<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>vir releases</title><description>Release notes for vir, an LLM Wiki for Claude Code in your Obsidian vault.</description><link>https://virwiki.dev/</link><language>en</language><item><title>vir 0.37.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**`vir drift [path]` finds where a repo&apos;s CLAUDE.md and README no longer match its code, with no LLM calls.**</description><pubDate>Tue, 17 Nov 2026 12:00:00 GMT</pubDate><content:encoded>**`vir drift [path]` finds where a repo&apos;s CLAUDE.md and README no longer match its code, with no LLM calls.**

- **`vir drift [path]`** reads the root `CLAUDE.md`, every nested `CLAUDE.md` and the root README, one doc at a time, and checks each claim against the code.
  - What it checks: paths (code spans, directory trees, relative links), npm scripts, the package manager, dependencies named in a Stack list or an `npm install`, the repo&apos;s own CLI commands and flags, and env vars written as `process.env.X`, `$X` or `X=`.
  - Only differences are listed, per doc: stale (the doc says it, the code has none) or changed (the doc says X, the code has Y). &quot;Changed&quot; is used only for a rename git recorded, or when one file with that name is left.
  - One undocumented list covers source modules and CLI commands no doc mentions. With nothing to report it prints &quot;No drift detected&quot;.
  - A path whose first folder isn&apos;t in the repo (the vault, `~`) is skipped, not reported. A section about a sub-package (`site/`) resolves against that package.
  - In a worktree, where CLAUDE.md is usually gitignored, the main checkout&apos;s copy is read and labelled as such.
  - Reports go to `drift/&lt;project&gt;.md` and `drift/&lt;project&gt;/YYYY-MM-DD.md` in the vault (Europe/Belgrade date; `--tz` changes it). `--report &lt;file&gt;` writes one more copy. `--stdout` and `--json` print without writing.
  - `drift/` is kept out of search and `vir lint`, like `journal/`.
  - On this repo: 678 claims across 2 docs, 1 finding (`src/dream/` missing from CLAUDE.md&apos;s Structure section).
- **`vir_drift` MCP tool** answers &quot;is this repo&apos;s CLAUDE.md still accurate?&quot; mid-session. It is read-only: no writes, no model call, and only repos under your home folder. It follows a worktree&apos;s main checkout only when that checkout is also under your home folder.</content:encoded></item><item><title>vir 0.36.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**Each project&apos;s CLAUDE.md / AGENTS.md block now opens with a line telling the agent to check vir before it starts work.**</description><pubDate>Mon, 16 Nov 2026 12:00:00 GMT</pubDate><content:encoded>**Each project&apos;s CLAUDE.md / AGENTS.md block now opens with a line telling the agent to check vir before it starts work.**

- **`vir sync-claude` opens every project block with a read-first line:** &quot;Before debugging or changing something in &lt;project&gt;, check what past sessions learned&quot; (`vir_query`, or `vir query`).
  - When `vir dream` has written the project&apos;s page, the line also points at `vir_project_summary` and says to prefer confirmed entries.
  - Why: Anthropic&apos;s agent-memory research found that telling a model to read memory before answering moved one model from 0.36 to 0.64. Having the memory available was not enough.
  - The line stays even when every gotcha was left out.
  - The sync diff shows it as `+ read-first line` until a file has it.
  - On the reference vault this is 25 project files; 9 of them also link a dream page.</content:encoded></item><item><title>vir 0.35.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**`vir dream &lt;project&gt;` rebuilds a project&apos;s reference page from its notes, with every entry quoted from its source and labelled confirmed or unconfirmed.**</description><pubDate>Mon, 16 Nov 2026 12:00:00 GMT</pubDate><content:encoded>**`vir dream &lt;project&gt;` rebuilds a project&apos;s reference page from its notes, with every entry quoted from its source and labelled confirmed or unconfirmed.**

- **`vir dream &lt;project&gt;` (and `--all`)** is a cross-session pass modelled on the &quot;dreaming&quot; in Anthropic&apos;s agent-memory research. One model call reads a project&apos;s notes together and groups the facts by area into `projects/&lt;slug&gt;.md`, which agents read through `vir_project_summary`.
  - Every entry must quote its source note verbatim, and an entry whose quote isn&apos;t in that note is dropped.
  - Each entry is labelled from your checks, not by the model: confirmed (you verified the note, or `vir audit --transcript` found it in its session) or unconfirmed.
  - Notes that disagree are listed under &quot;Needs review&quot;, and gaps the notes leave open under &quot;Gaps&quot;.
  - Flagged and contradicted notes never reach the prompt.
  - The page is capped at 40 entries, 12 KB and 0.35× its notes&apos; size, and is never used as a source for search or another dream.
  - An unchanged project costs nothing to rerun.
  - On the reference vault&apos;s `vir` project (38 notes): 26 entries, every quote found in its linked note, 3 entries dropped for a quote that wasn&apos;t there, 2 conflicts, 0.11× the source size.
- **`vir summarize &lt;project&gt;` and `--all` now run `vir dream`.** `--week` and `--month` are unchanged. The distill pass no longer overwrites a dream page with its old auto-summary.
- **`vir_project_summary` returns `kind`, `entries`, `confirmed_entries` and `stale_notes`** (notes newer than the page), and points at `vir dream` when a project has no page.
- **`vir audit --transcript` checks the notes agents retrieve most first**, then the rest oldest first. A `--limit`, and the daemon&apos;s 20 checks per pass, now go to the notes search actually serves.
- **A flag now covers every copy of a session.** The same session can be distilled twice, once from the main checkout&apos;s project folder and once from a worktree&apos;s. An audit or transcript check lands on one copy only, so the other could still reach CLAUDE.md / AGENTS.md through `vir sync-claude`, or reach `vir dream`. Flags are now matched by session id. A note you verified still goes in.
- **`vir prune --duplicates` keeps one note per session copied under two project folders.** It shows what it would change unless you pass `--apply`. The copy with an audit or transcript verdict stays, then the one with a lesson, then the newest. The others go through the usual prune flow (note to `.rejected/`, `vir prune --restore` undoes it). On the reference vault: 7 rows. `vir run` no longer distills a session a second time when another path already serves its note.</content:encoded></item><item><title>vir 0.34.1</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**`vir sync-claude` no longer adds a second VIR block when a line starts with inline code.**</description><pubDate>Sun, 15 Nov 2026 12:00:00 GMT</pubDate><content:encoded>**`vir sync-claude` no longer adds a second VIR block when a line starts with inline code.**

