[P1.58] Cut 0.9.4: PyPI 0.9.3 is 53 commits behind main, and CHANGELOG Unreleased omits most of the fixes

[frankbria]

Problem. pyproject.toml:7 is still 0.9.3, and git log v0.9.3..HEAD shows 53 commits. The published package lacks:

• [P2.35] CredentialManager blocks forever when the OS keyring backend is present but unresponsive #1181: unbounded keyring calls hang forever on headless, SSH or WSL hosts; v0.9.3:credentials.py:708 has no timeout, and this blocks cf pr * and cf auth *.
• fix(tokens): CallType had no verification_fix, so every self-correction record was dropped #1176: every self-correction token record is dropped, so costs are under-counted.
• [P3.28] cf proof run exits 0 when it verified nothing #1253, [P3.29] cf pr merge still uses workspace-global requirement scope #1254, [P3.30] PROOF9 requirement scopes can't express an absolute path, and silently match nothing #1258: PROOF9 exit code, merge-gate scoping and absolute capture paths.
• [P3.18] Escape the str-typed Typer arguments that error messages echo back #1206: markup crashes.

The CHANGELOG Unreleased section lists none of #1181, #1254, #1258, #1255, #1206, #1197 or #1160.

Impact. uv tool install codeframe-ai installs known-fixed bugs, and the README on main describes behaviour that is not released.

Acceptance criteria.

• The P0/P1 blockers in this audit land first.
• The CHANGELOG is brought up to date.
• scripts/quickstart-cleanroom/run.sh --source . passes, including step 6.
• The version is bumped and tagged, and the cleanroom check is re-run against the published 0.9.4.

---

Severity: medium · area: release · source: SaaS launch review 2026-09-25 (plan in tasks/todo.md).

Blocked by: #1262 (P0.34), #1263 (P1.44), #1264 (P1.45), #1265 (P1.46), #1266 (P1.47), #1267 (P1.48), #1268 (P1.49), #1269 (P1.50), #1270 (P1.51), #1271 (P1.52), #1272 (P1.53), #1273 (P1.54), #1274 (P1.55), #1275 (P1.56), #1276 (P1.57)

[frankbria]

#1262 merged in 31edbd3 (PR #1307), with its CHANGELOG Unreleased → Fixed entry. It is the README-install BUILD blocker: every cf work start --execute on uv tool install codeframe-ai ended in a blocker. It only reaches users once this release ships. Cleanroom evidence: scripts/quickstart-cleanroom/artifacts-1262 (--source) passes 6-work-start, where every published-package run failed it. Re-run the harness without --source after publishing 0.9.4 to confirm.

[frankbria]

P1.50 (#1269) is done in 64fb07a (#1312): the telemetry default is now telemetry.codeframe.sh and it's in CHANGELOG Unreleased. Note for the release: the collector host isn't live yet (#1311, needs-operator), so opted-in telemetry is dropped until it's deployed. That no longer blocks the release on security grounds.

[frankbria]

Status 2026-10-02. Release prep is in PR #1361: version 0.9.4, CHANGELOG brought up to date, README banner updated. The last blocker, #1276, merged as #1359.

Parked: scripts/quickstart-cleanroom/run.sh --source . needs a real ANTHROPIC_API_KEY, and the agent environment has none. Once that check passes and CI is green, the maintainer has approved merging #1361, pushing v0.9.4, and re-running the check against the published package.

#1348 (the Sonnet 5.5 defaults) ships separately in 0.9.5.

[frankbria]

Operator steps: tactical, in order (2026-10-06)

Priority 2 of 7. Blocked by PR #1361's CI going green. It was refreshed onto main today (9c9799e), with all 75 Unreleased entries folded into [0.9.4]. About 45 minutes, most of it two cleanroom runs. Costs a little Anthropic usage.

Facts already verified: the release workflow's model guard uses the staging environment's ANTHROPIC_API_KEY, which works (claude-code smoke passed on Oct 4). Publishing is PyPI Trusted Publishing, so there's no token to manage.

1. Confirm the PR is green.

gh pr checks 1361 --repo frankbria/codeframe --watch

2. Check out the release branch.

cd ~/projects/codeframe && git fetch origin
git checkout feature/issue-1277-release-0.9.4 && git pull --ff-only
grep '^version' pyproject.toml            # expect: version = "0.9.4"

3. Load the key without echoing it.

read -rs ANTHROPIC_API_KEY && export ANTHROPIC_API_KEY     # paste, then Enter

4. Run the cleanroom against the source tree. About 10 minutes. Needs Docker.

scripts/quickstart-cleanroom/run.sh --source scripts/quickstart-cleanroom/artifacts-0.9.4-source
# --source is a flag (it mounts this checkout). Do NOT write "--source .": the "." becomes the output folder.

