diff --git a/.agents/skills/import-upstream-pr/SKILL.md b/.agents/skills/import-upstream-pr/SKILL.md
new file mode 100644
index 000000000000..2430aa5cbad2
--- /dev/null
+++ b/.agents/skills/import-upstream-pr/SKILL.md
@@ -0,0 +1,156 @@
+---
+name: import-upstream-pr
+description: Evaluate, port, validate, and deliver an open-unmerged or closed-unmerged pingdotgg/t3code pull request into LastCode while preserving exact upstream provenance. Use when LastCode wants an existing upstream PR before it merges, after it closes without merge, or when re-evaluating a force-pushed upstream candidate. Do not use to author a new upstream contribution; use upstream-fix. Do not use for a fork-only change with no upstream PR; use lastcode-pr.
+---
+
+# Import Upstream PR
+
+Adopt one existing T3 Code PR as an independently owned LastCode change. Treat
+the upstream PR as a pinned candidate and evidence source, never as a substitute
+for LastCode review, validation, or merge authority.
+
+## Establish the Boundary
+
+1. Read the repository `AGENTS.md` and `docs/lastcode/fork-conventions.md`.
+2. Read [references/intake-and-evidence.md](references/intake-and-evidence.md)
+   before capturing the candidate.
+3. Keep the product port separate from notes, workflow documentation, and skill
+   changes. Use separate branches or worktrees.
+4. Do not infer permission to push, open a PR, use a browser, or merge. Obtain
+   the authority required by the repository instructions for each operation.
+
+## Pin and Inspect the Candidate
+
+1. Capture the upstream PR URL, state, observed date, title, author, base, exact
+   `headRefOid`, ordered commits, changed files, checks, reviews, review threads,
+   linked issues, and closure reason when closed.
+2. Force-fetch `pull/<number>/head` into a dedicated remote-tracking ref so a
+   previously cached ref cannot reject a legitimate upstream force-push.
+   Require the fetched SHA to equal the captured `headRefOid`. Never import a
+   moving branch name.
+3. Derive the complete, oldest-first commit list from the pinned base and head
+   Git objects. Record its count and require its final SHA to equal
+   `headRefOid`. Do not use GitHub's PR commits response as the source of truth;
+   that endpoint is capped at 250 commits even when paginated.
+4. Inspect metadata and the full diff before installing dependencies or running
+   any code controlled by the PR.
+5. Check current `upstream/main` and `origin/lastcode/main` for the same behavior,
+   a replacement, or an incompatible design.
+6. Treat a force-pushed upstream head as a new candidate. Range-diff it against
+   the prior pinned head and repeat the applicable review and validation.
+
+If `origin/lastcode/main` already contains the exact behavior or an accepted
+replacement, stop before creating an evaluation branch and report the candidate
+as already adopted or superseded. If only `upstream/main` contains it, decide
+whether normal nightly reconciliation is sufficient before creating a manual
+port.
+
+## Decide Eligibility
+
+- For an open, unmerged PR, decide whether LastCode benefits enough to adopt it
+  now instead of waiting.
+- For a closed, unmerged PR, establish why it closed. Inactivity, contribution
+  policy, or maintainer bandwidth can be acceptable. Incorrectness,
+  supersession, or rejected direction requires an explicit LastCode divergence
+  decision.
+- Pause for hidden dependencies, unresolved correctness findings, unclear
+  product value, unexplained closure, or scope too large to validate
+  proportionally.
+
+Record the adoption decision. Upstream CI is supporting evidence only.
+
+## Build an Isolated Evaluation
+
+1. Fetch `origin` with pruning and create a clean worktree on
+   `pr/upstream/<number>-<slug>` from the exact `origin/lastcode/main`.
+2. Apply the pinned Git graph's ordered commit list with `git cherry-pick -x`.
+   Preserve authorship and commit boundaries when the stack is coherent. If
+   the range contains merge commits, stop and plan an ancestry-aware import or
+   a reimplementation instead of flattening it blindly.
