What's changed

Fixed

• CI/Release workflows: Swatinem/rust-cache was silently missing on every run because no shared-key was set, causing each workflow/job combination to use a separate cache namespace. Added shared-key: build-<target> to all cross-target build jobs (shared between ci.yml and release-crates.yml) and shared-key: test-ubuntu-latest to all Linux test jobs (shared between ci.yml, pr-ci.yml, and release-crates.yml publish-crate). Also parallelised publish-crate (no longer waits for build-binaries), removed brew update from macOS protoc install, and dropped the redundant cargo publish --dry-run pass.

What's changed

Fixed

• unlost ingest: Panic on non-ASCII characters (e.g. box-drawing chars ├, └, ─) caused by byte-indexing into a multi-byte UTF-8 string in storage.rs. Fixed to use char-boundary-safe truncation.
• unlost reindex: Crash on capsules with null next_steps or symbols fields. Added #[serde(default)] to those fields in JsonCapsule. Also hardened conn_id and exchange_seq against missing values in old records.

What's changed

Fixed

• Release: Re-release of 0.20.0 with correct CI-built binaries (0.20.0 tag was corrupted by a manual release attempt).

What's changed

Added

• --no-llm flag on all LLM-using commands: recall, brief, reflect, explore, challenge, and pr-comment now all accept --no-llm to skip the LLM narrative and print raw capsule data instead. Allows full use of unlost memory commands on machines without an LLM configured. query, trace, thread, and init already had this flag; this change brings the remaining commands in line.

Fixed

• unlost reindex: Tolerate capsules with null string fields (category, intent, decision, rationale, request_path, source) instead of crashing with invalid type: null, expected a string. Adds #[serde(default)] to those fields so older or partially-written records are skipped gracefully rather than aborting a full reindex mid-run.

What's changed

Fixed

• unlost config llm anthropic --sso: The login flow was waiting on a
  local TCP listener for a redirect that Anthropic never makes — the
  authorize URL pointed at console.anthropic.com/oauth/code/callback,
  which is the manual copy-code page, not a real redirect target. Switched
  to the documented copy-code flow: the page now displays an authorization
  code that the user pastes back into the terminal. Added a 120 s timeout on
  the token exchange (the endpoint can take 40–60 s during platform issues
  and the previous default would silently fail). Switched the token request
  to application/x-www-form-urlencoded per Anthropic's current spec.
• Account selection guidance: Added an explicit warning before the
  browser opens that the OAuth flow uses whatever claude.ai session is
  already active. Users with a wrong account (e.g. personal vs team) need
  to log out of claude.ai first or open the printed URL in a private /
  incognito window — there is no prompt=select_account for this PKCE
  flow.
• Regression in v0.18.1: that release was cut from a branch that had
  not yet been merged with the second-brain export feature, so the
  unlost export command was missing from the 0.18.1 binary. v0.18.2
  restores it.

Full Changelog: v0.18.1...v0.18.2

What's changed

Fixed

• unlost config llm anthropic --sso: The login flow was waiting on a
  local TCP listener for a redirect that Anthropic never makes — the
  authorize URL pointed at console.anthropic.com/oauth/code/callback,
  which is the manual copy-code page, not a real redirect target. Switched
  to the documented copy-code flow: the page now displays an authorization
  code that the user pastes back into the terminal. Added a 120 s timeout on
  the token exchange (the endpoint can take 40–60 s during platform issues
  and the previous default would silently fail). Switched the token request
  to application/x-www-form-urlencoded per Anthropic's current spec.
• Account selection guidance: Added an explicit warning before the
  browser opens that the OAuth flow uses whatever claude.ai session is
  already active. Users with a wrong account (e.g. personal vs team) need
  to log out of claude.ai first or open the printed URL in a private /
  incognito window — there is no prompt=select_account for this PKCE
  flow.

What's changed

Added

• unlost export: New command that produces a high-signal second-brain
  vault from your capsule history, consumable by Obsidian, Logseq, Hugo, or any
  markdown-aware harness. The vault has three entry points: a root
  decisions.md (chronological log of every substantive decision plus a
  failure-signal table), a symbols/<project>/ tree (one knowledge page per
  significant source file, aggregating every decision that primarily touched
  it, with Obsidian-style wikilinks to related files), and a categories/
  tree of 13 README-only summaries grouped by a fixed taxonomy (debugging,
  architecture, refactoring, testing, etc.). Multi-workspace aware: all
  registered workspaces are exported together, with per-project scoping for
  symbol pages.
• unlost config export-dir <path>: Persist a default export directory
  in ~/.config/unlost/config.json so unlost export works without a
  --dir flag on every invocation.

Changed

• Aggressive noise filtering on export: about 47% of capsules
  (check-ins, replays, empty extractions, low-signal meta/other buckets)
  are silently dropped from the vault. Substantive capsules are those with a
  meaningful decision and rationale, or any non-none failure mode. The
  long tail of single-mention symbols is also dropped (≥ 8 capsules required
  for a symbol to earn a page).
• Symbol attribution uses position-3 cutoff: the LLM lists symbols in
  rough relevance order; hub files like src/main.rs appear in 1000+
  capsules at average position 11 because they're touched as boilerplate.
  Restricting attribution to the first three positions drops src/main.rs
  from 17 k lines to 1.1 k and total symbol pages from ~660 to ~60 — all
  signal, no boilerplate.
• Symbol filter is git-tracked-set-aware: instead of a hardcoded
  blocklist of build-artefact paths, the filter uses git ls-files per
  project. A path-shaped symbol earns a page only if it canonicalises to a
  tracked file under the workspace root. Self-tunes to whatever each
  project's .gitignore excludes — no per-toolchain assumptions.

What's changed

Fixed

• unlost config llm anthropic --sso: The login flow was waiting on a
  local TCP listener for a redirect that Anthropic never makes — the
  authorize URL pointed at console.anthropic.com/oauth/code/callback,
  which is the manual copy-code page, not a real redirect target. Switched
  to the documented copy-code flow: the page now displays an authorization
  code that the user pastes back into the terminal. Added a 120 s timeout on
  the token exchange (the endpoint can take 40–60 s during platform issues
  and the previous default would silently fail). Switched the token request
  to application/x-www-form-urlencoded per Anthropic's current spec.
• Account selection guidance: Added an explicit warning before the
  browser opens that the OAuth flow uses whatever claude.ai session is
  already active. Users with a wrong account (e.g. personal vs team) need
  to log out of claude.ai first or open the printed URL in a private /
  incognito window — there is no prompt=select_account for this PKCE
  flow.

Added

• unlost config llm anthropic --sso: New SSO flag on the existing anthropic subcommand. Opens the browser, runs an OAuth 2.0 PKCE flow against console.anthropic.com, exchanges the temporary access token for a permanent sk-ant-... API key via Anthropic's create_api_key endpoint, and stores it in config — no manual key management required. --api-key and --sso are mutually exclusive.

Changed

• docs/index.html: Updated landing page to cover 0.15.0 and 0.16.0 changes.