CHANGELOG.md

Changelog

All notable user-visible changes to @tikoci/rosetta are recorded here. Format follows Keep a Changelog; project uses Semantic Versioning.

Agentic rule. Any change with a user-visible effect (CLI, MCP tool shape, DB schema, CI behaviour, install flow) adds an entry under Unreleased in the same PR / commit. Promoting [Unreleased] to a dated version header is a manual step, done by hand alongside the package.json version bump before dispatching a latest-channel release (see MANUAL.md "Release Workflow") — CI no longer auto-bumps versions or auto-promotes CHANGELOG on any channel. Prerelease (alpha/beta/rc) release runs never promote [Unreleased] either, since they don't represent the next stable version. CI-only auto-bumps and pure refactors with no external effect are intentionally omitted — git history is authoritative for those.

Not a git log. Don't list every commit. One bullet per behaviour change, grouped under Added / Changed / Fixed / Removed / Deprecated / Security. Details and rationale belong in DESIGN.md; the "what's next" backlog belongs in BACKLOG.md.

Unreleased

Added

• CLI-Reference overlay (cliref_* tables + rosetta export datasets, issue #124). A source-faithful, version-less overlay of manual.mikrotik.com/docs/cli-reference/*: cliref_pages / cliref_entries / cliref_fields / cliref_flags retain each page's byte-exact Markdown (+ SHA-256), verbatim entry/field/flag descriptions, occurrence order, and line spans. Entries carry the manual-only facts a CHR /console/inspect cannot self-report (package / conditions / syscap). A stored cliref_entry_schema_links crosswalk resolves each entry to its schema_nodes command node (exact / single-internal-segment alias / manual-only), and the cliref_field_inspect_links view derives the zero-to-many field→argument mapping. rosetta export gains seven cli-reference/* outputs (six TSVs + byte-exact source/<slug>.md). Build via make extract-cliref + make link-cliref.
• schema_nodes.inspect_type. Preserves RouterOS' raw /console/inspect class (path | dir | cmd | arg) that the existing type column normalizes (raw path → dir). Additive; existing type consumers are unchanged.
• Released DBs now carry the CLI-Reference overlay, with CI gates (issue #127). release.yml's build runs extract-cliref + link-cliref, so published DBs and the cli-reference/* export datasets are populated (previously the overlay only built locally). Two new blocking gates run in qa.yml db-content against the built DB: V-cliref-link-drift byte-compares the entry→inspect crosswalk against a committed, reviewable cli-reference-links.tsv (24 alias + 120 manual-only rows) and rejects any alias naming a segment outside the known allowlist; V-cliref-db-integrity asserts structural + semantic invariants (integrity/FK/orphans/enums/occurrence-order uniqueness, and the field view exposing only settable Argument→arg mappings). The db-content floor now also requires populated cliref_* tables + crosswalk. Regenerate the baseline after a corpus refresh with make link-cliref-baseline.

Fixed

• Property tables wider than two columns stored the wrong cell as the description (issue #132). parseProperties hardcoded cells[1], so the five | Property | Type | Default | Description | tables in the corpus (Apps, VETH) had their Type cell stored as the description with type/default_val left NULL — 52 rows, e.g. /interface/veth/add address returning "IPv4/IPv6 address" instead of its prose, at high confidence. Columns are now resolved by header name (Property/Parameter, optional Type, optional Default, Description), and the Markdown pipe escape | is decoded once the cell is split (page_tables.raw_markdown stays byte-exact). Requiring an explicit Description column also drops 14 rows that were never properties — a device-mode feature matrix (| **Feature / Property** | **Home** | ... |) whose "descriptions" were the Home column's Yes/No. Apps' two auto-update records are retained: they only looked duplicated because the shift replaced both descriptions with the same Type cell.
• CLI-Reference discovery silently dropped every branching menu's own Directory page (issue #137). MikroTik decomposed the CLI Reference from module pages into per-command leaf pages, and serves a branching menu (app, caps-man/interface, interface/ethernet/switch/port, …) as a trailing-slash category URL in sitemap.xml — which has no Markdown of its own. The menu's Directory entry is published at <dir>/<basename>.md and listed only in llms.txt, so sitemap-only discovery lost 256 of 1,070 pages — and the field-heavy quarter of them: cliref_fields came out at 5,944 instead of 10,938. Those menus, their gates, and their argument tables were absent from the DB, from cli-reference/* export datasets, and from anything reading the overlay. Discovery is now the union of both inventories, and the new blocking V-cliref-discovery gate fails the build if any sitemap category dir stops contributing its Directory leaf. Rebuilt corpus: 1,070 pages / 1,077 entries / 10,938 fields / 976 flags / 967 crosswalk links (was 228 / 1,051 / 10,118 / 948 / 931).
• CLI-Reference crosswalk linked three commands to the wrong inspect node (issue #136). KNOWN_ALIAS_SEGMENTS treated "this word is a doc-internal module name" as a global property, but it is positional: poe and qos are real menu segments published elsewhere in the same source. interface/ethernet/poe/monitor was linked to /interface/ethernet/monitor (PoE-out fields vs SFP/link fields — disjoint sets), and interface/ethernet/switch/qos/port{,/reset-counters} to their non-QoS namesakes. The failure was silent because the alias branch only runs when the exact path is missing from schema_nodes — precisely the hardware-gated commands nobody can check from a CHR. A segment is now never dropped when the prefix ending at it is itself a published entry; all three become correctly manual-only. With the source reshape above, the crosswalk is 967 exact / 0 alias / 110 manual-only.
• CLI-Reference extractor truncated <ArgTableRow> attributes at a > inside a quoted value. A typ like iface_enum { <l2tp>:0xfffffffe } (e.g. on interface/pppoe-server) contains a literal >, which the row parser mistook for the tag close — the field's raw_type was silently stored empty with a garbled description. The tag matcher now skips over whole quoted attribute values, so those fields carry their full type.

Changed

• Schema version 10 → 11 for the CLI-Reference overlay tables/view and schema_nodes.inspect_type.
• routeros_lookup_property confidence now grades the row, not the query branch (B-0024 step 4). Previously high meant only "this row is on the page commands.page_id links to" and low meant "the scoped branch found nothing" — the label described which SQL ran, so every property on a linked page was equally high and a correct row found by the global fallback was always low. Each row is now scored against the requested menu using the menu paths named in its own documentation section, and results come back best tier first: high = the section is about that menu (it names the menu, and of the menus it names that one owns most of its properties) and the command tree does not contradict it; medium = the section names a neighbouring menu, names the requested one only in passing while documenting another, or only the page matches; low = nothing but the property name links the row to the menu. Without command_path every row stays medium — there is no menu to align to.
• routeros_lookup_property no longer lets a mislinked page hide the right answer. Candidates are drawn from every page documenting the property rather than only the page commands.page_id links to, with page alignment applied per row. Off-page rows are kept only when their tier is at least as good as the best the linked page offers, and linked-page rows win ties within a tier, so a correct link returns the same tight, same-ordered result as before. Previously /interface/bridge/host + vid returned two rows from the IGMP-snooping page it is mislinked to and never surfaced the "Static Entries" section that names /interface/bridge/host outright; the same suppression hit /interface/wifi/provisioning + radio-mac. Measured over the 14,832 (menu, property) pairs the command tree says are real: 31.5% of the labels that shipped as high survive, 2,295 demote to medium, 1,940 rows escape a wrongly-low label (835 to high), and 153 rows that were suppressed entirely are now returned (80 high). Regenerate with bun run src/eval/command-prose-join.ts. This is what makes /interface/bridge/port + pvid rank the two bridge-port sections above the unrelated Apps row (#131) and /ip/dhcp-server + address-pool rank the DHCP server table above the DHCPv6 one (#58).

0.11.1 — 2026-07-20

Fixed

• hAP ax2 and hAP ax3 linked to the wrong /hardware manual pages. MikroTik shifted the slugs by one upstream: the hAP ax² page moved /hardware/hap-ax → /hardware/hap-ax-2, and the hAP ax³ page moved /hardware/hap-ax-2 → /hardware/hap-ax-3. The committed device map still held the pre-shift slugs, which were correct when generated, so afterwards hAP ax3 resolved to /hardware/hap-ax-2 — a URL that now serves hAP ax² — and hAP ax2 pointed at a slug that no longer exists. Refreshing the hardware assessment corrects both rows.

Changed

• YouTube transcript corpus refreshed — 658 → 722 videos, ~2,316 non-empty transcript segments. The per-video duration cap rose from 1500s to 1600s so longer first-party channel videos qualify; recent MikroTik feature videos run past the old 25-minute cut, including the "Hermes explained" video (1528s) that discusses running an AI agent against RouterOS. MUM_TITLE_PATTERNS remains the direct filter for conference talks — the duration cap is only a coarse backstop for the untitled long tail. A residual handful of videos still fail per sweep on transient YouTube throttling rather than missing captions, and converge over subsequent reruns; they are deliberately not recorded in transcripts/known-bad.json, which is reserved for videos with genuinely no English captions.
• Product matrix snapshot refreshed to matrix/2026-07-20/ — no products added or removed; Cube 60Pro ac and CubeSA 60Pro ac correct PoE out from 18-48 V to 18-57 V.
• Hardware overlay artifacts re-scraped from live sources (ros-hardware-assessment.json, ros-www-assessment.json, device-map.tsv), clearing the drift the scheduled Device Map Refresh reported. The ros-www-assessment.json diff is large but content-neutral: the same 288 products with identical values, reordered because the committed copy predated the candidate-sorting change — regenerating with current code canonicalizes it once rather than repeatedly.

