Skip to content

release: 0.12.2, provider tokens scrubbed by format - #439

Merged
blitzcrieg1 merged 2 commits into
masterfrom
release/0.12.2
Nov 16, 2026
Merged

blitzcrieg1 merged 2 commits into
masterfrom
release/0.12.2

Conversation

@blitzcrieg1

Copy link
Copy Markdown
Owner

A privacy release ahead of the detection pass. Both fixes matter only with command logging or full-argument capture on (both off by default).

  • Provider tokens scrubbed by format: sk-proj-, sk-ant-, github_pat_, AWS STS keys and 16 more formats were stored as typed. Same patterns in the hook and the recorder, with a test that fails if they differ. From the detection pass (a125b46), cherry-picked.
  • Prefixed assignments (HF_TOKEN=..., fix(redaction): scrub secrets assigned to prefixed names #432): on master since 2026-11-15, not yet released.
  • CONTRIBUTING: two stale paths and the lint command.

No detection rule changes; ruleset fingerprint 15846a0915769d4a. The freeze is untouched: no frozen file is in the diff.

Checked locally: ruff clean; benchmark 0/0; full suite 2516 passed, 7 skipped, plus the two tests this machine's .env breaks (pass with it set aside); a fresh wheel in an empty venv reports 0.12.2, ships the dashboard, scrubs the new formats identically in both copies, and runs agentmetry demo to exit 0.

After merge: tag v0.12.2, release.yml publishes to PyPI, then gh release create v0.12.2 --verify-tag --latest.

🤖 Generated with Claude Code

0.12.1 on PyPI stores some typed secrets as typed whenever command logging
or full-argument capture is on. The OpenAI pattern stopped at the second
hyphen, so sk-proj- and sk-ant- keys passed through, and twenty other
provider formats had no pattern at all. Prefixed assignments such as
HF_TOKEN=... were fixed on master in #432 but never released. An audit on
2026-11-16 confirmed the gap from the published wheel's bytes.

This ships both now, ahead of the detection pass, because the first
outside tester is about to turn capture on. Both scrubbers, the hook's
inline copy and the recorder's, carry the same twenty provider patterns,
taken from the detection pass (a125b46). A test fails if the two lists
differ. The test that cross-checks them against the detection rules'
provider table skips until that table exists, so the file is identical
on both branches.

No detection rule changes; fingerprint 15846a0915769d4a. Also corrects
two stale paths and the lint command in CONTRIBUTING.md.

Checked: ruff clean; benchmark 0 missed, 0 false positives; 2516 passed,
7 skipped, plus the two tests this machine's .env breaks, which pass with
it set aside; a wheel built from this tree, installed in an empty venv,
reports 0.12.2, carries the dashboard, scrubs the new formats identically
in hook and recorder, and runs `agentmetry demo` to exit 0.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
blitzcrieg1 added a commit that referenced this pull request Nov 16, 2026
The redaction patterns shipped first, in 0.12.2 (#439), from this branch's
a125b46. That release needed the provider-table cross-check to skip where
the table does not exist yet, and its comments name 0.12.2 rather than
0.13.0. The same edits here keep the two copies of these files identical.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This machine collects two extra tests from gitignored launch entries, so
the count taken here was two high (CLAUDE.md, Commands).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@blitzcrieg1
blitzcrieg1 merged commit e65e6c2 into master Nov 16, 2026
10 checks passed
blitzcrieg1 added a commit that referenced this pull request Nov 16, 2026
Brings in 0.12.2 (#439) and the CLA bot removal (#440). The pass had stopped
getting CI because #435 conflicted with master on the README's test count;
it now says 2844, what pytest collects on the merged tree.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant