The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
v0-uvai	Ready	Preview, v0	Aug 2, 2026 5:40pm

🔍 PR Validation

⚠️ Large PR detected (586 lines changed)

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.

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

Dependency Review

Snapshot Warnings

⚠️: No snapshots were found for the head SHA a124bde.

Scanned Files

Agent Completion Truth Gate: BLOCKED

Reasons: invalid_payload

{
  "details": {
    "invalid_fields": [
      "policy.agent_login",
      "policy.run_id"
    ]
  },
  "reasons": [
    "invalid_payload"
  ],
  "verdict": "blocked"
}

Workflow evidence

Status at head 2355d04

Code checks are green. lint-python, lint-frontend, guards, validate, bandit, python-safety, npm-audit, gitleaks, Security Scan – javascript, dependency-review, and Canonical issue and evidence all pass; the Vercel preview is Ready; test/build/trivy/Security Scan – python/coverage are still running and green so far. Locally: 86 focused tests and 520 across the code-generator / video-processing / deployment / AI-generator suites pass; ruff clean; mypy adds zero new diagnostics.

The one red check is agent-completion/truth-gate (invalid_payload). Its verdict flags issue.number, policy.agent_login, policy.run_id — i.e. a missing agent-lock provenance manifest. This branch is claude/*, so the gate classifies the PR as agent-authored and requires a frozen pre-dispatch intent plus a trusted terminal agent result linked through that manifest.

I am not hand-filling that manifest to turn the gate green. Doing so would fabricate provenance the gate exists to verify, which the repo's policy explicitly forbids (REAL_MODE_ONLY; "do not weaken or impersonate the gate", as noted on #810). Even a well-formed payload would then fall to missing_agent_result — the same governance blocker currently on #810, #1242, and #1251 — because the frozen-intent and trusted-result artifacts are produced by the orchestration/human sign-off path, not by this fix.

What this PR is: it resolves CodeRabbit's blocking cancellation finding on #1251 — a request cancelled mid-scaffold could leak a uvai_project_* directory once the writes moved off the event loop. The write hops now shield their worker and wait for it to settle before propagating cancellation, and generate_project removes the orphaned scaffold on cancellation/failure.

Next step is human: review, then decide the truth-gate/merge disposition to protected main. I won't force the provenance gate.

Generated by Claude Code

@coderabbitai review

This PR resolves the blocking cancellation finding you raised on #1251: a request cancelled mid-scaffold could leak a uvai_project_* directory once the writes moved off the event loop. Please verify specifically that:

1. _run_offloop shields the worker and waits for it to settle before propagating CancelledError, so cleanup never races a live writer;
2. generate_project removes the orphaned scaffold on both cancellation and failure, before re-raising;
3. happy-path output is unchanged (byte-for-byte equality from perf: batch project scaffolding disk writes off the event loop #1251 preserved).

Generated by Claude Code

#!/bin/bash
set -euo pipefail

printf '%s\n' '=== PR diff ==='
git diff --find-renames --unified=80 'origin/main...HEAD' -- \
  src/youtube_extension/backend/code_generator.py \
  tests/unit/test_code_generator.py

printf '%s\n' '=== implementation outline ==='
ast-grep outline src/youtube_extension/backend/code_generator.py --items all --view expanded

printf '%s\n' '=== test outline ==='
ast-grep outline tests/unit/test_code_generator.py --items all --view expanded

printf '%s\n' '=== relevant definitions and call sites ==='
rg -n -C 5 -P '\b(_run_offloop|generate_project|_apply_write_plan)\s*\(' \
  src/youtube_extension/backend/code_generator.py \
  tests/unit/test_code_generator.py

printf '%s\n' '=== output/equality assertions ==='
rg -n -i -C 4 'byte|equal|equality|expected|write.plan|generate_project' \
  tests/unit/test_code_generator.py

#!/bin/bash
set -euo pipefail

printf '%s\n' '=== locate focused test file ==='
fd -a -t f '^test_code_generator\.py$' .

printf '%s\n' '=== generate_project and off-loop helper ==='
nl -ba src/youtube_extension/backend/code_generator.py | sed -n '25,265p'

printf '%s\n' '=== current write plans ==='
nl -ba src/youtube_extension/backend/code_generator.py | sed -n '275,460p'

printf '%s\n' '=== baseline write sequences ==='
git show origin/main:src/youtube_extension/backend/code_generator.py \
  | nl -ba \
  | sed -n '175,445p'

printf '%s\n' '=== test names and cancellation/equality coverage ==='
fd -a -t f '^test_code_generator\.py$' . -x sh -c '
  printf "\n--- %s ---\n" "$1"
  rg -n -i -C 8 "mkdtemp|cancel|orphan|offloop|byte|equal|verbatim|write.plan" "$1"
' sh {}

@groupthinking I found one blocking cancellation leak.

_run_offloop correctly shields each write worker and waits for it to finish before it re-raises CancelledError. This prevents _safe_rmtree from racing an active _apply_write_plan worker.

However, generate_project creates the cleanup scope only after this statement completes:

temp_dir = await _run_offloop(tempfile.mkdtemp, prefix="uvai_project_")

If cancellation occurs while the mkdtemp worker runs, _run_offloop waits until mkdtemp creates the directory and then raises CancelledError. temp_dir is not assigned, and the inner try/except that calls _safe_rmtree is never entered. The newly created uvai_project_* directory leaks.

Please make the mkdtemp hop own cleanup when cancellation occurs after directory creation, or otherwise retain its result long enough to remove it before propagating cancellation. Add a regression test that blocks mkdtemp, cancels the generation task, releases mkdtemp, and asserts that the created directory no longer exists.

The diff preserves the existing write order and uses equivalent content generation on the normal paths. I could not independently verify the focused test file because tests/unit/test_code_generator.py was not present in this checkout.

Fixed in a124bde. Both CodeRabbit and Vercel VADE independently flagged the same real leak: the cleanup try began only after temp_dir = await _run_offloop(tempfile.mkdtemp, ...), so a cancellation delivered during the mkdtemp hop created the directory and raised CancelledError before the assignment — the path never entered the cleanup scope.

The mkdtemp hop now owns that window via _make_scaffold_dir: it shields the worker and, on cancellation, removes any directory the worker created before propagating CancelledError. generate_project obtains the scaffold path through it, so both the creation window and the write phase are leak-safe.

Verification: added a regression test that blocks mkdtemp, cancels the task, releases mkdtemp, and asserts the directory is gone — prove-fail confirmed it fails against the prior leaky path and passes now. 87 focused tests pass; ruff clean; mypy adds no new diagnostics.

The agent-completion/truth-gate remains the only non-code blocker (provenance/human sign-off, as noted above) — unchanged by this commit.

Generated by Claude Code

Superseded by #1251, which carries both defects this PR fixes plus the off-loop conversion of the 28 scaffolding writes.

Full rationale and a side-by-side comparison are in #1251 (comment). Summarised: the two branches converged independently on the same shield-and-drain design derived from _run_sync_rpc, which is good corroboration. #1251 is kept because _discard_project_dir catches Exception whereas _safe_rmtree here relies on rmtree(ignore_errors=True), which is not total — a NUL byte in a path raises ValueError before the error callback is consulted. On the failure path that could mask the original exception.

Credit where due: this PR was ahead of #1251 on the generic-exception path, and that gap was real. It is now closed there in d178e210a.

Please reopen if #1251 stalls.