0.11.0 — 2026-07-20

Added

• The resolved DB now reports whether it can be trusted to ground claims about the code querying it (issue #94, B-0022 export-audit umbrella). routeros_stats gains a provenance block — resolved db_path, invocation mode, the db_meta stamps (release_tag/source_commit/built_at/schema_version), and a grounding verdict (ok / schema_mismatch / internal_inconsistent / tag_behind / unstamped) — so one call self-checks the DB against the checkout instead of shelling into sqlite (the TUI stats screen shows it too, via shared getDbStats). A new make db-doctor / bun run db:doctor prints the same verdict and exits non-zero when not ok (CI/pre-commit-ready); make db-sync / bun run db:sync fetches the latest CI release DB — the documented grounding source of truth — into the resolved path (atomic, safe while a server holds the file open). In a dev checkout, MCP startup now emits a loud but non-fatal banner when the served DB has drifted (e.g. a corpus whose PRAGMA user_version was bumped in place so its stamped provenance no longer describes its bytes); it never fetches, so a contributor's local build is never clobbered. The load-bearing new signal is internal_inconsistent (db_meta.schema_version ≠ pragma), which every prior check missed. Durable guidance lives in .github/instructions/local-db-grounding.instructions.md.

• New rosetta export <dir> command writes a DB-only dataset directory for local audit (issue #101, the B-0022 export track). Produces manifest.toml plus seven flat TSVs — changelog.tsv, callouts.tsv, properties.tsv (joined to pages with a resolved section_anchor), videos.tsv (metadata + per-video transcript segment/word/byte counts), commands.tsv, and the paired pages.tsv (one row per page) / sections.tsv (one row per section, pivot-able) — from the runtime SQLite database alone — dataset generation reads only the resolved DB, with no caches or re-extraction (startup may still download/refresh the DB, as every command does). The serialization contract is a reversible Postgres-COPY-style escape (\t/\n/\r/\\ backstopped on every value, whole-field \N for SQL NULL as distinct from the empty string) with stable ordering so a rebuild on the same DB is byte-identical, and manifest.toml carries db_meta provenance, per-file row counts, and honest disclosures of what the DB cannot provide (transcript provenance #21, per-version architecture #91). A column the DB cannot produce is omitted and disclosed, never recovered from a source artifact.

• rosetta export now emits tables.tsv, an inventory of every captured doc table (issue #111, under the #104 B-0022 export track). One row per page_tables record (855 on the rc.99 corpus) with page/section provenance, the table's shape (column_count/data_row_count/is_ragged), source size (raw_bytes, UTF-8 length of the stored Markdown), and a table_url that deep-links to the section containing the table (url#anchor; the bare page URL when no section resolves, since Docusaurus has no per-table anchor). This is the "what tables exist and where" list — it answers most table-audit questions without exporting the cell data itself (the per-fragment cell files remain E4). Reuses the settled TSV/manifest contract and is pruned/owned the same way as every other dataset.

• tables.tsv gains an is_property_source flag and reorders its columns human-readable-first (issue #118, rounding out the 0.11.0 export track for B-0077). is_property_source is 1 if and only if a row in the properties table (the data routeros_lookup_property surfaces) has a source_table_row_id resolving back to that table — i.e. extraction actually produced properties from it, not merely that it looks property-shaped, so a property-headed table the extractor gates skipped (issue #100) is honestly 0. This lets a human reviewer see at a glance which captured tables are already recognized as property sources versus candidates B-0077 still needs to pick up. Columns are reordered so identity (title, source_heading) and the shape/size stats sit up front and the two long URL columns (table_url, url) sit last; since tables.tsv first ships in this same 0.11.0 release, no already-released consumer is affected.

• Every Docusaurus pipe table is now retained as normalized, queryable data in schema v10 (issue #92). page_tables keeps raw Markdown plus page order, nearest h1–h6 heading, enclosing h1–h3 section_id, width, and raggedness; page_table_rows/page_table_cells preserve every header/data cell at its actual width with escaped pipes decoded. Table-derived properties now point to their exact source row through source_table_row_id, while bullet-derived and historical properties remain honestly NULL. Property output is byte-for-byte unchanged; the shared fence-aware parser also corrects the B-0022 census counting error caused by an ordered-list code fence in zerotier.md (the rc.98 corpus contains 855 tables / 8,287 data rows, not 852 / 8,258).

• README.md and DESIGN.md now describe the live Docusaurus corpus as current, not a pending migration (issue #77). README's intro and "What's Inside" table dropped the "still needs a major migration away from the retired Confluence export" framing and refreshed prose-corpus figures for the actual Docusaurus build (363 pages/~653K words, 4,402 properties, 943 callouts), and gained a "Hardware overlay" row for hardware_catalog/device_aliases (255 devices incl. legacy/EOL + accessories, ~750 alias mappings) that was previously absent from the feature list entirely. DESIGN.md's "Corpus Snapshot" section is retitled "Historical Corpus Snapshot (legacy Confluence)" to stop reading as current-state to a skimmer; its already-caveated legacy figures are otherwise untouched.

• YouTube transcript cache refreshed — 538 → 658 videos, ~2,090 non-empty transcript segments. make extract-videos picked up a large backlog of previously-unswept videos (not just the handful published since the last 2026-07-07 sweep), converging over several incremental reruns as transient yt-dlp/YouTube request failures cleared on retry. Three genuinely non-English videos (Finnish, Russian, Japanese) that fail every run because YouTube has no English auto-captions for them are now recorded in transcripts/known-bad.json so future sweeps don't keep re-attempting them.

• Retrieval-quality net hardened for the 0.11 corpus swap — golden set 24 → 35, plus a direct-surface eval matrix (issue #53). The Phase 0 golden set (fixtures/eval/queries.json) grows by 11 hand-verified queries: +2 natural-language (pppoe, dns), +4 property, +3 changelog (fixed-CVE / feature-introduced / version-scoped), +2 video — closing the zero-coverage gap on property/changelog/video retrieval that the Confluence→Docusaurus swap left untested. src/eval/retrieval.ts gains a per-query surface field: search (default) still drives the durable routeros_search recall/MRR/classifier gate (the entry point most exposed to the swap), while property/changelog/video dispatch to the dedicated tools (lookupProperty/searchChangelogs/searchVideos) and feed a separate, informational surface_matrix (per-surface hit@5) kept out of the gated aggregates so those stay sensitive — a starting coverage board to grow over time. Golden _thresholds recomputed for the larger set (0.88/0.82/0.75/0.90; the old 0.85/0.70/0.65 had drifted loose over 27 search queries — the 2pp regression check vs baseline.json remains the tight forward guard). Both baseline.json and the stale April-2026 Confluence-era self-supervised-baseline.json regenerated against the live Docusaurus corpus. Grounded audit (0.10.0 vs current compare-and-contrast via the new one-off src/eval/corpus-compare.ts, blocking self-eval rehearsal, and an ETL Bug Ledger) recorded in briefings/B-0020-0.11-retrieval-quality-audit.md: no net retrieval regression from the corpus swap (14/16 topic-hit on both corpora; existing golden set 91.7% Recall@5 on Docusaurus), and #26's two "regressions" confirmed to miss on 0.10.0 too (pre-existing query-core weakness, not swap-induced). Queries that legitimately fail today stay pointed at the correct answer and fail loudly — no fixture was loosened to green the board.

• Hardware overlay Phase 2A — hardware_catalog/device_aliases now drive device lookup and enrichment (MCP + TUI). routeros_device_lookup and routeros_search resolve devices from free-form input far more accurately: an alias stage (exact-normalized, authoritative — never fuzzy) maps product codes, old names, and www/hardware slugs the matrix name misses to the canonical device (cap_ac→cAP ac, RB750Gr3→hEX), reported as mode: "alias" with the matched_alias. A single matrix-linked result now carries a compact hardware overlay block — rosetta_device_id (stable, persistable key), category, discontinued, also_known_as (name-like aliases only; slug/table artifacts excluded), genuine non_default_ips, and reconstructed product_page_url (mikrotik.com/product) + hardware_page_url — whose note steers agents to the device tool rather than fetching a page. Non-matrix entities (accessories, series, legacy/EOL) surface in a new catalog array as labeled thin rows with a kind field (accessory/series/discontinued/device), so an agent never mistakes a GPeR for a router. The classifier gains catalog-prominent families (Chateau, Audience, OmniTik, PowerBox, KNOT, Cube, DISC) plus a whole-input alias probe in searchAll for bare codes the regexes miss; routeros_search's related.devices now carries category/discontinued; the browse TUI device card shows the overlay and a thin-row variant for catalog entities, and its stats screen lists the two overlay tables. No /hardware prose enters the FTS corpus (fields only). Reads the clean data from the #47/#48 cleanup. See briefings/B-0019-hardware-overlay-phase2-mcp-tui-surfacing.md and issue #49.

• New Docusaurus /docs prose extractor (extract-docusaurus.ts) replaces extract-html.ts as the default prose source. Discovers pages via sitemap.xml, fetches raw Markdown from manual.mikrotik.com, and populates pages/sections/properties/callouts — 360 in-scope /docs pages as of 2026-07-07 (CLI Reference and /hardware remain out of scope, tracked as follow-up work). make extract/make extract-full now run it by default; the legacy Confluence pipeline survives as make extract-legacy-confluence for rebuilding historical release DBs.

• New pages.rosetta_id column (schema v6) gives Docusaurus-sourced pages a stable, URL-derived identifier alongside the existing integer id — see DESIGN.md and briefings/B-0012-docusaurus-manual-migration.md "H7 — Identity / rosetta-id design".

• npm prerelease dist-tag channel. Testers can now opt into an in-progress build via bunx @tikoci/rosetta@next (newest prerelease of any stage) or @alpha/@beta/@rc (pinned to one stage), without moving the default latest channel. Channel is driven entirely by package.json's committed version (a -alpha/-beta/-rc suffix means prerelease); OCI image tags (:alpha/:beta/:rc/:next) mirror the same scheme, and the bare :latest OCI tag now never moves on a prerelease release run. See README.md "Prerelease channels" and MANUAL.md "Release Workflow".

• bun test --coverage now runs on every release build, summarized in the workflow's step summary and uploaded as a coverage-lcov artifact — informational only, not a gate.

