feat(swift-ios): show workspace images inline

[saphid]

Observed problem and reproduction

In the SwiftUI chat transcript, Markdown image references to workspace files were rendered as alternative text instead of as images. A representative reproduction is ![chart](out/render.png) in a message whose thread workspace contains out/render.png.

Cause

At the PR's original base, the native Markdown model had no image block and the transcript had no workspace-aware image resolver/rendering seam.

Change and integration boundary

The original contribution added standalone-image parsing, safe workspace-relative resolution through the existing signed asset route, bounded thumbnail loading, hosted-cell context, failure UI, and focused regressions.

For this conflict pass I selected a rebase onto the live target 22b22f1463b83963d851bd0574a36a20f224a28d. The first commit conflicted in MarkdownDocument.swift and ThreadDetailView.swift. Inspection showed that Theo's newer target commit cd2b2f8fef42e3762d03425e9d43f5a3255d8625 independently added the same capability end to end, including broader inline/direct-image behavior and focused source tests. The rebase was aborted. Upstream remains byte-for-byte unchanged and the Saphid head remains 60c3e1c56a7d0acb05d6d4fbd40d57fd8e70eadc; this PR is superseded rather than rebased into a duplicate.

Non-goals and deferred work

This pass does not alter Theo's newer parser, image loader, workspace resolver, or transcript composition. It does not port the old PR's review-fix series into another contributor's implementation. No web, desktop, React Native mobile, server, contract, provider adapter, or connection behavior is changed.

Affected areas

• Client and entry point: SwiftUI mobile chat transcript Markdown rendering.
• Platform: iOS and iPadOS native client.
• Providers: none; image rendering is provider-independent.
• Contracts/runtime: the existing signed workspace-asset route is reused by the superseding target implementation; no schema or server change is introduced here.
• Connections: the implementation is transport-agnostic, but this conflict pass did not exercise local, remote, relay, or tunnel runtime behavior.
• Reverse/failure states: blocked sources remain hidden or textual according to the target implementation; load failures retain its placeholder behavior.

Validation

• Ownership freeze: gh pr view 7378 --repo pingdotgg/t3code ... confirmed author/head owner saphid, head branch feat/issue113-inline-workspace-images, maintainer edits enabled, old head 60c3e1c56a7d0acb05d6d4fbd40d57fd8e70eadc.
• Live target fetch: git fetch origin refs/heads/t3code/rebuild-mobile-app-swift:refs/remotes/origin/t3code/rebuild-mobile-app-swift resolved 22b22f1463b83963d851bd0574a36a20f224a28d.
• Rebase: git rebase refs/remotes/origin/t3code/rebuild-mobile-app-swift stopped at the original feature commit with only the two overlapping conflicts above; git rebase --abort exited 0.
• Supersession inspection: target source contains MarkdownImage, MarkdownImageSource, workspace/direct image loading, transcript image context, and focused MarkdownDocumentTests introduced by cd2b2f8fe.
• Focused target test attempt 1: xcodebuild ... test -only-testing:T3CodeTests/MarkdownDocumentTests exited 65 before tests because the simulator code-signing subsystem reported an internal error.
• Focused target test attempt 2: build and signing completed, but Xcode stalled at Launch T3CodeTests while disk headroom fell to 366 MiB. The exact owned process group was terminated (exit 143), the corrupt incomplete result bundle was identified by xcresulttool exit 64, and zero tests are claimed. The private DerivedData and corrupt bundle were removed after preserving logs and receipts.
• No source or branch bytes changed, so there is no new contribution head to validate or push.

Risks, limitations, and known gaps

The superseding target implementation was verified by source and focused target coverage inspection, but this pass did not obtain a successful new local test execution because of the two recorded Xcode/disk infrastructure failures. Historical old-head tests and media do not prove the live target or a hypothetical rebased head.

Known gaps: full-size current-head light/dark before-and-after proof is queued in saphid/t3code-personal#150.

No review is requested because this PR is being closed as superseded.

Evidence

The following images are historical evidence for old PR head 60c3e1c56; they are explicitly not current-target or post-rebase proof.

Historical after, light appearance

Historical after, dark appearance

There is no current-head before/after or loading video in this handoff.