- **`vir sync-claude` no longer adds a second VIR block when a line starts with inline code.** Any line beginning with three backticks was taken for the start of a code block, so a line like ```` ```` ```json ```` occasionally ```` hid every VIR marker after it and sync-claude appended a new block instead of replacing the old one. Code blocks are now detected the way Markdown does: a backtick fence can&apos;t have backticks after its opening run, and only a bare run of the same character, at least as long, closes it. Running `vir sync-claude` once on a file that already has two blocks keeps the first and drops the rest. The diff preview now reads the same block that gets written.
- **`vir lint --links` no longer skips links after such a line.** It used the same check and now uses the same detector.</content:encoded></item><item><title>vir 0.34.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**Search results say how far each note can be trusted: verified, transcript-supported, unbacked, contradicted or unchecked.**</description><pubDate>Fri, 13 Nov 2026 12:00:00 GMT</pubDate><content:encoded>**Search results say how far each note can be trusted: verified, transcript-supported, unbacked, contradicted or unchecked.**

- **Search results say how far each note can be trusted.** `vir_query` (MCP) returns a `status` for every source: `verified` (you approved it in `vir review`), `transcript-supported` (`vir audit --transcript` found every claim in its session), `unbacked`, `contradicted`, or `unchecked`. The synthesized answer sees the same labels and is told not to state a contradicted note&apos;s claims as fact, and `vir query` prints a warning under a contradicted or unbacked source. A check on an older version of a note doesn&apos;t count. A prompt whose notes were never checked is unchanged. On the reference vault, an agent asking about TS2339 errors now gets the note that miscounted them labelled `contradicted`. `vir query --json`, the Obsidian plugin&apos;s contract, is unchanged.</content:encoded></item><item><title>vir 0.33.1</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**`vir queries` counts dead weight only in note folders.**</description><pubDate>Fri, 13 Nov 2026 12:00:00 GMT</pubDate><content:encoded>**`vir queries` counts dead weight only in note folders.**

- **`vir queries` counts dead weight only in note folders.** &quot;Never surfaced&quot; listed files vir never writes (a stray `Untitled.md`, the `decision.md` hub Obsidian creates when a Category link is clicked) and `projects/` hub pages; on the reference vault those were 24 of the 140 entries. Only patterns, gotchas, decisions, tools, topics, articles, PDFs and accepted insights are counted now.
- README: the test badge and the Project status table were stale (1411 and 1073 tests, version 0.24.2).</content:encoded></item><item><title>vir 0.33.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**`vir lint --links` finds links to notes that no longer exist, and `--fix` repairs Related lists in place.**</description><pubDate>Thu, 12 Nov 2026 12:00:00 GMT</pubDate><content:encoded>**`vir lint --links` finds links to notes that no longer exist, and `--fix` repairs Related lists in place.**

- **`vir lint --links` finds links to notes that no longer exist.** A session note&apos;s Related list is built from its nearest notes when it is written, and nothing updated it when one of those notes was later retitled, rejected or pruned. On the reference vault that left 310 of 1427 links pointing at nothing. Each dead link is reported as retitled (the session&apos;s note now has another title), removed (the target is only in `.rejected/` or `archived/`), or unknown. Project and category header links are not counted, and a link resolves the way Obsidian resolves it: by file name anywhere in the vault, or by alias.
- **`vir lint --links --fix` repairs Related lists in place.** It drops each dead bullet and fills the slot with the nearest existing note not already listed, from the note&apos;s stored embedding (no model or embedding calls). Live links and everything outside Related stay as they are. A dead link is never repointed by session id: a re-distilled session can come back under an unrelated title. In topic pages and journal days it only drops links whose target is gone; `vir compose` and `vir journal` rebuild those. On the reference vault: 310 dead links down to 38 in under a second, with no live link lost.</content:encoded></item><item><title>vir 0.32.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**`vir sync-claude` no longer writes notes your audit flagged into CLAUDE.md or AGENTS.md.**</description><pubDate>Tue, 10 Nov 2026 12:00:00 GMT</pubDate><content:encoded>**`vir sync-claude` no longer writes notes your audit flagged into CLAUDE.md or AGENTS.md.**

- **`vir sync-claude` leaves out notes your audit flagged.** A gotcha with a fresh `vir audit` verdict of reject, verify or merge, or a fresh `vir audit --transcript` contradiction, no longer becomes a CLAUDE.md or AGENTS.md lesson; the next-best gotcha takes its slot, and each file&apos;s diff says how many were left out. These blocks load into every session, so a doubtful note there misleads every session, not one search. Approving a note in `vir review --audited` (`verified: true`) lets it back in, and a verdict on an older version of the note doesn&apos;t count. On the reference vault this left out 56 flagged notes across 14 files, including a snapshot (&quot;Codebase structure: 304 TypeScript source files…&quot;) that would have gone into growthq&apos;s CLAUDE.md.</content:encoded></item><item><title>vir 0.31.1</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**`vir lint --stale` shows whether agents still retrieve each stale note, and `vir queries` stops counting a note&apos;s lookup of itself.**</description><pubDate>Tue, 10 Nov 2026 12:00:00 GMT</pubDate><content:encoded>**`vir lint --stale` shows whether agents still retrieve each stale note, and `vir queries` stops counting a note&apos;s lookup of itself.**

