Lore 0.5.1

Everything since v0.4.0 — the v0.5.0 feature set, which was never released, plus four fixes.

If you tried to install Lore before today, it did not work. marketplace.json was missing its required owner field from the very first commit, so /plugin marketplace add aoc81/lore failed for everyone with Invalid schema: owner: Invalid input. It went unnoticed through five versions because developing the plugin in its own repo loads it as lore@inline, which skips that validation entirely. Fixed, and guarded by tests/test_manifests.py.

Bootstrap from git history — /lore:mine

Cold start is the failure mode this whole idea dies of: empty store → recall surfaces nothing → nobody captures → still empty. /lore:mine reads the material already sitting in git — reverts, fix commits, bodies that explain a why, churn hot spots — scores it (English and Spanish signals; noise, sweep and store-only commits filtered), and marks any candidate an existing entry already covers using the recall scorer, so a second pass never duplicates the first. The threshold sits deliberately above "fix keyword + focused change": a bare Fix X with no body is the most common commit in any repo, and showing all of them buries the rest.

Personal store

The capture gate catches user preferences, which until now landed in the committed, team-shared — effectively published — store. personalStoreDir is the private half: recall reads it alongside the team store and flags it [personal], but it is never indexed, linted, or secret-scanned, and the skill routes preference-shaped learnings there rather than quietly writing them to the shared store. A path inside the team store is refused, since it would be committed anyway, and the deleted-ref freshness flag is skipped for entries stored outside the project, whose files: may name another repo.

Team-wide secret guard

The blocking pre-push scan only ever protected the machine that ran /lore:init, so a teammate without the hook could still push a key into a published store. A workflow template runs the same scan_secrets.py and verify_refs.py --strict on every PR, from the committed .lore/ copies, so CI and the local hook cannot drift apart and nobody else has to install anything. /lore:init offers it; Codex gets --ci.

Read telemetry

The recall log recorded what was shown, never what was used. A PostToolUse(Read) hook logs store reads, and /lore:stats separates the two — "surfaced 12x, read 0x" is a misleading title, not a lesson. With no read events logged at all, the line reports "none logged yet" rather than accusing every entry of it.

Codex parity

Codex has since gained PreToolUse and SessionStart(compact) with the same additionalContext contract, so edit-time recall and the post-compaction reminder now run there too — 4 hooks registered, up from 2.

Fixes

• The secret scanner missed every prefixed secret name. The hardcoded-assignment rule wrapped its keyword in \b, and _ is a word character, so \b never fires between an underscore and a letter: db_password, my_api_key, STRIPE_SECRET and aws_secret_access_key all walked past a blocking guard. Those are the dominant real-world spellings — snake_case and SCREAMING_SNAKE in Python, YAML and .env content, which is exactly what gets pasted into a learning. client_secret hit only because it is spelled out as its own alternative, which disguised the gap. The placeholder filter still neutralises db_password = changeme, and backtracking stays linear.
• The linter was dead in a personal-store-only project. verify_refs bailed whenever the team store directory was missing, even with personalStoreDir set and populated, so /lore:lint, /lore:stats and /lore:sweep were all silently gone while recall and capture kept working. It now gives up only when both stores are absent. Related: the check mode no longer reports "all entries have valid file refs" after examining zero entries.
• marketplace.json gained its required owner (see above), plus a manifest test suite that checks the declared versions agree and every hooks.json command points at a script that actually ships.

Upgrading

If you installed from the default branch during the window when the manifests said 0.5.0, you may be running a build with the old secret scanner under that same version string. claude plugin update lore@lore picks up 0.5.1.

Lore 0.4.0

Recall now honors your config everywhere, stops repeating itself during a refactor, and works for prompts that aren't in English.

