Reference

Release notes

What's new in each SQUAT release, most recent first — written for the person running rounds, not the person reading commit logs.

v0.156.0 — retire and bring back squad members, and attach files to a support case (September 2026)

You can now retire a squad member you are no longer using, bring one back, and swap one for another, without losing the work they took part in. A project you closed can be reopened the same way. Your plan's limits are checked as the change is saved, so two changes made at the same time can no longer take you past your limit.

Support cases accept files. If your client can send attachments directly it will; if it cannot, SQUAT gives you a secure upload link for that case instead. Links work only while the case is open to replies.

The workspace card now always counts your whole workspace — projects, squad members, squads, scenarios and rounds — whether or not you have a project selected. Selecting a project changes the actions offered and the roster shown, never the counts.

v0.155.1 — close a resolved support case yourself (September 2026)

If you opened a support case and it is marked resolved, you can now close it yourself when no further help is needed. Closing keeps the case history and explanation intact. It does not reopen the case, change its category, or affect any other case.

v0.155.0 — more of your saved work arrives as a source-backed card (September 2026)

More SQUAT views now arrive as cards that carry the same source-backed facts, exact links, and available actions whether your client shows the rich card or its Markdown version. Research, round, issue, journey, artifact, and report paths keep their references tied to the specific record they describe, so a view cannot quietly point an action at a different item. No saved workspace data is rewritten and no customer action is required.

v0.154.0 — send feedback directly, and clearer seat handling (September 2026)

You can now send us feedback from your account page on the website, without opening a support case. Feedback goes to the SQUAT team as feedback, kept apart from support cases, and it does not carry your name, email or workspace, so we cannot reply to it. If you need an answer, open a support case instead.

If you're a member of a workspace without an assigned seat, you're now told exactly who to ask — the workspace owner's name and email — instead of a generic refusal. And if an owner frees up your seat, your access key keeps working until it's separately revoked; only the seat assignment changes.

If your workspace has been with us since before we added an explicit owner field, a few actions that were incorrectly refused for those older workspaces — retiring a workspace, moving a pack, and a sign-in seat check — now work correctly.

If you use an MCP-connected assistant to work with SQUAT, it can now close a support case it opened on your behalf, and upload a small artifact in a single step on hosts that could not do the previous multi-step upload.

One note for administrators: running an overdue data-retention job by hand from your own session is now limited to the workspace owner; a workspace admin can no longer trigger it that way. Retention still runs automatically on its normal schedule for everyone.

v0.153.0 — workspaces you belong to, not only the ones you own (September 2026)

If you are a member of a workspace you do not own, you can now list it and switch into it, from the account area and from an assistant connected over MCP. Previously only an owner could do either, so a member with a perfectly valid seat could not reach their own workspace without help. Membership is what grants this: you still see only workspaces you belong to, and switching into one you are not a member of is refused.

The workspace summary figures an account needs are now served to the account area directly, rather than only through an operator tool.

Your seats page now shows who holds each seat. What is metered and what you are billed for do not change.

Error messages you see when something goes wrong are now a single fixed message rather than the underlying technical text. This is deliberate: that text could describe internal detail, and it should not be shown to customers. It also means an error is less specific than it used to be, so if you need to know what actually failed, quote the time and what you were doing when you contact support — the detail is recorded on our side.

One note for administrators: the member listing above relies on a database index that is deployed as part of this release. Until this release is live in an environment, that environment continues to show owned workspaces only, rather than failing.

v0.152.1 — announcements marked critical stand out (September 2026)

An announcement marked critical now carries an alert marker in the messages card you see at the start of a session, so an urgent notice is distinguishable at a glance from an ordinary one. The marker is shown only while such an announcement is outstanding and disappears once you dismiss it. Nothing else about messages, announcements or dismissal changes.

v0.152.0 — updated legal documents and a legal-status check (September 2026)

The End-User License Agreement, Terms of Service and Privacy Policy are updated and published as “Prerelease version v3 - effective 9/9/2026”. Because these are a new version of the documents, everyone is asked to read and accept them again the next time they sign in; previously recorded acceptances of the earlier version do not carry over. No other account setting, entitlement or stored content changes.

The account area can now ask the server directly whether your acceptances are current, so it can tell you what is outstanding and how long you have, rather than blocking you without explanation. Nothing about how or when consent is enforced changes in this release.

v0.151.0 — key reprotection includes persona learnings (September 2026)

Managed key reprotection now includes saved persona-learning observations, keeping them readable when they are moved to a newer workspace data key. This release does not start a rotation or retire a key, and requires no customer action.

Workspace retirement now distinguishes a missing owner from a missing home-workspace record. Access-state updates clear stale retirement or suspension labels, and provisioning removes obsolete workspace-pack entries while preserving the account’s dedicated workspace-pack record and other add-ons. The previously documented limitation when unlock races a license revocation remains.

v0.150.0 — choose where packs land and move them between your workspaces (September 2026)

If your account owns more than one active workspace, a new pack increase now goes only to the workspace you selected. You can also move one whole existing pack at a time between owned active workspaces from the account surface. Moving a pack does not buy, cancel, split, or change its kind, and does not change the total held by the account.

SQUAT refuses the move if the source does not hold the pack, either workspace is inactive or belongs to another account, the stored billing keys are ambiguous, or the source's currently measured seats, personas, or projects already exceed the reduced allowance. Because that usage measurement happens immediately before the move rather than inside the same database transaction, a simultaneous new use can still produce an overage that normal entitlement enforcement must handle afterward. Device-pack moves use the existing per-licence device rule: a lower cap is enforced when a device next connects rather than invented from a workspace-wide count.

v0.148.0 — read your support thread from your assistant, with safer intake and unlock handling (September 2026)

Your assistant can now read the customer-visible conversation on a support case owned by your workspace. It sees the original description, your replies, replies the operator actually sent, status updates, and attachment references. Draft answers, internal operator notes, unsent replies, and cases hidden as spam or duplicates stay out of that view. If a case is missing, belongs to another workspace, or is hidden, the answer is the same and does not reveal which situation applied.

An assistant can also look up an attachment named in that thread, but the result is deliberately limited to its name, media type, and size. It does not include the file itself, a storage location, or a download link.

The website and assistant support forms now use the same basic email-address check, so an entry such as notanemail is refused before a case is stored. This checks the address shape; it does not prove that the mailbox exists. Separately, unlock no longer reactivates a license that was already revoked when the unlock read it. A revocation that happens concurrently after that read remains a separate repair.

v0.147.0 — honest device allowances, safer refresh retries, and a reply address on every new support case (September 2026)

Your workspace allowance now describes devices the way SQUAT actually governs them: per licence, rather than showing a made-up workspace total of zero. The number of devices you may use and the active-device count still come from the licence’s device-registration path; this release corrects the broader entitlement view so an operator or assistant cannot mistake one unit for the other.

If a client refreshes its sign-in and loses the response to a network race, it can retry the immediately previous refresh token for thirty seconds and receive the same already-issued pair, provided the retry reaches the same running server and that pair has not moved or been revoked. SQUAT does not store the raw token to make this work. A retry after a restart, on another server instance, after another rotation, or for an older token keeps the existing fail-closed response.

Creating a support case from an assistant now requires a contact email, just as the website form already does, and a missing address produces a clear message explaining what to provide. This prevents new assistant-filed cases from being saved without anywhere for support to reply. Existing support cases are unchanged.

v0.146.0 — terms changes come with a grace period, issues have a real lifecycle, and support cases can add a seat holder (September 2026)

When a legal document you have already accepted is updated, SQUAT no longer stops you at the door. You get a grace period — thirty days by default, set by your platform operator — to accept the new version, and your assistant is told exactly when that window closes. A document you have never accepted still has to be accepted before you go on, as before. The reply your assistant receives is now structured as well as written out, so it can tell you what changed without guessing. One thing is not in this release: telling SQUAT “not now” is not yet recorded anywhere; that arrives with the website side of this work.

Issues raised in a round now move through a proper set of statuses, kept separate from what you decided to do about them, and the “No active issues” line on a journey card clears only when an issue’s status is actually changed. Recruiting a persona is now driven by the persona itself rather than by a page, with a readiness checklist and a scored readiness read your assistant can ask for. A support case can now have another licence seat holder added to it, and who may see a case is checked against the seat as it stands today, not as it stood when the case was opened.

Operators get two new tools: a setting for the grace period, and a bounded workspace diagnostic that checks protected content without decrypting it. A full backup now refuses to complete if a UAT Report cannot be fully captured, and asks you to retry, rather than producing a backup with a hole in it. Still arriving as the server recorded them, unchanged by this release: data-subject records, the expiry fields on trash purge operations, and the legal-hold activation strings; completion certificates stay exactly as written, for the reason given under v0.144.0.

v0.145.0 — your profile changes are reported honestly, and the test suite is now type-checked (September 2026)

When you change more than one thing about your profile at once — say your preferred name and your time zone together — and only one of them is saved, SQUAT now tells you exactly which one landed. Before, that situation was reported as though nothing had been changed at all, and the field that had been saved was echoed back as empty. Nothing was ever lost on the server; what was wrong was the answer you got about it. Relatedly, changing only your name no longer discards the time zone SQUAT already had for you, so times keep displaying correctly without you re-entering anything.

The larger part of this release is invisible from the outside and worth one sentence anyway: the project's test files are now type-checked on every build, and the check runs in the deployment pipeline rather than only on a developer's machine. Test files had never been type-checked at all, which meant a whole class of mistake could pass every gate. Nothing about how SQUAT behaves changed as a result — this is the machinery that catches problems before a release, not a change to the product.