- **`vir lint --stale` shows whether each stale note is still retrieved.** Each entry now says &quot;never retrieved&quot; or &quot;retrieved N×, last &lt;date&gt;&quot;, from `~/.vir/queries.jsonl`, and never-retrieved notes are listed first. Age alone was a weak signal: on the reference vault, 28 of the 31 notes flagged as stale had been retrieved since, one of them 36 times. Nothing is hidden on this basis, since the query log is a thin sample.
- **`vir queries` no longer counts a note as surfacing for itself.** A query whose text is the note (the Obsidian sidebar&apos;s related-notes lookup) returns that note at rank 1, which put whatever note was open at the top of &quot;most surfaced&quot;. Those hits are now left out and counted on their own line (24 on the reference vault). Both reports share one per-note usage count (`src/search/noteUsage.ts`).</content:encoded></item><item><title>vir 0.31.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**`vir audit --transcript` checks each note against the transcript it was distilled from, and the daemon can run it for you. The journal carries unfinished tasks forward.**</description><pubDate>Tue, 10 Nov 2026 12:00:00 GMT</pubDate><content:encoded>**`vir audit --transcript` checks each note against the transcript it was distilled from, and the daemon can run it for you. The journal carries unfinished tasks forward.**

- **`vir audit --transcript` checks notes against the transcript they came from.** The plain audit judges note text alone, so it catches generic and snapshot notes but not wrong ones. This mode sends each note with its transcript (tool-filtered, scrubbed and trimmed as the distiller saw it) and records each note as supported, unbacked or contradicted. Only contradicted notes join `vir review --audited`, as verify with the contradiction as the reason, and a fresh reject or merge verdict is never downgraded. On real notes, contradictions were genuine errors (an error count of 78 where the transcript showed 23, a reminder recorded as added that was only offered), while unbacked was mostly claims from truncated tool output. Notes whose transcripts the agent has pruned are counted as &quot;transcript gone&quot; and skipped, so the check is worth running soon after `vir run`. `--dry-run`, `--project`, `--limit` and `--all` work as in the plain audit, and notes are taken oldest first, since their transcripts are pruned next.
- **The daemon can run the transcript check for you.** With `&quot;checkTranscripts&quot;: true` in `~/.vir/config.json`, each `vir run --daemon` pass that is not halted checks up to `checkTranscriptsPerRun` (default 20) notes after distilling, oldest first, inside the same pipeline lock, and logs one line to `daemon.log`. A quota halt or failed call is logged and retried next pass; it never fails the run. Off by default because each note is one paid call priced like a distill.
- **The journal carries open loops forward.** Each day gets a **Carried over** section per project with tasks left unfinished on an earlier day and not touched since, each with its task pointer, its state and how many days it has been open. A loop stops carrying once it is completed or deleted, once a later session touches a task with the same subject (a resumed session or a re-plan), or after 7 days untouched.</content:encoded></item><item><title>vir 0.30.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**`vir journal` writes a daily record of what you did, per project, from your session transcripts. No LLM calls.**</description><pubDate>Thu, 05 Nov 2026 12:00:00 GMT</pubDate><content:encoded>**`vir journal` writes a daily record of what you did, per project, from your session transcripts. No LLM calls.**

- **`vir journal [date]`** writes `&lt;vault&gt;/vir/journal/YYYY-MM-DD.md` with each project&apos;s Done, In progress, Blocked and Open work, plus links to that day&apos;s decision notes. `--since &lt;date&gt;` writes every day with activity from that date on, `--all` rewrites every day, `--stdout` prints instead of writing, and `--tz &lt;zone&gt;` sets the day boundary (default: this machine&apos;s time zone).
- **Every line has evidence.** A Done line points to a commit sha or a completed task. A commit counts once git printed its sha; a task counts once its tool result confirmed it. Rejected commits and unanswered calls don&apos;t count. Tasks come from TaskCreate/TaskUpdate, TodoWrite and the Claude desktop app&apos;s task list.
- **The journal outlives your transcripts.** Claude Code deletes transcripts after about 30 days; what they prove is copied into two new SQLite tables first. The files are regenerated from those, so deleting the folder and re-running brings it back identical. Edits to the files are overwritten.
- **Search and lint skip the journal.** It&apos;s a record of work, not knowledge: `vir_query` and `vir lint` ignore the `journal/` folder.
- **Known limits:** Claude Code sessions only (not Codex yet); `git commit -q`, merges, rebases and commits made outside an agent session aren&apos;t seen; a task appears only on days it was touched.
- **npm shows a funding link** (Ko-fi).</content:encoded></item><item><title>vir 0.29.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**Search returns more relevant notes: MMR is off by default, and verified notes can no longer push past better matches.**</description><pubDate>Sat, 17 Oct 2026 12:00:00 GMT</pubDate><content:encoded>**Search returns more relevant notes: MMR is off by default, and verified notes can no longer push past better matches.**

