only user turn followed by a plain-text user turn unmerged. After tool-exchange repair this shape (assistant tool_use -> tool_result -> injected notification) produces two adjacent user messages, which strict Anthropic-compatible backends reject with HTTP 400. Switch to the asymmetric predicate isToolResultOnly(last) || !isToolResultOnly(converted): a tool-result-only running message absorbs whatever user turn follows (parallel tool_results or a trailing text), yielding a valid [tool_result, ..., text] message; a plain-text running message still only absorbs plain text. [tool_result, text] is valid for both native Anthropic (which concatenates anyway) and strict backends. * test(agent-core): pin micro-compaction flag in the shrunk-suffix probe The 'does not clear recent tool results when projecting a shrunk suffix' probe is an it.fails that only documents a real defect while micro-compaction is active. It inherited the ambient KIMI_CODE_EXPERIMENTAL master switch, so its pass/fail flipped with the runner: green locally (master switch on) but a hard failure in CI, where the flag defaults off and MicroCompaction.compact() is a no-op that leaves the tool result intact. Enable KIMI_CODE_EXPERIMENTAL_MICRO_COMPACTION explicitly for this probe so it deterministically exercises the micro-compaction path regardless of the environment. * fix(agent-core): harden full compaction against in-flight races, unbounded shrink, and media loss Three compaction-path fixes surfaced by review, each flipping its documenting it.fails probe to a passing it: - Append race (CMP-02): after the summarizer returns, the post-summary history check only compared the compacted prefix. A live step appending to the tail while a manual/SDK compaction was in flight slipped through — an appended assistant/tool turn is neither summarized (the summary covers only the snapshot) nor kept (the rebuild keeps user input), so it vanished. Now cancel when the appended tail contains a non-user message; an appended user message is still kept (rebuild picks it up), preserving the existing 'keeps messages appended while compacting an unchanged prefix' behavior. - Unbounded empty/truncated shrink: an empty or truncated summary dropped the oldest message and reset retryCount, so a model that kept returning empty could issue ~one request per history entry. Bound the shrink attempts by MAX_COMPACTION_RETRY_ATTEMPTS, mirroring the overflow-shrink counter. - Media dropped on truncation (CMP-07): truncating the oldest kept user message replaced its whole content with one text block, discarding any image/audio/video. Keep the non-text parts and spend the remaining budget (maxTokens minus their cost) on truncated text. * fix(vis): mirror legacy compaction tail in the model-mode projector For a pre-rework context.apply_compaction record (no keptUserMessageCount), agent-core's ContextMemory restore and the transcript reducer keep the old [summary, ...history.slice(compactedCount)] tail — a verbatim recent tail including assistant/tool. The vis model-mode projector always applied the new kept-user selection, so opening an older compacted session in model mode hid the assistant/tool tail the resumed agent still holds (and surfaced a pre-compaction user message the agent dropped). Branch on a missing keptUserMessageCount with compactedCount < history length and reproduce the legacy shape, matching the agent-core restore. * fix(agent-core): cancel compaction on any droppable user-role tail The in-flight append guard cancelled only when the tail grew with a non-user role. A user-role message that compaction would still drop — a background-task notification, hook/cron reminder, or shell-command output — slipped through: appended after the summary snapshot (so absent from the summary) and dropped by the all-user rebuild (which keeps only real user input), vanishing silently. Key the guard on the same predicate applyCompaction uses (!isRealUserInput) so it cancels whenever the appended tail holds anything compaction would drop. A real user message is still kept, so a live user turn racing a manual/SDK compaction continues to complete. * fix(agent-core): exclude pre-clear prompts from legacy folded length The transcript reducer's legacy fallback (records predating keptUserMessageCount, compacted with no verbatim tail) re-derived the kept-user count from the whole transcript, including messages before the last context.clear. Live ContextMemory rebuilds _history from post-clear messages only, so counting pre-clear prompts overstated foldedLength; MessageService then saw context.history.length <= foldedLength and skipped appending unflushed live tail messages, dropping recent output from the messages endpoint for old sessions compacted after a clear. Derive only from entries at or after clearFloor to match the live context. * fix(agent-core): drop media when truncating the oldest kept prompt Revert the media-preserving truncation: keeping non-text parts on the truncated boundary message overshot the kept-user budget when the media alone exceeded it, and reordered interleaved text/media parts. Both codex (no media-aware truncation) and Claude Code (strips media at compaction) decline to preserve media on a