CHANGELOG.md

Changelog

2026-08-17

markdown_dir gained exclude_patterns, and BACH's own knowledge base is now observed alongside its existing bach-facts/bach-lessons/bach-working memory-table sources.

• exclude_patterns on markdown_dir/remember_files drops filenames matching an fnmatch pattern back out of what patterns/glob already matched -- e.g. a generated help directory that ships one canonical language plus several machine-translated siblings per key, where patterns alone cannot express "*.txt but not *_en.txt" (fnmatch's [!seq] only excludes a single character, not a suffix).
• Six new BACH observe-sources: bach-wiki and bach-skills (sqlite_table against wiki_articles/skills in ~/.bach/bach.db, read-only), bach-root-docs/bach-system-docs/bach-docs (markdown_dir over BACH's README/architecture/changelog/roadmap files), and bach-help-de (BACH's generated per-command help directory, German only via the new exclude_patterns). BACH itself is never written to.

2026-08-16 [0.4.0]

• Discoverability, README-Design, Badges & Metadata Parity:
  • Synchronized badges in README.md & README_de.md across Test Suite (108 passed, 100% green), Version (0.4.0), Python (>=3.10), License (MIT), ellmos-ai Ecosystem, open-bricks Umbrella, and llms.txt Discovery.
  • Added interactive bilingual Mermaid architecture diagrams detailing the gardener CLI/Python API/Search GUI layers, Core Engine (FTS5 BM25 search, Materialize & Run execution engine, federated observe engine with secret redaction and cloud-alert gate), Dual-Database Substrate (gardener.db system / user.db user space), and federated observe sources (markdown_dir, remember_files, sqlite_table, agent_transcripts).
  • Added complete sibling tools matrix linking ellmos-ai, dev-bricks, and open-bricks ecosystem modules (ellmos-core, clutch, BACH, USMC, Rinnsal, ellmos-controlcenter-mcp, ellmos-filecommander-mcp, ellmos-codecommander-mcp, ellmos-clatcher-mcp, n8n-manager-mcp, skills, DevCenter, open-bricks).
  • Implemented automated metadata and discoverability parity test suite in tests/test_metadata.py (5/5 assertions passed).
  • Updated llms.txt Last-checked header to 2026-08-16 and test count to 108 passing tests.

2026-08-13

find() can be restricted to one observe-source, so a small source is no longer buried by a large one.

• --source <id>[,<id>] / find(source=...) filters on the observed/<source-id>/… namespace as a WHERE condition -- before the ranking, and in all three stages of find(): exact FTS, the multi-word OR fallback, and the LIKE fallback. Wiring it into only the first stage would have dropped the filter silently on any multi-word query.
  • Why it was needed: source sizes differ by three orders of magnitude. On the reporting machine codex-sessions holds 260,623 transcript lines against 518 in usmc-working, out of 290,902 observed entries total. BM25 hands the entire first page to the transcripts, so a subject-matter search returns nothing usable and reads as "not in the index".
  • A source id matches a whole path segment, so --source usmc does not pull in usmc-working. A leading observed/ is optional, several ids are OR-ed, and _/% in an id are escaped instead of acting as LIKE wildcards.
• --source without a query lists the source, newest first. FTS5 needs a term to match; "show me everything from this source" has none, so that case takes a plain WHERE/ORDER BY updated path instead of returning nothing.
• --type and --limit are now reachable from the CLI. find() already accepted both; only the command line did not pass them through.
• The find command parses its own flags. The CLI has no argparse and joined everything after find into the query, so --source would have become a search word. _parse_find_args() splits options from search terms, accepts --flag value and --flag=value, and reports unknown options and missing values instead of silently searching for them.
• Documented in README.md and README_de.md, including the pre-existing workaround for older versions (pass the source id as a search word -- entry names are in the full-text index) and the boundary against extra_tags, which labels a source at registration time rather than narrowing one search.
• Test suite grew from 86 to 108 tests.

Read-only change: no schema migration, no new index, recall() untouched, and existing find() callers keep their behaviour (the new parameter is appended and defaults to None). That includes the edge case find("") without a source: it still falls through to the LIKE stage, where %% matches everything and the call acts as a browse. Only the combination of an empty query and a source takes the new listing path -- asserted by a test, because guarding the three search stages on a non-empty query would have turned that browse into an empty list without anyone noticing.

