[Brian Krabach (bkrabach)]

Summary

Implement frozen P1 Gemini media-v1 projection: exact documented model/position facts, original-position user inline_data and call-bound functionResponse.parts, private count/send plan identity, per-attempt media validation and pre-emit pixel-safe diagnostics. Runtime imports Core only: no Core changes or Foundation/app/peer-SDK helper imports. No host activation.

The initial two-model allowlist was incomplete. The current catalogue follows public vendor evidence audited 2026-11-14, including the configured gemini-3.8-flash default. Exact per-model URLs, modality quotations, tool scope and route/availability caveats are in contracts/gemini-media-evidence.v1.json. Unknown positions are not qualified for this route, not asserted vendor-incapable.

Exact source and dependency holds

Base: accepted main e040c7cee5561c0b5b2df66a3dc8ac8cd2a4e043, not an unrelated local provenance branch.

DRAFT / DO NOT MERGE: parent106 final exact-head Astra/Fable review and integration against accepted F1/L1/host dependencies remain required. Offline pairing uses immutable archived Foundation candidate 31766b42efe71920386975cf10d136b38b85fc14, not a moving dirty checkout and not an accepted-graph claim. No DTU launch, credentialed model catalogue, live inference qualification, merge, release or live adoption is performed here. Integration owns the later accepted graph/installed qualification.

Structural seam: get_media_request_support(*,model,request_options=None) returns v1 positions/mime_types/evidence and optional reason. cap.validate(exact_private_request,purpose='count'|'transport',request_options=effective_options) gates every real attempt. Any cap-present request needs its host-owned private association, including text-only/unavailable projections. No witness marker or hydration is added to canonical metadata/history/suspension.

Current support matrix — offline wire, not entitlement/live vision

PNG/JPEG only under the frozen producer scope. Exact enumerated IDs, never a runtime prefix guess or automatic local model switch:

Exact IDs	User image	Call-bound tool image
gemini-3.8-flash, gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash, gemini-3.5-flash-lite, gemini-3.1-flash-lite, gemini-3.1-pro-preview, gemini-3.1-pro-preview-customtools, gemini-3-flash-preview	inline data	nested functionResponse.parts, explicit vendor G3 text-model scope
gemini-4-hydrogen	inline data	nested functionResponse.parts, named vendor example
gemini-4-carbon, gemini-4-argon, gemini-robotics-er-2-preview	inline data	unconfirmed scope; warning at original position before count
gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite	inline data	unqualified placement; warning at original position before count
Other/unresolved latest or dated aliases, retired/conflicting IDs	not catalogued for this route	not catalogued for this route

This is 16 user-image / 10 tool-image IDs. get_info and list_models vision tags use the same exact catalogue. System/developer/assistant images remain unqualified. No synthetic user image turn is added. Specialized image-generation models can genuinely accept images but have different output/function-calling contracts; Live, TTS, embedding, managed-agent, Decisions and video-output APIs are distinct. Existing specialized image-operation code is unchanged. No blanket Gemini-family capability claim.

Vendor routing notices (3.7-flash to 3.8-flash, 3.5-flash to 3.6-flash) and 2.5 existing-user restrictions/future December 1 routing are disclosed, not implemented locally.

Limits versus native measurement

Robotics standard ER2's exact public input/output facts are 131,072 / 65,536 tokens. get_info().defaults.context_window now publishes 131,072 rather than inheriting the unknown-model 1,048,576 fallback. Publication-only exact facts are separate from native-counter qualification; the separate Live endpoint and guessed Robotics suffixes do not inherit them.

Native budget counting remains unavailable for the three Gemini 4 IDs and Robotics ER2. Gemini 4 unified-window semantics are not guessed or changed. Tests distinguish actual mocked HTTP count on the previously qualified 3.x/2.5 set from wire-projection parity only on 4/Robotics; no fake provider_count or borrowed measured-fit claim.