Still arriving as the server recorded them, unchanged by this release: data-subject records, the expiry fields on trash purge operations, and the legal-hold activation strings. Completion certificates still arrive exactly as the server wrote them, for the reason given under v0.144.0 — their identifier is a fingerprint of their contents, and you verify one by recomputing that fingerprint with the identifier removed.

v0.144.0 — deletion, retention and re-keying times in your zone (September 2026)

Deletion requests now show their expiry and verification times in your time zone wherever they appear — when you prepare a request, when you authorize it, when you run it, and when you check its status — and the same is true of the retention tools that carry a deletion record inside them. The times a re-keying job reports also arrive in your zone: when the next batch is expected, and when its coverage was verified, on every tool that reports them rather than only on the status check. The decisions themselves still run on the stored values, so nothing about when data is actually deleted has changed.

One thing deliberately does not change, and it is worth knowing why. A completed deletion or purge gives you a certificate, and that certificate carries an identifier that is a fingerprint of everything in it — including the completion time. Showing that time in your zone would change the fingerprint and the certificate would no longer check out. So the certificate stays exactly as the server wrote it, and you can still verify it: recompute the fingerprint over the certificate with its identifier removed, and compare. This release adds a test that keeps it that way.

Still arriving as the server recorded them: data-subject records, the expiry fields on trash purge operations, and the legal-hold activation strings.

v0.143.0 — persona governance, retention and ledger times in your zone (September 2026)

Three more surfaces arrive in your time zone: the activation and update times inside a persona's runtime-governance block; an artifact set's retention expiry and its automation activation and deactivation times, in the set, in the list and in the update reply; and the start and completion times in a round ledger summary, in both the structured rows and the markdown table. The purge and retention decisions themselves still run on the stored values, so nothing about when data is deleted has changed. Behind the scenes, the four deletion-request tools are now driven through their registered contracts in the test suite, which found and fixed a test-harness gap rather than a product one. Still arriving as the server recorded them: data-subject records, the expiry fields on trash purge operations, the legal-hold activation strings, and the expiry and verification times on a deletion request; two smaller surfaces, retention preparation and protection re-keying, have times no test has yet checked and are covered next.

v0.142.0 — audition, trash, audit, brief, hold, template and export times in your zone (September 2026)

The next eight reads arrive in your time zone, each with a declared structure your assistant can rely on: when a persona audition was created; when a trashed round was deleted and when it will be purged (the purge decision itself still runs on the stored value); when an audit entry was recorded; how a scenario has been used across rounds; the prompt times inside a round brief; when a legal hold was placed; the template catalogue's dates; and every date and time inside a portable export. Your full backup is unchanged — it stays a byte-exact copy of what is stored, so it can still be restored and verified. Stored records and integrity checks are unchanged. Still arriving as the server recorded them, and next in line: the activation time inside a persona's runtime governance block, an artifact set's retention expiry, data-subject records, the expiry fields on trash purge operations, and the legal-hold activation strings.

v0.141.0 — report, persona and support-access times in your zone (September 2026)

More of what SQUAT reports now arrives in your time zone: when a support-access grant was created, expires or was revoked; when a UAT report or a research report was created, updated or invalidated, in the report itself and in the lists; and when a persona source was created, updated or withdrawn. Those responses also come with a declared structure your assistant can rely on. Stored records and integrity checks are unchanged. Still arriving as the server recorded them, and following in the next release: persona audition, trash-list, audit-log and scenario-usage times, plus a handful of smaller reads (terms acceptance, legal holds, round briefs, template browsing).

v0.140.0 — transcript times in your zone, and content packs install again (September 2026)

The one exception left by v0.139.0 is closed: the per-turn timestamps inside a dialogue transcript now arrive in your time zone too, whether you read them through the dialogue, the transcript, or the live panel. Stored records and integrity checks are unchanged. Installing, updating and revoking a content pack from your assistant works again for signed-in accounts and licence seats; those three actions had been rejecting every request since v0.35 because of how the requester was labelled internally.

v0.139.0 — every time shown in your own time zone (September 2026)

The dates and times SQUAT reports to your assistant — when a round completed, when a summary was saved, when an issue was raised, when a member joined — now arrive in your time zone rather than in the server's. One exception remains for now: the per-turn timestamps inside a dialogue transcript still arrive as the server recorded them, and follow in the next release. Set the zone once from your account page or by asking your assistant to update your profile with an IANA zone name such as America/New_York; a workspace zone stands in for anyone who has not set one, and UTC is the last resort, always labelled. Nothing stored changes: the server keeps its own clock, and only what you see is translated. The Round Summary card's date follows the same rule, and a workspace's bootstrap now names the person who purchased it.

v0.138.0 — reply to a support case from your assistant, and one id limit everywhere (September 2026)

You can now add to an open support case without leaving your assistant: ask it to reply to the case by its reference, and your message is appended to the case for the SQUAT support team, with your workspace shown as the author. The same release standardises the maximum length of every system id at 256 characters, so a card, a tool and the server can never disagree about which ids they accept, and an id you refer to is matched exactly as you send it rather than quietly tidied. Thirty-nine more tools describe the shape of what they return so your assistant can rely on it, and a handful of stored dates that had been saved in an internal format were repaired in place.

v0.137.0 — saved research reports are findable by name (September 2026)

Saving a research report now returns its title alongside its id, so the assistant can refer to your saved work by name. Two content-pack tools describe their results precisely, and an accounting fix makes the Research Reports card report exactly which rows it withheld and why.

v0.136.4 — research report lists stay in step with their reader (September 2026)

A guard was added so the continuation cursor a Research Reports card hands back can never be one the reader would refuse. No visible change unless you had hit that edge; it is now pinned by a test.

v0.136.3 — legacy Research Report links stay honest for unusable stored IDs (September 2026)

The Research Reports card now checks every stored report identity against the exact reader contract before it presents the row. A legacy record whose ID contains a character the reader cannot accept is omitted with an identity warning instead of appearing as a clickable report or being described as an ordinary missing-field record. Valid dotted legacy IDs remain readable, new reports keep their stricter modern ID format, and no saved data is rewritten.

v0.136.2 — saved legacy Research Reports open from every place they are shown (September 2026)

A Research Report saved under an older dotted record ID can now be opened from its list, card, continuation page, or @squat: resource, and an authorized owner can invalidate it when it is obsolete. SQUAT preserves the exact stored identity instead of displaying a clickable report and then rejecting the same ID when it is opened. New reports keep their stricter modern ID format, no saved report is renamed or rewritten, and no customer action or data migration is required.

v0.136.1 — generated issue keys are stored without surrounding spaces (September 2026)

SQUAT now normalizes accidental leading or trailing transport spaces before it stores a generated issue key in a UAT Report. The canonical key is then used unchanged when the report is rendered and linked to its tracked issue. Existing canonical keys are unchanged, and no workspace data is migrated by this release.

v0.135 — two approved legacy development rounds have a governed recovery path (August 2026)

SQUAT now includes a private operator recovery command for two specifically approved legacy development rounds. Publishing the release does not start the recovery: the operator must deliberately run each approved case after deployment. The process locks the affected round while it verifies and repairs the complete dialogue chain, can safely resume after interruption, and produces a final receipt. Saved summaries, drafts, reports, and report versions are not rewritten. No production workspace or other round is included, and no customer action is required.

v0.134 — support can post a clearly marked update without closing or advancing your case (August 2026)

A support operator can now send a mid-flight progress note or follow-up question while your case remains in its current open status. The conversation marks it as an operator update rather than a reviewed answer, and the delivery records the exact primary and CC recipients used. The notice contains no account or case link. A finished, duplicate, or spam case must be deliberately reopened before another update can be posted; reopening returns it to triage and never silently preserves a spam action. Your portal access rules do not change, and no customer action is required.

v0.133.4 — the bounded development preview recognizes one older receipt shape safely (August 2026)

The read-only operator preview for two specifically approved development rounds now recognizes an older append receipt that lacks one duplicated nested timestamp. Its outer timestamp and every identity, sequence, and integrity value still have to match exactly; the exception does not apply to the other approved round or any other receipt shape. The preview remains unable to reveal dialogue, repair data, rotate keys, or change a workspace. No production workspace is included and no customer action is required.

v0.133.3 — the bounded development preview recognizes receipt-free finalization markers safely (August 2026)

The read-only operator preview for two specifically approved development rounds now recognizes the two finalization markers that are intentionally stored without append receipts. That exception applies only when those markers carry no protected dialogue body; content-bearing or unrelated events still require exact receipts and fail closed when one is missing. The preview remains unable to reveal dialogue, repair data, rotate keys, or change a workspace. No production workspace is included and no customer action is required.

v0.133.2 — the bounded development preview accounts for saved research without changing it (August 2026)

The read-only operator preview for two specifically approved development rounds now lists every saved summary, draft, report, and report version that depends on the source, and locks that complete inventory into its review reference. The preview does not edit or invalidate those saved records, add a repair command, or change any workspace data. If a separately approved future repair makes an older authorization stale, the completion record will report that fact without requiring an acknowledgement. No production workspace is included and no customer action is required.

v0.133.1 — the bounded development preview reports the complete known legacy mismatch set (August 2026)

The read-only operator preview for two specifically approved development rounds now reports every saved payload whose recomputed integrity value differs from its stored value, rather than stopping after the first known legacy mismatch. It still refuses missing or malformed events, inconsistent saved stream metadata, or incomplete linked records; the repaired result would still have to pass the unchanged strict verifier. The preview reveals no dialogue and changes no workspace data. No production workspace is included and no customer action is required.