+3. Do not merge the upstream PR branch; that drags its base history into
+   LastCode.
+4. If the stack does not fit current LastCode, reimplement only the coherent
+   behavior and record the upstream PR URL, pinned SHA, and why cherry-pick was
+   unsuitable.
+
+Classify integration honestly:
+
+- changed-path overlap is a risk signal, not a conflict;
+- a clean auto-merge still needs semantic review in current LastCode context;
+- a textual conflict requires explicit resolution review;
+- a clean textual application can still have a semantic conflict.
+
+Treat every conflict resolution and downstream adaptation as first-party code.
+
+## Validate the Port
+
+1. Review every changed line against current LastCode and the linked problem.
+2. Walk the affected entry points, clients, providers, contracts, reverse
+   states, connection modes, performance concerns, and docs. Mark each
+   non-applicable surface explicitly.
+3. Complete dependency installation before starting checks. A partially
+   completed install is not evidence; require its zero exit and terminal
+   completion.
+4. Run focused behavior tests, targeted lint, the affected package typecheck,
+   and `git diff --check <destination-base> <port-head>` under the repository's
+   canonical toolchain. A bare `git diff --check` does not inspect an already
+   committed port. Record the toolchain command and versions with the receipt.
+5. Add focused regression tests for backend or automation behavior. Do not run
+   repo-wide checks merely for intake.
+6. For user-visible behavior, obtain browser/computer-use approval and use the
+   applicable repository app-testing skill against disposable state. Capture
+   matched LastCode-specific before/after evidence and identify the real client,
+   viewport, and any authorized fallback accurately.
+7. Keep one live app tab when shared browser storage can make cross-tab state
+   nondeterministic. Verify route, persisted state, and rendered state rather
+   than trusting a click result alone.
+
+When accepted, rename the branch to `port/upstream/pr-<number>-<slug>`.
+
+## Refresh and Deliver
+
+1. Immediately before delivery, fetch `origin/lastcode/main` again. If the port
+   parent moved, rebase the imported commits and rerun affected validation.
+2. Before the guarded push, require a clean worktree and record the local head,
+   destination base, and existing remote topic SHA (or its absence). Then push
+   in a clean environment so the pre-push `pnpm lastcode:ci:quick` gate sees
+   ordinary Git/SSH variables. Do not export `GIT_SSH_COMMAND` or inject an SSH
+   command through Git configuration around the push; those settings flow into
+   tests that intentionally control `GIT_SSH`.
+3. If Quick CI passes but the idle SSH transport subsequently dies, do not
+   treat its generic success line as an exact-head receipt. Require the local
+   head to equal the recorded head, the worktree to remain clean, the fetched
+   destination base to equal the recorded base, and the remote topic SHA (or
+   absence) to remain unchanged. Only then retry that recorded head with
+   `--no-verify` and an exact `--force-with-lease` tied to the recorded remote
+   topic state. Otherwise rerun the guarded push and its hook.
+4. Open a PR targeting `lastcode/main` only when explicitly requested. Include
+   the upstream PR and pinned head, observed state/date, import method,
+   adaptations, rationale, validation, closure/review context, and published
+   evidence.
+5. For review and merge, follow `lastcode-pr` and
+   `.agents/skills/_references/external-review-mechanics.md`: require a terminal
+   clean Codex result for the exact head, zero unresolved threads, and a full
+   `pnpm lastcode:ci` stamp for the exact head/current base. Merge only through
+   `pnpm lastcode:merge`.
+6. Verify the merged commit on `origin/lastcode/main`. For a squash merge,
+   compare stable patch IDs so provenance verification does not depend on the
+   topic commit remaining an ancestor.
+
+## Reconcile Later Upstream Movement
+
+- If upstream later merges an identical patch, let normal nightly reconciliation
+  remove the duplicate.
+- If upstream merges a changed version, compare it with the pinned imported head
+  and port only the desired delta.
+- If upstream closes unmerged, retain or remove the behavior according to
+  LastCode's product decision, not the state transition alone.