- **MMR reranking is off by default (`retrievalDiversity: 0`).** On vir&apos;s own retrieval eval (54 judged queries), MMR at 0.3 made results worse for both embedding models: nDCG@8 fell by 0.089 with nomic and 0.079 with bge. Under nomic it also pulled off-topic notes into the top results because they were unlike everything else. Existing installs are migrated once: a stored `0.3` (the old default, which `vir init` wrote into every config) becomes `0`. Any other value is kept, and if you set `0.3` again afterwards it stays. `config.json` gains a `configVersion` field to record the migration.
- **Verified notes win near-ties instead of jumping the queue.** The +0.2 boost for `verified: true` notes now applies only to notes already in the top 2×K by relevance, in both embedding and keyword search. Before, an accepted connect-pass rule ranked first or second for almost every query, including unrelated ones.
- **`vir embed` re-embeds accepted connect-pass rules.** `vir embed --force`, the command for switching embedding models, skipped them, so accepted rules dropped out of search until the next `vir run` restored them. Without `--force` it also fills in rules that are missing a vector or have one from another model, as `vir run` does.</content:encoded></item><item><title>vir 0.28.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>- **`vir sync-claude` writes lessons, not slugs.** A project&apos;s CLAUDE.md gets its top 5 gotchas, each as the one-line lesson from the note (`- Zod 4 toJSONSchema requires io: &quot;input&quot; when schemas contain .transform()`). Confidence scores and slugs are gone from the file. The global CLAUDE.md and AGENTS.md no longer get every project&apos;s notes: only one line pointing at `vir_query`, plus your global </description><pubDate>Fri, 16 Oct 2026 12:00:00 GMT</pubDate><content:encoded>- **`vir sync-claude` writes lessons, not slugs.** A project&apos;s CLAUDE.md gets its top 5 gotchas, each as the one-line lesson from the note (`- Zod 4 toJSONSchema requires io: &quot;input&quot; when schemas contain .transform()`). Confidence scores and slugs are gone from the file. The global CLAUDE.md and AGENTS.md no longer get every project&apos;s notes: only one line pointing at `vir_query`, plus your global rules. The first sync after upgrading shows the old lines as removals.
- **Gotchas get their lesson at distill time.** After titling, a gotcha note gets one more cheap classify-model call that states its rule in one sentence (`lesson:` in frontmatter and a new `lesson` column). `vir sync-claude` uses it, so a note whose first points are background no longer puts that background into CLAUDE.md. Older notes keep the extracted lesson; nothing is re-distilled. Other categories make no extra call.
- **`vir sync-claude --backfill-lessons`** gives those older gotchas a lesson too: one classify-model call per gotcha without one (asks first with the count; `--yes` skips the prompt), then the usual diff. Notes are not re-distilled or rewritten.
- **`vir sync-claude` no longer passes terminal escapes through.** Lines read back from an existing VIR block are cleaned before the diff prints them, so a CLAUDE.md you didn&apos;t write can&apos;t recolor or hide text in your terminal.</content:encoded></item><item><title>vir 0.27.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>- **Haiku 5.5 is the default Haiku.** `models.classify`, the `haiku` shorthand (`--force-model haiku`, `--model haiku`), new installs&apos; `distillFast` and the `vir init` pick are now `claude-haiku-5-5` ($0.10/$0.50 per MTok vs $1/$5). Configs that name Haiku 4.5 keep it. Haiku 5.5 thinks by default, which would eat vir&apos;s small output caps (a retitle gets 40 tokens), so API calls send `thinking: disa</description><pubDate>Wed, 14 Oct 2026 12:00:00 GMT</pubDate><content:encoded>- **Haiku 5.5 is the default Haiku.** `models.classify`, the `haiku` shorthand (`--force-model haiku`, `--model haiku`), new installs&apos; `distillFast` and the `vir init` pick are now `claude-haiku-5-5` ($0.10/$0.50 per MTok vs $1/$5). Configs that name Haiku 4.5 keep it. Haiku 5.5 thinks by default, which would eat vir&apos;s small output caps (a retitle gets 40 tokens), so API calls send `thinking: disabled` and `claude-cli` asks for `--effort low`. A refusal now fails the call with a clear error instead of an empty answer. Kie keeps serving Haiku 4.5.
- **`claude-cli` calls no longer think, and no longer see your Claude Code setup.** `claude -p` ran Haiku with extended thinking (3812 thinking tokens for a one-sentence answer, 30–75s per classify or retitle, on your quota) and loaded your global CLAUDE.md, skills, plugins and hooks into every prompt, which could leak into notes. Calls now set `MAX_THINKING_TOKENS=0` and pass `--safe-mode`. A one-sentence lesson dropped from 36.9s to 1.3s. Opus, Sonnet and Haiku 5.5 can&apos;t turn thinking off, so they still think.</content:encoded></item><item><title>vir 0.26.2</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>- **Codex sessions show a real short id** in `vir run --dry-run`, `vir cost`, `vir reconcile` and the logs (`TRAIN/d8e621fc`, not `TRAIN/rollout-`). Every short id now comes from the same helper as note filenames; a test keeps it that way. - **A Codex-only `vir init` no longer offers to register with Claude Code.** It prints the Codex config snippet instead (`codex mcp add vir -- vir mcp`). With b</description><pubDate>Wed, 07 Oct 2026 12:00:00 GMT</pubDate><content:encoded>- **Codex sessions show a real short id** in `vir run --dry-run`, `vir cost`, `vir reconcile` and the logs (`TRAIN/d8e621fc`, not `TRAIN/rollout-`). Every short id now comes from the same helper as note filenames; a test keeps it that way.
- **A Codex-only `vir init` no longer offers to register with Claude Code.** It prints the Codex config snippet instead (`codex mcp add vir -- vir mcp`). With both agents ticked you get the Claude Code offer and the snippet.</content:encoded></item><item><title>vir 0.26.1</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>- **Re-running `vir init` offers Codex once it&apos;s installed.** On a Claude-only install the wizard now pre-checks Codex too when `~/.codex/sessions` exists, as the README says, instead of only the agents already configured. - **A Codex desktop chat started in `~/Documents/Codex` itself** goes to `codex-scratch` like the dated chats under it, instead of a project named `Codex`. - **`vir doctor` no l</description><pubDate>Wed, 07 Oct 2026 12:00:00 GMT</pubDate><content:encoded>- **Re-running `vir init` offers Codex once it&apos;s installed.** On a Claude-only install the wizard now pre-checks Codex too when `~/.codex/sessions` exists, as the README says, instead of only the agents already configured.
- **A Codex desktop chat started in `~/Documents/Codex` itself** goes to `codex-scratch` like the dated chats under it, instead of a project named `Codex`.
- **`vir doctor` no longer says &quot;codex CLI not found&quot; when `codex login status` hangs or crashes**; it reports the actual error. Only a missing binary reads as not installed.</content:encoded></item><item><title>vir 0.26.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**vir reads Codex sessions too, and can distill them on your ChatGPT plan with no API key.**</description><pubDate>Wed, 07 Oct 2026 12:00:00 GMT</pubDate><content:encoded>**vir reads Codex sessions too, and can distill them on your ChatGPT plan with no API key.**

