Repository navigation
v0.16.0
Added
-
define_toolgives a tool'sexecutethe type that its schema validates to. AToolwhoseexecutewanted{ city: string }couldn't go intotools: [...], because aTool'sexecutetakesunknown, so every tool tookunknownand parsed its input a second time, though the engine had already validated it againstinput_schema.define_tool({ name, description, input_schema, execute })reads the type frominput_schema, hands it toexecuteand toneeds_approval, and returns a plainToolthat anytoolslist takes with no cast. The engine passes a tool only the input that passed its schema, and a call that fails goes back to the model as a tool error, so the type thatexecutesees is the type that it gets. A tool written as a plainToolkeeps working, and moving one over drops itsparsecall. The cookbook andexamples/tool-loopnow use it. -
checkpointtakeson_write_error, andgatekeeps a failed write fatal.on_write_error: 'record', the default, recordswrite_errorand returns the inner result, which is what a failed write now does (see Changed).'fail'records the same event and then throws the store's error.gatepasses'fail', so a gate whose store fails the write still fails the run before it suspends, as it always has, and never asks for approval of a result that its store didn't keep. A flow that builds the same thing fromcheckpointandsuspendpasses'fail'too, and the docs on checkpointing and human-in-the-loop now say so.flow_schema'scheckpointtakes the field, anddescribeshows a checkpoint the same way whichever value it has. -
BedrockCredentialProviderandBedrockCredentialsimport fromfascicle. Thebedrockprovider'scredential_provideroption takes aBedrockCredentialProvider, a function that resolves toBedrockCredentials, but neither type was exported, so code that wrote its own provider had to restate both shapes. Both import fromfasciclenow.
Changed
-
The
zodpeer takes any 4.x from 4.2.0, and the package no longer asks for a pnpm version. The peer was exactly4.4.3, so a project on zod 4.4.4 or later got a peer conflict, though every release from 4.2.0 on has the Standard JSON Schema interface that Fascicle reads, and 4.2.0 emits the same schemas that 4.4.3 does. The peer is now^4.2.0, and the repo still develops and tests against 4.4.3.engines.pnpmis gone as well. It shipped to every consumer, where it could only warn a pnpm user below 10.13 about a floor that matters only to someone working in this repo, and the floor was stale even there, since the workspace needs pnpm 11, whichpackageManageralready pins. A project that held zod at 4.4.3 to satisfy the peer can move it. -
turn_timeout_erroris exported, and the engine's types promise only what it does. A turn that outrunsturn_timeout_msbefore any chunk streams throwsturn_timeout_error, which reaches the caller wheneverretry_onleaves out'timeout', but the class wasn't exported, so code couldn't catch it byinstanceof. It now imports fromfascicle. Three things that nothing implemented are gone.ClaudeCliProviderConfiglosesskip_probe, which no code ever read.ClaudeCliErrorReasonlosesauth_missing,auth_expired,api_key_missing,engine_disposed, andparse_error, since the adapter reports an auth failure asprovider_auth_error, a missing key asengine_config_error, a disposed engine as anaborted_errorwhosereasonis'engine_disposed', and a line it can't parse as a trajectory event.FinishReasonloses'aborted', which no call finished with, since an abort throws. Code that setskip_probedrops it, and code that matched one of the removed reasons or'aborted'matches the error that the adapter throws instead.docs/cli.mdnow lists the reasons and the dispose errors as they are. -
flow_schematakes every composer and every field that the composer configs take. The JSON Schema for the YAML flow shape had drifted from the API. It had no key forgate,ensemble_step,improve, orlearn, so a YAML flow couldn't use one, and now it has a key for each of the four, with a key for every field that its config takes. Agate'sstoreis written by its name,improve'sbudgetis an object, andlearn'ssourceis an object whosekindpicksevents,paths, ordir. The other composers now have a key for each field that their configs take, soretrygainsmax_delay_ms,jitter,when,on_error, andproject,fallbackgainswhen,handoff, andproject, astepgains its metadata andarm,adversarial,ensemble,tournament, andconsensusgainproject,loop'sguardcan name a predicate, and every config that takes anameor adescriptionhas one. Achainentry takes the arm form (step,arm, andselect), and ascopetakes any node between itsstashanduseentries. Two changes reject what used to validate.looprequiresinitandfinish, asloopitself does, and auseentry names its function underfn, where it took a node underdo, thoughuseruns a function and not a step. A YAML flow that wroteusewithdowritesfnand the function's name instead. A test now pins every key against its composer's config, so a field that a config gains fails the gate until the schema gains it too. -
