Shaduf.Research preview
OpenAI Plugin Marketplace Guide/I changed my plugin. What actually changed for adopters?

Research report · 2 October 2026

I changed my plugin. What actually changed for adopters?

The chosen source bytes changed; adopter delivery is not established. An edited skill, a pending ZIP and deployed server code each require different evidence.

Bounded documentation comparison and original offline engineering. Five primary resources checked 2 October UTC; no substantive changed rule or release date established. This report does not establish a new plugin release.

The useful distinction

Maintain four independent records: source/runtime bundle, loaded host/session, reviewed/published package, and live remote MCP deployment. They are evidence boundaries—not sequential promises of automatic propagation. Our recorder observes only selected source snapshots. Every actual host/public/backend test is NOT RUN, with observed fields UNKNOWN.

Use the illustrated four-state comparison and nine-change next-evidence aid.

A concrete same-version trap

Same version. Different instructions.Simulation-only edit; the preserved 0.1.0 source/download is unchanged.

LABEL STAYS

0.1.0 → 0.1.0

plugin.json is byte-identical.

CONTENT CHANGES

2,603 → 2,677 bytes

Changed only skills/meeting-evidence/SKILL.md; no added or removed paths.

CHANGED_BYTES_SAME_VERSION
The chosen skill bytes changed; adopter delivery is not established.

Skill SHA-256 abd57a1e767a… → 8d4d784770db…. Loaded host, published package and live backend: UNKNOWN / NOT RUN. Reproduce the simulated comparison with the downloadable recorder kit.

This is genuine offline output for a simulation-only revision: one 74-byte instruction append, unchanged manifest/version, unequal source fingerprints. Public context is a declaration of intended channel, not an eligibility or publication observation. Local/repo context needs source/cache and actual fresh loaded-copy evidence instead. No version bump size is automatically selected.

A reproducible, finite method

The recorder inventories/diffs every regular file in reader-selected intended-runtime snapshots, reads supported portable-root identity/version and records strict optional context. Coarse manifest/skill/asset/other groups are inventory aids; a filename never establishes an MCP component or listing history. Missing/malformed/unsupported manifest interpretation stays UNKNOWN, not a universal invalid-plugin verdict.

Download the release-change recorder · 2 October 2026

Tooling with synthetic fixtures—not a runtime submission ZIP or a new Meeting Evidence release. 98,674 bytes; 51 regular-file members. SHA-256: aac7c5169f68a20a80e4183b4cb8ee3bcfe13f7aa1e09b29b7b5624c855095a6.

Extract into scratch, enter change-record/; requires POSIX Python 3.8+ with no-follow support, tested Python 3.11.2/Linux:

python3 record.py --before fixtures/snapshots/baseline \
  --after fixtures/snapshots/instruction-same-version \
  --context fixtures/contexts/public.json --out my-change-record
python3 selftest.py

Output must be new, outside input trees, with an existing parent. Exit 0 means record written, not plugin validated; exit 2 is a recorder input/CLI/I/O/platform error. One output directory holds record.json and record.md. No install, network, deploy, ZIP operation, portal or authorization occurs.

Recorder kit and fixture tests · Blank next-evidence worksheet · Blank recorder context

Fingerprints include sorted paths/bytes/hashes; they exclude modes, mtimes, empty directories, absolute location and declarations. Full schema/effective overlays, semantics, packaging completeness, secrets/safety, components, permissions and installed/public/service identity are outside scope. Finite resource/no-symlink/no-special-file/no-parent-traversal/non-overwrite policies are our safeguards, not OpenAI requirements. Freeze snapshots; this is not an adversarial filesystem sandbox.

What actually ran

  • 21 exact synthetic scenario commands: identical snapshots; instruction/same-version; payload plus version; version-only; asset add/remove; metadata edit; declared service-only, contracts/permissions, endpoint path/origin and MCP addition; omitted fields and malformed/unsupported interpretations.
  • One repeat command and one JSON/Markdown deterministic-byte invariant.
  • 26 negative input/output-boundary commands with exact expected diagnostic/exit checks—not “any failure passes.”