- **Codex sessions.** Set `codexSessionsDir` (or re-run `vir init`, which now asks which coding agents you use) and vir reads `~/.codex/sessions/` alongside Claude Code. Harness context Codex stores as user messages (environment, AGENTS.md, IDE and file wrappers, in-app browser state, agent-history replays) is stripped; only your request survives. Subagent and review threads are skipped like Claude Code sidechains, `codex exec` runs are skipped like SDK agents, and desktop chats outside a repo group as one `codex-scratch` project. A project used from both agents is one project. New Codex projects start undecided, so nothing is spent until you include them. Archived Codex threads are not read.
- **`provider: &quot;codex-cli&quot;` (experimental).** Distills through `codex exec` on your ChatGPT login: keyless, quota instead of dollars, with the same per-run cap and halt-on-limit as `claude-cli`. Calls run read-only and ephemeral, with Codex&apos;s shell, web search, apps, plugins and subagents switched off. Codex picks the model unless you pin one. `vir doctor` checks `codex login status`.
- **`vir sync-claude` also updates AGENTS.md** files that already exist (`~/.codex/AGENTS.md` and project roots, found the same way as CLAUDE.md) with the same VIR block. Rules still get their one y/n against CLAUDE.md; AGENTS.md carries the rules you approved. `--no-agents` leaves AGENTS.md alone, `--agents-only` updates only AGENTS.md (approved rules, no prompts). A missing AGENTS.md is never created or listed.
- **`vir mcp install --target codex`** prints the `~/.codex/config.toml` block (or the `codex mcp add vir -- vir mcp` one-liner); vir never edits Codex&apos;s config itself.
- **Security: `claude-cli` distill calls run with no tools.** `claude -p` loaded your tools, MCP servers and permission allowlist, while the prompt carries transcript text vir doesn&apos;t control. Calls now pass `--tools &quot;&quot;` and `--strict-mcp-config`: no built-in tools, no MCP servers, no claude.ai connectors.
- **Prompts.** Codex sessions are described as &quot;Codex session&quot; in the classify and distill prompts; Claude Code prompts are unchanged byte for byte.
- **Doctor and projects.** One sessions check per configured agent; `vir projects` shows a sources column once a second agent is configured; the &quot;prunes at ~30 days&quot; warning only appears when a pending session can actually be pruned.
- `vir connect` says &quot;1 candidate&quot;, not &quot;1 candidates&quot;.
- README: a short demo of the connect → review → sync-claude flow.</content:encoded></item><item><title>vir 0.25.3</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**Small fixes from the first real `vir connect` run.**</description><pubDate>Tue, 06 Oct 2026 12:00:00 GMT</pubDate><content:encoded>**Small fixes from the first real `vir connect` run.**

- **`vir sync-claude` asks again on a typo** at the per-rule `y / n / s` and the final `y / n` prompts, instead of treating it as skip or abort. In a terminal it re-asks up to 5 times; without one it asks once, as before.
- **No rule prompt for a CLAUDE.md that doesn&apos;t exist.** Such a rule could never be written and was re-asked every run; `sync-claude` now says once that it&apos;s waiting for that file.
- **Sources from merged duplicates are marked `merged`** in the rule hunk, the review screen and the rule file, so the same `[[note]]` appearing twice is explained.
- **`vir connect --dry-run` writes nothing,** not even the lesson-embedding cache, and its cost estimate uses the real output cap (1200 tokens), so &quot;up to $X&quot; is a true upper bound.
- **MCP `vir_query` keeps project-scoped rules under a `project` filter;** rejected-rule matching compares vectors from one embedding model only; rule files say which lines an edit keeps.</content:encoded></item><item><title>vir 0.25.2</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>- **Ships the 0.25.1 fix for real.** The 0.25.1 package was built a moment before the fix was merged, so it carried 0.25.0 code: `vir sync-claude` still skipped project folders like `pripremi.rs`. 0.25.2 is the same source as 0.25.1, built correctly.</description><pubDate>Tue, 06 Oct 2026 12:00:00 GMT</pubDate><content:encoded>- **Ships the 0.25.1 fix for real.** The 0.25.1 package was built a moment before the fix was merged, so it carried 0.25.0 code: `vir sync-claude` still skipped project folders like `pripremi.rs`. 0.25.2 is the same source as 0.25.1, built correctly.</content:encoded></item><item><title>vir 0.25.1</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>- **`vir sync-claude` finds projects whose folder name isn&apos;t already a slug.** A project folder like `pripremi.rs` or `My App` never matched its own project slug (`pripremi-rs`, `my-app`), so its CLAUDE.md was reported as missing and skipped — it never got a VIR block, and a connect-pass rule scoped to it could not be promoted. Folders under `~/projects`, `~/code` and `~/dev` whose kebab-cased nam</description><pubDate>Tue, 06 Oct 2026 12:00:00 GMT</pubDate><content:encoded>- **`vir sync-claude` finds projects whose folder name isn&apos;t already a slug.** A project folder like `pripremi.rs` or `My App` never matched its own project slug (`pripremi-rs`, `my-app`), so its CLAUDE.md was reported as missing and skipped — it never got a VIR block, and a connect-pass rule scoped to it could not be promoted. Folders under `~/projects`, `~/code` and `~/dev` whose kebab-cased name equals the slug are now matched, after every exact match, so projects that already resolved are unaffected.</content:encoded></item><item><title>vir 0.25.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**vir notices lessons you keep re-learning and proposes them as rules — cited, reviewed, and added to CLAUDE.md only with your yes. Plus a skill so any coding agent can use your notes.**</description><pubDate>Tue, 06 Oct 2026 12:00:00 GMT</pubDate><content:encoded>**vir notices lessons you keep re-learning and proposes them as rules — cited, reviewed, and added to CLAUDE.md only with your yes. Plus a skill so any coding agent can use your notes.**