STEP_KINDSlists only the kinds that a step can carry. It namedadversarial,ensemble,tournament, andconsensusas step kinds, but those four factories build acomposestep that carries the composite's name as its display name, so no step ever had one of those kinds. The four names are gone fromSTEP_KINDSand from theStepKindtype, andis_step_kindanswers false for them. A hand-built step that uses one of them as its kind now reads as a kind of its own, sodescribe.diagramlabels it by its id. Code that looked for a composite by itskindlooks forkind === 'compose'and the composite's name inconfig.display_nameinstead, which is what those steps have carried all along. -
The viewer answers only its own host, and the CLI takes posts only under
--listen. Bound to a loopback name, as it is by default, the server answers 403 to a request whoseHostnames anything butlocalhost,127.0.0.1, or::1, whatever the port, so a web page that rebinds a domain of its own to 127.0.0.1 can't read your trajectories through your browser./api/ingestreads onlyapplication/x-ndjson, answering 415 to any other content type, which a page on another origin can't send without a preflight, and it answers 413 to a body over 8 MB.fascicle-viewer <path>no longer mounts/api/ingestat all, though only--listenever said it would. To keep posting to a viewer that tails a file, pass--listenbeside the path. Open the viewer by one of those three names rather than by another hostname that reaches the machine, and sendapplication/x-ndjsonfrom a client of your own, ashttp_loggeralready does.start_viewerandstart_servergainingest, which defaults to true, so code that embeds the viewer meets nothing new but theHostcheck. -
map,loop, andensemblereject a config that can't run when they're built. Aconcurrencyor amax_roundsofNaNgot through every clamp, sinceMath.max(1, NaN)isNaN, somapstarted no workers and returned an array of holes, andloopran none of its rounds. Both now throw aTypeErrorwhen they're built.mapalso rounds a fractionalconcurrencydown, so2.5, which used to run three items at a time, runs two.ensemblewith no members threwno members produced a resultonce it ran, and now throws when it's built, astournamentdoes. Code that computes aconcurrencyor amax_roundschecks the number before it builds the step, and code that builds anensemblefrom a list that can be empty checks the list. -
suspendandgatehold their ids to the step id rule. Resume data reaches a gate under its id, butsuspendandgatetook any id, thoughstepandchainrefuse an empty id and one that isn't an identifier. Both now throw aTypeErrorwhen they're built, which names thesuspend idor thegate id, suggests an identifier, and says to put the label inname. A gate checks its id before itssuspenddoes, so the error names the gate. Every id in the docs and examples already fit. A flow whose id reads likeapprove-draftorapprove draftrenames it, say toapprove_draft, and moves the prose intoname. A durable run that's waiting at a gate can't be resumed once that gate's id is renamed, because it waits under the old id, and the rename changes the flow's shape, whichdurablerefuses by default withflow_changed_error. Let those runs finish, or delete them, before you deploy the rename. Repeated suspend ids still pass, anddocs/roadmap.mdlists that under Known Issues. -
checkpointkeeps the inner step's result when its store fails the write. Astore.getthat threw counted as a miss, but astore.setthat threw failed the run, after the inner step had run and been paid for, though a checkpoint is a cache and not the result. A failed write now records acheckpointevent withstatus: 'write_error', the error's message, the stepid, thekey, and thespan_id, as a failed read recordsread_error, and the step returns the inner result. The next run with that key misses and runs the inner step again. A run that used to fail on a broken store now completes, so code that counted on that failure passeson_write_error: 'fail'to keep it fatal, or looks forwrite_errorin the trajectory.gatepasses'fail'itself, so a gate still fails the run.CheckpointStatusgains'write_error'andis_checkpoint_eventaccepts it, so code that switches overCheckpointStatusneeds a case for it. The viewer shows no tick for it, as for a miss. -