• New hardware_catalog + device_aliases tables + device_overview view (schema v8) overlay manual.mikrotik.com/hardware and mikrotik.com/product/<code> onto the existing devices data. hardware_catalog is a superset of devices — it also covers accessories and legacy/EOL SKUs matrix.csv doesn't track, carries a never-null display name, an optional device_id link back (renamed from devices_id to match device_test_results.device_id) for the ~156 rows matrix.csv already knows about, and lands each device's non-default management IP and FCC/IC regulatory IDs into specs_json; devices itself is unchanged. device_aliases resolves every observed slug/code/name variant (matrix.csv, /hardware slug/link/table code, www requested/declared/compare code) back to one rosetta-curated rosetta_device_id, with priority-ranked, collision-counted assignment so no code is silently claimed by the wrong device. device_overview is the documented read surface (catalog + devices spec columns + alias counts). A www product is only accepted as a device's spec source when its own identity agrees with the device's code/slug family (killing cross-sell misattribution such as a mounting bracket's specs landing on a cube page), attaches to at most one row outside a tiny justified shared-kit allowlist, and a declared-code matching tier resolves ROSE Data server and both KNOT Embedded LTE4 rows. Built by the new src/extract-hardware-catalog.ts, now wired into make extract/make extract-full and the release.yml extraction pipeline (ordered after extract-devices, which wipes both tables) so the catalog + aliases ship in the published DB; a release is blocked if the DB carries fewer than 200 hardware_catalog or 600 device_aliases rows. It emits a committed, deterministic fixtures/hardware-catalog/catalog.json (sorted rows + aliases + drop ledger — the change-review diff gate), enforces hard output invariants (declared code present in own aliases; one www product per row; every input entity attached or dropped-with-reason), and fails loudly — DB untouched — on category-taxonomy drift, a www page-template change, matrix coverage regressing, a previously-resolved device disappearing, an alias-collision/drop surge, or a www 404-rate swing. See briefings/B-0017-hardware-overlay-device-resolution.md and issue #35. MCP/TUI surfacing of this data is deferred to a follow-up (B-0017 "Phased implementation plan," phase 2).

• New scheduled Device Map Refresh workflow (device-map-refresh.yml) watches for upstream hardware-page drift. Weekly cron (+ manual dispatch) does a LIVE re-scrape of manual.mikrotik.com /hardware and mikrotik.com /product via make assess-hardware assess-www device-map — the network path test.yml's per-PR drift gate deliberately never runs. It detects drift only: when the regenerated artifacts differ from the committed snapshot (or a device no longer resolves without a device-exceptions.toml entry), it fails the run (red check + scheduled-failure notification), uploads the regenerated artifacts, and opens/refreshes a single tracking issue with a diff-stat summary; the full diff is in the run logs. A human then resolves it through the normal flow — regenerate locally, review per briefings/B-0018 "How to audit", open a standard PR that runs the required device-map-check gate. Closes the "how do we notice MikroTik added/moved/removed a page" half of B-0018's audit loop that the committed-snapshot coherence check can't cover on its own.

• Reviewable device→URL map + drift gate (make device-map). device-map.tsv (one row per matrix device → /hardware + www URLs, with a needs_review column) and hardware-unmatched.tsv (the reverse audit view: /hardware pages with no matrix device), backed by a curated device-exceptions.toml and a blocking drift gate (make device-map-check) that fails when a device stops auto-resolving without an exception, a curated exception goes stale, or a committed artifact drifts from freshly computed output — now run on every PR via test.yml (no network; validates only the committed snapshot). Human/MikroTik-readable legend in briefings/B-0018-product-naming-three-source-map.md.

• New QA workflow (qa.yml) rehearses the release-locked quality gates without publishing (issue #40, B-0014 Option B). A workflow_dispatch + workflow_call workflow that runs the checks previously provable only inside release.yml — DB content floors (incl. the hardware-overlay hardware_catalog/device_aliases floors), MCP contract, golden + self-supervised retrieval eval, docs-count, device-map drift, and db_meta provenance — against a freshly built (db_source: local-build) or published DB, on any ref, with no npm publish, OCI push, or GitHub Release. Focused dispatch via test_scope (single gate or all), db_source, full_versions, and eval_self_blocking (flip the self-supervised eval to a hard gate on demand). This is the intended way to verify release-pipeline changes — such as the hardware-overlay extraction wired in this same release — before an actual dispatch. release.yml now calls this workflow directly (issue #42): its single build-and-release job was split into build → qa (uses: ./.github/workflows/qa.yml with db_source=artifact) → publish, so the gates that fence a release are byte-for-byte the ones a qa.yml dispatch runs against the exact DB about to ship — one definition, no drift (the old cross-file floor guard is gone). See MANUAL.md "Release Workflow" / "Rehearsing release quality gates without publishing" and VALIDATION.md V-qa-rehearsal.

Changed

• rosetta export now replaces a prior export safely — manifest-owned prune-then-publish + a --force gate (issue #108, part of #95). manifest.toml becomes the ownership record as well as the index: re-running into a directory that holds a manifest rosetta wrote rewrites the current file set and prunes any file the old manifest listed that the new run no longer produces (cleaning directories that empties), so a renamed page slug or dropped product from an earlier run no longer lingers — the stale-file hazard the variable products/** / pages/<slug>/** sets (E3/E4) introduce. It only ever deletes files a prior rosetta manifest named, so files you added yourself are untouched; a new/empty directory just gets written; a non-empty directory with no rosetta manifest is refused unless --force (or, on a TTY, an interactive confirm), so the command can never clobber a directory it did not create. [[files]].name is a directory-relative path, so the manifest already describes directory-nested files. Within a run the order is prune → write datasets → write the manifest last, so a name reused across runs transitions cleanly between file and directory and a crash never leaves a manifest naming half-written data (the directory is a regenerable audit surface, not a transactional store — a crash can still leave a mix of generations until the next run reconciles). Writes and prunes are contained to the export root both lexically and through symlinks, so a tampered manifest or a planted symlink can never act outside it. MANUAL.md documents the scheme; the DB-only boundary is unchanged (the export still reads only the DB handle, plus a single read-back of its own prior manifest).
• rosetta export output shape refined so the two page-level files read cleanly and the section rollup reconciles (issue #106, post-#103 per-TSV review; part of #95). The per-section file is renamed pages.tsv → sections.tsv and the per-page rollup pages_summary.tsv → pages.tsv, so each filename now matches its source table like every other dataset (changelog←changelogs, callouts←callouts, …); grouping sections.tsv by page_id in a spreadsheet pivots up to (near-)the pages.tsv word count, the small residual being heading-text lines that live in no fragment (disclosed in the manifest). callouts.tsv gains rosetta_id (page) and resolved section_anchor columns so a reader identifies the owning page/section by name without joining back to the DB — the same pairing properties.tsv already emits — with the long content field moved last. changelog.tsv moves sort_order before description for the same long-column-last reason. No schema change, no re-extraction, no MCP/TUI change; the export still reads the DB alone and rebuilds byte-identically. MANUAL.md updated.
• Section coverage of a page is now (near-)total — a synthetic lead ("H0") fragment captures pre-first-heading prose (briefings/B-0023-page-section-normalization.md, part of #95). parseSections split the page body only on h1–h3, so the lead-in prose before the first heading — an AI-summary blockquote, an intro paragraph, and the entire body of pages whose only heading is the title — belonged to no sections row: 11.7% of the corpus by word, and 40 whole pages, had no section to resolve to and didn't roll up to the page total. Each page with non-empty lead prose now gets a lead fragment (level = 0, reserved anchor_id = "_lead", heading = the page title) holding that content, so section words cover ~98% of page words (the residual is heading-text lines, which belong to no fragment) and every table/callout/property before the first heading resolves to a real section_id instead of NULL (measured on the local corpus: tables/properties/callouts with no section context all drop to 0). A page whose only pre-heading line is the title mints no lead fragment (no empty fragments). No schema DDL change (the sections table already stores level/anchor_id), no MCP tool-surface change; the rosetta export pages.tsv/sections.tsv rollup and the section_id disclosure reflect the fuller coverage.
• hardware-unmatched.tsv reframed as a classified inventory — new kind column. Every /hardware page with no matrix device is now labeled device / accessory / module / series-or-doc (67/12/8/15 of the current 102) by a shared, rule-based classifier (src/hardware-kind.ts), so the reverse-audit set is a typed inventory rather than an undifferentiated pile: kind = device isolates real off-matrix routers worth filing as matrix gaps, while series-or-doc/accessory/module are expected non-devices to skip. Classification carries a single per-slug override today (nray-series → device; one kit, not a family — see the wiring entry below), and the logic is shared so a future include-accessories-style filter or "include module" cross-link reuses it instead of re-deriving. Regenerated deterministically by make device-map; validated by the existing device-map-check gate. See briefings/B-0018-product-naming-three-source-map.md.
• hardware-unmatched.tsv now shows the ETL www-spec mapping status — new www_code/www_specs columns. Each off-matrix /hardware page now records which mikrotik.com/product code the catalog actually pulled specs from and how many real fields it captured, read from the committed fixtures/hardware-catalog/catalog.json so the file view matches the DB rather than approximating it (no DB or network dependency in the drift gate). A blank www_code means the ETL mapped no marketing product to that page: filtering kind = device + blank www_code yields the concrete spec-backfill worklist (15 real devices as of 2026-07-13 — dynadish-6, sxt-2, chateau-lte6, the ltap-lr8-* LoRa kits, …), each of which usually does have a product page assess-www simply never fetched. make device-map reads catalog.json when present (blank columns + a console note if absent); the feedback mechanism to close the gaps is tracked in #70. See briefings/B-0018-product-naming-three-source-map.md.
• Curated hardware-www-map.toml wired into the ETL — off-matrix spec gaps closed (device gaps 15 → 4; #70 first task). The maintainer answer-key of /hardware-slug → mikrotik.com/product codes the auto-resolver can't derive (e.g. sxt-2→RBSXTG-2HnDr2-168, dynadish-6→dynadish_6) now feeds two ETL points through a shared loader (src/hardware-www-map.ts): assess-www seeds every code into its candidate fetch, and extract-hardware-catalog force-attaches single-code products to their row past the identity-agreement gate (a human vouched for them). 14 off-matrix devices that shipped with no specs — sxt-2, dynadish-6, cube-60g-ac, chateau-lte6/-lte12, intercell, lhg-xl-2, wireless-wire-cube, … — now carry full spec sheets in hardware_catalog; the reverse-audit spec-backfill worklist (kind = device + blank www_code) drops from 15 to 4, the remainder being genuinely un-single-attachable (seed_only module/kit pages whose identity is owned by the kits that embed them, and series-style multi-variant pages — both still fetched for the deferred surfacing work). All 65 curated codes were cross-checked against mikrotik.com/sitemap.xml. nray-series reclassified series-or-doc→device (one kit, not a family). Baseline canaries updated for the +52 seeded products (dropped-www 30→67, all accounted in the drop ledger; matrix coverage unchanged). See briefings/B-0018-product-naming-three-source-map.md and issue #70.
• Hardware-overlay data-quality follow-ups (B-0017 items 1–2), before Phase 2 surfacing. Two extract-time filters clean device_aliases/hardware_catalog so any consumer can trust them without re-deriving classifications: (1) a hardware-link alias whose token names a dropped www product is dropped as cross-sell pollution (e.g. qm_x→sxtsq-5-ax, mant_lte_5o→chateau-lte12) — a device's /hardware page merely linking an accessory no longer leaves that accessory code as the device's alias; standalone series pages (hw-*-series) stay exempt since they legitimately claim member/kit codes (wap_60g→hw-wap-60g-series). (2) specs_json._non_default_ips now keeps only genuine subnet deviations (outside 192.168.88.0/24) — same-subnet .88.x secondaries are filtered, shrinking the field from 63 rows to 11 (the 192.168.188.1 embedded-LTE cluster + intercell + woobm-usb). Both are proven by new checkInvariants probes (#8/#9) and anchor tests; fixtures/hardware-catalog/catalog.json moves once (20 cross-sell aliases removed, 16 collisions cleared). See briefings/B-0017-hardware-overlay-device-resolution.md "Lingering data-quality items" and issue #47.
• /hardware ↔ matrix.csv matcher correctness fixes (feeds hardware_catalog). Canonical matching (canon/canonNoRev/canonForms) with code > slug > table precedence, cross-page table suppression, a shared-&-base guard, a bogus-accessory-link filter, and an own-slug collision guard (kept only when corroborated or title-agreeing). Fixes previously wrong binds — Chateau LTE7/LTE12 → chateau-lte6-us; R11e-LR8/-LR9 → wAP LR2 kit (a canonNoRev over-strip); /hardware/hap-ax-2 (titled hAP ax³) also claiming hAP ax2 — and regenerates ros-hardware-assessment.json + the hardware_catalog overlay accordingly. Anchor-tested in src/assess-hardware.test.ts. See briefings/B-0018-product-naming-three-source-map.md.
• Relative Markdown links inside property descriptions now resolve to live manual.mikrotik.com URLs instead of being left as broken relative paths once extracted out of their source page.
• release.yml CI now builds the DB from the live Docusaurus extractor, not the legacy Confluence HTML export. The html_url workflow input is gone; a new extract-docusaurus.ts --check-counts --strict step proves the docs-count invariant on every release run instead of only in manual/local runs. Rebuilding a historical pre-migration DB remains possible via the local-only make extract-legacy-confluence target — it is no longer reachable from CI.
• Work tracking moved from in-repo tasks/T-*.md files to GitHub Issues. Issues start as discussion and earn the agent-ready label once acceptance criteria settle; umbrella marks theme-tracking issues and blocked marks named waits. PRs close issues via Closes #N, and partial landings must spawn follow-up issues before merge (new issue-pr-linking.instructions.md). tasks/ is now a frozen archive; the 17 remaining active task files were closed with per-file disposition notes (superseded / won't-fix / migrated to an issue) in tasks/done/T-*.md. See issue #18.
• Release version bumps are now a manual step for every channel, including latest. CI's old bump-version job (blind PATCH + 1, auto-committed straight to main) is gone entirely — it couldn't reason across the new prerelease channels. A new preflight fails a latest-channel release if CHANGELOG.md lacks a ## [<version>] heading for the bare package.json version being released, so CHANGELOG promotion can't be skipped by accident. See MANUAL.md "Version bumps are a manual step".
• make lint (and test.yml's Lint step, now make lint) covers Markdown and spelling, not just Biome. Adds markdownlint-cli2 (config in .markdownlint-cli2.yaml, rules shared with the IDE via .markdownlint.yaml; LLM instruction files, briefings/, tasks/, fixtures/, and the manual/ cache excluded) and cspell (cspell.json + a seeded project-words.txt RouterOS/tikoci dictionary) as bun run lint:md / lint:spell, so local make lint and CI check the same three things. release.yml's fast-fail Lint stays Biome-only.