Owning issue and stack

Owning lane: saphid/t3code-personal#113. The PR targeted t3code/rebuild-mobile-app-swift, which remains related to #5178. The contribution is superseded by Theo's target commit cd2b2f8fef42e3762d03425e9d43f5a3255d8625 and is not stacked on #7345.

Implemented and reconciled by GPT-5.6 Sol high in the Codex harness.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. 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: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 87138f54-671d-463d-b58c-7e096e98662c

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.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This PR introduces a new feature - inline rendering of workspace images in Markdown messages. Despite being self-contained with good test coverage and proper path validation, the new user-facing capability warrants human review.

^{You can add or adjust custom eligibility rules. Learn more.}

[saphid]

Both High-severity findings against 90b3ac41f were verified and fixed in e45d52bcc:

• Macroscope — MarkdownImageReference.parse preserved backslash escapes, so ![plot](out/foo\(1\).png) requested the literal name and failed to load. Escaped ASCII punctuation is now unescaped in both the destination and the alternative text. New focused test covers escaped parentheses, escaped brackets, and a literal backslash in a path.
• Cursor Bugbot — MarkdownWorkspaceImageView rendered its @State image without checking it belonged to the current request, so a recycled UIHostingConfiguration cell could show another message's image. Loaded and failed state are now gated on request identity, exactly as the sibling FeatureRemoteAttachmentThumbnail does it, and the loader keys its cache from the same request. Cancellation is now handled with Task.checkCancellation() plus a CancellationError branch, matching that sibling too. This is a SwiftUI view-state repair, so it is verified by rebuild and review rather than by a unit test.

Re-run at the current head e45d52bcc, isolated temporary DerivedData, real exit statuses:

• Focused suites: exit 0, ** TEST SUCCEEDED **, 29 passed, 0 failed, 0 skipped.
• Full native unit bundle: exit 0, ** TEST SUCCEEDED **, 367 passed, 0 failed, 1 skipped (368 total).

The accepted on-device behaviour (Test 89, built from 90b3ac41f) was not changed to make these repairs. The GPT-5.6 Sol cross-provider review will run against e45d52bcc after the 2026-08-20T03:29Z quota reset; no Sol findings or approval are claimed yet.

[saphid]

Second round, against e45d52bcc.

Fixed in 9803983 — Macroscope Medium, quoted titles: a title holding an unbalanced parenthesis, as in ![plot](out/plot.png "generated (final"), made delimiter balancing miss the real closing ), so the line fell back to alternative text. Balancing now ignores quoted text while still accepting balanced parentheses inside a file name (out/foo(1).png still parses). Focused test covers both cases.

Acknowledged, deliberately not changed — Macroscope Medium, cache freshness: overwriting a workspace file at the same thread and path keeps the previously decoded bitmap until the cache evicts it. The reviewer is right about the mechanism. Invalidating it correctly needs a server-provided content version for a workspace asset, which is a wider change than this PR should make; dropping the cache instead reintroduces a signed-URL round trip and a re-download on every transcript cell recycle, which is the problem this cache was added to solve. A transcript is also a historical record of what a message referenced. Glad to follow up separately if a content version is wanted here.

Re-run at 9803983, isolated temporary DerivedData, real exit statuses:

• Focused suites: exit 0, ** TEST SUCCEEDED **, 30 passed, 0 failed, 0 skipped.
• Full native unit bundle: exit 0, ** TEST SUCCEEDED **, 368 passed, 0 failed, 1 skipped (369 total).

The two earlier High findings remain fixed in e45d52bcc; the on-device accepted behaviour (Test 89, built from 90b3ac41f) is unchanged by any of these repairs.

[saphid]

Third round, against 9803983. Both bots flagged the same defect and they were right — it was a regression from my previous fix.

Fixed in c8e303d77 — ignoring quoted text while balancing delimiters made an apostrophe in a file name, as in images/team's-logo.png, open a title that never closed, so the reference fell back to alternative text. A title is separated from its destination by whitespace, so only a quote in that position opens one; quotes and apostrophes anywhere else are ordinary path characters. Focused tests now cover an apostrophe in a path, a double quote in a path, and a real quoted title holding an unbalanced parenthesis, alongside the balanced-parentheses and backslash-escape cases.