+- Preserve the upstream URL and pinned SHA in the LastCode PR so future sync work
+  can explain the source.
+
+## Handoff
+
+Report the source PR state and pinned head, destination base and port head,
+import method and adaptations, validation and real-client evidence, PR/merge
+state, exact-head review result, unresolved-thread count, full-CI stamp, and the
+merged commit or remaining blocker.
diff --git a/.agents/skills/import-upstream-pr/agents/openai.yaml b/.agents/skills/import-upstream-pr/agents/openai.yaml
new file mode 100644
index 000000000000..c5dc80e0ea85
--- /dev/null
+++ b/.agents/skills/import-upstream-pr/agents/openai.yaml
@@ -0,0 +1,4 @@
+interface:
+  display_name: "Import Upstream PR"
+  short_description: "Evaluate and port unmerged T3 Code PRs"
+  default_prompt: "Use $import-upstream-pr to evaluate and import a pinned T3 Code pull request into LastCode."
diff --git a/.agents/skills/import-upstream-pr/references/intake-and-evidence.md b/.agents/skills/import-upstream-pr/references/intake-and-evidence.md
new file mode 100644
index 000000000000..057a67ef5130
--- /dev/null
+++ b/.agents/skills/import-upstream-pr/references/intake-and-evidence.md
@@ -0,0 +1,157 @@
+# Intake and Evidence Reference
+
+Use this reference to create a reproducible receipt before executing upstream
+PR code and to keep delivery claims tied to exact commits.
+
+## Candidate receipt
+
+Capture PR metadata before fetching or executing the candidate:
+
+```bash
+set -e
+
+repo=pingdotgg/t3code
+pr=<number>
+
+metadata=$(gh pr view "$pr" --repo "$repo" --json \
+  number,title,url,state,isDraft,author,baseRefName,baseRefOid,headRefOid,\
+  mergeable,mergeStateStatus,changedFiles,statusCheckRollup,\
+  reviewDecision,labels,body,closedAt,mergedAt,updatedAt)
+
+printf '%s\n' "$metadata"
+base_ref=$(printf '%s' "$metadata" | jq -r .baseRefName)
+base_sha=$(printf '%s' "$metadata" | jq -r .baseRefOid)
+head_sha=$(printf '%s' "$metadata" | jq -r .headRefOid)
+
+git fetch upstream \
+  "+refs/pull/$pr/head:refs/remotes/upstream/pr/$pr"
+
+fetched_head=$(git rev-parse "refs/remotes/upstream/pr/$pr")
+if ! git cat-file -e "$base_sha^{commit}" 2>/dev/null; then
+  git fetch upstream "$base_sha" ||
+    printf 'exact base SHA was not directly fetchable: %s\n' "$base_sha" >&2
+fi
+git cat-file -e "$base_sha^{commit}" || {
+  printf 'missing pinned base commit: %s\n' "$base_sha" >&2
+  exit 1
+}
+test "$fetched_head" = "$head_sha" || {
+  printf 'PR head changed while pinning: expected %s, fetched %s\n' \
+    "$head_sha" "$fetched_head" >&2
+  exit 1
+}
+
+if git ls-remote --exit-code --heads upstream \
+  "refs/heads/$base_ref" >/dev/null; then
+  git fetch upstream \
+    "+refs/heads/${base_ref}:refs/remotes/upstream/${base_ref}"
+  current_base_sha=$(git rev-parse "refs/remotes/upstream/$base_ref")
+else
+  current_base_sha=unavailable
+fi
+
+commit_count=$(git rev-list --count "$base_sha..$head_sha")
+final_sha=$(
+  git rev-list --reverse --topo-order "$base_sha..$head_sha" | tail -n 1
+)
+test "$commit_count" -gt 0 || {
+  printf 'pinned PR range is empty\n' >&2
+  exit 1
+}
+test "$final_sha" = "$head_sha" || {
+  printf 'commit walk did not end at pinned head: %s\n' "$head_sha" >&2
+  exit 1
+}
+printf '{"count":%s,"final_sha":"%s","current_base_sha":"%s"}\n' \
+  "$commit_count" "$final_sha" "$current_base_sha"
+git rev-list --reverse --topo-order "$base_sha..$head_sha"
+```
+
+Fetching the PR head is independent of fetching its named base branch because
+a closed PR can outlive a deleted target branch. The leading `+` refspecs allow
+cached PR and live-base refs to follow legitimate force-pushes. `baseRefOid`
+can predate the current base tip; require that exact commit object from the PR
+history or a direct SHA fetch rather than equating it with `current_base_sha`.