v0.133 — operators can preview a bounded legacy development repair without changing data (August 2026)

SQUAT now has a read-only operator integrity preview for two specifically approved legacy development rounds. It checks the complete saved dialogue and every linked integrity record, then produces a deterministic approval reference for later review. The preview cannot reveal the dialogue, repair or re-seal content, rotate a key, write an audit, or change any workspace data. No production workspace is included and no customer action is required.

v0.131 — new dialogue ledgers stay verifiable when protected content is re-sealed (August 2026)

New rounds now keep the protected envelope for each spoken turn separate from the immutable event-chain material, while committing that chain to the turn's exact revealed meaning. Transcript reads look the same. Before completion, SQUAT verifies both the chain and every semantic commitment, then refuses to close the round if the stream or its protected-content revision changed during verification. Existing rounds preserve their original bytes and are not silently migrated. This release performs no live reprotection or key rotation.

v0.129.2 — report integrity survives protected-content re-sealing (August 2026)

Saved Round Summaries, Research Reports, UAT Reports, and result evidence now use one semantic integrity check over their revealed meaning, so moving the same protected content to a new storage key does not look like a content change. The evidence check between an approved Round Summary draft and its permanent save follows that same boundary. Unreadable or mismatched report material stops repair and backup work before it can be accepted. Full backups verify UAT Report versions while preserving their original protected bytes. No workspace migration or key rotation is performed by this release.

v0.129.1 — key reprotection stops before touching an unsafe record (August 2026)

Before SQUAT begins moving protected workspace content to a newer data key, it now checks the complete selected catalog and stops the whole run if any record type cannot yet be rewritten without disturbing its integrity history. A stopped run creates no job or progress records and changes no protected content or stored object. This is a safety gate for the remaining integrity remediation; no customer action or data migration occurred in this release.

v0.129 — support replies can include a case-managed CC list (August 2026)

A support operator can now keep a bounded CC list on your case. The list is validated and normalized by SQUAT, cannot be supplied by customer intake or changed ad hoc on a send request, and automatically excludes the reply's primary recipient. Each delivery records the exact primary and CC recipients used at that moment, so a later case update applies only to later replies. Your existing support portal access rules do not change, and being copied on an email does not grant portal access.

v0.128 — support can reply to the verified holder of a submitting seat (August 2026)

If you open a support case from a connected AI without including a contact email, a support operator can now send the human-approved response to the current verified holder of the exact workspace seat that submitted it. That message is a complete SQUAT Support email with the approved reply and case reference. It deliberately has no case-page link and grants no portal access, because the portal remains limited to the verified requester email stored on the case. Cases opened with a contact email keep their existing email and portal experience. No customer action is required.

v0.127.1 — scheduled announcements arrive in the intended window (August 2026)

Maintenance and service announcements now use the actual instant in their activation and expiry times, regardless of which time-zone offset the operator supplied. An announcement cannot be created with an expiry equal to or earlier than its activation, and existing valid schedules continue to work without a data migration. No customer action is required.

v0.125 — long support conversations keep their newest replies visible (August 2026)

Opening a support case now starts with its newest customer-visible activity, even after the conversation grows beyond one page. You can continue backward to older replies without losing messages that share the same timestamp. The original intake remains visible on every page, while internal notes and processing steps stay private. No customer action or data migration is required.

v0.124 — record cards now come directly from your SQUAT records (August 2026)

Workspace, squad-member, round, issue, UAT Report, Research Reports, playbook, artifact-library, and messages cards now read the authorized SQUAT records themselves. Your AI no longer has to copy those records into a second card layout, so the displayed facts, actions, tallies, page limits, and continuation markers all come from one canonical source. Steering and Steward working views remain clearly labeled Agent-composed view.

Rich clients and direct MCP clients such as LM Studio receive the same SQUAT authoritative label and complete content. Research and message pages now say when another page exists and carry the cursor needed to continue safely; a changed message set is not mistaken for the next page of the old snapshot. If a client connected before this release reports an old {card: ...} input shape, reconnect or refresh its SQUAT tools. No workspace-data migration is required.

v0.123.13 — a saved owner seat no longer looks like a failed setup (August 2026)

If SQUAT successfully creates the owner seat for a new purchase or repair, a temporary problem loading the optional plan-limit note no longer makes the whole operation look failed. The saved seat remains the result, retrying finds that same seat instead of making another, and an operator repair sweep continues to the next account. Checks that happen before the seat is saved remain strict. No customer action is required.

v0.123.12 — an unavailable total is no longer shown as zero (August 2026)

The administrative overview now shows an unavailable inventory, growth, or prompt-version count as an em dash instead of reporting a convincing zero. A real empty result still shows zero, and the active-seat total remains strict because it controls billing state. This makes a temporary datastore or index problem visible without changing workspace data. No customer action is required.

v0.123.11 — plan limits stop safely when usage cannot be verified (August 2026)

If SQUAT temporarily cannot verify your workspace's active seats, personas, or projects, it now pauses a new addition and asks you to try again instead of treating the missing count as zero. Existing work remains available, removals are unaffected, and unlimited or intentionally ungated workspaces keep their normal behavior. This prevents a datastore or index problem from silently creating capacity beyond the paid plan. No customer action is required.

v0.123.10 — your seat count now follows active licenses (August 2026)

Workspace information now reports seats from active license records, the same source SQUAT uses to enforce your plan, instead of treating every member record as a billed seat. Revoked licenses and retained inactive members no longer inflate the number, while an active account-only seat still counts even when it has no key attached. The member list is unchanged. No customer action is required.

v0.123.8 — saved research resources keep their labels trustworthy (August 2026)

Round Summaries and Research Reports opened as MCP resources now keep saved titles, round names, and finding labels inside the part of the document where SQUAT placed them. Rich report bodies and finding observations still render as authored, but an observation cannot masquerade as a sibling SQUAT heading. Resource names remain normal text, and LM Studio and other direct MCP clients receive the same resource fields as before. No customer action is required.

v0.123.7 — imported round names stay exact and contained (August 2026)

Historical round imports now protect their optional display name the same way live rounds do. Ledger data still gives direct clients such as LM Studio the exact name, while the text table keeps that name inside its own row. Profile and consent confirmations apply the same one-line rule to names and service errors, and artifact metadata remains compact, parseable JSON with unchanged values. No customer action is required.

v0.123.6 — direct-client cards stay on one trustworthy line (August 2026)

SQUAT now handles every line-break character recognised by common client runtimes before it builds card text, journey maps, or live panel updates. A participant name, response, or round identity cannot make LM Studio or another direct MCP client interpret one result as a second heading, fence, or SQUAT authoritative label. JSON responses remain normal parseable JSON with the same values, and ordinary card and tool output is unchanged. No customer action is required.

v0.123.5 — journey maps keep their own labels contained (August 2026)

Journey maps now protect their text view at every point a step name is displayed, including the current-step subtitle, compact line shown by direct MCP clients, and fixed-width outline. This is defense in depth for the existing server-derived journey; approved step names and rich cards look the same, and no customer action is required.

v0.123.4 — live panel labels stay trustworthy (August 2026)

The live panel viewer now rejects expired or malformed access before reading any dialogue, and its text view safely displays the round identity even if a connected client bypasses its normal input checks. This is defense-in-depth for the existing SQUAT authoritative label; valid verified-account and direct-license sessions continue to work as before. No customer action is required.

v0.123.2 — authoritative cards prove their boundary (August 2026)

SQUAT now checks every card renderer against the exact composition label it is allowed to return. A prepared AI layout remains an Agent-composed view; a record-derived display remains SQUAT authoritative. Missing or inconsistent Round Summary counts stop the card instead of appearing as trustworthy zeros.

The live panel viewer now carries the same distinction. Its title and authoritative label come from SQUAT, not from the connected AI, and direct clients such as LM Studio receive that label before the viewer details and each live update. The viewer checks the complete labeled update before displaying new turns. The label means the dialogue and artifact references came from your authorized SQUAT records; it does not claim that every statement made by a participant is independently verified.

v0.123.1 — the connection approval page cannot be framed (August 2026)

The SQUAT connection approval page can no longer be embedded inside another website or client window. You still open it normally to sign in or enter a license key, but another site cannot place an invisible or misleading interface over that form. No customer action is required.

v0.123 — cards tell you who composed them (August 2026)

Every card now says whether it was derived by SQUAT from your stored records or assembled by your AI as a useful working view. SQUAT authoritative means the server read the record and calculated the display; Agent-composed view means the AI arranged available context and SQUAT has not verified that synthesis. The distinction appears in rich cards and in the complete text shown by direct clients such as LM Studio.

Round Summary cards now come directly from the saved draft or permanent summary. They separate persona results cited, findings written in the summary, and raw finding occurrences on the cited results, so a severity chart can no longer look like it is counting a different set. Journey cards now derive themselves too; clients no longer have to translate the journey checklist into a second payload.

If a client connected before this release reports that a Round Summary or journey card has the wrong input shape, reconnect it or refresh the plugin so it reloads the new tool definitions.

v0.117 — the account pages stop asking you to sign in twice (August 2026)

Signing in worked, and then the Workspaces page sent you straight back to the sign-in form — a loop with no exit. The cause was never your account: a security layer that verifies which app is calling (not who you are) had no way to recognize the SQUAT website's own backend, so it turned the website away, and the page misread that as “not signed in.” The website now identifies itself to the SQUAT server the same way our internal services do, and if that handshake ever fails again, the page says plainly that the fault is on our side — it will not blame your sign-in for a problem that isn't yours.

v0.116 — readiness sees everything you own (August 2026)

The readiness checklist could tell you to create a test scenario you already had. Older scenarios and projects, saved before a status field existed, were invisible to readiness while remaining visible everywhere else. Readiness now uses the same rule as every other view — anything not explicitly archived counts as active — so the checklist and your library can no longer disagree.

