4 Commits
Author SHA1 Message Date
n-papaioannouandPeter 92bb5d0a95 feat: audit any agent over chat (#363)
* chore: streamline the engine

Removes the record/replay bridge, the unused embedding classifier, the example scripts and the deprecated --profile alias (use --mode).

* feat: judging within your budget

* chore: align run stops and labels with main

* feat: audit any agent over chat

* chore: CI fix

---------

Co-authored-by: Peter <as.peter@ime.life>
2026-10-23 00:09:19 +03:00
stefyi-4355 6a01ffe22d feat: P14 judge truthfullness (#232) 2026-10-13 08:11:57 -07:00
n-papaioannou 9acc8209a6 Five-pillar grading, docs overhaul, verified fixture templates, traction cleanup (#66)
* Grade on the five core pillars only: the overall score uses a fixed GRADED_CATEGORY_WEIGHTS denominator, premium categories are reported without entering the grade, and P01 keeps its mandatory-minimum cap.

* Print n/a instead of crashing when every selected inspection is inconclusive and overall_score is None.

* Move the adoption charts into docs/traction.md, drop the unique-cloners chart and its apparatus, and archive all pypistats endpoints weekly before the rolling window drops them.

* Rewrite the docs in plain language and cut them to a third: three doc pairs merged into their natural homes, every inbound link rewired. Replace the open-source scorecards with before-remediation reconstructions of the Dragontail and Instagram incidents.

* Make the fixture guidance produce complete runs: an engine-verified template that clears every mandatory-minimum evidence floor in the doc, the plugin skill, and the scaffolded playbook, plus a per-inspection evidence-floors table and corrected field attributions.

* Trim .gitignore to generic patterns and ignore the plugin's working fixture file.

* Remove the website source links from both case-study scorecards.
2026-07-23 15:09:23 +00:00
n-papaioannou 4f9cf49775 feat: iFixAi 3.0.2 — Claude Code plugin + PyPI packaging (#47)
* feat: iFixAi 3.0.2 — Claude Code plugin + PyPI packaging

- Installable Claude Code plugin (skill, hooks, manifest) + multi-provider SUT/judge engine and interactive scorecard artifact
- pip-install ready: ifixai-diagnose entry point, ifixai.plugin package-data, version 3.0.2, tag-triggered PyPI release workflow
- Refreshed README + Diátaxis docs/

* plugin: model preflight and run-health gates for live runs

A per-model preflight aborts a live run on a bad model id before billing; a measurement-failure banner suppresses the grade when the agent or judge was unreachable; SUT transport errors now route to unscorable (base.py, b19/b20/b22) so a dead model scores INCONCLUSIVE rather than a false FAIL; adds a judge-health note and drops the dead b07 structural path.

* fix: gate platform-refusal to bridge mode, restore B01/P01 fail-closed, guard judge-timeout parse

Stops a legitimate live-API refusal that cites a usage policy from being dropped as unscorable (the wall check now runs only for bridge-routed SUT replies), reverts the B01/P01 capability-missing carve-out so an ungoverned target fails closed and caps at D per docs/scoring.md, and parses IFIXAI_JUDGE_TIMEOUT defensively so a bad value no longer crashes import. Also fixes the bridge str|None mypy error, drops dead _checkpoint_path, and dedups the grade constants into grading.py.

* fix: make plugin Windows-compatible (bootstrap.py + PS1 shim + exec-form hook)

Move provisioning logic into a cross-platform bootstrap.py (bin/ vs Scripts/,
python3/python fallback). Add bootstrap.ps1 for native Windows PowerShell.
Switch the SessionStart hook to exec form (python3 + python fallback entries)
so no sh is needed. Update SKILL.md with Windows paths and shim invocations.

* fix: Windows-compat follow-ups (UTF-8 console, LF gitattributes, dedup bootstrap hook)

* Set plugin author and marketplace owner to iFixAi

* docs: show unique-cloners chart in README (parity with main)
2026-06-23 18:47:45 +03:00