- **`vir connect`: recurring lessons become proposed rules.** It finds a lesson you&apos;ve recorded in at least 3 sessions spanning at least 7 days (clustering individual &quot;What Was Learned&quot; items by embedding), and asks the model to state it as one rule with a verbatim quote from each cited note. Invented quotes are discarded. Rules are written to `insights/rules/` as *proposed* and stay out of `vir query`/MCP until you accept them with `vir review --insights`. Accepted rules reach CLAUDE.md only via `vir sync-claude`, one y/n per rule, never under `--force`, `--dry-run` or a non-interactive shell. Rejections are remembered by session, so they survive note rewrites. `vir connect --dry-run` is free; a real run makes at most `connectMaxCandidates` (default 10) LLM calls and asks first. Requires an embedding provider. Thresholds are calibrated on a real 1,292-lesson vault (`docs/connect-pass-eval-2026-10.md`): precise over plentiful, so expect few proposals until your vault has real repeats. `vir status` shows proposed/accepted/awaiting counts. MCP `vir_query` gains `type: &quot;insight&quot;`.
- **Use vir from any agent.** New Agent Skills skill (`npx skills add djolex999/vir`) teaches Claude Code, Codex, Cursor and other skill-aware agents to check your vir notes before they work. It uses the MCP server when registered (with `synthesize: false`, so no LLM cost), else `vir query --json`, else the notes folder directly. It is read-only and never runs paid commands unless asked. A test pins every command, flag and MCP tool it names to the real CLI.</content:encoded></item><item><title>vir 0.24.3</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**Obsidian is optional. vir writes a folder of markdown, and Obsidian is one good way to read it.**</description><pubDate>Tue, 06 Oct 2026 12:00:00 GMT</pubDate><content:encoded>**Obsidian is optional. vir writes a folder of markdown, and Obsidian is one good way to read it.**

- **`vir init` asks for a notes folder, not an Obsidian vault.** If the usual vault (`~/Documents/Obsidian/MyVault`) exists it is still the default; otherwise vir suggests `~/notes` and offers to create it. Notes land in `&lt;folder&gt;/vir/` as before. Existing configs are unchanged.
- **The articles folder defaults to `raw/` inside your notes folder** instead of a hard-coded Obsidian path, and the prompt no longer assumes Obsidian Web Clipper.
- **README: new &quot;Without Obsidian&quot; section.** Notes are plain markdown with YAML frontmatter; `vir query` and the MCP server need no editor, and VS Code (Foam / Markdown Memo) resolves the `[[wikilinks]]`.
- **Internal:** transcript discovery, filtering and parsing now go through a session-source layer (Claude Code is the first source), groundwork for supporting more coding agents. No behavior change: `vir run --dry-run`, `vir projects` and `vir doctor` output is identical before and after.</content:encoded></item><item><title>vir 0.24.2</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**`vir reconcile` follows the same rules as `vir run`, and cost prompts show real numbers.**</description><pubDate>Mon, 05 Oct 2026 12:00:00 GMT</pubDate><content:encoded>**`vir reconcile` follows the same rules as `vir run`, and cost prompts show
real numbers.**

- **Reconcile no longer retries what `vir run` skips.** Rows that errored before the transcript and project filters existed (subagent and workflow transcripts, SDK-agent transcripts, sessions from excluded or undecided projects) were still retried, at full cost. Reconcile now applies the same filters and records those rows with their skip reason; they come back if you change the config.
- **Reconcile tells you when a transcript is gone.** Claude Code deletes old transcripts; those rows can&apos;t be retried and are now listed as &quot;transcript gone&quot; instead of &quot;recoverable&quot;. On claude-cli the retry cost reads &quot;quota&quot; instead of &quot;$0&quot;.
- **A failed reconcile retry counts toward the 3-attempt limit,** like a failed `vir run`, so a session that keeps failing is parked instead of retried on every pass (`--force` still retries it).
- **The `vir run` cost prompt shows a real estimate** (&quot;up to $X&quot;, &quot;subscription quota&quot;, or &quot;unknown&quot; for an unpriced model) instead of a fixed &quot;$1–5&quot;.
- **`$EDITOR` with arguments works in `vir review`** (e.g. `code --wait`).
- **`vir run --full` also re-processes articles and PDFs.**
- **Very large transcripts are trimmed to fit the model&apos;s context** (the start and the end are kept, and the trim is logged), instead of failing every attempt.
- **Linux home paths are scrubbed** like macOS ones before anything is stored.</content:encoded></item><item><title>vir 0.24.1</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**Four small fixes: an outage no longer crashes the run, and cost and config handling get stricter.**</description><pubDate>Mon, 05 Oct 2026 12:00:00 GMT</pubDate><content:encoded>**Four small fixes: an outage no longer crashes the run, and cost and config
handling get stricter.**

- **A provider outage no longer crashes `vir run`.** When the preflight probe failed (provider down, expired login), vir reported the outage and then Node killed the process with an extra stack trace in `daemon.log`. It now exits normally with the outage reported once.
- **`vir sync-claude` exits 1 when a `CLAUDE.md` could not be updated**, instead of printing ✗ and exiting 0.
- **`config.json` is never readable by others, even briefly.** It holds your API keys; it is now written to an owner-only temp file and renamed into place, instead of being written with the default permissions and tightened afterwards.
- **Spend on a model with no known price is reported as unknown, not $0.** With a model vir has no price for on your provider (for example `claude-sonnet-5` on Kie), calls were logged as free and `vir cost` under-reported. They are now counted as &quot;unpriced&quot; calls next to the total, cost prompts say &quot;unknown&quot;, and `vir doctor` warns about configured models without a price. Add a price under `pricing` in `config.json` to have them counted.</content:encoded></item><item><title>vir 0.24.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**Cost, safety and consistency fixes across the CLI, plus two refactors.** Nothing here changes what a normal `vir run` produces; it changes what vir spends, what it can race with, and how much of the code is checked.</description><pubDate>Mon, 05 Oct 2026 12:00:00 GMT</pubDate><content:encoded>**Cost, safety and consistency fixes across the CLI, plus two refactors.**
Nothing here changes what a normal `vir run` produces; it changes what vir
spends, what it can race with, and how much of the code is checked.