2026-08-05

Fixed: the never-index list missed Windows paths on non-Windows hosts.

• sources.is_excluded() split its argument with the host's own path rules, so on Linux and macOS a raw C:\_Local_DEV\CREDENTIALS\x.md arrived as a single segment and never matched credentials. Backslashes now count as separators everywhere -- the block list fails closed on any host. The CI runs on Linux and had been red on exactly these two cases since 2026-08-02.

2026-08-02 (later)

Secrets are now redacted on the way into the index, a credential found in a cloud-synced document raises an alert, and archived transcripts are read straight out of their zip.

• Secret redaction (sources.redact_secrets), applied in scan() -- the one gate every adapter's items pass through, so a future adapter cannot forget it. The pattern family stays readable (ghp_***REDACTED***), the value does not survive. Deliberate semantics: an agent that needs the real token must go to the source file. The index says where a credential lives, never what it is.
  • 13 families, following the documented formats used by GitHub secret scanning, gitleaks and Yelp detect-secrets: Anthropic, OpenAI (legacy and project/service/admin keys), GitHub PAT classic + fine-grained, AWS access key ids, Slack bot/user/app tokens, Google API keys, GitLab PATs, npm tokens, Authorization: Bearer headers, and PEM private-key blocks.
  • Every pattern anchors on a fixed length, a restricted character class and -- where the vendor provides one -- a literal marker (T3BlbkFJ, the trailing AA on Anthropic keys). That anchoring, not the prefix, is what keeps prose out: skalar, ghpx_…, AKIAA, AIzaX, npm_install and a bare "Bearer" in a sentence are all left alone (asserted).
  • AWS bodies match [A-Z0-9]{16}, not gitleaks' base32 [A-Z2-7]{16}: the narrower class would let a real key containing 0/1/8/9 through, and for a redaction step missing a live secret is the worse error.
  • Deliberately not included: entropy heuristics and keyword detectors (password=). Both are documented high-recall/low-precision and would black out hashes, UUIDs and ordinary config prose. A step that runs unattended must not guess.
  • Fingerprints stay computed over the original text -- they answer "has the source changed", and the source is the unredacted file. Rewriting them would invalidate every stored fingerprint and force a full re-index.
• Cloud credential alert. A signature found in a file under ~/OneDrive (override: GARDENER_CLOUD_ROOT) is a security finding in its own right -- the value has left the machine. One line per finding is appended to SECURITY-ALERT_TOKEN-IN-ONEDRIVE.md (GARDENER_CLOUD_ALERT_FILE) with date, source path and family -- never the value and never the surrounding text, because the alert file lives in the very folder it warns about. Idempotent: a finding already listed is not appended again. A signature in a local transcript (~/.codex, ~/.claude) raises no alert -- it never left. Findings are also reported in observe_sources()' stats and warned about on stderr.
• Zip archives as a transcript source. agent_transcripts reads JSONL members straight out of a .zip via zipfile; nothing is unpacked to disk. zip_inner selects members (default *.jsonl). Incrementality is per archive rather than per byte offset -- an archive is a finished thing, so an unchanged (mtime, size) skips the whole file unopened. Archives holding no matching member simply yield nothing.
• Refactor: the per-line work of scan_agent_transcripts moved into _transcript_item(), shared by the plain-file and zip paths so the two cannot drift apart in how they extract, name and cite a turn.
• Tests: +11 (redaction positives per family, look-alike negatives, prefix-stays-readable, redaction reaching the index through both a markdown and a transcript source, alert written/idempotent, local path raising no alert, zip indexing/incremental-skip/no-matching-member, and redaction inside archives). 74 -> 85.

Measured on this machine: retroactive sweep of the existing 284232 entries found 6 real credentials (4 npm tokens, 2 AWS key ids), all in local Codex transcripts, all pasted into sessions by hand -- context like //registry.npmjs.org/:_authToken=npm_… leaves no doubt they were genuine. All 6 redacted in place; 0 unredacted matches remain. Cloud alert initial sweep: 0 findings among indexed OneDrive documents. gemini-archive (the Antigravity conversation archives) indexed 3805 entries from 102 transcript.jsonl members in 4.3 s; second run 0.2 s.

2026-08-02

Every agent provider on the machine is now in one search -- and the three transcript presets added yesterday are corrected against the formats they actually claim to read. All three were written from assumption, not from the files; measured against real transcripts, two indexed nothing at all and one indexed mostly noise.

• codex preset rewritten. A Codex rollout carries the same conversation twice: event_msg (payload.type 'user_message' / 'agent_message', text in a flat payload.message) and response_item (the raw model exchange). The old preset read only the second one -- duplicating every assistant turn verbatim, pulling injected AGENTS.md/skill boilerplate in as "user" text, and missing the clean channel entirely. Now only event_msg is indexed.
• codex preset also drops sub-agent tool traffic. When Codex delegates, it wraps the sub-agent's tool calls and their output into ordinary agent_message events prefixed [external_agent_tool_call: Read] / [external_agent_tool_result]. The payload is a verbatim file dump, command stderr or diff -- prose everywhere else in this module is what gets indexed, and this is not it. Measured on a real archive: 49,286 of 309,883 indexed Codex turns (15.9%) were tool traffic.
• kimi preset rewritten. It looked for a TurnBegin wrapper and flat top-level role/content strings; neither exists in a real wire.jsonl, so the preset returned nothing for every line of every file. Kimi's wire log is an event stream: agent prose arrives as context.append_loop_event -> content.part (part.type='text', while 'think' parts are internal reasoning), user turns as context.append_message with message.origin.kind='user'. That origin check matters -- roughly two thirds of user-role messages are injected reminders, cron firings and hook results.
• gemini_antigravity preset narrowed. It treated any source == "MODEL" step as an assistant turn, which also matches VIEW_FILE, RUN_COMMAND, LIST_DIRECTORY, GREP_SEARCH and CODE_ACTION -- file dumps, command output and diffs, i.e. exactly the tool noise the Claude Code extractor skips on purpose. Only PLANNER_RESPONSE counts now.
• path may be a list of glob patterns, and agent_transcripts takes key_by: 'name'. Together they solve transcript rotation: Codex moves finished rollouts from sessions/ to archived_sessions/, and with a path-keyed state the same file came back as a new key, was re-read from offset 0 and landed in the index a second time under a second name. One source spanning both directories, keyed on the (globally unique) filename, keeps a moved file's identity.
• Never-index list (sources.EXCLUDED_PATH_SEGMENTS / EXCLUDED_FILENAMES / EXCLUDED_SUFFIXES, sources.is_excluded()), enforced per file inside the adapters, so an over-broad or mistyped glob cannot pull credentials into the index: CREDENTIALS/, .ssh, .gnupg, .gardener itself, node_modules, .git, .venv, __pycache__, and files like .npmrc, .env, auth.json, *.pem, *.key. Segments are matched whole, so a sibling named credentials-howto.md is not caught. gardener.py derives INTERNAL_SKIP_PREFIXES from that same list -- one list to maintain, and what a source adapter refuses to read, the home-folder walk refuses too.
• observe_sources() batches its writes. It used to spend three connections per item (a get, put's own, and put's return get), each opening the DB, ATTACHing the sibling and committing -- about 20 items/s, which is days for a six-figure transcript archive. It now holds one connection per source and commits every 2000 items. Same upsert, same FTS (trigger-maintained), same per-item fingerprint skip.
• New sources: codex-sessions (rollouts across sessions/ + archived_sessions/), codex-history (the flat cross-session prompt history), gemini-transcripts (one transcript.jsonl per brain/ session), gemini-automations (the automation.toml prompts), kimi-transcripts (wire.jsonl per agent per session). decisions-archive widened to *.txt -- the archived decision files are mostly .txt, so only the .md minority was being indexed.
• Deliberately not indexed, with reason: Antigravity's conversations/*.db (content columns are Protobuf BLOBs, not text), agyhub_summaries_proto.pb and annotations/*.pbtxt (binary/Protobuf), brain/*/.git (code snapshots, not conversation), and transcript_full.jsonl (same turns as transcript.jsonl, ~1.6x the bytes -- indexing both would duplicate every session).
• rinnsal-tasks stays registered and returns 0: both ~/.rinnsal/rinnsal.db and scanner_tasks.db exist, and their rinnsal_tasks table is genuinely empty. Nothing to fix.
• Tests: +6 (never-index list across segments/filenames/suffixes, the two adapters honouring it, the shared list reaching gardener.py, Gemini tool-step filtering, and a rotation test that moves a file between directories and asserts nothing is re-indexed). Corrected the two presets' tests, which had asserted the invented formats, and extended the Codex one to cover sub-agent tool traffic. 67 -> 73.

Measured on this machine after the change: 43 sources, everything 14293 -> 284232, user.db 57 MB -> 529 MB. The bulk is codex-sessions (260597) -- 4663 rollouts across sessions/ and archived_sessions/, 10.4 GB of raw JSONL. First scan 9.6 min; the second scan of the same 10.4 GB takes 0.7 s, because every unchanged file is skipped on offset+mtime without being opened.

2026-08-01 (later)

• markdown_dir/remember_files: new extra_tags (string or list), appended to every item's tags. type is always observed for anything an observe-source indexes, so a consumer going straight at the DB (rather than through recall()) has no way to tell a rule file apart from a rotating registry without it -- both are observed alike. extra_tags adds a source-level axis for exactly that distinction, without inventing a second type.
• 12 new observe-sources, all markdown_dir, all read-only:
  • .SYNC/_policies/library and /adoption (policy-library 4, policy-adoption 4), tagged policy.
  • Root-level pipeline steering docs (CLAUDE.md, README.md, MASTER-REGISTRY.md, POLICY-REG.md, STATUS_UEBERSICHT*.md) for six pipeline roots (pipeline-docs-topics 1, -ai 1, -research 13, -roblox 3, -software 4, -umbruch 2), tagged pipeline-doc. Root level only, deliberately not recursive -- a pipeline root can hold thousands of per-project files below it.
  • Root-level CHECKS-REG.md on the four pipeline roots where a plain (non-host-suffixed) copy actually exists (register-log-ai, -research, -roblox, -software, 1 each), tagged register-log. Deliberately excludes host-suffixed rotation copies (CHECKS-REG-<HOST>-<N>.md) and the large rotating CHECKS-LOG*.txt raw logs -- those are exactly the "thousands of files" scope this layer has always avoided.
  • AUTOMATIONS-MEMORY.md was searched for too, but not newly registered: the two canonical copies are already indexed by the pre-existing gemini-rules and gemini-antigravity sources.
  • everything count: 14257 -> 14293 (+36), matching the sum of what each new source reported indexed.
• Two disabled-by-default source-config templates for cross-host federation via a separate transit-sync mechanism that mirrors another host's databases to read-only ~/.republica/<host>/<namespace>.sqlite (Republica showcases) snapshots: usmc_replica_source_configs(host) (facts/lessons/working/ sessions, the same four-way split as this machine's own usmc-* sources) and gardener_replica_source_config(host) (the foreign everything table). Both raise for the current machine's own hostname -- a replica directory named after the current host is that host's replica of itself, and indexing it would duplicate every row under a second source_id. Neither is registered anywhere by default; arming one is a two-line call once a real other-host snapshot exists (see sources.py's module note above usmc_replica_source_configs).
• Tests: +10 (extra_tags on/off/single-string; the two template builders' self-host guard and disabled-by-default shape; a disabled replica, an enabled-but-absent replica, and an enabled replica against a real foreign snapshot are all clean, exception-free paths). Suite: 54 -> 64.