truncated message, since media cannot be partially truncated and keeping it whole breaks the budget. truncateUserMessage now keeps only the truncated text. Recent messages that fit the budget are still kept verbatim with their media; only the oldest, partially-overflowing boundary message loses its attachments. * fix(agent-core): make manual compaction and turns mutually exclusive A manual/SDK compaction could start while a turn was streaming, or a new turn could launch while a compaction was in flight. Either way the turn mutates the shared context (streaming content into an existing assistant message, or appending new messages) during the summarizer await, and that output is neither summarized nor preserved by the all-user rebuild — silent loss that object-identity checks can't detect (the streamed message is mutated in place). Guard both directions so the agent does one of {turn, compaction} at a time: begin() refuses a manual compaction while a turn is active, and launch() refuses a new turn while a compaction is in progress. Auto compaction is exempt — it runs from within the turn at a step boundary, which blocks the turn for its duration. * chore(changeset): consolidate compaction changesets into one * chore(agent-core): drop external-product references from compaction comments * test(agent-core): add Anthropic wire-compliance smoke tests for compaction Drive real compaction output and the compaction summarizer projection through the real Anthropic provider conversion and assert the wire request is well-formed: strict user/assistant alternation and every tool_use answered by an adjacent tool_result. Locks in the cross-layer guarantee (projector merge + Anthropic consecutive-user merge + adjacency repair + synthesizeMissing) that compacted sessions stay valid for strict Anthropic-compatible backends. * fix(agent-core): defer and replay inputs during manual compaction instead of rejecting Manual/SDK compaction runs outside a turn, so the earlier guard rejected prompts/steers that arrived while it held the context. That broke three things: a REST/web prompt got stuck 'running' (no terminal turn event), a background-task/cron steer was silently lost (null was read as 'buffered' but nothing was), and a follow-up prompt could land in the window after isCompacting cleared but before reminders were reinjected. Reuse the existing defer-and-replay model instead of rejecting: - steer() and launch() buffer into steerBuffer while a compaction is in progress (returning null = buffered), mirroring how an active turn defers input. - FullCompaction.compactionWorker keeps isCompacting true through refreshSystemPrompt + injectAfterCompaction (moving markCompleted and the completed event after reinjection), then replays the buffer via TurnFlow.onCompactionFinished — on success, on an A1 prefix/tail cancel, and on failure/abort. - onCompactionFinished flushes into an active turn if one exists, else launches a fresh turn from the deferred input. No PromptService change: a deferred prompt's eventual turn.started lets it associate the pending prompt and clear it on turn.ended. * fix(kosong): merge consecutive user turns for strict providers Gemini/Vertex require strictly alternating user/model turns and reject consecutive user turns with HTTP 400. They arise after compaction (kept prompts + user-role summary + injected reminders) and when a turn is steered in right after a tool result. Anthropic already merged them inline; the Google converter did not, so post-compaction requests failed. Extract the asymmetric merge into a shared mergeConsecutiveUserMessages helper applied at each strict provider's conversion boundary: refactor Anthropic to use it (behavior unchanged) and apply it at the Google converter's exit. A conformance suite drives every strict provider with the post-compaction shape and a steer-after-tool-result shape, asserting no consecutive same-role turns reach the wire, so a new strict provider cannot silently omit the merge. The provider-agnostic projector stays structure-preserving: lenient providers (OpenAI/Kimi) keep distinct turns for clearer message boundaries; only strict providers normalize, where the requirement lives.	2026-07-01 01:17:30 +08:00
_Kerman	dcf30754d0	feat: stream shell tool output (#676)	2026-06-12 11:10:20 +08:00
liruifengv	d7407b0ecf	feat: release experimental features (#569) ... * feat: release experimental features * refactor: remove redundant goal runtime gate * refactor: remove unused skill flag plumbing * feat: keep micro compaction opt-out	2026-06-09 11:42:50 +08:00
_Kerman	72c4b0adaa	feat: agent swarm (#424)	2026-06-08 14:26:56 +08:00
liruifengv	86a42a26a1	feat: Add persistent experimental feature toggles and a TUI panel (#420) ... * feat: Add persistent experimental feature toggles and a TUI panel * test(agent-core): isolate experimental flag tests * fix(tui): block experiments panel while busy * refactor(experimental): merge feature query APIs	2026-06-04 18:35:41 +08:00
_Kerman	2b74025302	feat: rework permission decision policies (#26)	2026-05-27 20:07:24 +08:00
_Kerman	c4dd1c7ff2	feat: flatten tool call records (#25)	2026-05-25 23:02:25 +08:00
Kaiyi	842e699a64	Kimi For Coding	2026-05-22 15:54:50 +08:00