Skip to content

Releases

ClAudit v2.13.0

Choose a tag to compare

@github-actions github-actions released this 27 Sep 22:36

Usage meter, live. The plan windows now refresh every 60 seconds (Settings, "Usage meter refresh", 30 to 600; config usage_interval) instead of every five minutes, and again immediately after any claude call ClAudit makes, which is when the numbers move. A snapshot counts as stale after 15 minutes regardless of the interval.

  • Project tab: a "Claude plan usage" panel at the top with one bar per window (5-hour, 7-day, each per-model weekly cap, extra usage when enabled), percent, reset countdown, and which one is the active limit. Green under 50%, amber under 80%, red past.
  • Tray icon mouseover: the watcher state plus one line per window with percent and reset time, refreshed with every snapshot.
  • Fix: a fresh window could start a usage fetch before its own first tick; the GUI smoke test caught it as a crash at teardown.

Install: pip install "claudit-cc[gui] @ https://github.com/sworrl/ClAudit/archive/refs/tags/v2.13.0.tar.gz" or download the wheel below.

ClAudit v2.12.2

Choose a tag to compare

@github-actions github-actions released this 27 Sep 02:56
  • Fix: the feedback endpoint answered 403 to anything that did not look like Claude Code. Sends now carry the same client headers as the CLI (x-app: cli, User-Agent: claude-cli/<version> (external, cli), anthropic-version, the session id). Verified live: the first queued draft came back with a feedback id.

Install: pip install "claudit-cc[gui] @ https://github.com/sworrl/ClAudit/archive/refs/tags/v2.12.2.tar.gz" or download the wheel below.

ClAudit v2.12.1

Choose a tag to compare

@github-actions github-actions released this 27 Sep 02:50
  • Fix: the feedback endpoint rejected every send with content: Field required. Claude Code wraps the report as a JSON string under content (plus session_id); ClAudit now sends that envelope. Verified against the live endpoint.
  • A rejected send is retried after six hours, or immediately after an upgrade (the failure is stamped with the version that produced it). --process-drafts always retries. An issue filed in an earlier pass gets the feedback id written in once the send succeeds.

Install: pip install "claudit-cc[gui] @ https://github.com/sworrl/ClAudit/archive/refs/tags/v2.12.1.tar.gz" or download the wheel below.

ClAudit v2.12.0

Choose a tag to compare

@github-actions github-actions released this 27 Sep 02:44

Claude Code's own bug reports, sent for you and tied to ClAudit's. Claude Code 2.1.28x drafts a bug report itself when a safeguards block (or its own mistake) derails a session, stores it in ~/.claude/feedback/drafts/, and waits for "2 to send". Five sat queued on the reference machine, one of them about the classifier.

  • process_drafts() runs every 60 s in both watchers. For each queued draft: a safety-classifier draft becomes a ClAudit issue on GitHub first (or is appended to the issue already carrying one of its Request IDs), then the draft is POSTed to api.anthropic.com/api/claude_cli_feedback exactly as the send key does (same payload shape, the CLI's own OAuth login, no transcript) with every related ClAudit issue URL in its description, then the feedback id and draft id are written onto the GitHub issue. Model-behavior drafts go to Anthropic only. A sent draft is copied to ~/.claude/claudit/drafts-sent/ and the card clears.
  • gh_create() now mirrors every ClAudit report to the same feedback endpoint with its GitHub URL and edits the feedback id into the issue body, so each report exists in both channels with a pointer each way. feedback_mirror turns it off.
  • Mute terms stop a draft from going anywhere. Filing honors the live window; sending does not. CLAUDE_CODE_DISABLE_FEEDBACK=1 stops all sending.
  • CLI --drafts and --process-drafts; Settings group "Anthropic feedback channel" with send_drafts, file_drafts, feedback_mirror.
  • Tests: draft parsing and classification, the exact feedback payload and headers, the two-way tie (new issue, existing issue, model draft, idempotence), mute and stale handling, the mirror on create. 126 total.

Install: pip install "claudit-cc[gui] @ https://github.com/sworrl/ClAudit/archive/refs/tags/v2.12.0.tar.gz" or download the wheel below.

ClAudit v2.11.0

Choose a tag to compare

@github-actions github-actions released this 27 Sep 01:42

The census stays on by default, and now nobody can miss it. The anonymous heartbeat is spelled out in red in five places and can be declined at install time:

  • README: a red CAUTION callout under the intro, in Install, in Quick start, and at the top of Census, each with the exact payload and every opt-out. The project page carries the same text in red.
  • The app shows a red notice once on first launch with "Keep it on" and "Turn it off" buttons (census_notice_shown in config). The Settings > Census description is red. The CLI watcher prints the notice in red at start, and --doctor reports the census state as a warning line while it is on.
  • --no-census on claudit_gui.py and claudit_scan.py turns it off and remembers it. CLAUDIT_NO_CENSUS=1 in the environment turns both heartbeats off for that run (pip installs have no installer to pass a flag to). The installers take --no-census (Linux, macOS) and -NoCensus (Windows) and write the config before the first run; without the flag they print the red notice.
  • census_settings(cfg) is the one place the three inputs (config, flag, env var) resolve. Tests cover the env override, the flag persisting, and the notice text.