Fixed

• A latent test-suite hazard that presented as a CI flake is now blocked structurally, not just caught on unlucky file order (issue #98). A *.test.ts that sets process.env.DB_PATH = ":memory:" but statically imports a module transitively reaching db.ts opens the real on-disk ros-help.db — the static import is hoisted above the env assignment, so db.ts's module-scope new sqlite(DB_PATH) runs before :memory: is set. query.test.ts's V-db-wipe-guard only trips on this when the offending file happens to load db.ts first, and bun's test-file order is not stable, so it hid across many green runs and did not reproduce locally. The live occurrence (a static import in extract-docusaurus.test.ts) was fixed in PR #97; new src/source-hygiene.test.ts (V-test-db-import-static-guard) now makes the extractor-import-side-effects rule enforced — it scans every DB_PATH-setting test file and fails at author time if any statically value-imports a db.ts-reaching module (statement-level import type is erased and stays allowed). Folded in a companion V-source-no-control-chars scan: no source .ts under src/ or scripts/ may carry a NUL or other disallowed C0/DEL control byte — the class where a stray NUL (also fixed in #97) makes file report the source as data and grep silently match nothing while tests/typecheck/biome all pass. Both are DB-less structural scans with no runtime cost.

• Video search now matches the video title, not just the transcript — correctly-titled changelog videos no longer get buried (issue #89). searchVideos() queried only video_segments_fts (chapter title + transcript), so a query word that lives only in the human-curated title — "changelog" is never spoken in the narration — was unmatchable, and a video like "RouterOS 7.23 changelog — VRF offloading, …" ranked #15 on changelog 7.23 routeros, below the display cutoff. It now also consults the already-trigger-maintained videos_fts (title + description; no schema change), ranking strong-before-weak: title-AND, then transcript-AND, then the OR fallbacks in the same order, deduped by video. On the live corpus that target video moves from #15 to #1, and changelog 7.NN routeros resolves to the matching release video for each version. Deliberately scoped to the FTS match surface: the shared extractTerms/stop-word handling that shreds dotted versions (7.23 → 7,23) and drops routeros is left untouched here — recognizing RouterOS version strings as first-class query tokens across routeros_search is tracked as separate, broader work.

• The shared version comparator now orders same-base prereleases instead of collapsing them (issue #107). compareVersions() in src/version-compare.ts stripped beta\d*/rc\d* wholesale, so 7.24beta1, 7.24beta2, 7.24beta3 all compared equal (likewise 7.20rc1 vs 7.20rc2) — deterministic only because callers (changelog.tsv, command_versions/routeros_command_diff ordering) bolt on a secondary sort key. The comparator now folds the prerelease number into the sort so the order is strict and total for one base: beta1 < beta2 < beta3 < rc1 < 7.24 (a bare beta/rc with no number sorts below its numbered siblings). Every caller improves at once; the module stays dependency-free so the DB-only export.ts boundary is unaffected. The honesty disclosure this bug forced into rosetta export's changelog.tsv manifest is removed, and a focused src/version-compare.test.ts pins the precedence.

• Agent-facing corpus counts no longer carry stale Confluence-era numbers (issue #88). src/mcp.ts's schema-guide resource and routeros_search's tool-description "Documentation: 317 pages from March 2026 Confluence export" line, plus MANUAL.md's Tables reference, still advertised frozen Confluence figures (pages 317, properties 4,860, callouts 1,034) and drifted command/version counts (command_versions 1.67M vs an actual ~2.2M) even though the live Docusaurus extractor has been the default prose source since the T-0035 cutover. Rather than re-baking exact numbers that drift every extraction, these surfaces now state order-of-magnitude counts, correct the Docusaurus provenance wording, and point at routeros_stats (already the live single source of truth) for exact figures — so the counts can't go silently wrong again. The #87 caveat note in MANUAL.md is dropped.

• 165 properties were silently destroyed on every build; properties and callouts now carry a real section_id (issue #90, schema v9). properties.section stored heading text, and UNIQUE(page_id, name, section) plus INSERT OR IGNORE dropped any property whose page repeated a heading — the shipped v0.11.0-rc.97 DB holds 4,416 of 4,581 parsed properties, and nothing logged the gap (the extractor's own "Properties:" line counted attempted inserts, so it reported rows the DB did not contain). ppp-aaa alone lost most of its user/profile properties to its four Properties headings. A corpus-wide measurement showed the obvious fix is insufficient: keying on section_id still destroys 87 rows, because the manual documents one property name several times within a single section (dot1x defines interface twice under Server — once for the server table, once for the client one), so section is not an identity at all. The constraint is therefore removed, not re-keyed; extractors now assert parsed == stored and fail the build on any gap (V-extractor-no-silent-drops). All 4,575 parsed properties are stored, and the 141 distinct properties the constraint had been eating — name = "Name used for authentication" and friends — are searchable for the first time. properties.section_id and the brand-new callouts.section_id (callouts previously had no section attribution whatsoever) resolve for 99.7% and 89% of rows respectively; the remainder are genuinely above any heading and stay NULL rather than being forced into a section. Property→section attribution rises from 72% resolvable to 99.7%. A property under an h4–h6 folds to its enclosing h1–h3 section, so sections remains the retrieval unit and routeros_get_page is unchanged; section keeps the raw h4 text, so the finer heading stays recoverable.

• routeros_lookup_property and routeros_search's related.properties now return section_anchor, which finally makes same-named properties tellable apart (issue #90). Storing every property surfaces a fact the old constraint hid by deleting it: 111 name/section groups (276 rows) are indistinguishable by section, because it holds heading text that repeats — a lookup_property(name="name") on PPP AAA returns three rows all reading "Properties", meaning the profile name, the login name, and the active-user name. section_anchor carries the section's anchor_id (properties, properties-1, properties-2), which is unique within a page and feeds straight back into routeros_get_page(page, section=…) to read that exact fragment — the workflow the tool description now spells out. Null when a property sits above any heading.