- **`vir_query` can skip the paid answer.** Pass `synthesize: false` to get only the matching notes. The tool description now tells the agent that synthesis is billed. Default unchanged.
- **New notes are embedded once, from the same text as every other note.** The vector computed while writing a new note used to be dropped (its database row did not exist yet) and the end-of-run sweep embedded the note again from the body only. The write-time vector is now stored, and the sweep and `vir embed` read the note file so every vector is built from the same text. Existing vectors are untouched until a rewrite or `vir embed --force`.
- **`provider: claude-cli` works under the daemon&apos;s minimal PATH.** vir now finds `claude` on PATH or in the usual install locations (next to the Node running vir, `~/.claude/local`, `~/.local/bin`, Homebrew, `/usr/local`), so scheduled runs no longer fail when `claude` lives outside launchd&apos;s PATH.
- **No more racing the daemon.** Terminal `vir review`, `dedupe`, `audit` (judging), `compose` and `summarize` now take the pipeline lock and exit with a clear message while another vir process holds it. Dry runs are unaffected.
- **Paid checks ask first.** `vir dedupe` (up to 30 calls) and the `vir lint` contradiction check (up to 20) now confirm before spending; `--yes` skips the prompt. Without a terminal and without `--yes`, the paid step is skipped with a note.
- **`CLAUDE.md` is written atomically.** `sync-claude` writes a temp file and renames it into place, so a crash can no longer truncate your file, and a symlinked `CLAUDE.md` stays a symlink.
- **Rejecting never overwrites an earlier rejected copy** of the same note in `.rejected/`; vir says so instead.
- **Under the hood:** tests and `eval/` are type-checked in CI (`npm run typecheck:all`); `vir run` and `vir reconcile` share one session-distill core; the large commands moved out of `cli.ts` into their own modules.</content:encoded></item><item><title>vir 0.23.3</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**A skipped re-run no longer hides a note you already had.** When a session with a distilled note came back through `vir run` (a resumed transcript, or `--full`) and that pass was skipped by the heuristic filter or by low classify confidence, the note was marked skipped. Its file stayed in the vault, but it dropped out of `vir query`, the MCP server, `sync-claude` and rewrites, and never came back</description><pubDate>Mon, 05 Oct 2026 12:00:00 GMT</pubDate><content:encoded>**A skipped re-run no longer hides a note you already had.** When a session
with a distilled note came back through `vir run` (a resumed transcript, or
`--full`) and that pass was skipped by the heuristic filter or by low classify
confidence, the note was marked skipped. Its file stayed in the vault, but it
dropped out of `vir query`, the MCP server, `sync-claude` and rewrites, and
never came back on its own. `vir reconcile` could do the same.

- **The existing note now stays served.** The new transcript is recorded as seen, so a low-confidence session is not re-classified (and billed) on every run. A note hidden by a failed re-distill is served again.
- **Notes already hidden this way are restored automatically** the first time 0.23.3 opens the database. Only rows that this bug can produce are touched.</content:encoded></item><item><title>vir 0.23.2</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**Two vir processes can no longer both hold the pipeline lock.** The lock file (`~/.vir/vir.lock`) was checked and then written in two steps, so a daemon run and a manual `vir run` starting at the same moment could both take it and distill the same sessions twice.</description><pubDate>Mon, 05 Oct 2026 12:00:00 GMT</pubDate><content:encoded>**Two vir processes can no longer both hold the pipeline lock.** The lock
file (`~/.vir/vir.lock`) was checked and then written in two steps, so a
daemon run and a manual `vir run` starting at the same moment could both take
it and distill the same sessions twice.

- **The lock is created atomically.** Exactly one process creates it; the others report that another vir process is running, as before.
- **Clearing a crashed run&apos;s leftover lock is guarded too.** Only one process at a time may remove a stale lock (via a short-lived `~/.vir/vir.lock.reclaim`), so a fresh lock is never deleted by mistake. Tested with 8 processes racing on a fresh and on a stale lock: exactly one winner every time.
- **A stale lock that cannot be removed** (for example, a permissions problem) now fails with a clear error instead of being silently overwritten.</content:encoded></item><item><title>vir 0.23.1</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**A claude-cli quota halt no longer loses articles and PDFs.** When `vir run` hit the subscription limit, the session loop stopped, but the article and PDF phases still ran into the same wall and recorded every new item as processed with an error. Those items were never tried again.</description><pubDate>Mon, 05 Oct 2026 12:00:00 GMT</pubDate><content:encoded>**A claude-cli quota halt no longer loses articles and PDFs.** When `vir run`
hit the subscription limit, the session loop stopped, but the article and PDF
phases still ran into the same wall and recorded every new item as processed
with an error. Those items were never tried again.

- **A quota halt now stops the whole run.** Project summaries, articles and PDFs are skipped after a halt, and a limit first hit inside the article or PDF phase halts there. Nothing is recorded, so the skipped items are simply new on the next run.
- **Items lost this way in earlier versions are retried automatically on the next run.** An errored article or PDF no longer counts as processed, so no manual step is needed.
- **Failing items stop retrying after 3 attempts,** the same bound sessions have. A changed file starts the count over. Articles and PDFs gain an `attempts` column (additive migration).</content:encoded></item><item><title>vir 0.23.0</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**`vir review --json`: review from the Obsidian plugin.** The terminal review loop now has a non-interactive twin with a stable JSON contract, so the Obsidian plugin can list the queue and approve, reject or restore one note without spawning a TTY. The interactive loop is unchanged.</description><pubDate>Tue, 29 Sep 2026 12:00:00 GMT</pubDate><content:encoded>**`vir review --json`: review from the Obsidian plugin.** The terminal review
loop now has a non-interactive twin with a stable JSON contract, so the
Obsidian plugin can list the queue and approve, reject or restore one note
without spawning a TTY. The interactive loop is unchanged.