CaseResultandBenchReportlose their input type parameter.CaseResulthad an_input?: Ifield thatbenchnever set, and the field existed only to give the input type a use, soBenchReport<In, Out>promised an input that no report holds._inputis gone, and both types now take<O = unknown, S = Score>.benchkeeps its input type for the flow, the cases, and the judges, and it returnsBenchReport<O, S>. Drop the first type argument, soBenchReport<In, Out>becomesBenchReport<Out>, andCaseResult<In, Out, S>becomesCaseResult<Out, S>. Don't leave a two-argumentBenchReport<In, Out>as it is: it still compiles, but it now reads as outputInand scoreOut, since the score type takes any type, so its cases'outputgets the wrong type.examples/swebenchdrops the argument too. -
ensemble,ensemble_step, andtournamentlet anundefinedresult compete, and a NaN score never wins. The three read a member's result as missing whenever it wasundefined, soensembleskipped that member without scoring it,ensemble_stepthrewwinner missing from resultswhen it won, andtournamentthrewmissing result for matchwhen it played. Each now reads its results by key, so anundefinedresult is scored, can win, and plays its matches, and anensemble_stepscore that is itselfundefinedcan win too. A NaN score inensemble, or a NaN rank inensemble_step, used to win whenever it came first, and now it never wins under eitherselect, wherever it falls, though it stays in the score map. A field whose every score is NaN throws anErrorthat says so and asks for a scorer that returns a number. A flow whose scorer can return NaN for every member catches that error or makes the scorer return a number, and a flow that counted on anundefinedresult being skipped, or on its throw, filters it out in the member or the scorer. -
An
acp_agentapp refuses a second connection. An app keeps one set of sessions and one client's capabilities, so a second connection to the same app overwrote the first client's capabilities and could reach its sessions by id, thoughdocs/acp.mdsaid that an app serves one. A secondconnectnow throws an error that says the app already serves a connection and to build an app for each one, and a secondinitializegets an invalid-request error back before it can touch the first client's capabilities. That second guard covers a server that shares one app among its clients, such as the SDK'sAcpServer, which runsinitializebefore any connect hook.serve_acpbuilds an app for its one connection and meets nothing new. A transport of your own that connected one app more than once builds one for each connection, and a client that sentinitializetwice on one connection sends it once.
Fixed
-
The viewer's scrubber and clock follow the newest run. In a trajectory file that holds more than one run, the canvas drew only the newest, but the timeline started with the first run. The scrubber and the header's T+ disagreed, an older run's failures were marked, and a scrub could land in a run that the canvas no longer drew. The timeline now starts at the newest run's
flow_structure, whether the viewer reads the file whole or follows it live, and a run that starts while you watch starts the timeline over and lets go of a scrub that was held in the old run. -
The viewer follows a trajectory file that appears late or gets rotated. The tail watched the file itself, so
fascicle-viewer <path>never read a file that didn't exist yet, though it said it would pick the file up once it was created. After a rotation that renamed the file and made a new one, it read the new file from the old one's offset. The tail now watches the directory while the file is missing or after a rename, and it reads a new file at that path from its first line. A missing file is something the tail waits for now, sostart_tailno longer reports it throughon_io_error, though a missing directory still reports there. -
The viewer's canvas keeps its elements from one event to the next. Every event rebuilt every node, segment, and label on the stage, and each rebuild restarted the element's animation, so the live march, the halo's pulse, and the bloom on any node that had ever emitted started over with every event. Each element now lasts as long as the glyph that it draws, so an animation runs through, and a node blooms once per emit.