• bunx/global-install DB staleness across same-schema releases (e.g. 0.11.0-beta.92 → beta.93, or any future patch that doesn't bump SCHEMA_VERSION) is now self-healing (issues #76, #23). ensureDbReady()'s redownload check widens from "schema version differs" to also compare db_meta.release_tag against the running package's own version — the new checkDbFreshness() (src/setup.ts) — since dbDownloadUrls() already resolves a version-pinned DB per exact release and stamp-db-meta.ts already stamps the matching tag; the missing piece was that ensureDbReady never looked at it. A local repro (documented on #76) confirmed bunx re-resolves the registry manifest fresh, unconditionally, on every invocation, so the running package version is always current — only the cached DB could go stale. Dev-mode checkouts are exempt (a locally-extracted DB's release_tag is expected to drift from package.json's in-progress version and must never trigger a network fetch). A release-tag-only mismatch (schema still current) that fails to redownload — e.g. offline — degrades gracefully to the existing, still-queryable DB with a warning instead of crashing startup; only a genuine schema mismatch remains a hard failure.

• New ROSETTA_OFFLINE=1 environment variable short-circuits all DB freshness network attempts instead of letting them time out. A release-tag-only mismatch falls back to the existing DB with a warning (same as an unintentional offline failure); a genuine schema mismatch, or no DB at all, still fails hard — offline can't fix either. See MANUAL.md "How Updates Work".

• New bunx-upgrade-smoke release CI job closes the gap bunx-smoke couldn't cover: it only ever installs onto an empty cache, so it never exercised the actual upgrade path. The new job seeds a real bunx-installed DB from the previously-published version on the same dist-tag, then invokes the just-published version bare (no --refresh) and asserts the startup banner shows the new release — proving checkDbFreshness actually self-heals a live install, not just its unit tests. Also exercises ROSETTA_OFFLINE=1 against that same stale DB to prove the graceful-degradation path end to end. Runs on macOS/Linux/Windows; skipped when there's no prior publish yet on that channel (issues #76, #23).

• Device search resolves concatenated product-name spellings — hapax3 now finds hAP ax³ (issue #67). The whole separator-free class (hexs, capax, chateaulte12, fiberboxplus, …) previously fell through every stage: device_aliases stored hap ax3/hap_ax3 but never the concatenation, and the searchDevices stage written for exactly this case (a slug-normalized LIKE over devices.product_url) was dead in production — no extractor populates product_url, only the test fixtures did. extract-hardware-catalog.ts now derives a canon()-collapsed alias (source collapsed, lowest rank, zero cross-device collisions) from every surviving name-bearing alias, plus a standalone hw-<slug> row's own page slug (its only identity — how chateaulte12 finds the catalog-only original Chateau LTE12); hardware-table codes and matrix-linked rows' page slugs stay uncollapsed so hAP ax³'s disambiguation slug hap-ax-2 can't claim hapax2 from the real hAP ax². Read side, both alias probes (searchDevices stage 1.5 and searchAll's whole-input probe) fall back to the collapsed query key, so any separator spelling (hap.ax3, hap-ax3) resolves as mode: "alias"; a new per-token probe in searchAll surfaces a device mentioned inside prose ("hapax3 wifi settings") for tokens that look like model spellings (digit or code separator — bare English words that double as product names stay regex-gated on their styled form). Derived spellings resolve lookups but never appear in also_known_as; the dead product_url LIKE stage is removed. MCP + TUI both inherit via the query core.

• Category/index pages no longer arrive content-free — leaked <DocCardList /> is now expanded into a real child-link list (issue #65). manual.mikrotik.com's Docusaurus .md (llms.txt) output emits the @theme/DocCardList component unrendered, so ~50 index pages (e.g. bgp.md) reached the corpus as a bare tag plus its import line with their entire child-page navigation missing — the one MDX leak that is real data loss rather than cosmetic scaffolding. extract-docusaurus.ts now reconstructs the list from the rosetta-id path tree it already holds (a pre-parse pass, so links flow into pages.text/sections like ordinary prose), rendering each direct child as - [Title](url) — summary sorted by title. Honest fallback: a <DocCardList /> with no discoverable children is left untouched rather than silently erased. Upstream MDX passthrough (imports, WideTable, Tabs, ::: admonitions) is characterized in #65 and intentionally left as-is — readable or no data loss. Reported to MikroTik; this is the local stopgap.

• Classifier no longer turns natural-language prose into a bogus command path (issue #59, BL-4 from briefings/B-0020). A no-slash question whose first word happens to be a top-level menu (port forward to an internal server, port 8291 access) was greedily pathified by the canonicalizer into /port/forward/to/an/internal/server at medium confidence, because detectCommandPath in src/classify.ts runs the (DB-less, tree-unaware) canonicalizer on any input starting with a top-level word. The no-slash branch now only treats the input as navigation when every trailing token is a bare menu segment (^[a-z][a-z0-9-]*$, not a prose stopword like to/an/from); a stopword or non-identifier token means it's a question and it's left for FTS instead. Legit space-separated navigation is unaffected (ip firewall filter, interface wifi, routing bgp, port). Pinned by a classifier-only golden anchor (cmd-path-nl-not-a-path, using a new expected_classified null = "must be absent" convention). Being a pure heuristic without command-tree access, two clean prose tokens can still slip through (port forwarding→/port/forwarding); a tree-validated post-check in query-core is the deeper fix, deferred. Retrieving the NAT/port-forwarding page for such verbose NL is the separate #26 FTS-candidate-pool weakness, not this bug.

• Command→page linkage now ranks candidates instead of picking an arbitrary one, fixing wrong-page property lookups on core menus (issue #58, BL-1/BL-2 from briefings/B-0020). link-commands.ts collected every page that merely mentions a command path in a code block, then linked commands.page_id to candidatePageIds[0] — the first in insertion order — so on the Docusaurus corpus 433 of 493 linked dirs (88%) pointed at an unrelated page (/ip/firewall/filter→l3-hardware-offloading, /ip/dhcp-server→hotspot, /ip/service→containers). The pick is now the authoritative page: the one whose own /docs/ slug (or breadcrumb) trails the command path (new pure, unit-tested src/link-ranking.ts — contiguous trailing-segment match anchored at the leaf, prefix-tolerant so /ip/dhcp-server↔.../dhcp still matches; property count only tie-breaks among aligned pages). A command with no aligned page is left unlinked rather than mis-linked, so routeros_lookup_property falls back to an honest low-confidence global search instead of resolving high-confidence to the wrong page. Effect: aligned links ~2× (60→111), and real properties that exist now resolve high-confidence to the correct page (/ip/dhcp-server lease-time, /ip/ipsec/peer exchange-mode, /ip/dns servers) while the existing clean anchors (prop-wireguard-listen-port, prop-vxlan-vni) stay green. Properties the new manual documents only as prose (firewall action, BGP distance) and dotted/nested wifi properties remain unresolved — a separate property-extraction/CLI-Reference-overlay concern tracked apart from linkage.

• Flaky CI: two timer/IO-bound src/setup.test.ts cases (waitForUsableDb returns true…, probeDb closes statements…) no longer trip bun's default 5s per-test timeout under GitHub shared-runner contention (issue #32). Both resolve in a few hundred ms locally but do real setTimeout/bun:sqlite file I/O; each now gets an explicit 15s bun-level timeout (well above its own 2s internal deadline) so runner jitter can't fail them without weakening any assertion.

• release.yml's first real 0.11.0-alpha dispatch (2026-07-09) failed at "Run tests (fast-fail)" despite 0 failing tests (649 pass / 0 fail). Root cause: bunfig.toml carried a long-dormant [test].coverageThreshold = { lines: 0.70, functions: 0.80 } (present since April 2026, never actually exercised because test.yml never ran bun test with --coverage) — the first time --coverage ran anywhere in CI was T-0037's new step, and real coverage (55.64% lines / 62.78% functions) tripped the dormant threshold, failing the whole release despite every test passing. bunfig.toml's threshold is removed — coverage reporting is informational only, per VALIDATION.md V-coverage-reported, and should never gate on its own. Also hardened release.yml's error reporting so a real future failure of this kind is self-explanatory instead of a bare "Process completed with exit code 1": the "Run tests (fast-fail)", "MCP contract tests (real DB)", and "MCP retrieval eval (Phase 0)" steps now each emit an explicit ::error:: annotation distinguishing an actual test failure from a nonzero exit with no failing tests reported.

• Phase 0 golden-query retrieval eval (src/eval/retrieval.ts, fixtures/eval/queries.json) now matches on pages.rosetta_id instead of the numeric pages.id. extract-docusaurus.ts re-mints pages.id as a fresh rowid on every run, so the golden set's old Confluence-era numeric ids could never match again after the T-0036 cutover — this is what failed release.yml's "MCP retrieval eval (Phase 0)" step on T-0037's first real dispatch (all 12 nl-question queries scored 0% recall) even though routeros_search was actually finding the right page for 10 of the 12. fixtures/eval/baseline.json regenerated against the live Docusaurus corpus. Two real ranking regressions the fixture fix uncovered (a firewall-filter query and a BGP-peering query both miss their real page) are left failing visibly in the eval's per-query output rather than papered over — tracked by tasks/T-0038-docusaurus-retrieval-ranking-regressions.md.

• Release skill and command-tree extraction now use authenticated GitHub API requests with retry handling. This avoids shared-runner unauthenticated rate-limit failures while fetching tikoci/routeros-skills and discovering restraml versions during release.yml; the extract-skills-from-cache Makefile comment now correctly describes it as a local/offline rebuild path, not the release CI path. The shared retry helper (src/github.ts) now also backs off on GitHub's secondary/abuse-detection 403 + Retry-After responses (not just primary-quota exhaustion) and no longer misreads a missing x-ratelimit-reset header as an immediate-retry signal; skill file bodies are fetched with the documented application/vnd.github.v3.raw accept header.