- **Queue mode.** `vir review --audited --json` prints the notes `vir audit` flagged, worst first, as `{&quot;items&quot;:[...],&quot;counts&quot;:{&quot;unaudited&quot;:n,&quot;stale&quot;:n}}` (`--project` filters). Each item carries the note path, session id, title, category, project, verdict and reason; `counts` says how many notes have no audit yet or a stale one.
- **Actions.** `--approve`, `--reject` and `--restore` with `--json` act on exactly one note and print the result as JSON. Use the one-token `--flag=&lt;path&gt;` form so a path starting with `-` is never read as a flag.
- **Reject stops serving immediately.** `--reject` moves the note into `.rejected/` and calls `markRejected`, so `vir query` no longer returns it before the next run.
- **Path guard.** Actions accept only paths inside the category directories (and `.rejected/` for restore), and only session notes; anything else, including `../` escapes, fails `invalid_args` before touching a file.
- **No racing a run.** The three actions take the pipeline lock and fail with `busy` while a lock-holding run (a normal `vir run`, `vir reconcile`, `vir audit --apply-rejects`) is in progress; `vir run --rewrite-only` and `--dry-run` take no lock.
- **Two new error kinds:** `busy` and `not_found`, alongside the existing ones. Errors follow the `vir query --json` contract: empty stdout, one-line error payload on stderr, exit 1.</content:encoded></item><item><title>vir 0.22.1</title><link>https://virwiki.dev/docs/changelog/</link><guid isPermaLink="true">https://virwiki.dev/docs/changelog/</guid><description>**`vir audit --apply-rejects` asks only about the notes it will move, and the audit&apos;s rejects were measured before being applied.** The one user-facing change is the preview fix; the rest is the note-usefulness eval under `eval/`, which never ships, and a byte-identical extraction of the query-synthesis prompt (`buildSynthesisPrompt`) that the eval keys its cache on.</description><pubDate>Mon, 28 Sep 2026 12:00:00 GMT</pubDate><content:encoded>**`vir audit --apply-rejects` asks only about the notes it will move, and the
audit&apos;s rejects were measured before being applied.** The one user-facing
change is the preview fix; the rest is the note-usefulness eval under `eval/`,
which never ships, and a byte-identical extraction of the query-synthesis
prompt (`buildSynthesisPrompt`) that the eval keys its cache on.

- **`vir audit --apply-rejects` previews only what it will move.** The confirmation listed and counted every fresh reject, including notes a human had approved, which the move step then skipped (&quot;move 18 notes?&quot; moved 17). The preview now uses the same approved-note check and shows approved notes as &quot;approved, skipped&quot;. The move still re-checks at act time.
- **Note-usefulness eval, second verdict: FAIL on noise, rejects applied on inspection.** A hand check of the first run found one of the 18 rejects was genuinely load-bearing: `208035d9`, the only note on pripremi.rs&apos;s legacy email-flag `$unset` with `strict: false`, which also had a title that did not match its body. It was approved and retitled &quot;strict-false-unset-migration&quot; (title from the production title prompt). The eval&apos;s reject set now skips human-approved notes, matching exactly what `--apply-rejects` moves; it had been testing a removal that could never happen. Re-run on the remaining 17 rejects (34 exposed questions): 32 answers unchanged; Δrecall −0.015, CI [−0.037, 0.000], passes (&gt; −0.05); Δcontradiction +0.015, CI [0.000, +0.037], fails (≤ +0.02). Every worse answer in both runs was checked by hand, and none traced to a fact held by a removed note: each came from an unrelated note swapping into the top 8 or from the answer being worded differently. At this sample size answer-wording noise alone can decide the gate, so the 17 rejects were applied on that inspection rather than on a PASS. They are reversible with `vir review --restore`.
- **`usefulness run --skip-none`** drops the report-only no-notes arm, which never feeds the gate: about a third fewer answer and grade calls. The run record carries `noneArm: false`, and the summary prints &quot;full − none recall skipped&quot;.
- **Note-usefulness eval (`eval/usefulness/`, never shipped): the first measured verdict on `vir audit --apply-rejects` is FAIL, narrowly.** `npm run eval -- usefulness mine|run|show` mines real later-session questions and the facts that session established (82 questions from 71 human sessions), answers each with production retrieval and `synthesize()` against the full vault, the vault with the audit&apos;s 18 fresh rejects removed in a DB copy, and no notes, then grades every fact as stated, missing or contradicted. The grader passed all four probes (oracle recall 1.00, null clean 1.00, negation 1.00, re-grade agreement 0.99). On the 38 exposed questions, removing the rejects left 34 answers unchanged, improved 2 and hurt 2: Δrecall −0.007, 95% CI [−0.053, +0.039], against a bound of &gt; −0.05; Δcontradiction +0.013, CI [0.000, +0.033], against a bound of ≤ +0.02. Both bounds miss, so non-inferiority is not shown. The mean effect is near zero and 38 questions leave a wide interval, so this says &quot;not proven safe&quot;, not &quot;the rejects are useful&quot;. Two rejects did carry facts a later session needed (popis snapshot/backup details, a pripremi.rs migration). Notes overall lift recall by only +0.05 over no notes, CI [+0.02, +0.09]. Run record: `~/.vir/eval/usefulness/runs/2026-09-28T13-04-40-271Z.json` (git abe6e8e, reject-set sha256 d0b4c5bc0f6c…).</content:encoded></item></channel></rss>