Cards also stop failing to load when you run many assistant sessions from one machine or office network. A safety limit meant for anonymous traffic was throttling legitimate signed-in work; the limit now distinguishes the two, and your license's own budget is what governs.

And support access grants can now run up to 7 days, your choice at grant time (previously 24 hours) — for the issues that take longer than a day to resolve. Everything else about grants is unchanged: exact fields only, fully audited, revocable by you at any moment. One note: an AI assistant connected before this release may still refuse durations past 24 hours — it is holding an older copy of the tool definitions. Reconnect the assistant (or toggle the SQUAT plugin off and on) and the full range is available.

v0.115 — the issue board, and three small fixes (August 2026)

Your workspace now has an issue board: ask for it and SQUAT tallies every open finding by severity and by the feature it's attached to, worst first, with the true total always stated alongside what's shown. It reads directly from your findings, so it can't drift out of sync with what verifying an issue already knows.

Three smaller fixes from the same night's testing. A configuration mistake meant some environments could silently load the wrong set of cards; that's now derived correctly. An assistant holding a stale copy of SQUAT's tool list used to tell you the terms had changed when they hadn't — it now says plainly that you should reconnect it, not re-accept anything. And the connecting line on a staged journey card, which had drifted loose from its dots, is back where it belongs.

v0.114 — your research plan gets a home (August 2026)

A playbook you build with your assistant — research questions, a recommended cast, a staged workplan, the decisions that matter — can now be saved and viewed as its own card, the same way a round or a scenario can. Stages, checks, and decisions are tracked in order, and the card is sealed the same way the rest of your workspace's content is.

v0.113 — confirming a full account purge, without your assistant ever seeing the code (August 2026)

Emptying your trash for good is the one action in SQUAT that can't be undone, so it now asks for a second, human-only confirmation: SQUAT generates a one-time 6-digit code, sends you to a page on squat.pro to see it and enter it, and your AI assistant never sees the code at any point — only you do. It expires in five minutes and can only be used once.

v0.111 — see your own support cases (August 2026)

Sign in at squat.pro and you can now see every support case you've opened — status, product area, and what you asked for — in one place, scoped strictly to cases tied to your own verified email.

v0.110 — attach a screenshot to a support case (August 2026)

A support case can now carry up to four attachments — screenshots, transcripts, or plain-text/markdown notes, 5MB each — so you can show us what you're seeing instead of describing it.

v0.109 — a guided way to build your research plan (August 2026)

SQUAT now offers a guided playbook flow: talk your assistant through an interview, or hand it evidence you've already gathered, and it drafts grounded research questions, a recommended cast (it will never create a persona on its own), a staged workplan, and the decisions in front of you along with what choosing wrong would cost.

v0.108 — exporting one person's data, not just deleting it (August 2026)

If someone asks what SQUAT holds about them specifically, you could already delete it end to end. You can now export it the same way — exactly the records tied to that one reference, nothing else, readable and byte-verified.

v0.105 — ask your assistant to open a support case (August 2026)

You can now open a support case directly from your AI assistant, no separate web form required. Each case gets a short reference number (SQUAT-1001, and up) you can quote back to us, and you can ask your assistant to check its status any time. Nothing is sent to you automatically — a human reviews and approves every reply before it goes out.

v0.104 — two method guides join your assistant's knowledge, and the legal documents update (August 2026)

Your assistant now has two official method guides to draw on directly — one for running a UAT round, one for running a qualitative panel — the same grounded process this manual describes.

Separately: the Terms of Service, EULA, Privacy Policy, and Acceptable Use Policy all moved to a new version this release (the privacy document is now a full Privacy Policy, not a notice). If you're prompted to accept updated terms, this is why.

v0.103 — importing a persona now tells you what didn't fit (August 2026)

Persona profiles used to accept anything you gave them, silently. They now check structure, not meaning: recognized fields are recognized, and anything else goes into a clearly labeled "extended" section rather than being dropped or guessed at — an unrecognized field is named back to you rather than failing quietly. And before your assistant imports a profile document on your behalf, it now reads what's in it and asks you about anything that looks out of place, rather than importing it blind.

v0.102 — the live catalog and alpha offer are ready (August 2026)

The ALPHASQUAD promotion is live: 50% off the eight paid products for three months, available for new redemptions through October 31, 2026, and limited to 100 redemptions. It deliberately applies to both subscriptions and packs; if both it and the 30% studio-volume pack discount are attached to the same subscription, they stack by design.

The production catalog now has its complete set of products and prices, each connected to the price the website displays.

v0.101 — the UAT Report becomes something you can see (August 2026)

The UAT Report, introduced last release, now has an actual card your assistant can show you and reach for on its own — the missing piece that made it real rather than theoretical.

v0.100 — clearer, more dependable results for your assistant (August 2026)

More SQUAT tools now describe the shape of what they return, so your assistant can rely on structured results instead of guessing from free-form responses. The contract is checked against the tools actually registered, while the small number of media responses that intentionally remain unstructured are called out explicitly. No action is required.

v0.99 — a new kind of artifact: the UAT Report (August 2026)

A Round Summary covers one round; a Research Report answers a question you asked. The UAT Report is different: it assigns work. It's built across many findings at once, always states its true total alongside what it shows — a report that quietly covered twelve of forty-six findings would be worse than no report at all — and every fix it recommends points you at running a new verification round, never at simply marking something resolved.

v0.98 — your draft survives the session (August 2026)

A Round Summary draft used to live only in the conversation you were having. Close the session and it was gone, and the round had to be written up again from nothing. Drafts are now stored against the round: come back tomorrow, ask for the draft, and it is there as you left it.

A draft is working material, and behaves like it. Saving again replaces it rather than stacking up versions, and there is no draft history — a draft is the thing you are still changing. History begins when you save the summary, which is what makes it permanent. Saving promotes the draft and clears it, so you are never left with two copies of the same analysis quietly drifting apart; if a save fails, your draft is untouched. You can also discard a draft outright, which only ever removes working material and never a summary you have already saved.

Your draft is encrypted in storage the same way a saved summary is — it is your analysis of your product, and being unfinished does not make it less yours.

v0.97 — you read the summary before it is saved (August 2026)

When a round finished, the next thing SQUAT offered you was Save round summary — which is the last step of writing one up, offered as though it were the first. The move on offer is now Draft round summary: your AI prepares the summary and shows it to you, you ask for changes and ask follow-up questions about the findings, and it saves only when you say so. Saving is what makes a version permanent, so it belongs at the end.

Related: a draft you were still working on couldn't be revised, while a summary you had already saved could — backwards, since a draft is exactly the thing you revise. Drafts now offer revision, so you can get a summary right before it becomes a permanent version rather than correcting one afterwards.

Drafts still live in your conversation until you save them. If a session ends before you save, the draft goes with it — the round is untouched and can be summarized again. Durable, resumable drafts are a separate piece of work.

v0.96 — a panel that runs short says so (August 2026)

Real panels run short. Someone is unavailable, a persona's source withdrew consent, a squad member was never activated. SQUAT has always let a round run with a narrower cast than the squad line-up — you don't cancel a panel of ten because one person is out. What it did not do was leave any trace of the shortfall: the round recorded who took part, and nothing about who was supposed to. A nine-person panel drawn from a ten-person line-up looked exactly like a panel that was always nine.

A round now stores the line-up it started from and every declared member who didn't take part, with a reason where one was given — unavailable, consent withdrawn, retired, not activated, or other. Ask about the round and it tells you in plain words: “Nine of the ten declared panelists took part; Marta Nilsson was unavailable.” Your round summaries should carry it too, because nobody should have to discover a shortfall by comparing two lists. Reasons come from a fixed list rather than free text on purpose — who missed a session is a scheduling fact about your panel, not research content, so it stays plainly readable to you instead of being sealed with your testers' responses.

How small is too small remains your call, and it should be: an internal design review and a regulated study have honestly different standards. Decide up front what your minimum panel is, what you'll do below it, and which absences actually matter to your findings, then tell your AI — it will hold rounds to that standard and say when one falls below it. SQUAT's job is to record what happened accurately enough that your standard can be applied to it afterwards.

Rounds recorded before this release don't carry these fields. An empty absence list on an older round means “not recorded”, not “nobody was missing”.

v0.85 — your words, sealed by default (August 2026)

SQUAT now decides what to encrypt by a single rule: if you could have typed it, it is sealed. Not our judgement of what looks sensitive — authorship. Project names and notes, scenario steps and tags, round missions and the build you were testing, squad names, issue titles, finding observations and suspected causes: all encrypted in storage under your workspace's own key. What stays readable is only machine-generated identifiers, timestamps, and SQUAT's own status values, because a handful of them have to be searchable for the product to work at all. None of those hold your text.

Behind that, key rotation became something we can actually run. Moving a workspace to a new encryption key now works through every record it owns, resumes where it left off if interrupted, and confirms afterwards that nothing is still relying on the old key — verified end to end on a full workspace before this release. Your files were checked byte-for-byte through the same path you download them by.