Recall correctness

• staleStatuses is respected everywhere. It only reached prompt-time ranking before; freshness flags and edit-time recall hardcoded superseded/obsolete/deprecated. A project-defined status like archived was penalized on prompt recall but treated as live guidance when you edited the file. Both paths now read the config.
• Edit-time matches are ranked before they're cut. maxRecall truncated in alphabetical path order, so a superseded entry could displace a current one. Live entries now sort first, ties keep store order.
• New staleAfterMonths (default 6) — one configurable freshness threshold, replacing the same constant duplicated in two units (STALE_AFTER_MONTHS = 6 in recall, STALE_AFTER_DAYS = 183 in the linter).

Edit-time recall dedupes per session

The PreToolUse hook fires on every Edit/Write, so a refactor touching one file ten times injected the same learning ten times. Surfaced entries are now remembered for the session in .git/lore-recall-seen.json (local, inside .git, never committed) and not re-injected. Keyed by the hook's session_id; a new session starts clean. Fails open — no session id, no .git, or corrupt state means the learning still surfaces.

Non-English prompts

• Unicode, accent-folding tokenizer. [a-z0-9_]+ split autenticación into autenticaci + n, quietly degrading recall for anyone prompting in Spanish (or any accented language). Tokens are now \w+ with diacritics folded, so an accented prompt also matches a tag written without accents (autenticación ↔ autenticacion).
• Spanish stop words. para, este, cuando, quiero, necesito and ~85 more reached the scorer and could prefix-match real vocabulary (para → parameters). Deliberately not stopped: todo/todos/algo, which are real English/domain tokens.
• New stopWords config key — words added to the built-in list, for domain noise (a product name in every prompt) or the three above.
• Docs: write tags in the language you prompt in — accents fold, languages don't.

Hardening

• Injected titles are capped at 140 characters and stripped of control characters. Recall's "never inject bodies" promise limits the prompt-injection surface of a shared store — but title: was injected verbatim and unbounded, so a malicious PR could smuggle an instruction payload through it.
• The secret scanner no longer double-reports an Anthropic key (the OpenAI sk- rule also matched it).

Performance & diagnostics

• iter_entries reads a bounded 8 KB head instead of whole entry files — it only ever needed the frontmatter, and it runs on every prompt and every edit. Constant cost per entry, with a full-read fallback so correctness never depends on the window.
• The linter names ignored .lore.json keys. load_config deliberately never fails, so a typo like maxRecal silently did nothing. /lore:lint and /lore:stats now report unknown keys with a did-you-mean suggestion, plus wrongly-typed values.

Tests

52 → 86. Closes the coverage gaps on run_pretool, the run_hook JSON envelope, --index, and recall-log rotation, and covers every change above (including a regression test that fails against the old English-only stop list).

No behavior change requiring action: every new config key defaults to the previous behavior.

Install / update: /plugin install lore@lore (auto-updates).

Lore 0.3.0

Sharper recall, a quieter capture nudge, near-duplicate triage, and a real test suite — all still hook-driven, staying in the lightweight recall/capture niche.

Recall

• Word-boundary scoring — an exact word hit scores 2, a prefix overlap scores 1, and an entry needs a score of 2 to surface. auth no longer matches author, and one weak overlap never injects noise.
• Short tags are findable — 2–3 char prompt words (ci, api, aws) now match against exact tags, where they're curated vocabulary. Previously the 4-char minimum made them impossible to hit.
• Directory & glob files: refs — a ref may be an exact path, a directory prefix (src/auth/), or a glob (src/*.py), and all three power edit-time recall, the linter, and drift triage. Path matching is case-insensitive on Windows.
• Local recall telemetry — every surfaced entry is logged to .git/lore-recall.log (local only, never committed) so /lore:stats can show which learnings actually get used and which never surface.
• New /lore:search <query> — query the same scorer the recall hook uses, on demand.

Capture

• captureNudge config (always | smart | off, default smart) — the smart mode skips the end-of-turn nudge on purely conversational turns (no tool use), so the LORE CHECK keeps its signal instead of becoming a mechanical reflex. Reads the transcript and fails open. Booleans work as shorthand (true→smart, false→off).
• Post-compaction reminder (new SessionStart hook) — right after context is compacted, a short lore reminder is re-injected — exactly the moment an uncaptured learning would otherwise die with the context. (PreCompact can't inject context, so the post-compact SessionStart is the supported channel.)
• Capture-time secret scan — the skill now scans the file it just wrote, so a leak is caught at write time, not just at the push boundary.