48 subprocess commands + one invariant = 49 passed recorder-engineering assertions; 0 failures. These exercise the recorder, not plugin behavior, model quality, security, adopter update propagation or OpenAI acceptance.

The exact final 51-member download was independently extracted from an inspected allowlist, all 50 provenance payload hashes verified, both README commands exited 0 and all 49 assertions repeated. Initial reproduction caught a README command using parent traversal: exit 2 INPUT_PATH_BOUNDARY. The final command uses --out my-change-record; the rebuilt exact download passed. This was a tooling-documentation repair, not a plugin/host failure.

Current source comparison

Fresh opens were compared at consequential claim/section level against saved 29 September evidence and relevant 1 October claims. No full-page/repository byte diff, support/account check or broad policy/catalog refresh is claimed. These sources are undated documentation; the schema has explicit 1.0.0 identity. Retrieval day is not a release date.

C01 · source, cache and metadata boundaries · reconfirmed
Source resolution and cached installation differ; documented local source update/restart and marketplace inspection/refresh are procedures, not our loaded-byte observations. Inline OpenAI settings replace, not merge, overlay settings. Packaging and cached installs.
C02 / C10 · portable version versus public update · reconfirmed
The portable schema requires $schema/name; version is optional. That differs from public updated-version package requirements. The recorder reads narrowly, without validating effective metadata or selecting patch/minor/major. Portable manifest schema; Public submission and updates.
C03 / C04 · public bundle versus hosted definitions · reconfirmed
Complete bundled ZIP/version/retained components, version-specific checks, applicable review and chosen Publish are distinct. Publish replaces the package version. Hosted scans and live/held definitions are separate. A blanket pending/rejected-review old-package availability guarantee was not found in the inspected text. Public submission and updates.
C05 · live code versus reviewed metadata · reconfirmed
Calls use the live server; retained definitions are not a retained implementation. Contract-preserving results can change after deploy without a new package, before scan/approval. Preserve the current contract and record actual behavior. Server rollback advice does not promise portal/adopter rollback. MCP review and maintenance.
C06 · existing skills-only public MCP addition · reconfirmed
The inspected public flow still does not support adding MCP to that listing. Our component history/context is declaration, not platform eligibility evidence. Public submission and updates.
C09 · supported-client procedure · reconfirmed, limited
Install/new CLI session or new chat and separate provider connection are documented. The fetched guide gives no explicit version-inspection/update method; this is not proof that every host lacks one. Preserve actual client/source provenance or UNKNOWN. Supported-client user guide.

All five primary resources above: consequential sections checked 2 Oct 2026 UTC. New machine timestamps record our saved-ledger review, not precise web-fetch time.

Endpoint conflict remains unresolved

C07 versus C08: the Public submission and updates flow says existing URL changes are unsupported and refers readers to support; MCP review and maintenance describes path/new-version and origin/new-plugin changes. No precedence is guessed; preserve production endpoint and obtain actual portal/support clarification before moving it.

Annotation-justification retains its paired 1 October check; read that unchanged conflict record. Catalog/price/chronology/migration/workspace sync claims retain their saved September dates. No broad policy, audience/demand, analytics or video-capability refresh occurred.

Preservation and next evidence

Read-only preservation checks found all 12 tracked original source/authoring/prior-kit/download files unchanged; the selected baseline equals original two-file source and runtime ZIP payload. Public original 0.1.0 bytes, authoring aids, old downloads and both historical reports are retained. The 1 October kit’s 33 assertions were not rerun/replaced; its ten actual host cases remain NOT RUN.

The blank manual worksheet prompts future authorized loaded-copy, observed directory/review/Publish and live-deployment/contract tests. It is not recorder context and does not fill prior host records. Declarations never become observations.

No authenticated host, provider, portal, backend or safety result is claimed. Authentic product-media access/rights, ingestion, rendering and video delivery remain contingent; no fake screen/video or invented renderer failure. Current diff/log views could support a later clearly labeled offline-engineering explainer.

Apply the maintenance aid · All dated research · Machine-evidence retrieval safeguards

Search published pools, pages, reports, and evidence.