Two boundaries were tightened while we were in there: a key rotation can now only ever touch the workspace that asked for it, and the background job that carries out the work can no longer be pointed at a workspace it was not authorised for. Both were already unreachable from outside; they are now refused outright rather than relied upon not to happen.

v0.84 — rotation that finishes on its own (August 2026)

Re-encrypting a large workspace used to need someone watching it. It now runs as a chain of background batches that hand off to each other until the work is done, retrying safely if a batch fails and never redoing work it has already recorded. Nothing about your data changes — this is the machinery that lets us move a workspace to a fresh key without downtime.

v0.80–v0.83 — protection you can point at (August 2026)

Every encrypted field is now registered in one place, so SQUAT can answer "what exactly is protected, and under which key version?" for your workspace rather than leaving it implicit. That registry is what makes safe key rotation possible, and it is checked automatically — a field that gets sealed without being declared stops the build rather than surfacing later as an unreadable record.

v0.79 — three defects found by our own squad (August 2026)

We ran SQUAT through its own methodology again, cold, and fixed what it found. The fixes came with tests that fail against the old behaviour, so the same defects cannot quietly return.

v0.78 — scored rounds can finish (August 2026)

The first scored round in a workspace would run all the way through — every tester answering, every result accepted — and then refuse to close. Nothing was lost when that happened; the round simply stayed open. It closes properly now, and a round that follows one still records what changed since last time.

Three smaller corrections found in the same testing. The connection check your assistant runs when it first reaches SQUAT no longer fails on a healthy workspace — it was looking for your plan in the wrong place and reporting the workspace as broken when it wasn't. Identifier fields now say what they'll accept before you get it wrong, instead of rejecting a perfectly reasonable name for an unstated rule. And a findings view for a round that hasn't finished yet no longer offers to save its summary: it says the round is still open, shows the findings you do have, and points you at completing it first.

All four were found by SQUAT's own testers, running against SQUAT.

v0.75.2 — your own retries stop counting against you (August 2026)

Some AI apps sign in fresh on every retry — and each fresh sign-in used to claim a new device slot, so a run of retries could fill your license's limit with copies of one client and lock you out. Now SQUAT recognizes the same client coming back — same address, same app — and quietly reuses its slot instead. Different machines and different apps still count separately, exactly as before: the limit still means what it says, it just stops charging you for your own retries.

v0.75 — checklists that tell the whole truth, and scenarios that know their home (August 2026)

Setup in SQUAT isn't a straight line — a tester can be recruited but not yet activated, a project can exist without its context saved — and the map now shows that accurately. Journey cards carry a checklist: each persona, project, or upcoming round listed with its real requirements checked off one by one, so "where am I?" answers with the whole picture instead of a single arrow. The workspace card gained the same treatment for protection posture, and your persona count.

Scenarios now know where they live: a scenario can belong to one project (usual for tasks written against a specific product) or sit in the workspace's shared library (usual for reusable discussion guides). Each project now sees exactly its own scenarios plus the library — no more neighbors' scenarios drifting into the wrong project's readiness. Running another project's scenario stays allowed for head-to-head comparisons; SQUAT simply says so when it happens.

We also ran SQUAT through its own methodology — sending fresh operators through setup cold — and fixed what they tripped on: guidance that pointed at a validation step that no longer exists now explains itself properly, two error messages name their exact fix instead of just the rule, the squad list no longer overflows on big rosters, and a few internal words ("artifact set") left the customer's view.

v0.74.1–2 — cards that count, and buttons that match the advice (August 2026)

The workspace status card now shows how many personas live in your workspace — not just how many are on the active squad — so "do I have testers yet?" is answered at a glance. When a card tells you the next step, it now hands you that step as its first button: "Create a project", "Add a scenario", or "Save the squad line-up", right where the advice is, instead of routing everything through the concierge. And the Compact view button only appears on cards where it actually changes something; cards with nothing to collapse no longer offer a button that does nothing.

v0.74 — recruiting unstuck, and your reports at your fingertips (August 2026)

Recruiting a tester no longer involves any validation ceremony — SQUAT now checks every piece server-side as it's saved and keeps its own receipts, so there's nothing for your AI client to prepare, prove, or get wrong. And when something genuinely isn't finished, SQUAT says so in plain words: instead of reciting internal record names, it tells you "Maya isn't finished recruiting — say 'continue recruiting Maya'" and you're moving again.

Wayfinding now reaches every client: the "where am I?" journey and the mid-session menu are available as slash commands even without the plugin (in Claude Code, type /mcp__squat__ and pick journey or menu). Saved Round Summaries and Research Reports can be pulled straight into your conversation on clients that support @-mentions — type @squat: and pick a report by name. Long transcripts now arrive in tidy pages sized to your client, continuing automatically instead of overflowing. And every SQUAT tool now tells your client accurately whether it only reads, writes recoverably, or deletes — so clients that paint warning labels reserve the scary ones for the handful of actions that deserve them.

v0.73 — ask "where am I?", and cards that chart (August 2026)

SQUAT can now show you the map: ask "where am I?" (or say /journey once your plugin updates) and you'll get your journey — the steps behind you checked off, the step you're on highlighted, branches shown accurately — in a picture where your client supports it and in plain text everywhere else. Cards also learned to chart: round-over-round trends, severity breakdowns, and a new review card that pins each finding to the exact step where it happened, with your squad members named. And SQUAT now collects de-identified, aggregate usage statistics — never your content, never your people — to help us improve the product; it's on by default and any workspace owner can opt out at any time, including right when accepting the terms.

v0.72 — name your keys, better device names (August 2026)

License keys and seats can now carry a friendly name — "CI key", "my laptop" — editable from the Seats page of your account, so two keys are never just two prefixes again. And a device that introduced itself with a connection probe's name now picks up your real client's name on its next connection; your own renames always win.

v0.71 — rename your workspace (August 2026)

Workspaces can be renamed — ask your assistant to update the workspace name (owners and admins). And framework prompts now enjoy the same automatic protection as your scenarios and personas: encrypted in storage, readable only through your workspace's roles, and visible to support only through a grant you create.

v0.70 — support access, spelled out (August 2026)

The manual now walks through letting SQUAT support look at specific records: an owner-created grant that names the exact fields, expires on a clock you set, and is audited end to end — you never share a password or key, and nobody outside your workspace and that flow can ever be granted access. The setup pages now carry SQUAT's real production addresses throughout, and the tier table reflects current hosted persona capacity.

v0.69 — one sign-in, one device (August 2026)

Connecting without any SQUAT software installed (the "connect by URL" path) no longer uses up a device slot on every reconnect — one sign-in now counts as one device, however many times you reconnect. If your device list filled up with duplicates before this fix, de-register the extras on your account's Devices page and you're set.

v0.68 — sharper cards, clearer results (August 2026)

Cards now lead with the SQUAT mark, and SQUAT's most-used tools describe their results in a structured form that AI clients understand directly — fewer misreadings of round results, issue lists, and ledger summaries in whatever client you use.

v0.67 — your purchase is your seat (August 2026)

Buying a plan now activates your workspace immediately — your seat is connected to your account sign-in from the moment the purchase settles. A license key is optional: add one whenever you want your agent tools to use the seat without your sign-in session. Revoking a key no longer touches your seat or your signed-in access, and safeguards prevent a workspace from ever losing its last owner by accident.

v0.66 — cards that name their target, and a friendlier front door (August 2026)

Card buttons can now say who or what they act on — "Recruit — Roy" instead of a generic "Recruit a member" — and a card can offer the same kind of action for two different targets side by side. Button text is still composed entirely by SQUAT from fixed templates, so nothing untrusted can ever become a clickable instruction; when a client suggests its own button text, SQUAT now says plainly that it composes the copy rather than swapping it silently.

SQUAT also greets you like a helpful colleague now: a couple of plain sentences about where things stand and a question about what you'd like to do — not version numbers, internal records, or technical constraints you didn't ask about. The details are all still there when you want them; they just stopped arriving first.

v0.65 — cleaner defect counts, and squad members that remember (August 2026)

When a test failure turns out to be the simulated tester's limitation rather than your product's fault, it no longer counts against your product. Those results are now set aside from product findings — failures attributed to the testing setup, or not yet attributable to the product, kept out of defect counts until reviewed. You can review the set-aside queue, mark each item (including promoting a misclassified one back into your real issues, history intact), and your issue counts and trends stay accurate.

Squad members also begin to remember: on plans that include it, each scored round can save what the evaluator observed about a persona, so future rounds can treat them as a returning user. On the free plan this covers the one squad member holding your learning slot; when observations aren't kept, SQUAT tells you plainly and shows them once instead of dropping them silently.

v0.59 — a clearer front door (August 2026)

• Saying /start now opens with a menu, not a wall of settings. SQUAT reads where your workspace is and shows your status, your roster, and the few moves that make sense next — with the fix offered first whenever something needs setting up. A brand-new workspace is pointed at a risk-free practice round to begin, rather than asked to fill in fields.
• The concierge is now a standing option. Reach it from the start menu any time — it's the place to go when you're not sure what to do next or how to get a job done, not just at the very beginning.

v0.47.1 — accurate project totals (August 2026)

• Project status now reports exact valid-round and panel totals. Earlier it could count only the most recent rounds it displayed, so a busy project could show fewer valid panels or rounds than it actually has. The totals are now exact regardless of how many rounds the project holds.

v0.47.0 — SQUAT-branded cards (August 2026)

• The cards SQUAT shows inside your client now carry the SQUAT wordmark. Every card — your workspace, squad dossiers, rounds, round summaries, issues, the steward, artifact and research libraries, and the steering menu — shows the mark above its title, as does the live panel viewer.
• It looks right in both light and dark. The wordmark adapts to your client's theme automatically.

v0.44.0 — reviewed plugin distribution (July 2026)

• The public plugin bundle is now reviewed before distribution. The downloadable package contains the provider-neutral skills, cards, schemas, and local launcher needed by supported clients.
• Local and hosted connection choices are clearer. The standard public package uses local stdio configuration; hosted connections can still be configured through the client that provides them.

