Claude Code First, specification version 0.1.0. The plain-language guide

Version 0.1.0. Grade R3/S2.

F105 Work that needs a person's judgement or hardware the machine lacks cannot be closed by the agent, and a bar the agent set itself blocked work

FieldValue
Statusactive
TopicPlanning, recording, and what the person decides
GradeR3/S2
Casespolyweave, winwright
Contrarynone
Harness factsnone

Claim

Where a task needed a person's look or hardware the machine did not have, work closed without them rested on cases that had never run, work set aside with the reason came back once the person or the hardware was there, and a margin the agent chose in place of a person's bar blocked a port for a day.

Mechanism

Some acceptance conditions cannot be computed: whether a render looks right at display size is a person's verdict, and a runner that needs an engine or a controller cannot run where neither is installed. An agent that must finish can substitute a proxy (a margin, a run on the hardware it has) and report the proxy's result as the verdict. A deferred store with a recorded reason makes "waiting on a person" a state the backlog can hold, so the next pick moves on instead of retrying.

Observations

Independence

Two origins. polyweave is the first, with several separate occasions (PW36, PW53, PW56, PW76). winwright's is independent: a different project, hardware and date, recorded as its own correction (rule 1). The deferred store is a roadkeep feature, a shared tool, and adds no origin; polyweave declared it in response to its own failure.

Counter-evidence

Both origins are projects whose product depends on physical or aesthetic inputs (a renderer, an engine, a controller, a person's eye), which the case studies name as particular to them; the other three projects record no such task, so the finding may be scoped to that property. polyweave's history is three days long, and the person who accepted PW76 is the project's single author, so "a person" here is always the same one.

Grade

R3: independent occurrences in two projects. S2: each observation is a dated commit recording the deferral, the return or the correction. The figures in them (109 samples, 0.3977 against 0.37, seven of eight deferred lines) measure the render or were counted from the file, not the phenomenon, so the grade is not raised to S3. Not S4: nothing checks that a task needing a person was not closed without one.