5. Read the verdict. The script exits 0 even when a step failed, so read the files. Statuses are OK, FAIL or TIMEOUT. Four of the five archived runs, including the 0.9.3 PyPI check, recorded 6-work-start FAIL behind a clean finish. This check fails closed: a missing results file is "NO RESULTS", never a pass.

A=scripts/quickstart-cleanroom/artifacts-0.9.4-source
column -t -s$'\t' "$A/timings.tsv"
if [ ! -s "$A/timings.tsv" ] || [ ! -f "$A/findings.tsv" ]; then
  echo "NO RESULTS in $A - wrong folder, or the run did not finish. NOT SAFE TO TAG"
elif awk -F'\t' 'NR>1 && $2!="OK"' "$A/timings.tsv" | grep .; then
  echo "NOT SAFE TO TAG (failing steps above)"
elif grep -qi critical "$A/findings.tsv"; then
  grep -i critical "$A/findings.tsv"; echo "NOT SAFE TO TAG (critical findings above)"
else
  echo "SAFE: all $(($(wc -l < "$A/timings.tsv")-1)) steps OK, no critical findings"
fi

Gate: it must print SAFE: all N steps OK, no critical findings. Anything else means stop.
If anything fails, post the evidence: gh issue comment 1277 --repo frankbria/codeframe --body-file "$A/findings.tsv"

6. Set the release date in the CHANGELOG heading, then wait for CI again.

sed -i "s/^## \[0.9.4\] - 2026-10-02$/## [0.9.4] - $(date -u +%F)/" CHANGELOG.md
git diff --stat                            # expect: 1 line in CHANGELOG.md
git commit -am "docs(changelog): 0.9.4 release date (#1277)" && git push
gh pr checks 1361 --repo frankbria/codeframe --watch

7. Merge.

gh pr merge 1361 --repo frankbria/codeframe --squash --delete-branch

8. Tag from main.

git checkout main && git pull --ff-only
grep '^version' pyproject.toml            # expect 0.9.4: the release job refuses a tag that disagrees
git tag -a v0.9.4 -m "codeframe-ai 0.9.4" && git push origin v0.9.4

9. Watch the release. It runs Unlocked resolution, then Build (with the model guard and a wheel smoke test), then Publish to PyPI. If GitHub asks you to approve the pypi environment, approve it.

sleep 10
REL=$(gh run list --repo frankbria/codeframe --workflow release.yml -L1 --json databaseId -q '.[0].databaseId')
gh run watch "$REL" --repo frankbria/codeframe

10. Confirm PyPI has it. It can take a minute.

curl -s https://pypi.org/pypi/codeframe-ai/json | jq -r .info.version     # expect: 0.9.4

11. Run the cleanroom against the published package. No --source, so it installs from PyPI.

scripts/quickstart-cleanroom/run.sh scripts/quickstart-cleanroom/artifacts-pypi-0.9.4
A=scripts/quickstart-cleanroom/artifacts-pypi-0.9.4
if [ ! -s "$A/timings.tsv" ] || [ ! -f "$A/findings.tsv" ]; then
  echo "NO RESULTS in $A - wrong folder, or the run did not finish. NOT SAFE TO TAG"
elif awk -F'\t' 'NR>1 && $2!="OK"' "$A/timings.tsv" | grep .; then
  echo "NOT SAFE TO TAG (failing steps above)"
elif grep -qi critical "$A/findings.tsv"; then
  grep -i critical "$A/findings.tsv"; echo "NOT SAFE TO TAG (critical findings above)"
else
  echo "SAFE: all $(($(wc -l < "$A/timings.tsv")-1)) steps OK, no critical findings"
fi

It must print SAFE: …. Anything else means the published 0.9.4 is broken: post $A/findings.tsv here and stop.

12. Close.

gh issue close 1277 --repo frankbria/codeframe --comment "Released 0.9.4 (release run $REL). Cleanroom green on source and on the published package."

If step 9 fails after the tag is pushed: if nothing reached PyPI, git push --delete origin v0.9.4 && git tag -d v0.9.4, fix, and re-tag. If 0.9.4 is already on PyPI, it's immutable: ship the fix as 0.9.5.

Next: #1348 (0.9.5), which depends on this release.

Edited 2026-10-06: step 4's command was wrong (--source . wrote the results into the repo root), and the step 5/11 check printed a pass when the results file was missing. Both are fixed: the check now fails closed.

[frankbria]

Released 0.9.4 (tag v0.9.4, on PyPI). Cleanroom on source: 17/17 OK (artifacts-0.9.4-source). Cleanroom on the published package: install, cf --help, cf --version and cf init OK; stopped by the operator during 4-prd-generate after 36 repetitive discovery questions, the unbounded-discovery bug tracked as #1443 (P1, ships in 0.9.5). The full PyPI cleanroom re-runs with 0.9.5.