v0.42.0 — clearer status and safer cleanup (July 2026)

• Project status says when its open-round list is complete. A status view now tells you whether more open valid rounds remain beyond the displayed page.
• Saved research can be removed safely. Owners can select Round Summaries and Research Reports for a reviewable deletion request. SQUAT includes their saved versions and supporting stored files, can resume after an interruption, and leaves source rounds and artifact sets in place.
• Expired sign-in state is cleaned up. Old authorization records are removed in bounded background batches while still-valid refresh records remain available for normal security checks.
• Optional familiar greetings. When a verified account name is available, SQUAT may use it once in a natural startup greeting.

v0.41.0 — move project history safely (July 2026)

• Local history can be prepared for SQUAT. When a project began in a local UAT setup, SQUAT can package its reviewed history, evidence, and artifacts for an intentional import while keeping credentials out of the package.
• Hosted work can be saved for review. You can create a portable local copy of current SQUAT workspace records and round/report metadata. Full backups remain available for complete recovery and deletion workflows.

v0.40.2 — safer local setup preservation (July 2026)

• Approved local dependencies stay available. Releasing SQUAT no longer removes an installed root dependency when the project declares it intentionally.

v0.40.1 — more reliable hosted updates (July 2026)

• Hosted updates retry cleanly. Repeating a deployment after its retention setup already exists no longer treats the existing configuration as a new error.

v0.40.0 — smoother startup and project saves (July 2026)

• One orientation read at startup. A connected session can use the workspace bootstrap read to choose a project and see the next readiness step without a long series of discovery calls.
• Read-only discovery is easier for supported clients. SQUAT labels safe reads for clients that use MCP read-only hints. Actions that change rounds, artifacts, reports, or data still require confirmation.
• Project saves are more reliable. Saving a project no longer fails when optional build/version information has not been supplied.

v0.39.1 — current-state guidance (July 2026)

• Research correction paths are easier to find. The manual now explains how to append a corrected panel reaction without changing the dialogue, and how to invalidate an obsolete Research Report without deleting its history.
• Panel evidence requirements are explicit. A result covers the panelist’s complete exchange, including follow-ups and redirects; the structured reaction cannot cite only a convenient subset.
• Protection language matches current operation. Connected SQUAT protects hosted content automatically; clients do not perform a separate sealing or workspace-key setup step.

v0.39 — stronger research evidence records (July 2026)

• Historical rounds can have useful, accurate Round Summaries. Older completed rounds that predate the dialogue ledger can be summarized from all of their submitted results. The summary clearly records that no aligned transcript is available.
• Panel follow-ups stay attached to the reaction. A panel result must include every answer in that panelist’s exchange, including follow-ups and redirects, before it can be accepted.
• An inaccurate panel reaction can be corrected without rewriting history. Owners and admins can append a corrected structured reaction while retaining the original version and the underlying dialogue.
• Invalid Research Reports are preserved safely. A report built from a bad source contract can be marked invalid, excluded from current project coverage, and kept for audit without returning unverifiable content.
• Large report compilations are more reliable. Research Reports can combine many saved Round Summaries while still detecting concurrent source changes.

v0.38.1 — faster project status and cleaner reader metadata (July 2026)

• Project status remains responsive as research history grows. SQUAT now reads only the recent Round Summaries shown in the status view while still returning the exact number of saved Research Reports.
• Research readers receive research content, not account identifiers. Round Summary and Research Report views no longer include the internal member, license, or request references used to audit who saved a version.
• Large summary libraries continue cleanly. Round Summary lists retain cursor-based pagination beyond 100 saved rounds.

v0.38 — Round Summaries and compiled Research Reports (July 2026)

• Each round can have its own saved write-up. A Round Summary records the analysis, findings, transcript coverage, and supporting evidence for one completed round. You can review the draft before saving it and add later revisions or addenda without replacing earlier versions.
• Research Reports can combine several rounds. Select saved Round Summaries from the same project to compile a broader report. Later report versions can add more summaries while preserving the earlier report.
• The review flow makes the distinction clear. Round Result cards show what happened operationally, Round Summary cards show one-round analysis, and the Research Report library shows the cross-round reports saved for the project.
• Empty report libraries have a useful next step. If no Research Report exists yet, SQUAT offers to select saved Round Summaries and compile the first one.

v0.37 — project ledgers and research-report libraries (July 2026)

• Project history stays in context. Choose the project you are working on and SQUAT keeps its rounds, scenarios, artifacts, and reports together for that task.
• Scored tests and qualitative panels have different ledger views. UAT rounds retain pass, fail, inconsistency, and regression fields. Panels show completion and participant reactions without implying a score.
• Rounds can have readable names. Names lead in history views while stable references remain available for precise follow-up.
• Saved research reports are easier to find. A dedicated report card lists report titles, versions, transcript coverage, source-round counts, and findings, with copyable references.
• Panelists receive the project context used when the round began. Moderator, panelist, and review steps use the same saved project-context version throughout the round.

v0.36.1 — optional build attribution (July 2026)

• Target builds are a choice. Scored rounds ask once and let you continue without setting a product version or build. Concept panels normally skip it.
• Cards stay focused on recorded facts. If a round has no target build, its card simply omits that field instead of presenting a warning or empty value.

v0.36 — question-bound panel artifacts (July 2026)

• Each panel question can carry its own exact artifacts. A moderator guide can pair one question with ordered images, audio, text, or another file type from one finalized project artifact set. A panelist receives only the files for the current question.
• Artifacts appear before the question. Compatible clients place the exact file content in the visible conversation before asking the paired question. The live panel viewer previews safe images, audio, and text, and provides a file tile for other formats.
• Panel completion checks the exchange. SQUAT records when the coordinator presented the required artifacts and when the panelist responded. Missing, reordered, or mismatched exposure evidence pauses the panel instead of accepting an incomplete result.
• Interrupted panels preserve the same rules. Recovery can restore an already-authorized artifact-bearing response only when its prompt, exact file references, exposure record, and response linkage still match the open round.

v0.35.1 — reliable repeat releases (July 2026)

• The daily retention schedule stays configured across repeat releases. This maintenance update corrects the deployment path used when the authenticated retention job already exists. No customer action is required.

v0.35 — verified accounts, live panels, and managed data lifecycle (July 2026)

• Account-backed sessions represent a verified person. When the catalog connection is enabled, SQUAT binds the signed-in account to its workspace membership and assigned seat. Direct license connections remain available.
• Live panels can be followed turn by turn. The first-person panel view shows moderator questions, panelist responses, and the exact visual artifacts shown during each exchange.
• Deletion and retention have a governed workflow. Authorized operators can prepare supported deletion requests, apply legal holds, execute approved removals safely, and retain a completion certificate without copying deleted content.
• Project artifacts work across compatible clients. Authenticated resource links let a capable host retrieve the exact finalized artifact without exposing storage coordinates. Local transfer remains the fallback for hosts that do not support the resource.
• A free guided tutorial reduces cold-start work. The built-in tutorial pack can install a starter persona, scenario, moderator guide, and squad through the same governed content workflow used by future packs.
• Readable reference codes are available. Optional study, panel, and scenario codes make common references easier to recognize and copy while stable system IDs remain the authoritative record.

v0.34 — clearer panel recovery and progress (July 2026)

• Panel setup tells you how to correct a mismatch. If you choose a task scenario where a panel guide is required, SQUAT names the steps to create one moderator instrument for the research purpose and recheck readiness.
• Restarted coordinators return to the last valid step. SQUAT reports the current panel checkpoint and the valid recovery action instead of letting a new session improvise the next question or result.
• Interrupted dialogue has a precise recovery boundary. A launcher-capable client may restore an already-authorized turn whose hosted append failed. If the server had not issued the next role brief, the panel pauses. Recovered turns remain visibly identified in the transcript record.
• Transcripts can be followed incrementally. Authorized clients can request only the canonical turns recorded after their last known sequence, which supports reliable resume and future live panel views.
• Research reports cite the evidence they used. Saved reports can point to exact dialogue turns and exact project artifacts while preserving the underlying transcript and file versions.

v0.33 — guided, verifiable panel execution (July 2026)

• Panels guide the moderator from one valid step to the next. SQUAT reports the next ready panelist, question, response, neutral review, result, or completion action from the live round record.
• Each response is tied to the question actually asked. A panelist receives the recorded moderator prompt, and the answer must follow that exact prompt in the canonical dialogue.
• Follow-ups and redirects remain flexible. Moderator wording is not scripted, but every new exchange is recorded in order and reviewed before the panelist result is submitted.
• Incomplete orchestration cannot be mistaken for a complete panel. SQUAT checks every selected panelist, every saved primary question, the complete dialogue references, the neutral review, and one valid reaction per panelist before closing.
• Research synthesis remains a separate final job. The cross-panel readout begins only after all individual reactions are complete and continues to use the versioned research-report workflow.

v0.32 — panel preflight and artifact library (July 2026)

• Panels can be checked before they start. SQUAT verifies the project, its default squad, the selected members, the moderator guide, and any finalized artifact set before creating the panel. If a task scenario was selected where a moderator guide is needed, the operator gets a clear correction first.
• Common lifecycle questions have direct answers. Ask whether a scenario has actually reached panelists, whether a round is open or complete, or whether a project is ready for its next round without searching old chats or local session files.
• Local attachments have a standard upload inbox. Launcher-capable clients can place selected files in the private SQUAT transfer folder and use the normal hosted artifact workflow. Local-only work remains available when the hosted service is unavailable.
• Artifact libraries are easier to browse and reuse. List artifacts by workspace, project, or scenario, then retrieve the exact finalized versions used in earlier valid rounds.
• Images appear in a slide-sorter view. Capable MCP Apps clients show responsive thumbnails, file fallbacks, details, and copyable reference IDs. Other clients receive the same inventory as readable Markdown.
• Names lead; IDs support. Projects, scenarios, squad members, rounds, and artifacts are described by readable names first. Stable IDs remain available as secondary copyable metadata when needed.