Install: pip install "claudit-cc[gui] @ https://github.com/sworrl/ClAudit/archive/refs/tags/v2.11.0.tar.gz" or download the wheel below.

ClAudit v2.10.0

Choose a tag to compare

@github-actions github-actions released this 27 Sep 01:05

Install census: how many nodes run ClAudit, on which version, and whether they are still up. Three sources, merged hourly into docs/nodes.json, an Installs block in the README, a card on the project page, and the Project tab in the app.

  • Anonymous heartbeat (on by default). Every 10 minutes the app or the CLI watcher POSTs {node, v, os, mode, event} to a Cloudflare Worker in telemetry/ (deployed at claudit-census.fogbank.workers.dev). node is a random id made on first run (~/.claude/claudit/node_id); the Worker keeps those fields and a timestamp for eight days and never reads the request IP. A clean quit sends stop. The Worker's cron sorts nodes into running (beat within 30 min), stopped (said stop within 24 h), and quiet (seen within 24 h, then silent: crash, sleep, offline), with version, OS, and install-mode breakdowns at /stats.
  • GitHub heartbeat (off by default, login-visible). Every 6 hours the node creates or edits one comment on #14 under the user's own login: version, OS family, install mode, a short non-reversible node hash, time. Quiet after 24 h without an edit. Delete the comment to leave.
  • Passive census. Every filed issue ends with the version that filed it; the aggregator counts distinct reporters, the version each last used, and 7/30-day activity. No heartbeat involved.
  • Settings gains a Census group with the two toggles (turning the anonymous one off sends a final stop; turning the GitHub one on posts immediately). claudit_scan.py --census prints the live numbers and exactly what this node sends. Config: census_anon, census_github, census_url.
  • Tests: node id persistence and payload shape (no PII fields, hostname and home path asserted absent), tick scheduling for both cadences and the stop, the POST body, the comment upsert (edit own, create when absent, skip when signed out), and the aggregator on fixtures. 120 total.

Install: pip install "claudit-cc[gui] @ https://github.com/sworrl/ClAudit/archive/refs/tags/v2.10.0.tar.gz" or download the wheel below.

ClAudit v2.9.0

Choose a tag to compare

@github-actions github-actions released this 27 Sep 00:24

Update monitor with in-app updates. Until now a git clone updated itself silently and a pip install got one tray toast with a command to paste. Both now go through one monitor (claudit_ui/updater.py, pure stdlib so it tests without Qt):

  • Every three minutes the app compares what runs with what GitHub has: for a clone, commits behind/ahead of origin/<branch>, the version on the remote branch, whether the tree is dirty, and whether the waiting commits touch code; for a pip install, the latest GitHub Release. Release notes are shown when the remote version has a release; otherwise the waiting commit subjects (poll-bot commits filtered out).
  • Header: an amber pill ("Update available: 2.9.0 (3 commits)", or "Update 2.9.0 waiting: the working tree has uncommitted changes") that is hidden when there is nothing to do. Settings: an Updates panel with the status lines, the notes, Check now, Update now, and an Auto-update toggle (config auto_update, default on, git clones only). Tray: Check for updates.
  • Update now opens a dialog: notes, Release page, Check again, and the update itself with its log streamed live. A git clone fast-forward pulls; a pip install runs pip against the release tarball in the same interpreter. The window restarts on the new code when the change touched code; a docs-only pull just refreshes the panel. Dirty or diverged checkouts are never touched; the panel says why.
  • The old auto behavior is unchanged with the toggle on. Off, the app only reports and waits.
  • Tests: check() in git mode (behind, clean, blocked by dirty tree or local commits, up to date), pip mode, the pip command, the release cache and offline path. Verified end to end on a scratch clone set three commits behind. 115 tests.

Install: pip install "claudit-cc[gui] @ https://github.com/sworrl/ClAudit/archive/refs/tags/v2.9.0.tar.gz" or download the wheel below.