2026-08-01

• README language parity: Synchronized README_de.md with the canonical English structure and restored byte-identical code blocks across the maintained EN/DE pair.
• Every agent's memory in one search: the observe-source layer now covers the other coding agents on the machine, not just Claude Code. Newly indexed (all read-only, originals untouched): Codex/GPT memories and rule file (codex-memories 4, codex-rollouts 256 run summaries, codex-rules 1), Gemini/Antigravity (gemini-rules 4, gemini-antigravity 3), Kimi's prompt history (kimi-prompts 515), and the USMC memory database (usmc-facts 9, usmc-lessons 8, usmc-working 95, usmc-sessions 4). All of it lands as observed, never as memory/lesson: foreign material belongs in find(), and must not crowd out what recall() was curated for.
• sqlite_table: content may now name several columns, joined in order. A row whose meaning is split over two text fields was only half searchable before — whichever column was not configured simply was not in the index. This was live: bach-lessons had been indexing solution while every lesson's problem text stayed invisible. Both BACH and USMC lessons are now indexed as problem + solution (174 BACH lessons re-indexed).
• agent_transcripts: new default_role for single-role archives that carry no role field at all. Without it such a file indexed nothing, because an absent role never matches the roles filter — which is exactly what a bare prompt history looks like. A missing role without default_role is still skipped rather than indexed under an invented one.
• Fixed a source pointing at a path that no longer exists: memoryhooker-docs still referenced the module's pre-2026-07-26 location. A source can go stale silently — it keeps reporting success while indexing an empty directory.
• Tests: +4 (multi-column content incl. an unknown-column refusal, default_role, and the unchanged no-role-no-default guard). Together with the search GUI landed the same day, the suite stands at 54.

2026-07-31

• Search GUI for humans (search_gui.py, gardener gui):
  • New dependency-free, read-only web UI over the FTS5 search core: search box, type filter, BM25-ranked results with match snippets and an entry detail view. Pure standard library (http.server), binds to 127.0.0.1 only, GET endpoints exclusively — no writes against gardener.db/user.db (privacy per design: local, read-only).
  • Follows the BACH unified_search pattern (FTS5 snippet() markers >>>/<<< rendered as highlighted matches in the browser).
  • find() gained an optional with_snippets=True parameter returning an FTS5 snippet() context per hit (LIKE-fallback hits carry no snippet; default behaviour unchanged).
  • CLI: new command gui [--port N] [--no-browser]; help text and translations (cmd.gui) added; pyproject.toml ships the new search_gui module.
  • 11 new tests (tests/test_search_gui.py): snippet API, index page, search/entry/status endpoints, type filter, 404 handling, read-only enforcement. Suite verified at 50/50 passing.

2026-08-01 (multi-agent transcripts)

• Multi-Agent Transcript Format Support:
  • Added native format extractor presets for Gemini Antigravity (gemini_antigravity), Codex (codex), and Kimi (kimi) in scan_agent_transcripts (sources.py).
  • Added support for indexing Gemini transcript logs (~/.gemini/antigravity/brain/*/.system_generated/logs/transcript.jsonl), Codex session & history JSONLs (~/.codex/history.jsonl, ~/.codex/archived_sessions/*.jsonl), and Kimi wire transcripts (~/.kimi/sessions/*/*/wire.jsonl).
  • Extended metadata mapping (session, uuid, timestamp, step_index) across all supported agent formats while retaining 100% backward compatibility for claude_code and generic.
  • Added 3 new unit tests in test_observe_sources.py (test_gemini_antigravity_format_preset, test_codex_format_preset, test_kimi_format_preset).
  • Suite nach Integration in den aktuellen Stand: 64 -> 67 grün. (Die Arbeit entstand parallel auf einem OneDrive-Checkout gegen einen 42er-Stand und wurde am 2026-08-01 im Zuge der Plan-D-Migration hierher übernommen.)

2026-07-30

• Multi-Word Query UX Improvement & FTS5 BM25 Ranking:
  • Implemented automatic FTS5 OR query decomposition (_build_fts_or_query) for multi-word queries in find() when strict AND search yields 0 results.
  • Multi-word searches (e.g. "Registry Mitgliedschaft") now match documents containing individual terms while automatically ranking documents containing all terms higher via FTS5 BM25 relevance. Preserves explicit quotes ("...") and boolean operators (AND, OR, NOT).
  • Hardened SQLite connection handling across find() and get() with strict try...finally: conn.close() resource protection.
  • Added unit test cases (test_multi_word_query_ux_or_fallback_and_ranking and test_build_fts_or_query_helper) verifying 39/39 passing tests green.
• Maintenance & Technical Hygiene:
  • Updated llms.txt Last-checked timestamp to 2026-07-30.
  • Re-verified full test suite execution (39/39 passed) and clean repository working tree.

