is_effects 0.9.0
is_effects: ^0.9.0 copied to clipboard
The shared vocabulary of intersyntactic's actions: what reading, writing, starting a program and reaching the world take and give back, and the drivers that carry them out.
Changelog #
0.9.0 #
- Breaking:
WebRequestOutcomeis nowWebResponse: it holds the site's HTTP response (status,created,refusedBecause), nothing about the request or the airlock's own outcome, and the old name collided withOutcome<O>, which it isn't one of.CarriesOutWebRequests.carryOutWebRequestnow returnsOutcome<WebResponse>.encodeWebRequestOutcome/decodeWebRequestOutcomeare nowencodeWebResponse/decodeWebResponse; the JSON shape (status,created,refusedBecause, no type tag) is unchanged. - Breaking:
ModelReplyis now a sealed class:FinishedReply(the model's whole text) orCutOffReply(the call ran and spent tokens, but the reply ended before the model was done — such as ChatGPT's ownincompletestatus). A cut-off reply is still recorded asRan, neverRefused, since tokens were spent and a reply did come back; the split exists so a switch overModelReplyhas to say what it does with a cut-off one, rather than leave that to a flag a reader could forget to check. Any code built against the old concreteModelReply(text, ...)needs updating toFinishedReply(text: ...)orCutOffReply(text: ...). - Both variants gain
usage(a newModelUsage:input,cacheRead,output, pluscacheCreationandcontextWindow, both null when a driver can't report them).ModelUsage.contextisinput + cacheRead + (cacheCreation ?? 0)— neveroutput, which leaves with the reply rather than staying in context. The wire stays additive where it can:encodeModelReplystill writes a bare string for aFinishedReplywith no usage, identical to every reply recorded before this version, anddecodeModelReplystill reads a bare string that way. Any other reply — usage set, or cut off — encodes as an object instead, which is breaking for a reader that matches the old bare-string shape directly rather than reading throughdecodeModelReply.
0.8.0 #
- Breaking:
ProgramMessage(and its two cases,ReportedandFinished) now takes a type parameter for the outcome:Finished<O>carries anOutcome<O>, the same wayencodeOutcome/decodeOutcomealready took an output codec.encodeProgramMessage/decodeProgramMessageeach take that codec too. This is a Dart-types- only change — the JSON bytes forProgramMessage<ExecResult>are unchanged (pinned by a golden-bytes test against lines captured from the deployedprogram_driversocket, infra-5e) — but it is still source-breaking for any exhaustive switch or decode call written against the old non-generic types. A driver can now carryFinished<ModelReply>down the same wire aRunsProgramsdriver already used forFinished<ExecResult>, instead of hand-rolling its own envelope (claude_code_driver did exactly that, flagged by web-driver-0b). - New, purely additive:
encodeModelCall/decodeModelCall, a codec for the{signIn, model, systemPrompt, request}envelope a caller sends anAsksTheModelservice, agreed field-for-field with claude-code-driver-9f so the airlock and a model driver can't disagree on the shape.systemPromptandrequestpass through byte-exact — no trimming or normalising.decodeModelCallthrowsFormatException, never another exception type, on anything it can't read, since a service on the other end treats that one type as "refuse the call" and anything else as a crash.
0.7.0 #
- Breaking:
ProgressgainsReplyArriving, a piece of the model's reply as it streams, never recorded. A new sealed-class case is breaking for any exhaustive switch overProgressoutside this package. - The model call is now an action, reversing the 2026-11-12 ruling that
it stayed on the agent's side: the loop that proposes a call holds no
model sign-in and reaches the model only through a driver (RULINGS.md,
"the model call goes through the airlock"). New, purely additive:
ModelSignInName(a model sign-in by name, never the secret, built likeSignIn),ModelName(a model id, a checked string),ModelReply(the model's final text, kept whole, recorded once as an outcome), andAsksTheModel, named for what it does: holds every model sign-in, soaskTheModel(signIn, model, systemPrompt, request, {required report})is the only way off this machine for a model call.systemPromptandrequestare each the exact bytes the airlock approved, verbatim — found during review (web-driver-0b, against a driver whose system prompt is a separate CLI argument, not part of one request body) that a singlerequeststring had nowhere for it.AsksTheModel's doc is now also a contract: whenrequestitself names a model, it must bemodel, or the driver refuses without calling, so the record can never show one model decided and another called (airlock-api-04).