ClAudit v2.8.0

Choose a tag to compare

@github-actions github-actions released this 26 Sep 23:59

Current block wording, the silent model fallback, Message IDs, and a live-filing window. A pass over the transcripts on the reference machine (244 sessions) against what classify() and the GUI knew:

  • Silent model fallback is now detected and filed. Since September Claude Code answers a safeguards refusal by re-running the request on a weaker model and leaving only a system entry (subtype: model_refusal_fallback, "Opus 5.5's safeguards flagged this session ... Opus 4.8 is answering instead"). The entry carries the refused call's Request ID, the API refusal category, both model IDs, and the uuid of the refused user message. fallback_event() reads it; the finding is keyed like any other block (category cyber, anything else aup), the prompt is looked up by that uuid rather than "the last thing typed", and the report gets a "Model fallback" section plus a triage note. 106 such events sat unseen on the reference machine.
  • Details: [cyber] decides the kind. Every block message now ends with the API's refusal category in that tag; classify() trusts it before any wording heuristic, so the next rewording of the prose cannot silently drop blocks to other again. The current wording ("intentionally broad safeguards ... Cyber Verification Program ... real-time-cyber-safeguards") is matched explicitly as well.
  • Message IDs. Blocks carry Message ID: msg_... next to the Request ID. Captured per occurrence, printed on the ID lines of the report, and protected from every scrub layer the same way req_ is.
  • Plan-cap messages are limit, not other: "You've reached your Fable limit. Run /usage-credits ...", "hit your session limit", "hit your weekly limit".
  • Live-filing window. Newly detected blocks older than 7 days (config max_live_age_days, --max-live-age N, 0 = off) are parked in the backlog instead of live-posted, in both the dwell and the auto path. A detector improvement like this one would otherwise post weeks of history in one burst; on the reference machine that was 37 fallback reports from August. They are counted in the backfill bar and drip out only if backfill is on.
  • Models. The Settings picker lists Haiku 4.5, Sonnet 5, Opus 5.5, and Fable 5.1 (a saved older id is kept as its own entry); the per-call cost table knows the new ids; model_label() turns claude-opus-5-5 into "Opus 5.5" for reports. The Issues tab's "Fable 5" filter became one "model: X" entry per flagging model seen on the board.
  • Tests: 12 new (current wordings, tag precedence, Message ID capture and protection, the fallback event end to end, the live window on both paths, model labels). 111 total.

Install: pip install "claudit-cc[gui] @ https://github.com/sworrl/ClAudit/archive/refs/tags/v2.8.0.tar.gz" or download the wheel below.

ClAudit v2.7.1

Choose a tag to compare

@github-actions github-actions released this 24 Sep 04:47
  • Fix: the PyPI publish job never ran, because a GitHub Release created by the workflow token cannot trigger another workflow. It is now a second job inside release.yml, after the release is published, marked continue-on-error until the pending publisher exists on pypi.org.

Install: pip install "claudit-cc[gui] @ https://github.com/sworrl/ClAudit/archive/refs/tags/v2.7.1.tar.gz" or download the wheel below.

ClAudit v2.7.0

Choose a tag to compare

@github-actions github-actions released this 24 Sep 04:45

The GUI is a package, the screenshots are generated, the counter was undercounting, and the release path reaches PyPI and Homebrew.

  • claudit_gui.py (170 KB, one file) is now claudit_ui/: common, widgets, workers, dialogs, main_window, app. The split was done by an AST pass that computed each module's imports from the names it uses, then verified by building the full window against live data. claudit_gui.py stays the entry point and re-exports every name, so import claudit_gui as g and the launchers keep working. The watchdog and self-update paths point at the entry script explicitly.
  • claudit_gui.py --screenshot DIR builds the window read-only (no watcher, no filing, no singleton lock), saves one PNG per tab, and exits. The four README screenshots were regenerated with it; they had been from July 1 and showed none of the usage meter, mute list, closure markers, or doctor.
  • Fix: the README counter and trend listed at most 800 issues, so it reported 447 closed while 1023 were closed. The limit is 5000 now; the next hourly run corrects the numbers and the trend line will step.
  • Umbrella issue #86940 gets its title refreshed along with its body: "Inactivity bot has closed 593 ClAudit false-positive reports since 2026-08-09 (18 sweep days)" instead of the first day's 83.
  • Cloud defend checks the PAT's expiry header every run: a warning from 30 days out, a failed run (which emails you) at 7.
  • PyPI: the distribution is named claudit-cc (claudit is a reserved placeholder owned by someone else). .github/workflows/publish.yml publishes on every GitHub Release through PyPI Trusted Publishing, so there is no token to store; it needs the one-time pending-publisher registration on pypi.org described in the workflow header.
  • Homebrew tap sworrl/homebrew-claudit exists: brew tap sworrl/claudit && brew install --HEAD claudit.
  • Ruff runs its full ruleset in CI now ([tool.ruff] in pyproject.toml, every ignored rule carries its reason). 35 findings fixed mechanically (implicit string concatenation in literals, unused noqa, chained-operator parentheses).
  • Settings: the closure defender label showed a literal &&; the header line lost its dash.

Install: pip install "claudit-cc[gui] @ https://github.com/sworrl/ClAudit/archive/refs/tags/v2.7.0.tar.gz" or download the wheel below.