Identity, retry, metadata and privacy

A weak request-scoped vendor plan binds the exact candidate/options/selected connection; SDK calls receive copies. Native signatures, provenance and call pairing are retained. Finalization tool_choice=none stays NONE. Host/loop single-refit remains authoritative; no new refit loop, compact_context or public iterator stream method is fabricated. complete owns the existing internal streaming transport.

Qualified media requests disable SDK/aiohttp inner retries, AFC, redirects and unqualified HTTP auth/request hooks/retrying transports without mutating the shared client. Received transient failures revalidate before retry; unknown sends, timeouts and partial streams do not automatically replay. Completion/error/cancellation/count failure/session close release provider-owned cache state.

Diagnostic COPY projection occurs before hooks.emit. Known image transports are omitted with explicit not-retained markers, not fabricated resolvable refs or raw-base64 fallbacks. Legacy-image retry/error diagnostics are protected too. No diagnostic ingest/publish or new canonical retention authority is added.

Verification and review

Existing isolated dependencies: Core1.6.1, google-genai2.30.1, HTTPX0.28.1, Pillow12.3.0. No manifest/lock changes. Both provider keys unset; real socket connects refused before pytest/provider imports, plus maintained autouse offline fence.

• Catalogue source at 66f92628... / final evidence wording 264b59ae...: 187 focused passes paired with immutable F1 candidate31766. Matrix: 64 valid PNG/JPEG user complete/SSE cases, 40 tool cases, 64 actual F1 canonical-to-private cases plus prior guards. Every literal declared model is covered by actual SDK MockTransport, decoded original bytes/SHA/order/MIME/options/privacy; warnings are frozen at the original call-bound block before count.
• Last unchanged broader standalone campaign at catalogue source: 1,041 passed, 1 optional F1-module skipped, 1 live test deselected, 38 warnings. Reused, not rerun or relabeled as a fresh full pass for the narrow Robotics fact correction.
• Current narrow Robotics published-limit/native-gate repair: 88 published-limit and native-count tests passed, 25 preexisting warnings; 8 Robotics SDK/F1 wire cases passed, 179 unrelated media cases deliberately deselected. Exact input/output publication, native count unavailability/no transport, unverified Robotics endpoint/suffix and unchanged Gemini 4 behavior covered.
• Independent worker accepted the catalogue source and reproduced187/1041. Manager-commissioned critique230921 accepted exact catalogue delta through264b (vendor/offline scope only). The current narrow correction receives its own changed-risk review; parent final Astra/Fable remains separate.
• Helper/changed test Ruff and git diff --check pass. Full init.py Ruff retains two proven baseline findings (unused DEFAULT_LIMITS, release_images lambda), not falsely green.

Retained failures / scope incident

The Robotics regression was first demonstrated failing (1048576 !=131072); failure log retained. Earlier catalogue run had one stale 3.8 no-vision expectation failure, corrected from exact documentation; prior valid-pixel assertions initially used standard rather than the SDK's URL-safe strict base64 alphabet, corrected while preserving exact original byte equality. Failed logs remain retained, not converted to passes.

An earlier initial focused command accidentally omitted the live-test exclusion and ran the existing real-API image test with an inherited credential; it failed, was disclosed, never repeated or used as qualification. Subsequent checks unset both keys, exclude live tests and enforce the pre-import socket fence. Initial wire-alias/Mock diagnostic-recursion failures and timeout/termination logs are retained. No later inference campaign is implied.

Evidence and cleanup

Task-owned untracked working-files retains exact result packets, sanitized actual-wire XML/JSON with decoded SHA/order, source/dependency joins, baseline/current/failure logs and immutable F1 archive. Source checkout and isolated .venv remain intentionally available for parent integration. No containers, DTUs or remote runtimes created; no unrelated teardown, automatic cleanup or live activation claimed.