+If the named base no longer exists, record it as unavailable instead of failing
+the already pinned head. Requiring the fetched PR ref to equal `headRefOid`
+detects a head race; restart the capture if it changed. Git's `base..head` walk
+is the complete source of truth because the GitHub PR commits endpoint returns
+at most 250 commits, while `gh pr view --json commits` exposes only the first 100. Record the oldest-first `--topo-order` list and its count. Also record:
+
+- observed date and timezone;
+- ordered commit SHAs and authors;
+- changed file list and diff stat;
+- linked issues and claimed behavior;
+- all formal reviews and issue comments;
+- all review threads, including unresolved and outdated threads;
+- upstream check results;
+- closure reason or superseding change when closed; and
+- current fetched `upstream/main` and `origin/lastcode/main` SHAs.
+
+Use the paginated review and thread queries in
+`.agents/skills/_references/external-review-mechanics.md`. Do not infer a clean
+candidate from an empty GitHub summary.
+
+## Duplicate and integration checks
+
+Before adoption, compare the candidate against both current destinations:
+
+- ancestor containment answers whether the exact commit exists;
+- stable patch IDs help find identical patches with rewritten commits;
+- code/search and linked-issue inspection find replacements with different
+  implementations;
+- path overlap identifies files needing closer review but does not prove a
+  conflict;
+- a merge-tree preview or the isolated cherry-pick establishes textual
+  applicability; and
+- source review establishes semantic compatibility.
+
+For multiple commits, preserve the pinned Git graph's oldest-first topological
+order. Check `git rev-list --merges "$base_sha..$head_sha"` before cherry-pick;
+if it is non-empty, plan an ancestry-aware import or reimplementation rather
+than flattening merge commits blindly. Use `git range-diff` when the candidate
+head changes or when validating a rebased port.
+
+## Validation receipt
+
+Tie every result to the exact port head and destination base. Record:
+
+- toolchain activation command plus Node and package-manager versions;
+- dependency install command and terminal success;
+- focused commands and test counts;
+- `git diff --check <destination-base> <port-head>` result for the committed
+  port range;
+- affected-surface matrix with explicit non-applicable entries;
+- integrated client, disposable-state boundary, viewport, route/state/rendered
+  assertions, and before/after artifact paths or published URLs;
+- any fallback browser system and the approval that allowed it; and
+- clean worktree, local head, remote head, and destination base at publication.
+
+Do not convert missing evidence into a product pass or failure. Mark the
+assertion `blocked`, explain the evidence gap, and either obtain authority for a
+fallback or hand it off.
+
+## LastCode PR body checklist
+
+Include:
+
+- upstream PR URL, title, author, observed state/date, and pinned head;
+- linked issue and LastCode adoption rationale;
+- `cherry-pick -x` or reimplementation method;
+- conflict resolutions and LastCode-specific adaptations;
+- unresolved upstream review or closure context;
+- focused and integrated validation; and
+- durable GitHub-hosted visual evidence for UI changes.
+
+Before merge, refresh the PR snapshot and require the same exact head across the
+clean Codex result, zero unresolved threads, full-CI stamp, and merge command.
+
+After a squash merge, verify stable patch equivalence:
+
+```bash
+git diff <port-parent> <port-head> | git patch-id --stable
+git diff <base-before-merge> <merged-main> -- <changed-paths...> | git patch-id --stable
+```
+
+Matching patch IDs show the validated port became the merged change even though
+the original topic commit is not an ancestor of the squashed result.