Re-run at c8e303d77, isolated temporary DerivedData, real exit statuses:

• Focused suites: exit 0, ** TEST SUCCEEDED **, 31 passed, 0 failed, 0 skipped.
• Full native unit bundle: exit 0, ** TEST SUCCEEDED **, 369 passed, 0 failed, 1 skipped (370 total).

Standing state of the earlier findings: the two High findings are fixed (e45d52bcc), the quoted-title Medium is fixed (9803983 plus this repair), and the cache-freshness Medium remains a documented, deliberate trade-off explained above. The on-device accepted behaviour (Test 89, built from 90b3ac41f) is unchanged by any of these repairs.

[saphid]

Fourth round, against c8e303d77.

Fixed in a91412b12 — Macroscope Medium, angle-bracketed destinations: CommonMark treats <...> contents as literal, so ![plot](<out/plot).png>) is a valid reference, but the parentheses inside were being counted and the destination ended early, falling back to alternative text. Angle-bracket contents are now skipped when they open the destination. Focused tests cover an unmatched parenthesis inside angle brackets and an angle-bracketed path containing spaces followed by a title.

Re-run at a91412b12, isolated temporary DerivedData, real exit statuses:

• Focused suites: exit 0, ** TEST SUCCEEDED **, 32 passed, 0 failed, 0 skipped.
• Full native unit bundle: exit 0, ** TEST SUCCEEDED **, 370 passed, 0 failed, 1 skipped (371 total).

Standing state: all High findings fixed (e45d52bcc); quoted-title, apostrophe-in-path, and angle-bracket parsing Mediums fixed (9803983, c8e303d77, a91412b12); the cache-freshness Medium remains the documented, deliberate trade-off described in the PR body, since invalidating it correctly needs a server-provided content version for workspace assets. The on-device accepted behaviour (Test 89, built from 90b3ac41f) is unchanged by any of these repairs.

Every parsing repair in this series degrades to prior behaviour rather than to a wrong image: a reference the parser cannot read stays alternative text.

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit a91412b. Configure here.}

[saphid]

Fifth round, against a91412b12. Both bots found the two halves of the same gap in my angle-bracket handling, and both were right.

Fixed in f61d3a359:

• Cursor Bugbot, Medium — literal mode only ran when < sat immediately after (. CommonMark allows whitespace there and the destination reader already trims it, so ![p]( <out/pad).png> ) still ended the destination early. Literal mode now starts wherever the brackets actually open the destination.
• Macroscope, Medium — the destination reader used a plain search for > while the delimiter scan used an escape-aware one, so ![p](<out/a\>b.png>) truncated the path to out/a\. Both sides now share the one unescaped scan, which is what made the two halves disagree in the first place.

Re-run at f61d3a359, isolated temporary DerivedData, real exit statuses:

• Focused suites: exit 0, ** TEST SUCCEEDED **, 33 passed, 0 failed, 0 skipped.
• Full native unit bundle: exit 0, ** TEST SUCCEEDED **, 371 passed, 0 failed, 1 skipped (372 total).

Standing state: all High findings fixed in e45d52bcc; every parsing Medium fixed across 9803983, c8e303d77, a91412b12, and f61d3a359; the cache-freshness Medium remains the documented, deliberate trade-off explained in the PR body. The on-device accepted behaviour (Test 89, built from 90b3ac41f) is unchanged by any of these repairs, and every parsing repair degrades to prior behaviour — a reference the parser cannot read stays alternative text rather than becoming a wrong image.

[t3dotgg]

This definitely needs before/after screenshots, ideally also a video of the loading in behavior (i hate layout shifts)

[saphid]

This definitely needs before/after screenshots, ideally also a video of the loading in behavior (i hate layout shifts)

added

[saphid]

Closing as superseded by Theo’s target commit cd2b2f8. The conflict pass selected a rebase, then aborted it when the live target proved to own the same Markdown image parsing, workspace/direct resolution, bounded loading, transcript context, and focused coverage. No branch bytes changed, no force-push occurred, and the newer upstream implementation was preserved exactly. The attempted focused local target run did not execute tests because of recorded Xcode code-signing and disk-headroom failures, so no validation pass or review request is claimed; the updated description contains the full evidence and limitations.