Versions match plugin/.claude-plugin/plugin.json and the git tag of the same
name.
- The key link and command last 30 minutes, up from 10.
- The agent's key never goes through chat: Claude Code sends it with a command, anywhere else the user pastes it on a one-use page.
- The agent search skips
.envfiles, and the key's name comes from the agent's code, never from.env. - A settings change keeps the agent's key. Only a refused key asks for a new one.
- The agent search also skips env files with other names, like
prod.envorapp.env.local.
- The report read-out says when a verdict came from the setup, not the agent.
- The recap after a build names the files it read and what they didn't show.
- A build without a connection saves nothing and says how to save it.
- The docs page speaks in findings with proof, and lists only the tools that exist.
- Sandbox choices before a run when the agent's tools never reached it. The sandbox address stays out of chat.
- Building the simulation environment asks nothing. No repo: the guide hands over iFixAi's brief.
- A run that stopped early opens on how far it got. The guide no longer says a left-out gating inspection counts as a failure.
- No package: read back what the grant holds.
- Dashboard link now points to app.ifixai.ai.
- Hosted only. No package: claim the free fast audit or redeem a promo code in the dashboard, then come back. A used-up grant gets the upgrade line. The guide and docs no longer offer a run on the user's machine.
- New describe prompt; Connect before Build.
- Grant workspaces can run. No inspection ids or totals.
- Privacy page redirects to ifixai.ai/privacy.
- No package: open source on your machine, or a hosted fast audit. Validate with ifixai 4.0.0.
- The guide names only the tools the connector still has. A paid run covers the whole package or whole bundles; a free run covers all 50, or the ids the user names.
- The guide is a third of its old length. The demo, the package rules, the sandbox question and "findings, never fixes" live in the MCP server's instructions now, so the guide no longer repeats them.
- The free plan writes its simulation environment from the step 4 headings and validates it; the long fixture rulebook is gone.
- After the simulation environment is saved, ask once for a test token per
role and store them with
set-role-logins.test-connectionnow reports which logins the agent accepted and whether it forwards the request id.
- Free path: a run is read out as findings, each failed and inconclusive inspection with its reason. The score, grade and percentages the terminal prints are never read out.
- Free path: the scope is offered as one of the six bundles, a pick of 15 or more, or all 50. A bundle runs as its categories.
- No em-dashes in the guide. Colon or full stop, as the house rule has always said; a test now fails on one rather than leaving it to review.
- A running audit is read out as counts: "N of T inspections run so far, P passed, F failed", never a percentage. On clients that render MCP Apps the same poll draws the Audit screen: the categories grouped into the six bundles, failing ones in red.
- When the audit finishes the operator says where else the two reports can be read: here, as saved HTML, or in the iFixAi dashboard after logging in. They are the same reports, so no surface is described as fuller than another.
- The simulation environment is built from the whole agent repo. The markdown description is a last resort, offered only after the user confirms there is no repo, and never because the repo is large or awkward.
- Every tool needs a sign-in now,
run-demo-auditandlist-inspectionsincluded. The guide says so up front: the first call raises the sign-in prompt, which is free and where the account is created. - The demo is for a signed-in user on the free plan or with no run yet. The signed-out path and the "locked rows unlock after sign-in" line are gone; every demo finding is open.
- The packages read out of
get-plan, and the same rows still ride onlist-inspections.
- The operator never reads a judge count out. Step 0 no longer lists "judges
included" among what
get-plannames, and Step 7 no longer says how many judges the package judges with. Both contradicted the guide's own rule. - The free suite is costed as "a few dollars on your own key" instead of naming judge models and per-model prices.
- The free run prints "the results", not "the scorecard": there is no scorecard on a customer surface any more.
- The gloss "(the fixture file, and the
*-fixturetools that write it)" now sits at the first mention of the simulation environment, not at Step 4, so an operator meets the two names together before either tool is called.
- The guide now speaks the product's own words. An inspection is the unit that passes or fails, never a "check". The simulation environment is what the agent is supposed to do, named as that in prose (the fixture file is still the fixture file, and the tools keep their ids). The flow is Connect → Simulation environment → Audit.
- The audit is introduced the way the product introduces it: "We stress test your agent inside the simulation environment. Every result is judged by AI models your agent never runs on."
- A report opens on its verdict sentence, then the counts: "The audit of has revealed critical findings." when a failure broke a hard rule, otherwise how many findings, otherwise none.
- The preview offers the selection as " inspections across categories", always the package's own numbers, and names the six bundles a customer picks from.
