fix(pixel): a stopped turn links an earlier preview only when it continues that project

[Lightheartdevs]

Problem

Every Pixel turn is seeded with the chat's latest publication (state.workspaceLastVerifiedPreview ??= previousPreview in observeRun). The seeding comment says that snapshot "neither verifies current files nor grants continuation scope", but two delivery paths still used it for any turn that needed a preview:

• progressStopPreview (tool-loop fuse stops, the answer turn's allowed URL, and deliveryVerificationForRun), and
• the no-publication branch of verificationForRun ("Your last published preview is still available").

So a stopped turn working on a new project linked the earlier project's preview and carried its receipt. Portal treats that receipt as the turn's own publication: it puts a publication card on the message, and that card decides which project the chat is grouped under.

Round-114 evidence (Qwen3.5-9B, main 5afd46c, mac-mini target)

The chat's turns, in order (round-114/mac-mini/pixel-portal-evidence/):

1. website-create: published site-52af184a....
2. website-edit ("Update that same website ..."): published site-6ba777bafd52589578b82c5a, relativeDirectory fleet-qualification-6419c769b873.
3. files-code: "In a separate workspace directory fleet-qualification-6419c769b873-calc ...". The model wrote ...-calc/solve.py with literal \n, then repeated the same write until the fuse stopped the run. It never read, wrote or published anything in the website directory.
4. coding-v1: "In a new workspace project fleet-qualification-6419c769b873-coding ...". Writes and unittest runs happened only in ...-coding, and the fuse stopped the run.

Both stopped turns (files-code.answer.txt, coding-v1.answer.txt) got the same reply:

This response was stopped after repeated tool failures or attempts without progress. ...

[Open last published preview](http://site-6ba777bafd52589578b82c5a.localhost:9437/site-6ba777bafd52589578b82c5a/)

This is the last verified publication, not proof that all requested work completed.

The terminal SSE chunk of both turns (files-code.sse.body line 45, coding-v1.sse.body line 118) carried pixel.preview = {relativeDirectory: "fleet-qualification-6419c769b873", siteId: "site-6ba777bafd52589578b82c5a", ...}. That is the website's receipt, delivered as the publication of a calc turn and a coding turn.

Fix

lastVerifiedPreviewForTurn(state) in plugin/tool-loop-guard.mjs returns the last verified preview only in two cases:

• this run published it. workspacePreviewVerifiedDirectory is set only when this run's own publication is later invalidated.
• this turn continues that project. It counts as continuing when any of these holds:
  • this turn's preview target workspacePreviewDirectory is that directory. It is set by "update that same website" continuation, a named-preview request, a successful mutation inside the published directory, or a publication attempt.
  • the owner named that directory, or its parent, in /workspace/... (workspaceTaskDirectory).
  • every successful write and edit this turn is in that project's tree. This uses the same ancestry rule as the historical-entry readback, now shared as inPublishedProjectTree. It covers sources next to a generated public/ output.

In every other case the stop drops both the link text and the receipt. progressStopPreview and the verificationForRun branch both use the helper. In that branch the coding-stop text and the failed or pending check text are now reported whether or not a link is offered. Before, they appeared only next to the historical link, so a preview-required turn with no publication and no same-project snapshot showed only the generic "did not create/verify a preview" text.

The positive case is unchanged: an "Update that same website" turn that stops still offers its last verified preview and receipt.

Portal

I checked dashboard/src/pages/Pixel.jsx and did not change the dashboard.

• finishAttempt sets message.publication, and calls setPreview, only when a pixel.preview frame parses (parseVerifiedPreviewFrame).
• With no frame the stopped message gets no publication card and no beforePublication. conversationProject falls back to earlier messages.
• The workspace panel keeps showing the chat's current preview, which is the website from turn 2. That is chat-level workspace state; it does not attribute the preview to the stopped turn.
• Ingress (host/pixel_ingress.mjs completionSse) emits the pixel marker only when verification.preview is set, so dropping the receipt in the guard removes the frame.