• extract-docusaurus.ts's parseProperties() now also parses bullet-list property definitions (- **name** (type[; default: x]) : description), not just Markdown tables. Characterizing the 32 pages a naive bold-bullet grep flags (issue #20) showed the overwhelming majority are non-property prose (command menus, chain/enum-value lists, JSON API fields, naming-convention explainers) that must stay excluded, not force-fit; the real gain is scoped to 4 pages whose bullet lists are genuine property definitions: queues.md (+51), scheduler.md (+6, including one real upstream doc typo now correctly flagged malformedEmphasis instead of silently mis-parsed), clock.md (+2), and queues/htb-hierarchical-token-bucket.md (+3) — +62 properties corpus-wide (4509 → 4571), with zero new false positives across all 365 in-scope pages.

• FTS snippet() excerpt highlighting no longer collides with native Markdown bold in the Docusaurus corpus (issue #24). Since T-0035, indexed prose is raw Markdown and can already contain **bold**; wrapping a match in the same ** marker produced visibly broken output like ****hAP****. All 7 snippet() call sites (pages, properties, callouts x2, changelogs, video_segments, dude_pages) now share one exported sentinel-marker pair (query.ts's EXCERPT_MARK_START/EXCERPT_MARK_END, >>>/<<<) — corpus-safe, since the only literal occurrences across all 365 in-scope pages are confined to fenced code blocks (api.md's wire-protocol notation), never in prose. Also fixed 3 adjacent TUI rendering bugs in browse.ts found in the same footprint: the pages-excerpt bold conversion was correct but dead code (query.ts still emitted **); callouts/changelogs/videos/dude-pages excerpt rendering had no bold-off pairing (bold never turned back off once triggered); and renderProperties accepted but never rendered its own excerpt field, always showing the plain unhighlighted description.

Removed

• CI's automatic package.json/CHANGELOG.md version-bump commit (the bump-version job in release.yml) — superseded by the manual version-bump step above.

0.10.0 — 2026-07-08

Changed

• v0.10.x will be last release(s) based on Confluence HTML doc extraction. v0.11.x and beyond will use new manual.mikrotik.com as source for MikroTik documentation pages.

Fixed

• Only changelog and version bump. Promote to an even-number release. Otherwise, identical to 0.9.3.

0.9.3 — 2026-07-07

Changed

• YouTube transcript cache refreshed for the final Confluence-corpus release. The committed transcript cache now covers 538 MikroTik channel videos through June 19, 2026, and remains the source release CI imports via extract-videos-from-cache.
• Product matrix snapshot refreshed for the final Confluence-corpus release. The default device extractor now consumes the July 7, 2026 browser-exported matrix snapshot with 156 products.
• Release docs now identify local-only source refreshes. MANUAL.md distinguishes cache inputs that CI consumes from live sources CI already refreshes, including the product-matrix browser export caveat.

0.9.2 — 2026-06-18

Changed

• Docs now flag MikroTik's manual.mikrotik.com migration. Project docs record that future RouterOS manual updates are Docusaurus-based, not Confluence HTML exports, and outline the extraction/MCP/TUI redesign options.

Fixed

• Schema-mismatch setup guidance now points MCP users at a fresh bunx resolution. The refresh message tells users to restart the MCP client or run bunx @tikoci/rosetta@latest --refresh when their cached package is older than the published DB.

Removed

• Stale ros-toc.json PDF-era artifact removed. The current HTML extraction pipeline is the source of RouterOS page metadata, and no runtime code consumed the empty-title TOC file.

0.9.1 — 2026-05-26

Fixed

• Release npm publish now fails fast on missing or unauthorized NPM_TOKEN. The release workflow verifies npm package availability and read-write package access before extraction, artifact publishing, OCI pushes, or GitHub Release creation; partial release retries can update existing GitHub Release assets before publishing to the explicit npm registry.
• Release skill extraction now authenticates GitHub API reads. The extract-skills.ts GitHub API calls use GITHUB_TOKEN/GH_TOKEN when available, and release CI passes github.token to avoid unauthenticated API 403s while fetching tikoci/routeros-skills.

0.8.13 — 2026-05-03

Added

• make verify target. Runs typecheck + lint + tests + MCP contract tests + Phase 0 retrieval eval in one command. Requires a populated DB (make extract first). Covers V-typecheck, V-lint, V-unit, V-tool-registry, V-tool-shapes, V-tool-budget, V-retrieval-floor. Skips the clean-tree check that make preflight enforces.

Changed

• CLAUDE.md is now a thin routing index, and agent rules live in narrow .github/instructions/*.instructions.md files. Canonical reference material moved to MANUAL.md and DESIGN.md, legacy broad instruction files became routing stubs, and Copilot-facing guidance now points at the scoped rule files instead of duplicating the whole project reference.
• TUI/MCP parity and CLI help/manual parity are now CI-enforced. New src/browse-parity.test.ts proves every MCP tool has a matching .routeros_* browse dot-command, and new src/cli-help.test.ts locks MANUAL.md's CLI Flags table to bun src/mcp.ts --help. The --help output now documents browse <cmd> [args], browse --once <cmd>, and TLS env var names explicitly.
• MCP contract tests and Phase 0 retrieval eval are now blocking in release CI. Both were previously continue-on-error: true pending a first green CI run; both have now passed — continue-on-error removed and step names updated (dropped (non-blocking) suffix).
• Work tracking restructured. BACKLOG.md slimmed to an inbox + triggers list. Active work now lives in tasks/T-NNNN-*.md (frontmatter: status, depends_on, conflicts_with, validation, acceptance). Research and decision notes live in briefings/B-NNNN-*.md. New VALIDATION.md matrix names every load-bearing invariant and the CI step that proves it. Three new .github/skills/ (pick-next-task, promote-idea, verify-task) wrap the conventions. CLAUDE.md and .github/copilot-instructions.md doc-rule tables updated to match.
• Task verification docs now distinguish current proofs from planned ones. tasks/README.md and the verify-task skill no longer assume a make verify target already exists, VALIDATION.md now points V-db-min-content at the real inline release step, and V-retrieval-self is recorded honestly as a tracked gap until release CI actually runs the self-supervised eval.
• The test workflow now exercises the real stdio MCP client path. .github/workflows/test.yml runs src/mcp-stdio-client.test.ts, which spawns bun src/mcp.ts through @modelcontextprotocol/sdk's StdioClientTransport, proves the 14-tool registry/resources surface over stdio, and catches stdout framing pollution.
• Release bunx-smoke matrix now includes windows-latest. Catches the EBUSY / readonly-WAL / temp-file class of bugs on Windows. Step uses RUNNER_TEMP instead of mktemp (not available in Git Bash) and sets shell: bash as the job default.
• Release CI now runs the Phase 1 self-supervised retrieval eval (non-blocking). After the existing Phase 0 hand-curated eval step, release.yml now executes src/eval/self-supervised.ts against the freshly built full DB and appends the pass/fail result to the workflow summary. Results are visible but non-blocking until a stable baseline is established.

Removed

• make release, make build-release, make bump-version removed from Makefile. All release artifact production now goes through the GitHub Actions release.yml workflow. The Makefile retains ETL targets and developer checks (make preflight, make verify).

Security

• CodeQL ignore scope now anchors the root skills/ cache explicitly. .github/codeql/codeql-config.yml now ignores /skills/** instead of an unanchored skills/**, keeping the committed .github/skills/ workflow docs distinct from the extracted root-level skill cache.

0.8.12 — 2026-05-02

Fixed

• Windows package-mode DB installation no longer renames a SQLite-opened temp file. Download validation now finalizes every SQLite statement before close, and replaceDbFile retries transient EBUSY / EEXIST / EPERM rename failures for up to 30 seconds to cover delayed handle release, antivirus, or indexers.
• Abandoned .tmp.* DB artifacts are removed immediately when no active download lock exists, so failed Windows installs do not keep accumulating 274 MB temp databases between launches.

0.8.11 — 2026-05-02

Fixed

• Stale .tmp.* cleanup now runs on every startup, not only when a download is triggered. This removes accumulated 274 MB temp files from previous failed downloads even when the DB is already healthy.
• Windows rename now handles EBUSY in addition to EEXIST/EPERM in the replaceDbFile fallback path, providing better defense against antivirus or indexer locks on the destination.
• Schema-mismatch recovery messages no longer reference bun pm cache rm or use shell && syntax. The actionable command is now bunx @tikoci/rosetta@latest --refresh, which works cross-platform and handles both package and DB refresh in one step.

0.8.10 — 2026-05-02

Security

• CodeQL + Dependency Review wired up. New .github/workflows/codeql.yml runs the security-and-quality suite (security-extended + code-quality queries) against javascript-typescript and actions on push, PR, and a weekly cron. .github/codeql/codeql-config.yml excludes vendored/generated content (box/, dude/, transcripts/, matrix/, skills/, fixtures/, dist/, images/) so scans focus on shipped/runtime TypeScript, extractors, the bin shim, release scripts, and workflow YAML; test/eval harnesses are excluded to avoid temp-file/file-race noise outside shipped code. New .github/workflows/dependency-review.yml blocks PRs introducing high-severity dependency advisories. New .github/dependabot.yml opens weekly grouped update PRs for github-actions and bun ecosystems. The Test workflow gains an "AI findings probe" step that polls candidate Code Quality endpoints and prints a CI notice (no-op until GitHub ships a stable API). Repo-level Dependabot security updates, secret scanning with push protection, and private vulnerability reporting are enabled. See SECURITY.md for the configured posture summary.

Changed

• Documentation/instruction cleanup: agent-facing instructions, release/extraction docs, and BACKLOG structure now match the current CI pipeline, MCP resource surface, and DESIGN.md source-of-truth for cross-tikoci command validation strategy.
• MCP search/property confidence metadata: routeros_search.classified now includes command_path_confidence, and routeros_lookup_property rows include confidence (high/medium/low) to distinguish scoped command-page matches from global fallbacks.
• CI release hygiene: the Release workflow input formerly named force is now republish_assets, making clear that it reuploads GitHub Release assets / OCI tags while skipping immutable npm publication. Release CI also runs bun test in the early fast-fail gate before downloading the HTML export while preserving the post-extraction DB-wipe guard.
• DB retention: release builds now run make gc-versions after command linking to prune schema_node_presence to active RouterOS channel heads (stable, long-term, testing, development). Full command-version history and changelogs remain untouched.
• Tool descriptions: routeros_stats and routeros_current_versions now follow the workflow-arrow (→) convention. stats suggests → routeros_search; current_versions suggests → routeros_search_changelogs with a from_version/to_version hint. The Phase 2 contract test's KNOWN_EXCEPTIONS allow-list is removed — every registered tool now carries a follow-up arrow.