[Brian Krabach (bkrabach)]

Advertised-model offline wire coverage follow-up

Test-only follow-up closes the review gap for the already-advertised gemini-2.5-flash user-image path. Production provider code, model declarations and dependencies remain byte-identical to reviewed e7f959e.

Exact follow-up head: 6c1cff46ff3fa372da53e39023a4d39c4d42e42f.
Independent read-only review accepted this exact test delta and independently
reproduced 29 focused / 943 standalone passes under the pre-import socket fence.
Reviewed test hashes: 8ab502ab7b841751613657f50e2bfef76277dcc724f2edfc274a4e46d708dfa8
and 457f1a48a89487e717a3ab5a2539ebb01a66be896b27fb8bf40b83aef6246e43.

Literal model/position	Valid PNG	Valid JPEG	Complete	Real SSE stream
gemini-3.1-pro-preview user inline image	pass	pass	pass	pass
gemini-2.5-flash user inline image	pass	pass	pass	pass
gemini-2.5-flash tool result	unavailable at original call-bound position before count	unavailable at original call-bound position before count	pass	pass
gemini-3.1-pro-preview tool result	existing fixtures retained	no new broader tool-MIME claim	existing fixture retained	existing fixture retained

The 8-case model/MIME/transport matrix uses actual google-genai serialization into HTTPX MockTransport, Pillow-verified original images, exact selected URL/model, decoded original bytes/SHA, user text/image order, count/send options parity, unchanged owner metadata and pre-emit privacy. It does not merely assert a support declaration.

The 4 additional Flash fixtures use the real SessionMediaStore/capability/projector from immutable archived Foundation candidate 31766b42efe71920386975cf10d136b38b85fc14. They assert the tool unavailable warning at its exact message/block/call position before any HTTP count, then exercise the actual HTTP count and SDK send paths with the same private object/options, original user image bytes and no extra synthetic user turn. This remains candidate-graph evidence only.

Verification uses both provider key variables unset and a socket.connect/connect_ex refusal installed before pytest/provider imports, in addition to the maintained autouse fence. No inference, live credentials/config changes or DTU occurred.

• Focused media + immutable Foundation candidate suite: 29 passed (17 retained + 12 added).
• Full standalone offline suite: 943 passed, 1 skipped, 1 deselected, 38 warnings. The optional Foundation module is the skip; live marked test remains excluded.
• Both changed tests pass Ruff; git diff --check passes.
• First matrix run: 12 failed, 17 passed because the new assertions used standard-alphabet strict base64 decoding while locked google-genai2.30.1 _common.encode_unserializable_types uses urlsafe_b64encode. The corrected assertions strictly decode the SDK alphabet with altchars=b'-_' and still require exact original bytes. Original failure log/XML retained; not relabeled green. Privacy checks reject both alphabet representations.

Prior failed/inadvertent-live incident evidence remains retained and unchanged. No runtime fixes or blanket capability-tag changes were made for this coverage repair. Parent exact-head Astra/Fable review, accepted F1/L1 graph integration and integration-owned qualification remain holds; PR stays draft.

[Brian Krabach (bkrabach)]

Full documented model catalogue correction (offline candidate)

Public Google documentation audited 2026-11-14 confirms the old two-ID allowlist was incomplete: it omitted even the provider's configured gemini-3.8-flash default. This follow-up expands exact documented input-model IDs, not a family-prefix guess. The checked-in contracts/gemini-media-evidence.v1.json records per-model URLs, modality quotes, narrower tool evidence, routing/availability notes, and specialized-route/unknown boundaries.

Final head 264b59aea12a1052d2f508facde070eb4c119ef9:
changed-risk source accepted by independent read-only reviewer at 66f92628...,
then one evidence-only wording correction clarifies that the Robotics Python
example uses an uploaded PNG URI while REST demonstrates inlineData.
Affected evidence assertion passes; reviewed production/tests unchanged.
Exact-head Python 3.11/3.12 CI and CLA pass:
run 43615458522.