-
DEFAULT_PRICINGmatches the current list prices. The table had no rows for the Claude models after Opus 4.8, none for OpenAI's aftergpt-4o, and Gemini 2.5 rows that had fallen behind Google's prices, with 2.5 Pro's output at half of its list price and 2.5 Flash's input at a quarter. Every row was checked against its provider's own pricing page on 2026-10-18, and a comment names the page. New rows price Claude Fable 5.1, Fable 5, Opus 5.5, Opus 5, Sonnet 5.5, Sonnet 5, and Haiku 5.5, OpenAI'sgpt-6-astra,gpt-6.1-sol, andgpt-6-luna, and Gemini 4 Argon, 3.8 Flash, and 3.5 Flash-Lite, and the Gemini rows gain a cached-input rate. A model whose price rises with the length of its prompt is priced at its base tier, and Gemini 4 Argon and 3.8 Flash are priced at the rates that hold until they double, on 2026-12-01 and 2027-01-01. A cost estimate for a Gemini 2.5 call now comes out higher, and a call to one of the new models now gets a cost where it used to recordpricing_missing. Code that registered its own price for one of these models keeps it, since an engine'spricingwins over the defaults. -
openainative sendsxhighandmaxas they are. The openai dialect clamped both tohighonreasoning_effort, from a time when OpenAI's Chat Completions field took onlylow,medium, andhigh. OpenAI's reference now lists every level fromnonethroughmaxfor it, and theai_sdktransport already passed both through, so the two transports sent different efforts for the same call. Effortxhighandmaxnow reach OpenAI unchanged ontransport: 'native'. Not every model takes every level, so a call that relied on the clamp to sendmaxto a model that tops out athighpasseshighitself.lmstudionative keeps the clamp, since a local server takes only the three. -
claude_clicounts its cached input ininput_tokens. The CLI reportsinput_tokenswithout the tokens that it read from or wrote to the cache, as the Messages API does, and the adapter copied that number as it came, thoughUsageTotals.input_tokenscounts the cache on every other path. Cost works out the fresh input by taking the cached parts back out ofinput_tokens, so a cached call's reported cost gave fresh input too small a share, or none, and the cache too large a one.input_tokensnow adds the cache reads and writes, so a result with 100 input, 40 read from the cache, and 20 written to it reports 160, withcached_input_tokensat 40 andcache_write_tokensat 20. The total cost doesn't change, only how it splits. Code that added the two cache fields to aclaude_clicall'sinput_tokensreadsinput_tokensalone. -
openrouternative counts the cache writes that OpenRouter reports. On a model with explicit caching, such as Claude, OpenRouter reports the tokens that a request wrote to the cache asprompt_tokens_details.cache_write_tokens, and the native transport read only thecached_tokensbeside them, so a call that wrote the cache showed no writes.cache_write_tokensnow carries them, as it already did ontransport: 'ai_sdk'.input_tokensstays as it was, sinceprompt_tokensalready counts the writes, as it counts the reads. A request to Claude Sonnet 5.5 that wrote 16,299 tokens to the cache reported 16,358 prompt tokens, and the same request sent again reported the same 16,358, with 16,299 of them read from the cache. The charge that OpenRouter reports doesn't change, only how it splits, so the writes now take their share ascache_write_usd, where they used to count as fresh input ininput_usd. Where OpenRouter reports no charge, a pricing row now prices the writes at its cache-write rate, so that estimate comes out higher. Code that adds up a call's input cost frominput_usdandcached_input_usdaddscache_write_usdtoo. -
filesystem_loggerdrops an event it can't write instead of failing the run. An event that JSON can't serialize, such as one that holds aBigInt, or an append that fails, such as on a full disk, threw out of the logger and into the flow, so a trajectory file could fail the run that it was recording, whilestderr_loggerandhttp_loggernever throw.filesystem_loggernow takes an optionalon_error, ashttp_loggerdoes, and hands it each failure. Without one, it writes a singlefilesystem_logger:line to stderr per failure, naming the file and the reason. Either way the event is dropped and the run goes on.docs/api-reference.mdnow states the rule that every adapter logger keeps. Code that caught that throw aroundrunpasseson_errorinstead. -
judge_llmreads a score that prose follows. The parser that pulls the JSON out of a judge's reply took the whole reply whenever it began with{, so{"score":0.8} That's my evaluation.failed to parse and the judge abstained, though the same object with prose before it parsed. It now takes the first balanced object in the reply and skips any brace that sits inside a string, so prose after the object, prose on both sides, a fenced block, and a}in the reason all parse. A reply with no object still abstains, and so does one whose first{never closes. A bench whose judge answered with JSON and then a sentence now scores those cases, where it used to drop them from that judge's mean. -
A durable gate gets the first datum that reaches it, even when two resumes race.