v0.31.1 — focused cards and record-level provenance (July 2026)

• Workspace cards focus on the work. Normal status cards show the workspace name, activity, roster, latest round, and next actions without repeating generic active/internal labels or account-plan details.
• Provenance appears where it is useful. Persona records retain their acquisition/source context. Dialogue records distinguish live-session turns, recovered appends, and preserved archive imports.
• Reports and results are easier to read. Cards, reports, issues, ledger summaries, and ordinary operator responses no longer repeat the same simulation banner.
• Card headings are host-dependent. The card itself uses its SQUAT title. Some clients may still show a humanized tool name in the collapsible frame above it; that frame belongs to the client.

v0.31 — protected workspace content and reusable artifacts (July 2026)

• Protection stays automatic across clients. Claude, ChatGPT/Codex, LM Studio, and other compatible clients use the same ordinary SQUAT conversation. You do not set up, carry, or recover a workspace content key.
• Dialogue and results are easier to pick up later. Spoken turns are the canonical round record, and authorized reads reconstruct the same aligned moderator-and-panelist exchange after a client reset. Results cite those turns instead of relying on a separate transcript upload.
• Attachments belong to the project. Images, files, and other panel inputs are saved in the project artifact library by default, so you can reuse them in a later round or attach the same exact item to a research report. Local files remain useful for local-only work or an interrupted hosted session.
• Artifact management is explicit. List project or workspace artifacts, rename their display labels, set retention policy, and delete an unreferenced set. SQUAT keeps the historical item reference intact when a round or report already relies on it.
• An authorized failed append can be recovered after a disruption. If a capable local client preserved a receipt-backed turn the server had already authorized, you can review and load it into the same open round. The batch is accepted as a whole or left unchanged; no new offline exchange is treated as live panel execution.
• Portable exports are readable. Current persona governance and runtime material, results, and research reports travel as an authorized readable package. Administrative access records, licenses, and service protection material stay out of it.
• Support access is deliberate. A workspace owner can authorize a named support case for exact protected fields and a limited time, review the grant, and revoke it. General platform administration does not grant content access.

v0.30 — automatic panel protection and recovery (July 2026)

• Connected panels no longer need encryption setup. Claude, ChatGPT/Codex, LM Studio, and other MCP clients send ordinary panel questions, guidance, dialogue, subjects, and reactions. SQUAT protects the hosted panel-critical content automatically; you do not create or recover a workspace content key.
• Role-specific reads stay simple. Actors receive only their spoken question and runtime material, moderators receive the private guidance they need, and evaluators receive the evaluation view. Storage and encryption details do not appear in the conversation.
• Interrupted panel dialogue can return to the same round. A launcher-capable client can preserve an already-authorized turn whose append failed and, after you approve, load the validated batch into the existing open round. The file is unchanged, identical retries do not duplicate turns, and the aligned transcript identifies the recovered origin.
• Panel files now have a hosted library. Create a reusable project artifact set, finalize an ordered version, and pin those exact files to a panel. Dialogue and saved research reports can cite the exact items shown. Small image/audio/file content works through standard MCP; larger files need the local launcher or another compatible host adapter.
• Artifact retention is explicit. You can rename sets/items and record an expiry policy without rewriting historical versions. Expiry is not automatic yet, and a referenced set cannot be deleted.
• Local recovery files still need care. They remain readable research content on your device until loaded. Chat history alone is not a recovery file; if a client cannot create durable capture, pause before the next question.

v0.29.1 — reliable cards (July 2026)

• Round cards no longer fail over customized button wording. SQUAT now supplies each card action’s label and follow-up message from its action id, so a moderator can present the right next steps without reproducing hidden fixed copy.
• Card headings describe the result. Embedded cards use labels such as SQUAT round result and SQUAT next steps. A client may still show its own tool-call label in the surrounding frame.
• Steering menus no longer show an inert view toggle. Rich/compact controls appear only when they change visible content, and a missing card now gets a clearer message without guessing why it was absent.
• Visual panels explicitly show the visuals. The coordinator loads attached images and includes the same ordered set in each vision-capable panelist call instead of assuming a separate agent inherited the attachments. If the selected model route cannot receive images, the panel stops rather than inventing a reaction.

v0.29 — canonical dialogue and reliable panels (July 2026)

• A project’s default squad is now authoritative. A round uses that exact saved line-up or an explicit subset. A different squad, an outsider, a retired member, or an archived project is rejected before the round begins. Panels and steward sweeps follow the same rule.
• Dialogue has one canonical record. New rounds append each spoken turn to an ordered event ledger and reconstruct the exchange from that chain. Ordinary turns remain inline; unusually large content uses the managed large-object path transparently. Historical inline and object-backed transcripts also remain seamless to read.
• Round coordination is recoverable and single-writer. An event-ledger round closes only with its active dialogue coordinator credential. If result_upload_finalize is missing from a current tool list, reconnect before uploading or finalizing because that MCP session is stale.
• Missing capture is explicit. A contract panel pauses for canonical dialogue recovery or invalidation. Eligible non-panel rounds may use a separately approved evidence-gap record instead of silently pretending a transcript exists.
• Panels require a moderator instrument. A versioned instrument supplies the spoken turns and private moderator guidance; an ordinary task scenario cannot substitute for it.
• Role bundles minimize data without overclaiming isolation. SQUAT-generated runtime, actor, moderator, and evaluator bundles exclude provenance and cross-role fields. SQUAT records which bundle it returned but does not attest that an external model host isolated contexts.
• Login secrets stay on your machine. A scenario records only a non-secret authentication policy. Use an existing UAT session, an approved local login helper when your operator has configured one, or type/password-manager-autofill directly while browser automation and capture are paused.
• Panels can end with a versioned Round Summary. The summary is the durable one-round analysis, including evidence, disagreement, hypotheses, and remaining uncertainty. Separate Research Reports combine selected Round Summaries across rounds and can add more summaries in later versions.
• Inconsistent results are easier to read. Ledger and review views now use Inconsistency for a result that needs a re-run before it can be trusted as a regression.

v0.27.1 — knowledge activation hardening (July 2026)

• Moderation guidance is activated as part of release. A release does not finish unless its governed interview knowledge is available to rounds.
• The v0.27 concept-interview guidance is active. Moderators can consistently retrieve the intended neutral-questioning and follow-up framework.

v0.27 — governed simulation (July 2026)

• Personas can be human-derived, documentary, synthetic, or combined. SQUAT records what material a persona was built from without pretending any source is the truth. Profiles preserve claims, inferred motivations, conditional behavior, tensions, uncertainty, and knowledge limits.
• Source material stays separate and protected. Complete human or documentary content lives in dedicated source records with allowed-use and withdrawal controls. Actors, moderators, cards, and round briefs never receive raw sources, permissions, provenance, or a real-versus-synthetic label.
• Hear the persona before using it. A communication capsule can produce a simulated writing or dialogue audition for review. Every audition is explicitly synthetic, cannot become evidence, and must pass fidelity and anti-caricature review before its exact runtime artifact becomes active.
• Pending personas cannot run. Recruiting or importing creates a draft. SQUAT requires eligible source pins, an approved audition, and an exact provenance-excluding runtime pin before a new governed persona can enter a round. Legacy personas remain readable and are labeled legacy-unverified rather than silently upgraded.
• Moderation has a governed instrument. Concept interviews and design panels now separate purpose, phase guidance, probes, and listen-fors from the spoken question. The actor sees only the current turn; the evaluator uses the exact pinned framework; the moderator never receives the persona's provenance.
• Provenance is recorded at its source. Persona records retain acquisition/source context, while canonical dialogue records identify live-session turns, recovered appends, and preserved archive imports. Operational cards and reports do not repeat the same note.
• One contract across clients. Claude, ChatGPT/Codex, LM Studio, and other MCP clients use the same tools, governed bundles, and v2 MCP Apps card semantics. Rich UI is optional; complete Markdown remains the universal floor. No custom ChatGPT extension is required.

v0.26 — interactive cards, one standard (July 2026)

• Rich where the host supports it, complete everywhere. Workspace status, dossiers, round results, issues, steward reports, and steering menus now use standard MCP Apps resources. A capable host renders a compact interactive card; Claude, ChatGPT/Codex, LM Studio, terminals, and other clients still receive the full Markdown result with the same facts and next actions.
• No client-specific extension. Buttons send an ordinary user message through the shared MCP Apps interface. They never mutate data directly, and SQUAT does not depend on a custom ChatGPT global, Claude-only branch, or LM Studio workaround.
• Cards show only the information appropriate to the view. Operational cards use bounded fields and keep sensitive source material, credentials, and hidden methodology out of presentation payloads.
• Accessible fallback by design. Cards use semantic headings, tables, native buttons, visible keyboard focus, narrow-screen layouts, and light/dark host themes. If the host cannot send a card action, the same next steps appear as numbered text.

v0.25 — safer data lifecycle (July 2026)