2026-07-25 (later)

• Documentation clean-up (repo after-care):
  • Removed references to internal, non-resolvable directory names from the public documents (ROADMAP.md, docs/decisions/knowledge-index.md, locales/translations.json). They meant nothing to outside readers.
  • ROADMAP.md is English again: the trailing sections had drifted back into German and carried internal migration notes. Replaced by a short, publicly meaningful "Gardener as a memory module" section, and the German counterpart added to ROADMAP_de.md — both roadmaps now cover the same ground.
  • Architecture tree in both READMEs completed (sources.py, i18n.py, locales/, tests/ were missing); header now states the version instead of a stale date.
  • Corrected observe-source test count 15 → 17 in both roadmaps (counted at the source); full suite verified at 37 passing.
  • Module manifest visibility set to public (the repository has been public for a while); _after-care/ added to .gitignore.

2026-07-25

• Maintenance & Technical Hygiene:
  • Added [tool.pytest.ini_options] with pythonpath = "." in pyproject.toml for standard pytest resolution.
  • Updated llms.txt Last-checked header to 2026-07-25 and test suite count to 37 passing tests.
  • Added Shields.io status badges and LLM integration note callout to README.md and README_de.md.
  • Verified 37/37 unit and integration tests passing green.

2026-07-23

• New (v0.3.0): Cross-source federated index. observe()'s read-only, "look outside" principle is extended to knowledge that lives in other tools, not just Gardener's own home folder. New module sources.py with four adapter kinds:

  • markdown_dir -- a directory (or wildcard glob of directories) of markdown files, one entry per file.
  • remember_files -- .remember-style note files below a root, via recursive glob.
  • sqlite_table -- a single table in a foreign SQLite database, opened strictly read-only (mode=ro); path/table/column-mapping come entirely from config, so it can index any foreign schema without Gardener knowing it in advance. Table and column names are whitelisted against the live schema before use in SQL.
  • agent_transcripts -- JSONL chat transcripts, indexed line by line, text turns only (tool calls/results and "thinking" blocks are skipped). Ships a built-in field mapping for Claude Code's own transcript format, plus a generic dotted-path role/text mapping for other line-based JSON transcripts. Large, growing files are tailed from a saved byte offset (~/.gardener/observe_sources_state.json) -- a refresh never re-reads bytes it already indexed.
  • Every indexed entry carries a source_ref in meta (file/DB path, table+row, or transcript line+uuid) so a search hit always cites back to where it actually lives. find() already searched gardener.db + user.db in one query, so cross-source hits (stored as ordinary observed entries in user.db) show up alongside your own entries automatically -- no new search API needed.
  • New Gardener methods: observe_source_add, observe_source_remove, observe_source_list, observe_sources. New CLI: gardener observe-source add/list/remove/refresh. Configuration lives in config.json under observe_sources; nothing is hardcoded to a specific machine, user, or tool.
  • Deliberately out of scope for this release: adapter presets for the Codex/Gemini/Kimi transcript formats (only Claude Code ships a built-in mapping; other formats route through the generic role_field/text_field mapping) and the v0.2 decay/usage-tracking items (unrelated roadmap section, not touched here).
  • Added 15 regression tests with synthetic fixtures (test suite: 19 -> 34), covering all four adapters, incremental refresh behavior, federated search across own + observed entries, and observe-source config CRUD across a simulated restart.

• New (v0.3.1): patterns config for markdown_dir. The markdown_dir observe-source adapter can now match more than one filename pattern per directory via an optional patterns list in config (default ["*.md"]), e.g. patterns=["*.md", "*.txt"] to index plain-text notes alongside markdown in the same source. Files matching more than one pattern are only indexed once. Backward compatible: the older singular glob key keeps working unchanged for existing configs; patterns takes precedence if both are set. List-valued config like patterns has to go through the Python API (af.observe_source_add(...)) -- the CLI's plain key=value form only accepts strings, not JSON.

  • Added 3 regression tests (test suite: 34 -> 37) covering the default markdown-only behavior, the new patterns list, and the legacy single-glob backward-compatibility path.