Health / freshness

• /lore:lint --dupes — near-duplicate triage: entry pairs sharing title/tag vocabulary (e.g. two teammates capturing the same gotcha on parallel branches) plus category-name variants (CI vs ci).
• Richer /lore:stats — now reports recall activity (most- and never-surfaced entries), the near-duplicate count, and a note when the project's .lore/ hook copies fall behind the installed plugin version.
• Faster drift triage — --report now computes drift in a single streaming git log pass instead of one subprocess per referenced file.
• .lore/VERSION stamp — /lore:init stamps the hook copies so the linter can flag stale ones.

Robustness

• _TEMPLATE.md and _/.-prefixed files are no longer counted as entries (fixes the stats count and the generated index).
• .lore.json values are type-validated — a wrongly-typed key is ignored instead of crashing a hook.

Tests

• A 52-test stdlib unittest suite (scorer, ref matching, config, transcript-based nudge gating, secret scan, git drift) and a CI workflow running it on 3 OSes × Python 3.9/3.13.

⚠️ Behavior change

The end-of-turn capture nudge now defaults to smart (skips tool-less turns). To restore the previous every-turn behavior, set "captureNudge": "always" in a project's .lore.json.

Install / update: /plugin install lore@lore (auto-updates).

Lore 0.2.0

Four new capabilities — all hook-driven, staying in the lightweight recall/capture niche.

Recall

• Freshness flags inline — every surfaced learning now shows cheap staleness signals (a deleted-file ref, or "verified N months ago") so a possibly-stale entry is never trusted blindly. The costly git-drift check stays in --report, out of the hook.
• Edit-time recall (new PreToolUse hook) — when the agent is about to edit a file, any learning whose files: names that path surfaces right then. The gotcha shows up exactly when you touch the code. Verified live in the VS Code extension.

Capture

• Deterministic overlap check — before writing, the skill runs recall.py --query to list already-related entries, so a new learning updates one instead of duplicating it.

Health / freshness

• /lore:stats — store-health snapshot: counts by status/category, drift backlog, long-unverified entries, and a soft dangling-[[link]] count.
• Richer --index — a status summary line in the generated README.
• Fixed two pre-existing --report crashes (empty git log output on untracked files; dict comparison on a gap tie).

Docs

• A "See it in action" demo in the README — an animated SVG rendered from the real hooks (demo/make_cast.py), showing recall appearing on its own and the LORE CHECK closing the turn.

Install / update: /plugin install lore@lore (auto-updates).

Lore v0.1.0

First tagged release of Lore — a self-maintaining knowledge store for Claude Code and OpenAI Codex: capture non-obvious learnings, recall the relevant ones into context, and keep them fresh.

Highlights

The loop

• Recall (UserPromptSubmit hook) surfaces relevant past learnings into context — matches title + tags only, never echoes bodies.
• Capture (Stop hook + lore skill) records a learning when one is both non-obvious and reusable.
• Freshness — /lore:lint ref-check + drift triage + index regen, with an optional pre-push hook.

Security hardening (this release)

• Blocking secret scan at the push boundary — the store is committed/pushed by default and capture is autonomous, so scan_secrets.py runs in the pre-push hook and aborts the push if a learning contains a likely key, token, private key, JWT, credential URL, or hardcoded secret. Redacted output, placeholder filtering, and allowlisting (lore:allow-secret, secretAllow in .lore.json, LORE_SCAN_BLOCK=0, --no-verify).
• Capture-time prevention — the skill now references secrets instead of quoting them.
• New /lore:scan command to run the guard on demand.
• README Security section documenting the posture, the recall anti-injection design, and pinning the plugin to a reviewed tag.

Install

Claude Code:

/plugin marketplace add aoc81/lore
/plugin install lore@lore

Then /lore:init per project. OpenAI Codex: python3 codex/install.py.

Everything is stdlib Python — no dependencies, no network calls.