Fixed

• bunx/package startup no longer races the shared ~/.rosetta/ros-help.db. Package-mode DB preparation now uses a sidecar lock so concurrent MCP clients wait for the first download instead of competing to rename the file on Windows. Waiters no longer probe-lock the canonical DB while another process is replacing it, startup aborts instead of falling through to a schema-only empty DB when recovery fails, probes no longer create a missing canonical DB as a side effect, and stale .tmp / -wal / -shm artifacts are cleaned up instead of accumulating in ~/.rosetta/.
• Video transcript VTT cleanup: malformed cue markup is dropped without leaking tag fragments into extracted transcript text.
• Release workflow npm propagation log: the bunx smoke-test polling loop now reports the correct attempt number while waiting for the npm registry.

Added

• routeros_explain_command MCP tool: read-only CLI command explanation with canonical path/verb, argument property matches, warnings, docs, changelogs, version check, and TUI dot-command parity.
• canonicalize.ts: pluggable verb resolver, extractMentions(), per-command confidence flag (issue #5 — H4, H6, H8).
  • CanonicalizeOptions { isVerb?: (token, parentPath) => boolean } lets callers plug in a path-aware verb classifier. rosetta wires a DB-backed resolver against the commands table so /interface/wifi-qcom/info, /system/script/run, and other menu-specific verbs classify correctly instead of falling back to bare navigation. The resolver supplements the curated universal verb heuristic (it does not replace helpers like find, which are not enumerated everywhere in the command tree).
  • extractMentions(input, cwd?, options?) — surfaces every distinct path the input references, including bare navigation with no verb (e.g. /ip/firewall/filter standing alone in prose). Superset of extractPaths(). ParseResult also carries a new mentions: string[] field for callers that already use canonicalize() directly.
  • CanonicalCommand.confidence: 'high' | 'medium' | 'low' — high for well-formed CLI (absolute path with directly-identified verb), medium for relative-with-cwd or pure navigation, low when the verb was inferred from a trailing path segment (looser/prose-shaped input). Lets consumers filter prose-extracted results when they need higher precision.
• src/canonicalize-resolver.ts — DB-backed isVerb adapter for rosetta's commands table, with per-resolver in-memory caching. Wired into searchAll() via a ClassifyOptions { isVerb? } pass-through on classifyQuery, so MCP routeros_search and TUI s benefit automatically when input contains a path with a menu-specific verb.

Fixed

• Changelog version lookup and bridge VLAN retrieval. routeros_search / routeros_search_changelogs now keep exact patch-version lookups exact, but fall back from an absent major.minor changelog (for example 7.22) to its patch rows (7.22.*). Generic "what changed in X.Y" questions now populate related.changelogs, and bridge VLAN filtering searches treat "switch" as context so the dedicated Bridge VLAN Table page ranks in the top results.
• canonicalize.ts robustness — markdown / prose / common-verb gaps. Tokenizer now strips a leading U+FEFF BOM and treats backticks (`) and zero-width space (U+200B) as whitespace in both the outer and word loops, so inputs from markdown fences, doc snippets, and BOM-prefixed files extract cleanly instead of embedding the noise into the first path segment. GENERAL_COMMANDS gains four verbs that are universal in the rosetta commands table but were missing: clear, unset, reset-counters, reset-counters-all. Cross-checked against the DB to confirm zero path collisions — info/warning/error/debug are intentionally NOT added (/error is itself a top-level cmd; info is a dir at /interface/wireless). Menu-specific verbs need a path-aware resolver (tracked as H4 in the audit). New src/canonicalize.fuzz.test.ts documents both the shipped behaviour and the still-on-the-books H1–H8 hardenings.

0.8.9 — 2026-04-23

0.8.8 — 2026-04-22

Changed

• CI: bump-version now auto-promotes [Unreleased] → [VERSION] — DATE and prepends a fresh ## [Unreleased] skeleton after every release. No manual CHANGELOG fixup is needed — agents and developers only write to [Unreleased]; the version heading is filled in automatically.
• CI: Phase 0 retrieval eval runs on release builds (non-blocking). release.yml now executes bun run src/eval/retrieval.ts against the freshly built full DB after extraction and writes the report to the job summary. Non-blocking while the baseline adapts to the real-DB corpus — flip to blocking after one green real-DB run refreshes the baseline.
• CI: Phase 2 contract checks run in a dedicated real-DB step on release (non-blocking). release.yml executes bun test src/mcp-contract.test.ts after the full bun test suite so the token-budget and shape-invariant blocks run against the freshly built full DB in a fresh process (the shared bun test run pins the DB singleton to :memory: before this file loads, so Blocks B/C would otherwise skip). Non-blocking: continue-on-error: true while we observe the step green across a few rebuilds; test output is written to the job summary. test.yml intentionally does not get a dedicated step: a clean CI checkout has no ros-help.db, so B/C would skip regardless and the step would be redundant with Block A in the main run.

Added

• MCP behavioural eval framework (Phases 0–2) — three new surfaces for validating that the MCP tool layer keeps doing what we expect, with no LLM cost in the default flow:
  • Phase 0 (make eval) — 20 hand-curated golden queries in fixtures/eval/queries.json, scored on recall@k / MRR / classifier accuracy with baseline regression gating (2pp tolerance).
  • Phase 1 (make eval-self) — ~170 auto-generated queries from section headings, property names, and page titles using deterministic seeded sampling. Per-strategy thresholds + 5pp baseline tolerance.
  • Phase 2 (bun test src/mcp-contract.test.ts) — frozen tool registry test, workflow-arrow (→) convention check, token-budget guardrails on 10 canonical queries, and response-shape invariants for 5 representative queries (portable across DBs of varying richness). Runs inside bun test.
  • See BACKLOG.md "MCP Behavioral Testing — research + roadmap" for the full 5-phase plan.
• Tool-surface change ritual documented in CLAUDE.md: adding, removing, or renaming an MCP tool requires updating both src/mcp.ts and the EXPECTED_TOOLS array in src/mcp-contract.test.ts, plus a CHANGELOG.md entry under [Unreleased].

Fixed

• Phase 1 self-supervised sampling is now deterministic on full DBs. The cmd-path strategy no longer uses SQL randomness; it samples from a stable ordered set using the same seeded shuffle as the other strategies, so self-supervised-baseline.json stays reproducible across runs.

• CHANGELOG.md (Keep a Changelog format, back-filled from v0.1.0) with an agentic "update [Unreleased] on every user-visible change" rule in CLAUDE.md + CONTRIBUTING.md.

• TUI: view / v command. Re-renders the current context (page, results, sections, etc.) without popping the navigation stack the way b does. Useful after exiting the pager to re-read what you were looking at.

• TUI: bare page re-renders current page. When already in a page or sections context, page with no args re-renders the current page instead of erroring.

• CI: fast-fail quality gate. release.yml now runs typecheck + lint immediately after bun install, before the ~2-minute extraction pipeline. Tests continue to run post-extraction as the DB-wipe guard.

• CI: bump-version rebase-retry. Back-to-back release runs no longer fail with ! [rejected] HEAD -> main (fetch first). The job fetches + rebases onto origin/main and retries the push up to 3× (safe because the commit only touches package.json).

• routeros_search_tests: 512-byte rows surface first when no packet_size filter is set. 512B is the conventional mid-size benchmark RouterOS admins compare on, so within the LIMIT they now precede 1518B "best case" rows that previously crowded them out. Pin packet_size to override.

• TUI dot-commands print usage on missing required args. Calling e.g. .routeros_get_page with no args now prints the args, brief description, and TUI equivalent instead of silently returning null.

Fixed

• TUI device detail benchmark truncation now always keeps all 512B rows. When compacting long per-device test lists, the renderer now preserves every 512-byte result (the common comparison size) and only truncates non-512 rows.
• Tests/CI: importing extract-test-results.ts no longer opens the DB or runs extraction side effects at module-load time. The extractor now runs only under import.meta.main, and extract-test-results.test.ts sets DB_PATH=:memory: before dynamic import to prevent cross-file DB singleton contamination that could make query.test.ts fail depending on test order.
• extract-test-results: throughput values with thousands separators now parse correctly. Values like 7,112.3 Mbps were truncated to 7 because parseFloat stops at a comma. The extractor now strips commas before parsing, so the DB will contain correct figures after the next re-extraction.
• TUI pager: navigation keystrokes no longer bleed into the REPL prompt. Pager ran in raw mode while readline's data handler was still active, so each keystroke (1, 4, q, etc.) accumulated in readline's internal line buffer and reappeared echoed after the next prompt (e.g. > 1432q). Fixed by clearing rl.line/rl.cursor before re-prompting after dispatch.
• TUI: [p] and [cal] page hints now work on pages with sections. Pages with headings push ctx.type = "sections" (not "page"), so the p/prop and cal/callouts context-scoped handlers were silently falling back to "no page, show usage" even while a page was showing. Both handlers now check for sections context too, so all five footer hints ([N], [p], [cmd], [cal], [b]) work correctly regardless of whether the page has headings.
• TUI help text mentions [N] section navigation. The post-pager hint line now reads [N] = go to section N alongside [p] / [cal] / [b].
• TUI pager: digits open the listed result. In a results pager (search, devices, callouts, videos, properties, changelogs, sections, command tree, dude), pressing 1..N (where N is the number of visible results) now opens that result and exits the pager. Previously digits were always interpreted as page jumps, so users had to quit the pager (q) and then type the number — wasted keystrokes on the most common path. Page-jump still works for digits beyond the visible result count.
• routeros_search_changelogs X..Y version range is now inclusive on both ends, normalises reversed ranges (7.21..7.20 → 7.20..7.21), and returns entries chronologically (oldest first).
• Build: missing compareVersions import in src/browse.ts — was failing typecheck on both test.yml and release.yml.

