Changelog

lt-suite records notable changes using Keep a Changelog 1.1.0. Versions follow Semantic Versioning 2.0.0.

[Unreleased]

Changed

• Update fonttools to 4.68.0 in uv.lock and the skill requirements.

Fixed

• Accept Dependabot's sign-off in the pull-request DCO check, only on pull requests that Dependabot opened.

[0.2.4] - 2026-11-02

Added

• Report blocked mutator rebinds in script receipts as guard_rebinds_blocked.
• Name the commit mutators that got the start-up placeholder in LtGuard.status().late_define_blocked and in script receipts as guard_late_define_blocked.
• Count replies read only in part (past 32 levels or 50,000 values) as resp_truncated in net_diag.
• Add binding: "unbound-legacy" to a fill report item's receipt when the receipt carries no source_payload.
• Add fill-notes.schema.json, a template that emit_fill.py payload writes, and make_fill_report.py --check-notes, which lists every problem at once and writes nothing.
• Add emit_fill.py repin, which moves a payload's answers.json pin after a change that leaves every written draft identical, checks the question digest, and never writes answers.json.
• Add run_notes to fill notes and reports, and source_check_plan to reports.
• Print MANUAL MODE pages, held for approval counts and each single-attempt item's captured state ("answered, not submitted") in plan_fill.py.
• Optional loopback body server (browser.body_server, off by default): on a page-cache miss the loader fetches the body by SHA-256 from the host and checks it before running it.
• tools/install_body_server.py copies the server into a host-only folder after checking packaging/host-files.sha256, and writes a systemd user unit it never enables.
• ltjs.py render cache_status reports which page scripts the origin's cache holds, without running them.
• Add browser.devtools_socket so lt-chrome reaches a DevTools port through a Unix socket from a sandbox with its own network.
• Report the socket's state and the chosen route in doctor.

Changed

• Hold single-attempt drafts without an approval in the plan; the report lists them skipped, "yours to check and submit", with no note needed.
• Refuse a stale page-kinds.json at payload time with or without a RAT candidate page; the report applies the same rule.
• Default started_at to the earliest bound payload's generated_at and ended_at to the last receipt; a notes time that differs needs times_reason.
• Replace fill-report.json only after the validator passes the new report; a rejected report goes to fill-report.rejected.json.
• Rename the capture summary field table_author_grids to table_given_grids; the capture skill reads either name in stored receipts.
• Return 16-hex source_payload_sha16 and upload sha16 values in browser receipts in place of full sha256 values.
• Refuse record --supersede on a stored row without verify_receipt_sha256, and archive the fields that changed.
• Every need:'payload' reply names its reason.
• The page cache keeps entries in least-recently-used order within 524,288 characters; a hit refreshes its entry.
• Cached bodies no longer carry the lt-suite version, so a release keeps unchanged scripts cached.
• With a body server, loaders name a plan policy longer than 256 characters by SHA-256 instead of carrying it.