resumeread a gate's inbox and wrote its datum only when the inbox was empty, but two resumes that read before either wrote both found it empty, so the last write won, and a run could resume on the second webhook's data rather than the first's. Each resume now writes under a claim on the gate's inbox key, and the resume that loses the claim drops its datum, as it would on finding one already there. The claim lasts the driver'slease_msand comes off right after the write, so a resume that dies between its claim and its write holds that gate's data back until the claim lapses, and data sent for the gate in that window is dropped. A resume in that window getssuspendedback at the gate it was meant to move, and the same call sent again once the claim lapses lands as usual. -
A late durable
resumedrops its datum once the run has passed the gate. A resume that read the run just before another resume moved it past a gate still wrote its datum into that gate's inbox, where nothing read it until the run was deleted. If the run was deleted in that window and a new run started under the same id, the new run took the stale datum up at that gate. Under its claim on the gate's inbox key, each resume now reads the inbox and then the run's record, and it drops the datum when the run has already passed the gate or is gone, which is whatdocs/human-in-the-loop.mdalready said it did. A resume whose run was deleted under it still rejects withrun_not_found_error. Two narrower windows stay open until the store contract gains a compare-and-set: a resume that stalls past its claim between those reads and its write, and adeletethat lands between the record read and the write. -
A SIGINT or SIGTERM aborts only the runs that installed the signal handlers. Every run joined the set that a signal aborts, whatever it passed, so once any run had installed the handlers, a Ctrl-C also aborted the runs started with
install_signal_handlers: false, though opting out says the host forwards cancellation itself. A run that opts out now keeps going through a signal, and the handlers come off once the last run that installed them settles. A host that counted on another run's handlers to stop its own runs aborts them throughabort, whichfascicle/acpandfascicle/mcpalready do. -
A
model_callordefine_agentstep that starts after an abort throws the reason the run was aborted with. Both steps check the signal before they call the engine, and that check threw anaborted_errorof its own, which dropped the caller's reason.model_callalso made up a{ signal: 'abort' }reason and a tool-loopstep_indexof 0. Both checks now keep the rule that the rest of core keeps, so anErrorthat the run was aborted with comes back as that same instance, and any other reason arrives as thereasonof anaborted_errorthat has nostep_index. A barecontroller.abort()aborts with the platform'sAbortError, which is anError, so it comes back as it is too. Code that caught this check byinstanceof aborted_erroralone also catches the error that it aborted with, or theAbortError. -
The permission prompt and the cards know a workspace tool by more than its name. A tool of the flow's own that was called
write_fileoredit_file, one from an MCP server included, got the workspace prompt, with a diff worked out from its input and Allow all edits this session, which switched the session to Accept edits. One calledrun_commandran without asking whenever its command matched a grant that the realrun_commandhad earned. The prompt now gives that treatment only to a call whose input the session's own workspace tool saw first, so a tool that shares the name gets the plain prompt, and no grant covers it. A card opens with nothing but the tool's name, so cards key on the names that the session'sworkspace_toolsbuilt, and in a session that never built them, every tool gets the plain card. Within a session that did, the six names are reserved, anddocs/acp.mdsays so. A flow that gave a tool of its own one of those names renames it, and a wrapper that replaces a workspace tool'sneeds_approvalgets the plain prompt. -
An "always allow" for a command covers that exact command and nothing else. The grant used to key on the program and its first argument that wasn't a flag, so allowing
find . -name '*.ts'also allowedfind . -deleteandfind . -exec rm, andpnpm --filter web testkeyed aspnpm web. A grant now covers the command line it was given, with the whitespace at its ends trimmed and everything between compared exactly, sorm "a b"andrm "a b"are two grants. The option's label names the whole command. A re-run of the same command still runs without asking, and any variant asks again, so a user who relied onpnpm testcoveringpnpm test --watch=falsegrants each one. -