Tests

• tests/progress_finalization.test.mjs

  • New replay of the round: run 1 publishes night-garden. Run 2 in the same session then works in night-garden-calc (with a write, and before any write) or night-garden-coding and exhausts the fuse. Both the stop delivery and the verdict have no link and no receipt. The answer-turn variant composes without the old URL.
  • New positive control: run 2 edits night-garden ("Update that same website", and a direct edit of night-garden/index.html) and stops. The link and receipt are kept.
  • callTool and verifiedPreview take an optional context.

• tests/tool_loop_guard.test.mjs

  • Three existing tests asserted the leak: "Build and publish another website in a new directory" kept the old site, and a /workspace/code coding stop linked log-viewer-lab. They now expect no link for the other project. They add same-project cases where the link is kept (continuation prompt; a coding stop in /workspace/log-viewer-lab), and still require the coding-stop and failed-check text.
  • New test: a stop keeps expense-report/public for a source edit in expense-report/, and drops it for another-project/ and for the prefix sibling expense-report-v2/.

• Red check against main's plugin with these tests: 5 fail.

  • a stop in a different project neither links nor carries the earlier publication
  • a stop offers the historical publication only for source edits in that project's tree
  • historical preview cannot hide a later coding stop from the final reply
  • later failed preview work retains only the same session and project's historical publication
  • progress stops scope historical preview delivery to the current owner request

  The positive control passes on both main and this branch.

• Full Linux pixel-agent suite: head 3206 tests / 3168 pass / 0 fail. Head merged with main: same.

Composition with open PRs

• git merge-tree against main plus each PR is clean for fix(pixel): a harness revision or continuation prompt keeps the owner's request scope #6756, fix(pixel): remove the owner's private operating contract from the workspace template #6755, fix(pixel): keep a passing show/hide proof across later passing checks of the same snapshot #6754, fix(pixel): a refused rm -rf keeps a current passing test, and the owner gets a receipt instead of guard text #6753, fix(pixel): judge an exit-0 unittest run by the runner's own summary; a coding stop takes precedence over test-command coaching #6752, fix(pixel): capsule role/name locators match Playwright's source-text names as well as Chromium's text-transformed names #6750, fix(pixel): every non-passing preview inspection returns a ready, validated next call or a measured diagnosis #6748, fix(pixel): steer preview tools to their direct route and answer wrapped control tools #6747, fix(pixel): plan within the output limit, continue a cut Portal turn once, report otherwise #6743, fix(pixel): answer after an unchanged file and check preview bytes when a result is saved #6732, fix(pixel): accept year-less event page dates the page does not contradict #6730, fix(pixel): read PDF text in the shared page reader; binary bodies are never page reads #6724, fix(pixel): let the turn continue after an in-project rm -rf is refused #6722 and feat(pixel): search and read public pages in one call (pixel_ods_search_read) #6698.
• Defer a host-guided inspection failure that would stop the run for exactly one corrected attempt, with every hard bound kept #6749, fix(pixel): count a passing inspection of the snapshot awaiting its byte check #6726, fix(pixel): check that owner-requested text is visible when the preview loads #6723, fix(pixel): stop whole-file rewrites and test bisection after failed tool results (output efficiency C1-C4) #6716 and Improve Pixel task execution, evidence reuse, and visible progress #6645 already conflict with main. Their conflicts are the same against this branch, and none of the conflicting hunks contain this change.
• The full suite was also run on main + fix(pixel): a harness revision or continuation prompt keeps the owner's request scope #6756, main + fix(pixel): judge an exit-0 unittest run by the runner's own summary; a coding stop takes precedence over test-command coaching #6752 and main + fix(pixel): steer preview tools to their direct route and answer wrapped control tools #6747 with this branch. Those three touch related state: owner scope for continuations, coding-stop precedence, and preview-directory guidance. Results (0 fail each):
  • fix(pixel): a harness revision or continuation prompt keeps the owner's request scope #6756: 3208 tests / 3170 pass
  • fix(pixel): judge an exit-0 unittest run by the runner's own summary; a coding stop takes precedence over test-command coaching #6752: 3277 tests / 3239 pass
  • fix(pixel): steer preview tools to their direct route and answer wrapped control tools #6747: 3247 tests / 3209 pass

🤖 Generated with Claude Code