2026-07-11

• Release hygiene: i18n.py now carries built-in German/English CLI help fallbacks, so non-editable installs that miss locales/translations.json still show readable help text instead of raw translation keys.
• Added a regression test that runs gardener.py from a wheel-like copy without the locales/ directory.

2026-07-03

• Security: materialize() sanitizes filename/original_name from entry meta to their base name. Previously, meta set via put() could contain .. or absolute paths and make materialize() write outside the destination directory (path traversal).
• Security docs: new "Security Model" section in README/README_de documenting that run() and the seeded shell tool execute code without a sandbox, and that any layer exposing put()/run() must bring its own authorization.
• sync() in always_absorb mode no longer absorbs and deletes its own config.json (which silently reset the mode to selective on the next start). config.json is now part of the shared internal skip list.
• _is_internal() compares whole path segments instead of string prefixes: sibling names like .absorber-notes.txt or .outputs/ are no longer wrongly skipped; internal dirs are now also skipped at any nesting depth.
• observe()/sync() build observed/... entry names with POSIX separators (rel.as_posix()), so the same file yields the same entry name on Windows and Unix (previously Windows produced observed/sub\file.txt, causing duplicates in cross-system setups).
• absorb() raises a clean FileNotFoundError for directories instead of crashing later in _hash_file() with IsADirectoryError/PermissionError.
• CLI: stdout/stderr are reconfigured to UTF-8 with replacement errors in main(), so umlauts no longer crash on Windows consoles without PYTHONIOENCODING=utf-8.
• CLI: gardener absorb <path> prints a clean error message for missing or unreadable files instead of an unhandled traceback.
• CLI: renamed the task loop variable that shadowed the i18n translation function t.
• Added 6 regression tests for the above (test suite: 13 -> 19).

2026-06-22

• Hardened entry deserialization so invalid meta JSON is normalized to an empty object instead of leaking as a string and crashing recall() sorting.
• Added a regression test for recall() on memory entries with invalid meta JSON.

2026-06-12

• Removed the never-populated blobs table from the schema: blob metadata (blob_path, blob_hash, size, mimetype, original_name) deliberately lives in the entry's meta JSON, which is what absorb()/materialize() and the design docs already use. Deliberate decision, see DESIGN.md/KONZEPT.md.
• absorb() now stores original_name in meta (was only original_path), matching what materialize() reads and what the design docs document.
• observe() now skips the internal runtime dirs .absorber/, .output/, .gardener/, __pycache__/ via a skip list shared with sync() (previously it skipped a stale export prefix and indexed absorber/output files).
• Tasks are now sorted by semantic priority (critical > high > normal > low) instead of alphabetical string order.
• find() now preserves FTS5 relevance (bm25 rank) for full-text hits; LIKE-fallback results are ordered newest first. Previously the final sort discarded the rank and listed oldest entries first.
• consolidate() no longer decays or forgets pinned entries.
• Documentation: corrected the local data directory to ~/.gardener (env GARDENER_DATA) in README, README_de, KONZEPT and DESIGN; the previously documented AppData/Local/Gardener/ path was never used by the code.
• Added regression tests for all fixes above (test suite: 5 -> 10 tests).
• Added a minimal pyproject.toml (distribution gardener-os, since gardener is taken on PyPI; console script gardener = gardener:main, requires-python >=3.10, zero runtime dependencies). Verified with an editable install in a throwaway venv.
• Replaced romanized German umlaut spellings in seeded user-facing knowledge and bridge-tool descriptions with real umlauts.
• Updated German runtime error messages for tool execution failures to use real umlauts.
• Added a regression test that verifies seeded German texts no longer contain the old ae/oe/ue spellings.

2026-06-11

• Added README and llms.txt discovery context for the canonical ellmos-ai/gardener repository path.
• Added audience, preferred search phrases, disambiguation, and Last-checked: 2026-06-11 metadata to llms.txt.
• Fixed llms.txt documentation links to use the repository's actual master branch.

2026-06-06

• Updated the Gardener test workflow to actions/checkout@v6 and actions/setup-python@v6.
• Documented the CI hygiene refresh without changing runtime behavior.