write_fileandedit_filerefuse writes into.gitandnode_modules. In Accept edits, the model could write.git/configor a file innode_modules/.bin, and what it wrote would run on the user's nextgitorpnpmcommand without ever showing ingit diff. Both tools now refuse, in every mode, a path that passes through a file or a directory named.gitornode_modulesanywhere under a root, a nested repository's.gitand a worktree's.gitfile included, whether the path gets there as written or through a symlink. The error names the directory and points the model atrun_command, and the permission prompt shows it in place of a diff. A flow that wrote there through the file tools usesrun_commandinstead.docs/acp.mdnow says that an edit to any other file that git ignores never shows ingit diffeither, so Manual is the mode for a checkout whose ignored files matter. -
The viewer stays up when it can't read a file. A read that failed while the viewer tailed a path, as one that names a directory does, crashed the process with an unhandled rejection, and now reaches
on_io_error. A trajectory or an asset that passesstatbut can't be opened now costs that one response rather than the server. The publishedfascicle-viewerbin also prints an error as onefascicle-viewer:line and exits 1, as the in-repo script always did, where it used to print a raw stack. -
The viewer warns about any host that reaches beyond the machine, and prints an IPv6 address that it can open. The no-auth warning fired only for
0.0.0.0, so--host 192.168.1.5or--host ::exposed your trajectories without a word. Any host but127.0.0.1,localhost, or::1now warns, and the warning names the host. An IPv6 host gets brackets in the URL that the viewer prints and opens, which used to readhttp://::1:4242and open nothing.ViewerServergainsport, so code that binds port 0 reads the port without parsingurl. -
tee_loggerkeeps each sink's spans apart. When a sink threw fromstart_span, the span ids of the sinks after it each moved up a place, so each of those sinks ended a span that belonged to its neighbor. Each sink now keeps its own id, and the id that the tee returns is the first one that a sink gave back. -
http_loggerhands every failure toon_error. It serialized an event before its promise chain began, so an event that holds aBigInt, or a customfetchthat throws before it returns a promise, threw out of the logger and into the flow, though the docs say that a failure reacheson_error. Both run inside the chain now, so neither one fails the run. -
mcp_clientcloses a server that it started when listing the tools fails. IflistToolsrejected, the client thatmcp_clienthad opened stayed open, so a stdio server that it spawned outlived the call. The client closes before the error propagates now, and a client that you passed in stays open, as before. -
OTel metrics keep each model call's attributes when calls overlap. The metrics dropped the newest call's provider and model whenever any call ended, so with two calls in flight, the one still running reported the attributes of the one that had finished. Each call's context now comes off by its own span id.
-
run_stdiosends its console to stderr, asserve_acpdoes. A flow that calledconsole.logunderrun_stdiowrote the line to stdout ahead of the result, which broke the parent's parse, and the run still exited 0.run_stdionow movesconsole.log,console.info,console.debug,console.dir, andconsole.dirxmlto stderr for the life of the process before it reads stdin, through the same code thatserve_acpuses.execute_stdioleaves the console alone, since it doesn't own the process, and a write that goes straight toprocess.stdoutstill lands there. Console lines now share stderr with the trajectory, where the viewer skips each one with a warning, so a flow that logged through the console to report progress reports it throughctx.emitinstead. -
serve_acpkeepsconsole.diroff the protocol stream. It sentconsole.log,info, anddebugto stderr, since stdout carries the protocol, butdiranddirxmlwrite to the console's own stream, so a flow that called either one wrote into the protocol and could break the session. Both go to stderr now.run_commandalso stops arming a grace timer for a call that an abort has already settled, so no timer outlives the call. -
claude_clikeeps a structurednullasnull. With aschemawhose reply parsed tonull,contentheld the reply's raw text, the string"null", because the adapter fell back to the final text whenever the parsed value wasnullorundefined. A call with aschemanow takescontentfrom the parsed value every time, and only a call without one takes the final text. Code that compared a nullable result'scontentagainst"null"checks fornull. -
