fix: preserve multimodal evidence through deep workflows

[seonghobae]

Buyer-visible gap

An authorized, non-identifying five-image runtime check completed transport and stored five captions but zero OCR characters. The conducted workflow had converted the source pixels into a text marker before thinker, worker, verifier, and synthesizer execution, so successful transport was not successful evidence extraction.

Behavior

• Normalizes Responses input_image URL blocks into the existing Chat image_url representation without provider-specific routing.
• Preserves the original image blocks beside every evidence-bearing workflow instruction while Conductor access lists continue to isolate prior model outputs.
• Selects and fails over image work only through agents explicitly tagged vision; it fails before provider I/O when the pool has no such agent.
• Keeps text-only routing unchanged and updates multimodal HTTP fixtures to declare their synthetic mock capability honestly.
• Records the decision in ADR 0018 with APA 7 references to Fugu, TRINITY, Conductor, and the official OpenAI API contracts.

Verification

• python -m pytest -q -> 1471 passed in 539.13s
• Focused multimodal regression -> 6 passed
• Focused HTTP compatibility regression -> 30 passed
• python -m compileall -q contextual_orchestrator tests/test_multimodal_workflow_evidence.py
• git diff --check
• codegraph sync && codegraph status -> healthy, up to date

Stack contract

This PR is intentionally based on #765 at exact head d3f9a9b96523ed572b908c8abba1afa527eb49dc. It does not modify #765's gateway/reasoning work or #776's inbound-framing work. ADR 0018 avoids the concurrent ADR 0017 allocation in #776.

Remaining runtime acceptance

Repository behavior and contracts are verified here. A follow-up private LineageWeave backfill must restore non-empty OCR and region evidence before this is represented as live product recovery; only aggregate, non-identifying evidence will be reported.

---

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 0b4dfe7f-1df3-4fcf-8131-c6b52e0c6092

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[seonghobae]

@opencode-agent Please review the exact current head 6e0daf8. Verify that every evidence-bearing workflow step retains source image blocks, Responses input_image normalization stays provider-neutral, and image failover never selects a text-only agent.

[seonghobae]

@opencode-agent Review exact current HEAD 6e0daf8 only. Verify that image source blocks survive Responses-to-Chat normalization and every evidence-bearing workflow step, VISION-tag failover never selects text-only agents, no-VISION pools fail before I/O, access lists remain intact, and the ADR explicitly separates transport/evidence preservation from the pending LineageWeave OCR backfill. Submit a formal exact-head verdict without modifying the branch.

[seonghobae]

Validated the exact stack head 6e0daf8: multimodal/HTTP focused suite 30 passed, full suite 1471 passed in 531.30s, compileall and git diff --check passed. The implementation preserves typed source images for each evidence step and explicitly leaves non-empty OCR/backfill as a separate LineageWeave acceptance gate. No provider-specific routing or model-name inference was added. Await the independent exact-head review and protected Checks.

[seonghobae]

Runtime acceptance update for exact HEAD 6e0daf859813e600e6c5fdef4f948a781dfd7169: after the consuming LineageWeave bootstrap removed Docker env-file outer quotes, auto discovery produced 7 VISION-tagged agents and 0 empty-model agents. A generated synthetic PNG containing SYNTHETIC TEST 42 completed through the real provider-neutral endpoint with 96 caption characters, 17 OCR characters, 3 tags, and the expected TEST/42 text present. No provider model was named or selected locally. The authorized five-image aggregate backfill remains the final downstream acceptance gate.

[seonghobae]

Current-head review refresh: the remote branch has advanced beyond the previously reviewed 6e0daf8; current exact head is 918654711e9d70ff327c9368e57ec0129866c5b7. Revalidate multimodal source-image preservation, vision eligibility/failover, Responses normalization, and LineageWeave recovery on this SHA only. The stacked base has no independent required Checks until its parent is integrated.

[seonghobae]

OpenCode review requested for exact current head 49c6d93. The new boundary regression exposed that an explicit text-only primary could bypass required_tags=(vision); the shared failover candidate builder now excludes capability-ineligible primaries and fails closed when no matching agent exists. Focused multimodal and provider-reliability verification: 21 passed.

[seonghobae]

Private consumer-runtime aggregate evidence for the image propagation path (non-identifying): the authorized LineageWeave run obtained image analysis for all 5 source images and persisted 776 image-OCR characters, 5 Markdown table signals, 31 visual regions, and 773 region-OCR characters. This runtime used the image-propagation implementation from 6e0daf8; exact-head 49c6d93 additionally closes the explicit-primary required-tag bypass and is verified by 21 focused multimodal/reliability tests.

[seonghobae]

Exact-head verification update for 49c6d93: the complete local suite passes, 1472 passed in 527.42s. A correctly configured private consumer structure request also returned complete cardinality (23 input semantic units, 23 structured decisions); no private content or identifiers are included.

[seonghobae]

Remote exact head advanced to 9aa99731f8ee242f10114b16f082608d9411191f. Review found 14 repository-wide Ruff defects in this stacked snapshot (mostly stale __main__ test calls plus one unused import/duplicate test key); fixed them at the shared test baseline. Focused multimodal/evidence/API/security suite: 94 passed; Ruff and diff-check pass. Revalidate this SHA; prior 49c6d93 Checks are stale after the push.

[seonghobae]

[seonghobae]

@devin review exact current HEAD 915f780. This branch was updated onto current parent b225e13; the merge preserves multimodal evidence changes, parent capability/temperature/provider changes, and inbound framing/deadline. Integrated focused suite passes (58 passed) with changed-file Ruff/diff clean. Revalidate this exact head only.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 1 additional finding.

[seonghobae]

@devin review exact current HEAD f696daa. Parent is now 61337d9; this stack preserves multimodal evidence, provider negotiation, deadline-boundary correctness, and static-check cleanup. Integrated focused suite passes (61 passed) with full Ruff/diff clean. Revalidate this exact head only.

[seonghobae]

Current exact HEAD: 915f780aec8d7d0ae6a35141a221aec9b9cbc40c.

• Devin Review: no issues found; CodeRabbit status is the configured base-branch skip.
• Exact-head local proof: multimodal/evidence/provider focused suite 55 passed; full suite 1506 passed in 547.06s; Ruff, compileall, and diff checks pass.
• The branch is reconciled with the current stacked base without force-push.

Normal auto-squash merge is now armed; no approval or bypass is used.