Model/position matrix

• 16 user-image IDs: gemini-4-carbon, gemini-4-argon, gemini-4-hydrogen, gemini-3.8-flash, gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash, gemini-3.5-flash-lite, gemini-3.1-flash-lite, gemini-3.1-pro-preview, gemini-3.1-pro-preview-customtools, gemini-3-flash-preview, gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite, gemini-robotics-er-2-preview.
• 10 nested tool-image IDs: the 9 explicitly enumerated G3 text-model IDs above (vendor guide explicitly scopes multimodal function responses to G3 series), plus gemini-4-hydrogen (named GenerateContent nested-image example).
• Unconfirmed/unqualified tool positions: Carbon/Argon, the three 2.5 endpoints, and standard Robotics ER2. They preserve an unavailable warning at the original call-bound position; this is not a claim of vendor image incapability.
• Specialized image-generation models genuinely accept image inputs but have image/text output and no function calling. Live/TTS/embedding/managed-agent/Decisions/video-output routes have distinct contracts, not implied text-provider parity. Existing specialized image-operation code is unchanged. No blanket gemini-* vision tag.
• Unresolved *-latest/dated aliases and retired/conflicting IDs remain unknown for this route, not silently revived/resolved or called image-incapable.

get_info/list_models vision tags and get_media_request_support user positions share the same exact catalogue. The structural API remains unchanged; its existing optional reason field now states route/availability caveats. Unknown-position diagnostics say not qualified for this route, not fictional vendor incapability.

Important limits retained

• Google documents 3.7-flash -> 3.8-flash and 3.5-flash -> 3.6-flash routing at its service. The provider does not reroute locally. Existing-user restrictions on 2.5 and future December 1 routing changes remain disclosed.
• Gemini 4's unified input/output window and Robotics ER2's smaller window are not qualified by the existing native-budget adapter. Native count remains unavailable on those four IDs; tests do not report a wire projection as successful HTTP provider_count. No unrelated limit/thinking-level/long-decode redesign is hidden in this media change.
• PNG/JPEG only per frozen producer scope. System/developer/assistant images remain unqualified.
• Runtime still imports Core only. Retry, expiry/revocation, signature/provenance, private association and pre-emit omission mechanics are unchanged by this catalogue repair.

Offline verification

Keys unset and socket.connect/connect_ex refused before pytest/provider imports, plus maintained autouse fence. Existing isolated deps: Core1.6.1, google-genai2.30.1, HTTPX0.28.1, Pillow12.3.0. No installs, credentialed API listing, inference, config changes, DTU or activation.

• 187 focused passes against immutable archived F1 candidate 31766b42efe71920386975cf10d136b38b85fc14 (candidate-only, not accepted graph).
• 1,041 standalone passes, one optional F1-module skip and one live-test deselection, 38 warnings.
• Actual SDK complete/SSE matrix: 64 user cases (16 IDs x PNG/JPEG x complete/SSE), 40 tool cases (10 IDs x formats/transports), 64 real F1 canonical -> private -> count/send cases (16 IDs), plus prior retry/unknown-send/stream/privacy guards.
• Literal test catalogue is independent of production declarations; tests require evidence JSON/vision tags/actual wire matrix to agree. Each case verifies valid original pixels/decoded byte SHA/order/MIME/options and diagnostic omission. Native count-unavailable cases are distinguished from mocked real HTTP count cases.
• New helper/test lint and git diff --check pass. Full-file preexisting lint findings remain baseline, not relabeled green.
• First standalone run after expansion: 1 maintained expectation failure (3.8-flash formerly had no vision tag), 1,040 passed. Updated that exact documented-model expectation; failed log retained. Earlier live-test incident/failure packet remains unchanged and is not qualification.

Parent exact-head Astra/Fable review, accepted dependency graph/host integration and integration-owned installed qualification remain holds. PR stays draft; no merge or release.