0.8.2 – 0.8.3 — 2026-04-22

Changed

• TUI polish round-2. Dot-command aliases (.s → .routeros_search), back navigation re-renders, page calendar rendering, Markdown → ANSI sweep across skills/pages.
• CI: bump-version decoupled from bunx-smoke — a smoke regression no longer blocks the next version from being available for the fix release. Force-mode runs also skip the npm publish step (npm versions are immutable).
• Lint rule sharpened. bun run lint must be zero errors repo-wide, not just on touched files.

Fixed

• browse CLI args now route through the normal TUI dispatcher, so every TUI command (not just s) works when passed at launch.
• Resolved noNonNullAssertion lint errors in canonicalize.test.ts that were blocking CI.

0.8.0 – 0.8.1 — 2026-04-21

Fixed

• bunx install path is now rock-solid on macOS. The last { readonly: true } DB open (in mcp.ts::ensureDbReady and setup.ts::dbHasData) was removed. Freshly-written WAL-mode SQLite DBs with no .shm sibling cannot be opened readonly on macOS, which caused Validated … | Still incompatible after re-download (DB=unreadable) for v0.8.0 users. Added a structural anchor test that forbids { readonly: true } on DB opens.

Added

• Cross-platform bunx smoke job in CI. release.yml now runs a bunx-smoke matrix on macOS + Linux after npm publish, pinning the just- published version and exercising --refresh, --version, and the full MCP server boot path. Linux-only CI had green-lit v0.8.0 before this was added.

0.7.5 – 0.7.8 — 2026-04-21

Added

• db_meta table (schema v5). Database provenance — release_tag, built_at, source_commit, schema_version. Stamped at release time, shown in the startup banner.
• Auto-update story for bunx. DB download URL pins to the running package version (releases/download/v<VER>/ros-help.db.gz) with latest as fallback. Atomic .tmp.<pid> write, magic-byte + size + schema probe, then renameSync. Stale .db-wal / .db-shm siblings are cleaned up in the same step. Schema mismatch is a hard error with an actionable message.
• TUI usability: MCP probe via dot-commands. .routeros_search, .page, .device etc. invoke the same code path as the MCP server tool and dump raw JSON. .help lists all 13 dot-commands. Contract: "a human can always see exactly what the agent would receive."
• Hunger-knob related caps. routeros_search.limit scales callout / video caps proportionally via relatedCaps(limit).
• Glossary in related. Short queries that match a glossary term/alias surface the definition in related.glossary.

Fixed

• CI DB-wipe regression (v0.7.6). extract-dude.test.ts had imported extract-dude.ts (which loads db.ts) before any DB_PATH=:memory: was set; query.test.ts:beforeAll then DELETE FROM …'d the CI-built DB, shipping a 3-page release. Fixed with DB_PATH=:memory: hoisting, a query.test.ts hard-fail if the singleton isn't :memory:, and a release.yml DB content gate (pages ≥ 200, commands ≥ 1000, devices ≥ 100, properties ≥ 1000) that runs before publish.
• extract-html.ts exits non-zero if 0 pages are extracted.
• probeDb and ensureDbReady open the DB read-write so WAL-mode init doesn't fail on macOS.

0.7.0 – 0.7.4 — 2026-04-20

Added

• North Star — unified routeros_search. New pre-search regex classifier in src/classify.ts (pure module, 42 table-driven tests) detects command path, version, topic, device model, command fragment, and property-name candidate. searchAll() in src/query.ts wraps searchPages and runs classifier-driven side queries in parallel, returning { query, classified, pages, related: {command_node, properties, devices, callouts, videos, changelogs, skills, glossary}, next_steps }.
• Glossary table. Seeded at DB init. Resolves RouterOS domain jargon (product codes, abbreviations, subsystem names).
• Known-topics table. Union of changelog categories and command path segments for soft topic routing in the classifier.
• Changelog range expansion. buildChangelogVersionSet includes channel head versions and latest long-term patches.

Removed

• routeros_search_callouts and routeros_search_videos — folded into routeros_search.related. Tool count: 15 → 13. The underlying searchCallouts() / searchVideos() functions remain in query.ts as internal helpers used by searchAll() and getPage() TOC mode.
• routeros_search_properties — previously removed (useless without command-tree context); internal function retained for TUI.

Changed

• routeros_get_page is budget-aware: TOC mode surfaces top properties, related videos, and callout summary inline, so small-budget callers rarely need a second tool call.

0.6.4 – 0.6.9 — 2026-04-13 → 2026-04-20

Added

• schema_nodes table + multi-arch import. deep-inspect.json from tikoci/restraml is now the preferred source. Dual-arch (x86/arm64) trees, _completion data (11K+ args with valid values + 17 style types), schema_node_presence flat junction, _attrs JSON catch-all. The commands + command_versions tables are regenerated from schema_nodes for backward compatibility.
• desc_raw decomposition. Parsed into data_type, enum_values, range_min/range_max, max_length at import time.
• Completion data in browseCommands() / browseCommandsAtVersion().
• RouterOS agent skills as MCP resources. rosetta://skills (listing) and rosetta://skills/{name} (per-skill content) with provenance header noting community/AI-generated/human-reviewed status.
• CLI flag support for DB path. Explicit --db <path> overrides all discovery modes.
• Section-level excerpts in search + server-wide instructions surfaced via SERVER_INSTRUCTIONS.
• RouterOS CLI path canonicaliser. src/canonicalize.ts maps any input form to { path, verb, args } tuples (61 tests covering subshells, blocks, navigation).
• Release workflow version resolution. release.yml reads package.json for version when workflow input is blank.

Fixed

• dude_pages: stripped Wayback / wiki chrome from extracted text; removed stub entries; routeros_dude_get_page accepts max_length.
• browseCommands arch filtering corrected and tests added.
• Removed { readonly: true } from early DB validation in setup (repeat regression trail — finally closed in 0.8.1).
• Stop words + compound terms counts corrected in tool descriptions.

0.5.x – 0.6.3 — 2026-04-09 → 2026-04-13

Added

• MCP Registry metadata. server.json manifest + CI validation job.
• MCP dataset resources. rosetta://datasets/device-test-results.csv, rosetta://datasets/devices.csv, rosetta://schema.sql, rosetta://schema-guide.md.
• routeros_command_diff. Structural diff of command trees between two RouterOS versions.
• PRAGMA user_version written at DB init; MCP server validates on boot.
• Sitemap-based device slug resolution for 100% product-page coverage; AKA / alias matching via dash-split + slug-normalised LIKE.
• Changelog extraction: legacy version support with CI verification; version-set building tests.
• ensureDbReady function in mcp.ts — hard validation before the server starts serving.

Fixed

• Per-session HTTP transport routing (each MCP client session gets its own McpServer + transport).
• .dockerignore added to slim the build context.
• OCI smoke test via docker pull (not docker load); container entrypoint restored in Docker build context.

0.4.x — 2026-04-04 → 2026-04-09

Added

• routeros_search_tests. Cross-device ethernet + IPSec benchmark search with mode, configuration, and packet-size filters.
• Device test results + block diagrams. Scraped from mikrotik.com/product/<slug>: 2,874 measurements across 125 devices, 110 block-diagram URLs.
• Experimental TUI (browse). Interactive terminal browser — REPL with paging, OSC 8 links, context-scoped navigation.
• Video transcripts via yt-dlp. 518 MikroTik channel videos, ~1,890 chapter-level segments with timestamps. NDJSON cache in transcripts/ makes CI reproducible without a yt-dlp dependency.
• Unicode superscript / subscript normalisation in product names.
• Auto-bump patch version after release (Makefile + CI).
• Security policy documentation (SECURITY.md) + build-script hardening against shell injection.

Fixed

• HTTP transport test stabilisation; lint sweep.
• search_tests response slimmed to reduce context bloat.

0.3.x — 2026-03-31 → 2026-04-01

Added

• Streamable HTTP transport via --http flag. Built on Bun.serve() + WebStandardStreamableHTTPServerTransport, stateful per-session routing, optional --tls-cert / --tls-key for direct HTTPS. Defaults to localhost; --host 0.0.0.0 logs a warning.
• OCI image publishing (ammo74/rosetta on Docker Hub, ghcr.io/tikoci/rosetta on GHCR). Multi-arch linux/amd64 + linux/arm64. Smoke-tested in CI via docker pull.
• get_page smart budgeting. max_length default 16000, compact callout summary in TOC mode.

Fixed

• Replaced crane with Dockerfile + docker buildx for OCI builds — several crane approaches all failed identically on Docker 28 with containerd image store.
• Per-session HTTP transport routing.

0.2.x — 2026-03-30

Added

• npm distribution. bunx @tikoci/rosetta as canonical install. Runtime version resolution (import.meta.dirname + package.json read) so --version shows a real number. Claude Desktop full-path PATH workaround documented in --setup output.
• Changelog extraction from download.mikrotik.com/routeros/<ver>/CHANGELOG; routeros_search_changelogs tool with version range + category + breaking filters.
• Markdownlint configuration (.markdownlint.yaml, .markdownlintignore).

Fixed

• CI release workflow: pass HTML dir to extract-properties; tolerate Confluence zip absolute-path entry; lint issues; TypeScript dev-dependency for typecheck.
• inspect.json fetched from restraml GitHub Pages (removed ~/restraml dependency).

0.1.0 — 2026-03-26

Initial public release.

Added

• Core MCP server (src/mcp.ts) with 8 tools: routeros_search, routeros_get_page, routeros_lookup_property, routeros_command_tree, routeros_device_lookup, routeros_command_version_check, routeros_current_versions, routeros_stats.
• HTML extraction pipeline (317 pages, 4,860 properties, 1,034 callouts, 2,984 sections) + command tree (46 RouterOS versions, 1.67M command-version junction rows) + product matrix (144 devices).
• SQL-as-RAG with FTS5 (porter unicode61 for prose, plain unicode61 for device model numbers), BM25 ranking, compound-term recognition, AND→OR fallback.
• Compiled single-file binaries for macOS arm64/x64, Linux x64, Windows x64 via bun build --compile.
• --setup flow. Downloads DB from GitHub Releases, prints MCP client config snippets for Claude Desktop, Claude Code, VS Code Copilot, Copilot CLI, Cursor, Codex.
• DB_PATH env override + three-mode DB path resolution (compiled / dev / package at ~/.rosetta/).
• Bun tests for the query planner + schema health.