Fixed

• Record each page's real Lt route in captures. A Background page gets /run/background and flow page N after it gets /run/page/N-1. Capture takes each number from the page's dataKey, checks it against the tab's URL and painted page, and leaves url null when they disagree. New captures carry route_source.
• Refuse page routes from every capture without route_source, whatever its titles or boundaries. Nothing else in it reliably marks a Background page, after which its URLs were one page late and could pass a live Checkpoint. check-url and the guard report recapture-needed, and media, fill and check plans drop those URLs. Export and answer drafting still read such a capture.
• Refuse a plan without the route_source key, which was built before measured routes and holds no titles or source URL to check. Media, fill and check plans always write the key, null allowed.
• Map /run/background only through a captured url, never by default to flow page 1.
• Reject plans in which two pages claim one route or a URL number exceeds the pages before it.
• Accept Background pages, unconfirmed routes and route_source in the module schema and validator, and check that URLs match page_no.
• Let Lt's start-up call removeCommittedCheckpoints and removePagesAfterUncommittedCheckpoints before it paints a page. The guard locks both once a page paints or render_wait_ms runs out, and runs no script body before then.
• Give commitClicked or setCheckpointCommitState, when Lt has not defined it yet, a non-configurable placeholder accessor on $root from the moment the guard installs. A raw call throws at once. Lt's plain assignment lands without throwing and never becomes callable. Lt's Object.defineProperty on the name throws inside Lt, so a late definition can no longer replace the accessor and run before the next check. The two page-list mutators Lt has not defined lock on $root when start-up ends.
• Check the route again after the guard's wait, so the page Lt painted meanwhile decides.
• Wait up to render_wait_ms for a painted page before the fill location rail and fetch_media read current(). Refuse with lt-not-painted when none paints; the fill receipt carries guard_last.
• Identify Lt's run-data PUT by typed panelKey, dataKey or uuid fields, so 3- and 4-digit panel keys can be acknowledged.
• Never acknowledge a run-data save on a bare 2xx: require a per-key success or a top-level "success": true reply.
• Fail keys whose single-key reply is HTTP 200 with {"success": false}.
• Leave every run-data key pending when a success reply carries a failure marker at any depth, such as {"success": true, "data": {"success": false}} or a nested status: "failed".
• Fail a reply item (an object with an identity field) when a failure marker sits anywhere inside it, whatever its own success says, so {"results": [{"panelKey": 209, "success": true, "result": {"success": false}}]} fails panel 209. Sibling items keep their own outcomes.
• Confirm no key of a request whose reply carries a failure marker outside every item, lists a failed value that names no key of the request, or runs past 32 levels or 50,000 values. Inspection used to stop at 8 levels and accept the rest.
• Leave a generic single-key save unconfirmed when its reply carries any failure marker, such as {"success": true, "data": {"success": false}}.
• Sign the Lt client build by the URLs, query strings included, of the core bundles loaded before DOMContentLoaded, so page chunks loaded later no longer drop calibration and /core.js?v=2 no longer keeps a calibration made on /core.js?v=1.
• Drop save records bound to the old c1- script signature once, with the reason in the calibration history.
• Keep the save record when a fill page shows no build marker.
• Record keys_weak, net_matched and save_key_source in fill receipts, and keys_weak and net_matched in saved waits.
• Compute save_diag model_sync per request row, and add run-data identity counts to save_diag.
• Name the hash each browser receipt bound by, and the local payload file check, in report items.
• Name every --payload path that record cannot use, and a payload path outside the run in report binding refusals.
• Print each stored receipt row that record keeps because only its binding differs.
• Treat any [BLOCKED: value in a receipt's full payload hash as redacted.
• Reject a fill report in a run folder whose fill-js/check-plan.json it predates.
• Rebuild every job of a payload from the current answers.json before make_fill_report.py trusts an answers-repin.json record, and refuse on any difference. repin records the --anchors and --approvals it used.
• Compare rebuilt and written jobs in repin and in the report by sha256 of their content (page, panel, text, each table cell's row and column, selections and placements). The 32-bit receipt hash used before let a text with a colliding hash, or a table cell moved to another row, re-pin and verify.
• Report an ordinary item filled-unverified, never filled, when its latest receipt carries no source_payload: such a receipt is tied to no executed payload and not to today's draft. Only a receipt bound to an executed payload whose jobs match today's draft verifies.
• Start the default run window at the earliest bound payload's generated_at. When no receipt binds a payload it starts at the earliest payload in the run folder, and at the first receipt only when the run folder holds none. Refuse a receipt observed before its own payload was built.
• Name a body server miss only from a fixed set of reasons, so reply text never reaches a receipt.
• Accept /run/background in the fill location rail and the panel probe for the plan's Background page, placed by current().
• List fill-notes.schema.json and references/fill-notes.md in the release allowlist, so release archives build.
• Read a Signed-off-by written below a --- line, as Dependabot writes it, in the public history check.
• Warn in doctor when the stable lt-chrome launcher is missing and the config serves lt-chrome tools.

[0.2.3] - 2026-10-26

Changed

• Inline fill calls reuse one cached runner with current frozen arguments and guard policies.
• Download handoffs use paced groups of 8 and report missing media keys by page.
• Media calls use a configurable 20 second budget and return pending keys for resumed downloads.

