Deepbom

- Publish the maintainer-requested channel usage update under the explicit
stable-release interval exception. Agent contract 1.0.2 is a review candidate
for the expanded website/privacy notice; evidence contract 1.0.0 is unchanged.
- Add consented Web analysis usage alongside ChatGPT, channel/date filters,
weekly browser estimates, unfinished-run counts, and explicit unmeasured
coverage for Claude, CLI and local MCP. Keep legacy ChatGPT history and
channel-specific browser identities; never sum identities across channels.
- Separate cached public npm download trends from opt-in analysis evidence.
Preserve missing dates and provider outages rather than displaying zeros.
CLI/local MCP execution remains offline with no new telemetry.
- Raise runtime/verification budgets by 32 KiB, docs/development tooling by
8 KiB and the app entry by 2 KiB for channel summaries, bounded registry reads,
browser controls, regressions and recorded release qualification.

- Add Google sign-in directly to the administrator console, returning to the
console after authentication and offering account switching to non-admins.

- Publish the coordinated web, CLI, and MCP update at the maintainer's explicit
request. This requested release uses the stable-release interval exception.
Prepare Agent contract 1.0.1 for the optional usage/privacy change while
retaining evidence contract 1.0.0. Source validation and server deployment
do not assert OpenAI approval of this new contract candidate.

- Add optional ChatGPT usage statistics with explicit browser consent, random
pseudonymous identifiers, UTC revisit estimates, test exclusion, withdrawal,
and deletion. Retain only allowlisted events and remove expired records daily.
Keep model files, names, hashes, and conversation text out of usage storage.
- Add administrator-only 7/30/90-day dashboards, format/export counts, daily MCP
operational counts, and aggregate CSV/JSON evidence exports. Distinguish
consenting browser estimates from verified people and downloads. Update the
privacy notice and review preparation files; keep CLI contracts unchanged.
- Raise handwritten runtime, verification, and documentation ceilings by 48,
32, and 16 KiB respectively for the collector, consent controls, administrator
dashboard, SQLite/browser regressions, and usage/review documentation. Add
8 KiB to development tooling for the associated checks, build metadata, and
measured release qualification records.

- Add an explicit Save via ChatGPT action for prepared exports in hosted
widget revision 20260916.6. Use the host's file upload, temporary HTTPS
download URL, and external navigation APIs when iframe downloads are blocked.
Offer Send link to chat and retain file IDs across URL failures. Local export
preparation does not upload anything; the separate save action shares only
the generated export, never the original model. Keep unsupported host APIs and
upload failures visible instead of claiming a file was saved.
- Raise the verification source ceiling by 16 KiB for sandbox-blocked download,
exact JSON file handoff, schema/hash, upload failure, and URL retry regressions.

- Make hosted widget revision 20260916.5 work with `worker-src blob:` by
downloading the self-contained classic Worker bundle before starting it.
Bind WASM asset resolution to the deployment URL at build time, without a
remote module import, runtime eval, or main-thread parser fallback. Exercise
the stricter CSP inside nested sandboxed iframes and report analyzer-loading
failures separately from attachment failures. Clarify that opening the panel
does not establish successful or ongoing analysis when the panel has failed.

- Fix hosted ChatGPT TFLite analysis in a different-origin sandbox by starting
a local Blob Worker that imports the deployed, CORS-enabled worker bundle.
Preserve deployment-relative WASM loading and verify cross-origin MobileNet
analysis with the actual production asset paths. Hosted widget revision
20260916.4 retains the 1.101.0 analyzer and CLI contracts.
- Raise the deployment ceiling by 2 MiB for the self-contained ChatGPT worker
bundle; keep cross-origin access limited to the declared public runtime assets.

- Publish this coordinated web, CLI, and MCP release immediately at the
maintainer's request to restore accessible visualization file downloads.
This explicitly requested release precedes the usual stable-release interval.
- Move ChatGPT visualization and file controls above long evidence tables and
report component height changes. Keep SVG, 300-DPI PNG, and Word-ready ZIP
Save links available after download requests. PNG handoff now requests an
inline image and a real ChatGPT-issued download link when available.
- Add browser-local CycloneDX 1.7 and SPDX 2.3 JSON export buttons. Reuse the
existing CycloneDX artifact-evidence projection; represent SPDX artifacts as
FILE-purpose packages with measured SHA-256, explicit license unknowns, and
evidence annotations. Validate against pinned official schemas and exercise
actual file downloads, narrow-panel visibility, and optional URL failures.
- Raise the reviewed development-tooling source ceiling by 8 KiB for the
pinned SPDX schema validator and this release's channel qualification records.

