Skip to content

Codemble v0.24.0 — evidence that stays coherent

Latest

Choose a tag to compare

@udhawan97 udhawan97 released this 07 Oct 18:14
· 1 commit to main since this release

Codemble v0.24.0 — evidence that stays coherent

Codemble v0.24.0 repairs eleven correctness, recovery, and concurrency defects
found by an evidence-led development review. The visible learning loop is the
same: explore a parser-owned galaxy, study real source, and light a system only
by passing graph-derived checks. This release makes the facts behind that loop
more trustworthy when source, requests, or saved state change at awkward times.

Study now reads, hashes, and decodes one source snapshot. If a file changed
since the folder was parsed, Codemble refuses the mixed result before cache or
provider use and tells the learner to reopen the folder. Python call resolution
also respects lexical ownership instead of attaching a call to an import that a
parameter, assignment, closure, comprehension, or definition-time expression
has shadowed.

Checks no longer infer execution order from a call edge's source line. A direct
call question asks for every parser-certain target and scores it as a
multi-answer fact. Existing file-hash-based amber understanding remains saved;
historical passes are not retroactively re-evaluated by the new question schema.

Progress writes are serialized across threads and processes, stale graph and Map
work cannot republish after invalidation, and mode-save failures reconcile to
the server-confirmed state. When a two-file preference update cannot confirm its
own rollback, the interface says the result is uncertain and asks the learner
to reload instead of making a false saved/not-saved claim. Launch retries retain
the choices still visible in the dialog, and obsolete clear responses cannot
commit into a replacement project session.

Local Ollama requests bypass ambient HTTP proxies after the URL is validated as
loopback. Narration has its own bounded worker pool; timed-out blocking workers
retain their admission slot until they actually finish, so repeated timeouts do
not starve parser, Map, Study, or check routes.

Public sharing remains disconnected. v0.24.0 adds no upload, provider, account,
or public cloud route and makes no claim that operational deletion or media
erasure has been proven. Independent deployment receipts remain mandatory before
public sharing can be connected.

Known limits

  • A preference update spans a project file and the learner-default file. Normal
    second-write failures are compensated, but the pair is not crash-atomic; an
    uncertain outcome requires a reload.
  • A blocking provider thread cannot be killed by Python. Provider socket
    timeouts are finite, admission stays bounded, and shutdown stops new work
    without waiting for an in-flight call.
  • Saved understanding remains keyed to parser file hashes. The direct-call
    question correction does not re-test historical passes automatically.
  • Public share delivery remains disconnected pending the M20 operational gate.
  • The original unaided learner-acceptance issue remains open; automated evidence
    does not substitute for learner testing.
  • The 3D Galaxy requires WebGL. The complete Canvas 2D Map remains available
    when WebGL is unavailable.