- The operator reads out what each package holds: the tagline, a one-sentence summary of its inspections and audits a month per agent, and, from Growth up, that the customer may say their agent has been audited by iFixAi. Still no prices, no judges, no recommendation, and no agent counts: how many agents a package covers is what the call with us is for.
- Four packages now, matching the pricing page: Startup, Growth, Enterprise, Agentic Enterprise.
- A free user gets that offer once after the demo or a free run, and on any
package refusal, with
request-accessas the way to ask for one. Signed out the same rows ride onlist-inspections, and every tier reads the same card. - Audits are counted per agent:
get-plansays how many each connected agent has left this month, and a refusal names the agent that is out and whether a sibling still has room.
- The connector reports findings, not grades. A hosted run opens on the counts ("N of T checks passed, F failed, I inconclusive"), a failed check carries a severity word instead of a number, and no letter grade or percentage reaches the reply. The demo audit opens the same way.
- Wording: a hard rule becomes its own check, the panel judges rather than grades, and a gating inspection that is not run counts as a failed check.
- The judges are not the user's to pick: the package's panel grades, the preview says how many, and a run that fails on iFixAi's side says so in one sentence and counts no audit.
- The self-judging note applies to the free plan, where the user supplies the judge key.
- Fixture authoring no longer points at files outside the published package.
- Dropped the coverage caveat: checks that do not dial the endpoint grade the setup we read and authored, so calling them unreached was misleading.
- Privacy: the report no longer names the judge, so the processor table says iFixAi selects the judges rather than promising they are named.
- Hosted sandbox: the operator points the test copy's tools at the
connection's
rest_url, checkslast_call_atbefore starting a run, and readstool_calls(what the agent did) off the report. - A "no" to the test-target question runs nothing: on a paid workspace the
operator walks the user through a test copy (Step 0b: hosted sandbox or a
fake backend, test branch, tunnel) first. No local
ifixai sandboxcommand in that walkthrough yet: no public release has it (PyPI tops out at 3.4.1); it returns when the free recipe's pin moves to 3.5.0.
- Packages, not credits. Step 0 reads the package off get-plan (audits used of allowed this month, judges, agents) and offers request-access when there is none; Step 7 is "Preview, then wait for yes" with preview-run, which prices nothing; cancel-run yields no report and counts no audit. The wallet, quote and credit-request tools are gone.
- The demo opens every first conversation:
run-demo-auditbeforeget-plan, a demo audit of a fictional neobank support bot (Kestrel Bank Assistant, Grade C), read out as the grade, three impact cards, then "Ready to test yours?"; skipped when the user opens by naming their own agent.
- Step 4 authors from the repo in two calls: select-repo-files with the file list, then author-fixture with the files it asked for. Step 5 recaps with the citation behind each value and the assumed ones as questions.
- Step 4 fixes the description's headings, order and source priority (the shape author-fixture is trained on), and gives the free path the reference fixture's per-section rules, so the same agent authored twice gets the same fixture.
- Operator asks whether the target is a test deployment with sandboxed
backends before every run, and passes
targetIsSandboxedto run-inspection.
- Skill compressed ~40% (288 to 182 lines): same rules, half the prose.
- Free runs accept any judge key (OpenRouter, OpenAI, Anthropic, Gemini): the operator asks which the user has and the recipe matches it.
- Skill tightened and made client-neutral (no Claude-Code-only tool names), since the connector now serves it to every MCP client as a prompt.
- Free runs: the operator asks suggested set vs whole suite before running, and states judge cost per model (Qwen, Gemini Flash) without naming vendors.
- Free runs: the operator now offers both ways to execute the recipe, the user's own terminal or the assistant's shell, and asks before running.
- Free validate commands pin engine 3.4.1 (was 3.3.0), matching the connector's recipe. The free roster is now 50 inspections.
- The skill leads with the no-account demo audit (
run-demo-audit) and explains that sign-in happens in the chat window, not on a website.
- Free tier. Signing in is now enough: the 45 inspections that ship in the open-source engine run on your own machine with your own LLM key, nothing billed and nothing stored with us. A paid workspace still adds the 57 iFixAi-only inspections, hosted runs, saved history and the report.
- Upgrading works mid-session. The tier is resolved per call, so the tool list never changes and no reconnect is needed.
searchandfetchadded, so the connector also works in clients that require them.- The pro roster is no longer visible to a free account: names, descriptions and ids of the iFixAi-only inspections are filtered server-side. Only the count of what exists is shown.
- Repository renamed to
ifixai-plugin. The install command is unchanged.
- First release: the
ifixaiskill and the remote MCP server atmcp.ifixai.ai, for hosted, credit-billed audits.