0.6.0 #
- Breaking:
encodeStart/decodeStartare nowencodeProgramToRun/decodeProgramToRun; the error-label default'start'is now'programToRun'. Flagged after 0.5.0 retiredStartsPrograms/startas fossil words still naming this codec pair and its own error messages.ProgramToRunoverRunProgram: this encodes data (the program and its args), not an instruction. No JSON key changes —'start'was only ever the error-message field label — so this needs no coordinated redeploy. No deprecated alias kept.
0.5.0 #
- Breaking:
StartsProgramsis nowRunsPrograms; itsstartis nowrunProgram. Nick's ruling:startnamed one of five steps (feed stdin, stream throughreport, enforce the time limit, return only on exit or kill);runcovers all of it.CarriesOutActionsandReadsAndWritesFilesfrom 0.4.0 stand as they are. No deprecated alias kept.
0.4.0 #
- Breaking: the four driver interfaces are named for what they can
do, not as a thing — Nick's ruling, after
implements WebDriverread as "WebDriver is an api?":Driver→CarriesOutActions,FileDriver→ReadsAndWritesFiles,ProgramDriver→StartsPrograms,WebDriver→CarriesOutWebRequests. No deprecated alias kept.
0.3.0 #
- Breaking:
WebDriver.sendis nowcarryOutWebRequest;Sentis nowWebRequestOutcome. Nick ruled out tiny words for complex things on 16 Nov (sendandrenderernamed as examples): "longer names are ok". No deprecated alias kept, so the old word doesn't survive through the types every session reads.
0.2.3 #
- Depends on
is_path_confinementsoFileDriver's doc comment can point atisInsideTreeconcretely, instead of describing confinement only in prose. Purely additive:FileDriveris still just an interface, no behaviour changes.
0.2.2 #
ProgramMessage: what a program driver sends back about one run,Reported(Progress)while it is underway, then oneFinished(Outcome<ExecResult>).Progressis unchanged.- Codecs for a program run across a wire:
encodeStart/decodeStart(whatProgramDriver.startis given),encodeProgress/decodeProgressandencodeProgramMessage/decodeProgramMessage. For the airlock ↔ program driver service. Purely additive. - Fixed
RelativePath.folderon a path made entirely of..segments (e.g.RelativePath('..')): it was returning root instead of one level further up again. Shipped broken since 0.1.0, but unreachable until the rename let an escaping path be constructed at all; no known caller hits it, fixed now to close the gap before one does.
0.2.1 #
- JSON codecs for every type:
RelativePath,FileText,ProgramName,Args,ReadResult,Wrote,ExecResult,Origin,Method,WebPath,SignIn, headers,Request,Sent,Outcome<O>. Encode to a plain JSON value; decode re-runs the type's own validation and throwsFormatExceptionon bad input. For the agent loop ↔ airlock and airlock ↔ driver wire (is_airlock_api0.5.0'sCarry). Purely additive, no breaking changes. isCredentialHeaderexported alongsidevalidatedHeaders, so a caller that needs to check a header without catchingArgumentError(anError, never caught by this repo's own convention) can.
0.2.0 #
- Breaking:
Request.pathis now aWebPath, not a plainString.WebPath.parserefuses anything without a leading/, a//authority, a..segment, or an@— ways a path could make a request reach somewhere other than the origin it was decided for.WebPath.rootreplaces the old default of''. - Pulled the credential-header check out of
Request's constructor into an exportedvalidatedHeaders, sois_airlock_api'sPostcan apply the same guarantee to its own headers instead of re-declaring the forbidden-header list. - From issue #1147's design discussion (
enspyrco/intersyntactic#1147):WebPathis web-driver-0b's finding,validatedHeadersis shared with effects-98's.
0.1.3 #
- Genericize destination examples in
web.dart's doc comments (smba.trafficmanager.net,hart-bot): a shared vocabulary shouldn't carry one consumer's deployment details in its examples.
0.1.2 #
- Fix
refusedBecause's doc comment: a refused sign-in (an AADSTS code, nothing sent) isRefused, notSent.refusedBecause, which is only for a destination's own refusal of something that did go out.
0.1.1 #
SentgainsrefusedBecause, the destination's own reason for a non-2xx reply (a Teams 400 body, an AADSTS code), so a failure's why reaches the record instead of staying in the driver's own logs.- Reworded
Request.bodyandSent's doc comments: no body is kept because what a reader needs back is already elsewhere in the record, not because of a by-reference ruling (there is none).
0.1.0 #
- First release: the vocabulary of effects and the driver interfaces, moved
out of
is_airlock_apiso a driver depends on effects alone.PathInRepo,ProgramName,FileText,Args,ReadResult,Wrote,ExecResult(Exited,Killed),Outcome(Ran,Refused),Progress(Printed), andFileDriver,ProgramDriver.