You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The package ships a primer for coding agents, as dist/AGENTS.md. An agent in a project that installs Fascicle knows little or nothing about it from training, and the primer it needs lived only in the repository, at docs/learn/agent-primer.md, which a project could reach only by cloning. The build now copies it into the package, with each link into the docs pointed at the same file on GitHub at the release's tag, so a CLAUDE.md can import @node_modules/fascicle/dist/AGENTS.md and get the copy that matches the installed version. The README says where it is.
Changed
A call with a schema shows the model the schema. The engine asked for "a single JSON value that conforms to the expected schema" and never said what the schema was. Only an ai_sdk adapter with the structured_output capability hands the schema to the provider, and only on a call without tools, so everywhere else the model guessed the field names, and the repair loop taught them one failed reply at a time. The instruction now carries the schema as compact JSON Schema, closed with additionalProperties: false as a tool's schema is, so the first reply can fit. On OpenRouter's Claude Haiku 5.5, a three-field extraction with no example in its prompt took two round trips before and one after, at about half the cost, and a 4B model on Ollama that never found the right fields now does. A schema that JSON Schema can't express, such as one that holds a Date, keeps the bare request and still validates its replies. The .describe() text on a zod field reaches the model the same way, which is what the blueprint's advice to keep field rules there assumed.
branch returns the union of its arms' outputs. Both arms had to return the same type, so a branch that answered null on one side and a number on the other failed to compile until you wrote branch<Pair, number | null>(...), and the course had to teach that spelling before it taught anything else about branch. The otherwise arm now has its own output type, defaulting to the then arm's, and the branch returns either, so the union is inferred. Naming one output type, as in branch<Pair, string>, still holds both arms to it, so every existing call compiles unchanged.
describe.diagram draws a named composite as the steps it was built from.adversarial, consensus, ensemble_step, gate, and improve are made of scopes, stashes, and loops, and the diagram drew all of it, so an adversarial over a draft and a check took fourteen rows of stash and snapshot around the two steps that the caller wrote, and a lesson that taught the pattern had to use loop instead to stay legible. compose takes parts, the steps inside it that a reader should see, by role, and records the path to each, so the diagram draws the composite as one row over its parts (build and critique, the members and the score, the gated work and its approval), while describe.json keeps the whole tree that the viewer maps spans onto. { internals: true }, or --internals on fascicle-diagram, draws the insides as before. A part that isn't inside inner throws at construction, since the diagram would have nothing to draw for it, and roles that collide, such as a member named score beside the judge, get a numbered suffix.
Fixed
A chain opened without its input type says how to fix it. An unannotated chain('pair') typed its input as never, so the first field read failed with Property 'a' does not exist on type 'never' and run with an argument that wasn't assignable to never, and neither said what to do. The input now defaults to a branded type that nothing satisfies, so the same mistakes fail where they did, as Property 'a' does not exist on type 'ChainInputUndeclared<"call .input<T>() on the chain to say what it takes">'. A chain without an input type in a parallel now constrains the parallel's input as well, where never constrained nothing.
fascicle-viewer labels a loop's guard guard. A loop wraps a guard function in an anonymous step whose id it makes up, as loop_1_guard, and the canvas labelled the puck with that id, which describe.diagram never shows. An anonymous step now reads as the role its parent gives it, guard for a loop's second child, and otherwise as its kind, which is how the diagram reads one.
checkpoint_store_conformance no longer fails a correct store's renewal when the machine stalls. The renewal check claims a key, renews it halfway through its lifetime, then has a second owner try for it after the original lifetime is up, and a store that restarts the clock on renewal refuses. The check skipped judging when it woke too late, but it read the clock before the second claim, so a machine that stalled inside that claim reached the end of the renewed lifetime, the store rightly handed the key over, and a correct store failed. It reads the clock after the claim now. It flaked this way under the load of the repository's own full check.
fascicle-viewer counts each place in a flow apart, and labels a step or a composer by the name you gave it. The canvas joined every span to its node by step id alone, so steps that shared an id pooled their runs. Every chain closes on a step called output, so in the course's commit tool, a chain inside a loop inside another chain, a run of three rounds drew the inner output as ×4 and the flow's last step as TERMINUS ×4. A chain binding that shares the id of the step it wraps doubled the same way. A span now joins the place that its parent span's node holds it in, the canvas keys every place apart, and a suspension or a checkpoint lookup, which names its step by id, lands on the place whose span is open. A group label reads the composer's name before its id, so a loop named draft_until_clean is DRAFT_UNTIL_CLEAN where it was LOOP_1, and a step's name labels its puck, since core resolves names in that order.
parallel takes an input that every member accepts, even when a member accepts a union. The input type pooled every member's input into one union and then intersected its parts, so a member that took string | readonly Message[], as a model_step does, beside a member that took a string made the input string & readonly Message[], which no string satisfies, and a member that took either of two object shapes demanded both. Each member's input now stays whole while the intersection is taken across the members, so the first case takes a string, and the second takes either shape along with the other member's fields. A member whose input reads as never, which is what a zero-parameter function's becomes once a generic wrapper loses it, still constrains nothing.
A stub's inline content function gets its parameter types.StubResponse['content'] was typed unknown, which absorbs a function type, so content: (opts, call_index) => ... written inline failed strict mode with an implicit any on both parameters, the example in docs/testing.md included, and every function needed (opts: GenerateOptions, call_index: number) by hand. content is now a StubContent, which is StubContentFn alongside any other value, so an inline function takes its parameter types from StubContentFn, and a value typed unknown is still accepted. StubContent is exported from fascicle/testing.
A sequence with a step out of order reports the joint that broke. TypeScript 7 reports a call that matches no overload against the last overload alone, and the last one was the fallback for runtime arrays, so a step out of order got two errors that only said each step had to be a Step<T, T>. The joint-checked signature now comes last as well, and the same mistake gets one error, on the step that can't take its input, naming the joint, the type that arrives there, and the type that the step accepts: SequenceJointMismatch<"after children[1]", string, number>. A call that compiles resolves exactly as before.