- Use the documented stable-release interval exception for a P0 public-tool
correctness failure: the advertised ChatGPT `Send PNG to chat` path could
fail in the host sandbox after the static audit had succeeded. The release
gate now exercises the complete render-to-upload handoff. The additive Model
Summary ships in the same corrected engine rather than publishing a second
immediate release.
- Add `deepbom.model_summary.v1`, one deterministic Common Model IR projection
for operation, reversible block, storage, and identity-only views. The CLI,
browser workbench, local MCP, ChatGPT component, and Claude MCP App reuse the
same source-bound rows without inferring framework trainability or runtime
order. Publish the JSON Schema and cross-format regression coverage for
TFLite, ONNX, GGUF, SafeTensors, Core ML, ExecuTorch, GraphDef, SavedModel,
Keras, PT2, HDF5, and safe PyTorch checkpoint envelopes.
- Fix ChatGPT visualization PNG export by rasterizing the validated render
model directly to an opaque monochrome canvas. This removes the sandbox-
dependent SVG Blob decode step while retaining canonical SVG as the source
and preserving the PNG derivation manifest.
- Raise the reviewed handwritten-runtime ceiling by 32 KiB for the shared
Model Summary projection and sandbox-safe rasterizer, the documentation
ceiling by 16 KiB for its public schema and contracts, and the verification
ceiling by 16 KiB for cross-format, MCP, and browser handoff regressions.
- Keep `Report in chat` explicitly user-initiated and reusable after the
browser-local audit completes. ChatGPT can acknowledge a component follow-up
request without rendering a new turn, so the widget no longer reports
delivery as complete or disables the only recovery control. The bounded
result prompt still identifies the Model IR SVG, PNG, and Word-ready exports.
- Raise the reviewed documentation source ceiling by 16 KiB for the official
ChatGPT submission import, portal values, demo protocol, icons, and explicit
plugin privacy disclosures. Raise the development-tooling ceiling by 24 KiB
for the corresponding schema and asset checks; runtime and generated-data
ceilings are unchanged.

- Use the documented same-day stable-release interval exception for a P0
release-contract correctness fix: engine-only updates were incorrectly
coupled to the reviewed Agent surface. The bounded
`check-agent-contract-release` mutation suite proves that engine-only drift
preserves the reviewed fingerprint while OpenAI, Claude, and evidence
contract drift are classified separately. This record is a release-policy
exception, not evidence of platform approval.
- Separate the frequently changing analyzer engine version from the stable
`deepbom.agent_contract.v1` and
`deepbom.artifact_evidence_envelope.v1` contract versions. Public plugin
manifests, host-evaluation cases, and Skill behavior no longer change merely
because an engine patch is released.
- Make the public Skill resolve a local or explicitly downloaded engine by
declared Agent/evidence compatibility, then bind the exact observed engine
version with a self-test before analysis. Registry access remains opt-in.
- Add a deterministic reviewed-surface fingerprint for OpenAI MCP metadata,
UI resource policy metadata, plugin and Skill contents, privacy/transfer
boundaries, the Claude remote surface, and the evidence-contract identity.
Ordinary engine releases now fail closed if any of those contracts drift.
- Add release classification tests proving that an engine-only version change
does not request a new platform review, while OpenAI metadata, Claude remote
metadata, and evidence-contract mutations produce their distinct required
actions. Source checks continue to distinguish a prepared candidate from an
account-observed approval or listing.
- Add the Agent contract gate to both the web deployment preflight and the
installation-channel publication boundary, with a public operations guide
for deliberate contract revisions.
- Keep the publication-repository gate executable without exporting the
private remote-MCP worker implementations: private release validation
captures live metadata, while the exact public export recomputes its public
Skill, manifest, policy, endpoint, and listing surfaces against the pinned
live-metadata hashes.
- Raise the reviewed documentation source ceiling by 8 KiB for the public
contract-versioning policy; executable and generated-data ceilings are
unchanged.
- Raise the verification source ceiling by 32 KiB for the contract snapshot,
classification, mutation, and release-gate checks introduced in this release.
- Raise the development-tooling source ceiling by 24 KiB for the deterministic
contract-snapshot builder; runtime and generated-data ceilings remain unchanged.

- Pin npm 11.6.0 on Windows installation-channel runners before invoking the
official MCP Inspector. This avoids the reproducible npm 11.6.2
`ECOMPROMISED: Lock compromised` regression while leaving the tested
Inspector version and DEEPBOM runtime contract unchanged.