• Back up exactly what you intend. Portable exports are a bounded current-state package. Full backups explicitly include history, results, reports, trash, and project artifacts, with a durable manifest SQUAT can verify before a purge. Credentials, licenses, audit records, and administrative grants stay out of every export.
• Irreversible purge can recover from interruption. SQUAT confirms the exact targets and your backup choice, then resumes safely from its last completed step if deletion is interrupted.
• Purpose-bound support access, only when you grant it. A workspace owner can authorize a short-lived support case for an exact resource and purpose, review its use, and revoke it.
• A plainly stated revocation limit. Removing access stops later authorized retrieval, but it cannot erase content someone already exported or copied while authorized.

v0.24 — every version, retrievable (July 2026)

• See all versions of a squad member or scenario. SQUAT has always archived every prior version on save — now you can read them back. Ask “show me all versions of Maya” for the full timeline, ask for a specific version to see exactly what it said, or ask what changed between two versions and get a comparison of the real archived documents. See Recruiting testers.

v0.23 — the scaffold release (July 2026)

• Better tester briefings, automatically. The template that turns a squad member's profile into their in-round briefing is now served content — improvements reach every client without updates, each tester records which version compiled them, and the rules guarantee what the depth work promised: real specifics survive, real gaps stay gaps, and what a tester learned in past rounds reads like memory, not a changelog.

v0.22 — install and authorize (July 2026)

• The connector ships in the plugin. Installing (or updating) the SQUAT plugin is now the whole setup in the Claude and ChatGPT apps: the app prompts you to connect, your browser opens SQUAT's sign-in page, you authorize once, and you're set — no URLs to paste, no config files. The classic launcher setup remains for LM Studio and anywhere else that needs it.

v0.21 — the cards release (July 2026)

• One look, everywhere. Dossiers, workspace status, round results, and issues now follow a single visual contract in every client — the same card in the Claude app, in ChatGPT, or as clean text in a terminal. A status card that reports a problem puts the useful next action first.
• Every key ships with a seat. When your license key is issued, your seat in the workspace comes with it — your identity for roles and system-message choices. See Roles & security.

v0.20 — the connector release (July 2026)

• Connect from the Claude app, no config files. SQUAT is now a proper connector: add it by URL, a sign-in page opens in your browser, paste your license key once, done. The app holds expiring, revocable passes — your key never lives inside any AI client, disconnecting revokes access immediately, and revoking a key kills every pass it issued. This also fixes the “SQUAT keeps disconnecting in the Claude app” problem at its root. LM Studio and Codex keep their existing setup unchanged. See Setup.

v0.19 — the identity release (July 2026)

• The identity triple. Every persona now centers on three distinct things: a name (personas are named by default — an unnamed archetype is a deliberate choice, not a fallback), an archetype (“The World-Builder” — the self-image that explains why they do things), and a practice (“API/Developer” — the competence that explains how). Dossiers lead with all three.
• Alter-egos and thinking styles. Optional depth, recorded only when you actually know it: the person's non-obvious second identity (the accountant who's a weekend musician — it explains delight the day job doesn't predict), and a thinking-style read on how they approach problems, from any framework you like. Neither appears in the actor's script as labels — both are translated into concrete behavior. Maintenance sweeps now also read thinking styles across the roster and tell you which kinds of findings your cast is structurally blind to.

v0.18 — the depth release (July 2026)

• Persona depth check. Rich personas are what make SQUAT rounds worth running — a profile that could only describe one real person. The recruiting methodology now grades every profile's depth (automatically after imports, on request, and during maintenance sweeps), shows you what's strong and what's missing, and offers a short enrichment interview to fill the gaps from your memory of the real person. Doc-generated personas are accurately labeled drafts until enriched — the model never invents detail to fill a profile.

v0.17 — the scenario gate (July 2026)

• The front door is now /start. Some tools show skills without the plugin-name prefix, which made the old front door display as the doubled “Squat: Squat.” Renamed to start — it's where you begin: workspace status, setup, your roster, and directions to everything else. Saying “/squat” still gets you there.
• Missing scenarios now block imports — for real. Importing history that references scenarios you haven't created used to print a hint and proceed, but the server rejected those rounds anyway. Now the import checks scenarios up front, exactly like testers: anything missing blocks before a single round lands, and you're offered the choice to enter the scenarios right then (from your scenario-library file, or by describing each one), map them onto scenarios you already have, or skip the affected results and import the rest. See Importing existing data.

v0.16 — local file reading (July 2026)

• Read local files anywhere. Clients without file access (LM Studio especially) can now ask SQUAT to read a local text file by its full path — a scenario library, a persona document, product notes — and save its contents with the right tools. No more pasting files into the chat.

v0.15 — the housekeeping release (July 2026)

• Trash & restore. Rounds can now be deleted — softly. A deleted round goes to the trash for 30 days, fully restorable, and nothing is destroyed until you explicitly purge. Purging is forever and uses a separate typed confirmation. The protocol requires the operator to enter the phrase in a later turn; MCP itself cannot prove who typed text. See Maintenance.
• Start clean, safely. “Delete all my rounds” is a deliberate sequence: SQUAT shows the exact targets, offers a backup first, records the operator's choice, and requires the operation-bound confirmation returned by the prepare step before it acquires destructive locks.
• Integrity check. A report-only scan for loose ends: issue evidence pointing at deleted or missing rounds, testers referenced by rounds but absent from your squad. It reports in plain language; it never changes anything.
• Scenario hints on import. Importing historical rounds that reference scenarios you haven't created yet no longer leaves silent gaps — the import proceeds and prints a hint naming each missing scenario ID, so you can create it and keep the ledger's links resolving. See Importing existing data.

v0.14 — the messaging release (July 2026)

• System messages. SQUAT can now bring you notices at the start of a session — maintenance announcements for everyone, and account messages like a first-session welcome or a billing notice for your workspace. Each one offers exactly two choices: remind me later (it comes back next session) or dismiss (it never shows again). Your choice is yours alone — a teammate dismissing a message doesn't dismiss it for you.

v0.13 — the migration-hardening release (July 2026)

• Import pre-flight gate. Historical-round imports now check every referenced tester against your squad before committing anything — matching by ID, then by name (an import saying maya_chen matches your member “Maya Chen” automatically). Anything unmatched blocks the whole import, with four ways forward: import the missing testers first, map them explicitly, skip the unmatched material and import the rest, or accept the orphans knowingly. See Importing existing data.
• Large imports work transparently. Very large persona documents and project contexts no longer hit the earlier import ceiling, and exports still carry their full content.

v0.12 — the operator release (July 2026)

• Test logins stay outside SQUAT. Prepare a test-account browser session before the actor starts, or pause automation while the operator signs in. A persona never receives a password. See Running rounds.
• The concierge. A hands-on assistant for jobs that span the product: migrating existing personas and history, first-time setup end to end, and plain-language triage when something errors. It shows you the plan first, you confirm the decisions that are yours, and it carries out the supported steps.
• This manual. The full user manual you're reading now — quick start through glossary, printable, shipped with the product.
• Steward joins your rounds. When a project has an intent registry, starting a round asks one extra question up front: include a steward sweep this round? Say yes and the steward reviews the surfaces your testers touched and adds its report card — a letter grade per check, an overall grade, trend arrows against the last sweep — to the round's report. See Steward & panels.
• Instrumented accessibility checks. Where the browser tool can run scripts, the steward backs its accessibility grades with axe-core, the industry-standard scanner — measured WCAG violations instead of impressions — and can suggest a deeper Lighthouse audit when it smells performance trouble.

v0.11 — the orchestrator release (July 2026)

• Issues that remember. Findings now roll up into tracked issues that persist across rounds — first seen, worst severity, who found it, your call on each one (fixed, planned, investigating, won't-fix…). A claimed fix isn't believed until a verify round re-tests it — ideally with the same tester that found it — and a fix that doesn't hold comes back as REGRESSED, the loudest alarm in the ledger.
• The steward. An expert review of your live product as your proxy, judged against your recorded intent registry — prescriptive findings, recommendations, and questions for you. See Steward & panels.
• Design panels. Put a wireframe, mockup, or copy draft in front of the squad before it's built — each persona reacts in character; reactions are reported per person, never averaged, and never pollute the scored trend.
• Maintenance sweeps. A periodic housekeeping pass: stale testers and scenarios, the issues queue, trash, retention, and backups. See Maintenance.
• Returning testers acknowledge fixes. When a tester who reported an issue returns for a later round, they're briefed in character on what got fixed since their last visit.

v0.10 — the measurement release (July 2026)

• Structured findings, scores, and coverage. Beyond pass/fail: severity-graded findings (S0–S3), five experience scores per result, per-feature coverage states, and a priority formula that ranks what to fix first by severity, how many distinct testers hit it, and whether it's on your product's core loop. See Findings & scores.
• Invalidation, not deletion. A compromised session (deploy landed mid-round, harness error) is flagged and excluded from trends — visibly, with a reason — never erased.
• Harness notes. Operator memory: lessons about running SQUAT itself, read before every round.
• Complete exports. Export bundles now carry everything — full project records, context history, missions, notes, issues. Portability is the promise.

v0.9 and earlier — foundations (July 2026)

• Recruiting interviews and persona imports (verbatim or as a baseline); the dossier.
• Rounds with the actor/evaluator split, missions vs. scenarios, browser-driven rounds, and screenshot evidence.
• Hosted workspace protection and role-scoped access controls.
• Roles (owner / admin / runner / viewer) with role-filtered tools; audit trail.
• Starter content: eight generic scenarios and two scoring frameworks; served methodology with per-tier entitlements.
• Multi-client support: Claude Code, the Claude desktop app, LM Studio, and Codex CLI via the launcher; multi-model routing with advisory model policy.
• Historical imports: bring your existing rounds and personas with original dates intact.