Fixed

• Show local citation paths by file name in exported notes and PDFs.
• Name each failed receipt binding check and reserve the RAT prefix for RAT receipts.
• Preserve hyphens at wrapped line ends when checking PDF content.
• Accept bounded save request diagnostics in fill reports and export validation.
• Use a neutral receipt binding message for ordinary fill receipts.
• Keep payload discovery inside each run and resolve manifest paths from their containing folders.
• Bind fill records to the executed payload, including payloads generated in sibling folders.
• Refuse conflicting receipt hashes and archive corrected rows with record --supersede.
• Refuse draft status text in answer segments and payload HTML through configurable patterns.
• Check the answer file capture path and hash binding before generating fill payloads.
• Wait for the Lt view model and refuse calls while commit guards remain unlocked.
• Include compact request diagnostics in fallback save receipts and saved wait records.
• Measure table occupancy over target cells while reporting other filled cells separately.
• Name stale page classifications and the command that refreshes them before checking receipt bindings.
• Add short payload hashes for browser tools that redact full hashes and bind unique matches.

[0.2.2] - 2026-10-26

Changed

• Inline browser tools receive gzip compressed page scripts with SHA-256 verification.
• Page cache loaders reuse verified code with current frozen arguments and guard policies.
• Chunked payload snippets verify numbered transfers before the stub assembles its body.

[0.2.1] - 2026-10-26

Added

• Opt-in export.citations.transcript_page_from_filename links recording citations to notes pages named after recording files.

Fixed

• Table cells no longer repeat the answer's citations when cell provenance references are absent.

0.2.0 - 2026-10-25

Added

• Separate fonts for page PDFs, diagrams and micrograph labels, with Noto and system fallback chains.
• Palette-matched micrograph copies bound to their upload and captured media for PDF and notes exports.
• Configurable Chromium selection, profile folders and loopback DevTools attachment through lt-chrome.
• Setup questions and doctor checks for output fonts, browser attachment, RAT sheets and Check Answer.
• Quarantined RAT answer-sheet folders, matching and shared page classification for permitted sheet entry.
• Opt-in Check Answer retries for eligible multi-attempt items, with attempt records and reconciled exports.

Changed

• Exported SVGs and interactive diagram pages carry font fallback stacks for other machines.
• RAT filling and Check Answer share current page classifications; Check Answer refuses RAT pages.

Fixed

• Windows text encoding, line endings, interpreter pinning and path handling.
• Check Answer refuses changed live classification inputs and requires recapture and reclassification.
• lt-chrome rejects unknown browser keys and nested paths before selecting startup defaults.
• Font discovery handles system filenames and checks every rendered face before font modification.
• Micrograph exports reject malformed, stale or edited palette-copy manifests.
• Fill receipts retain the executed sheet identity and reject changed classification bindings.
• Unread Check Answer verdicts require a fresh document before further checks.

Security

• Font embedding and modification respect licence metadata and embedding restrictions.
• lt-chrome validates loopback addresses and isolates npm startup from workspace configuration.
• RAT sheets and Check Answer feedback stay outside cited research and curriculum reads.

0.1.0 - 2026-10-24

Added

• Six skills: lt-suite-setup, lt-module-capture, lt-module-answer, lt-module-fill, lt-module-export, and mechanistic-diagram.
• Accessible defaults with Noto Sans and Noto Sans Mono, measured palette selection, and configurable alternatives.
• Event-driven save waits that observe acknowledgements and retain a bounded fallback.
• Automatic per-origin calibration after a verified first page, followed by a verified two-page trial.
• Batches of up to five pages after calibration; missing saves restore single-page operation.
• Multi-agent packaging for supported plugin clients and standalone skill archives.
• Synthetic examples, privacy checks, licence notices, and a Developer Certificate of Origin policy for pull requests after release.

Changed

• Browser navigation waits for save evidence and verifies writes after a fresh reload.
• Institution settings, output paths, palettes, and fonts use configuration.

Removed

• The permission gate and its configuration, scripts, and call sites.