iFixAi Files

Pro Offer

The first interactive ifixai run on a machine asks once whether you'd like to claim a free fast audit at https://www.ifixai.ai/ (#223).

• No (preselected, just press Enter) — the run starts.
• Yes — opens the site, then asks whether to continue the current run or abandon it. Abandoning exits 0 before anything happens: no provider calls, no telemetry, no run directory.
• Shown once per machine, whatever the answer.
• Never shown in CI, when stdin or stdout is not a terminal, or on --dry-run, so scripted and agent-driven runs are never blocked.
• Skip it with --no-promo or IFIXAI_NO_PROMO=1.

Important Fixes

Scores that were too generous

• Judge verdicts are read as real booleans. A verdict of "false" sent as text used to count as true and inflate grounding scores (#137, [#170]).
• A judge reply must cover every expected claim exactly once. Missing or extra verdicts are rejected instead of graded (#138, [#206]).
• Truncated judge replies and truncated grading inputs stay unscored instead of being graded on partial text (#149, [#181], [#209]).
• An inspection whose consistency cap pulls its score below threshold now fails. It used to keep its pass (#133).
• HTTP provider: retrieval evidence comes only from the current query. A failed retrieval used to reuse sources from an earlier chat (#124).
• Confidence intervals are computed over the probes that were actually scored (#189).

Judge budget

• Judge calls no longer run twice (#121).
• --judge-budget holds under concurrent evaluations (#135).
• Probes skipped because the budget ran out stay unscored, and a deterministic veto still applies once the budget is spent (#169, [#179]).

Safety and credentials

• Generated fixtures no longer grant destructive tools to non-admin roles whose name merely contains "admin" (#131).
• Gemini: concurrent requests keep their own API key (#190).

Reproducibility and resume

• The run manifest fingerprints the inspection files and rubrics actually used, across every series (#152, [#171]).
• --resume refuses to reuse results when the system under test changed (endpoint or system prompt), and re-runs damaged checkpoint entries instead of skipping them (#187, [#130]).
• Checkpoints and reports are written atomically. A failed write keeps the previous file (#148, [#154]).

CI gates and reports

• --min-score rejects nan, inf and values outside 0–1 before the run starts (#196).
• Unscored results no longer display a failing grade or a false verdict, and scorecards show when a mandatory minimum was not part of the run (#134, [#203]).
• The CI example fails on partial or unscored runs instead of passing them (#136).

Providers

• Exhausted quota is reported as such instead of being retried as throttling (#197).
• Configured endpoints are honoured for Gemini, Hugging Face and judge ensembles (#193, [#172], [#207]).
• ifixai install keeps your own text in an existing AGENTS.md (#212).

Behaviour Changes

• Scores can move. Runs that were over-credited on 4.0.0 may score lower or show more unscored probes.
• The first interactive ifixai run on a machine asks one question before starting. Automated runs are unaffected.
• An invalid --min-score is now a usage error.
• A resumed run whose endpoint or system prompt changed is refused.