claude_clipassessetting_sourcesandplugin_dirsto the CLI. The adapter built the CLI's arguments from an empty config, so both options were dropped without a word. Both reach the command line now. -
claude_clitells a missing sandbox apart from aclaudebinary that the sandbox can't see. Asandboxwhosebwraporgreywallwasn't onPATHthrewclaude_cli_errorwithreason: 'binary_not_found', thoughdocs/cli.mdsaid that a missing sandbox binary throwssandbox_unavailable. Abwrapthat ran but found noclaudebinary inside its sandbox threwsandbox_unavailable, because its error,bwrap: execvp <binary>: No such file or directory, namesbwrapand a missing file. Each case now gets the other's reason. A missing wrapper throwssandbox_unavailable, and aclaudebinary thatbwrapcan't find throwsbinary_not_found, with a message that names the binary and the paths that the sandbox binds. Code that read either reason to tell a user what to install reads it the other way round for these two cases.docs/roadmap.mdlists the binds that leave out aclaudein your home directory under Known Issues. -
Native transports fail a request for the reason that it failed. Each one built its request body inside the
tryaroundfetch, so aprovider_capability_errorthat the mapper threw, for a part that the wire can't carry, was retried as a network failure and reached the caller as one. The body is built before the fetch now, so that error arrives typed on the first attempt. A response body that fails partway through its read is now a network failure that the engine's retry policy retries, as a failed stream read already was. A body that arrives whole but isn't JSON throws aprovider_errorthat quotes what came back and asks you to checkbase_url, where it used to throw a bareSyntaxError. -
ollamanative reports an error that the daemon sends mid-stream. When a request fails after its stream has started, the daemon writes a frame with anerrorfield, which the adapter skipped, so the turn failed as truncated output and the daemon's message was lost. That frame now throws aprovider_errorthat carries the message. -
A thrown
undefinedcounts as a failure in a tool and inmap. The tool loop andmapeach took a failure to be a thrown value that was defined. A tool that ranthrow undefinedwas recorded as a success with no output, and amapitem that did the same left a hole where its result belonged while the map went on. The tool's call now fails, so it goes back to the model as a tool error, or throwstool_errorundertool_error_policy: 'throw', andmapstops and rethrows, as each one does for any other throw. -
improvekeeps no lessons whenlessons_capacityis 0. A cap of 0 kept every lesson, becauseslice(-0)returns the whole array. -
learnthrows the caller's abort reason as it is. It wrapped anErrorthat the run was aborted with in a newaborted_error, while core andbenchthrow that same instance, so code that aborted with its own error couldn't catch it by identity.learnnow keeps the rule that core keeps. -
describeshows afallback'shandoffand adefine_agentstep's description.fallbacklefthandoffout of its describe metadata, anddefine_agentnever passed thedescriptionfrom an agent's frontmatter to its step, sodescribeshowed neither. -
fascicle-diagram --exportreads only the module's own exports. It checked the name within, so a name such astoStringmatched a member ofObject.prototypeinstead of failing as a missing export. A name that the module doesn't export now fails as any other missing export does. -
Three errors say what to do.
stash()anduse()outside ascopenamed neither the key nor the fix, and now each one names its keys and says to wrap it inscope([...])with the steps that share them.tee_loggerwith no loggers now shows how to pass the sinks.
Internal
- The docs now say that a
fascicle/uistream doesn't stop its run when the client disconnects, anddocs/human-in-the-loop.mdshows how to passrun.streamanabortthat fires with the client, both from a route handler and fromnode:http. - A sweep brought the docs, the site, and
VALUES.mdback in line with the code, and swapped the opaque decision tags in the source, scripts, configs, and rules for the rules that they name. examples/pr-improvedefaults to OpenRouter, the examples build their tools withdefine_tool, and the smoke suite runs the six examples that need no key.- The configs, rules, and scripts match the current layout. A new ast-grep rule checks doc comments in
.tsxfiles, thetunneloverride is gone, and.mcp.jsonpinsast-grep-mcp. - The
claude_clisandbox test hidesbwrapfrom the child'sPATH, so it passes on GitHub's Ubuntu 26.04 runners, which shipbwrap.