Changelog

0.2.9 — 2026-09-19

Symlinked-root patch. Closes #35 via #37.

• Engine parity holds on a symlinked root (#35). rg prints resolved paths, and search() passed the caller's root through unresolved, so _rel could not make the rg hits relative. It fell into except ValueError and returned the full absolute path with a doubled leading slash, and the two engines then reported different paths for the same file. That is the parity contract 0.2.8 shipped to guarantee, and the caller could not tell the wrong path from a real one. The root is the cause, not _rel: both engines derive their paths from it, so search() resolves it once. One syscall per call, not one per hit.
• _rel no longer catches ValueError. The old branch did not recover, it returned a wrong value.
• Not macOS-only. /var is a symlink to /private/var there, so every tempfile root reproduces it. A symlinked checkout or a container bind mount is the same shape on Linux. CI has neither, which is why it shipped.
• New tools/ci-local.sh mirrors every step of .github/workflows/ci.yml. Its first check parses every workflow and fails when the glob finds none. PKC #82 was an unparseable ci.yml that failed at startup with zero jobs, and a workflow cannot check itself.
• Tests: test_symlinked_root_agrees_across_engines searches through a real symlink.

0.2.8 — 2026-09-19

Retrieval-ladder patch: parse once, engine parity, fail-closed rg override, rg test coverage. Closes #30, #31, #32 via #33.

• rkc_search.py scan path parses each file once (closes #30). iter_okf already parsed every file to yield it; search() parsed each again (measured 2.0×, ~2.9 s vs 99 ms for rg on a 3k-file tree). The scan rung is what hosts without ripgrep run.
• rkc_validate.py parses the tree once per run; validate() and spine_issues() share it (load_tree).
• Search parity (#31): the filename stem no longer enters the haystack (hay_title, hay_id). A titleless node whose filename matched the query was found by scan but invisible to the rg prefilter. Filenames are not content; display still falls back to the stem.
• find_rg() (#32) honours RKC_RG_PATH / OKF_RG_PATH / SECOND_BRAIN_RG_PATH and fails closed when the override is unusable — the same rule as research-graph and PKC, so one env var means one thing across plugins.
• Scan search skips unparsable files instead of aborting, matching the rg path.
• Tests: tests/fixtures/fake_rg.py (ported from PKC) and rg-vs-scan parity tests; the rg engine had no coverage.

0.2.7 — 2026-09-12

Query-time retrieval, PKC/SAC/DEKC parity.

• agents/research-retriever.md — isolated retriever. Scores and deep-walks off the parent. Returns a summary card only.
• /research-retrieve skill + command. Parent spawns the retriever; do not pack inline for Q&A.
• rkc_pack.py --summary — compact card-friendly pack (seed, hops/nodes/tokens, spine counts, lead nodes, gaps). --tiny is hops=1 / max_nodes=8. --max-tokens fail-closed (default 32000). Seed may be an id, path, or stem.
• rkc_search.py + /research-search — git-native rg → scan ladder. AND terms, type filter, limit 5, --json. No Chroma, BM25, or Kuzu. Layer 1 may feed seeds later.
• research-capturer delegates Q&A retrieval to research-retriever.

0.2.6 — 2026-08-25

• --dry-run does not create research/catalogs/ingest-index.json or its parent directories. Closes #25.
• Empty, whitespace-only, and frontmatter-only sources are skipped and reported. --min-source-bytes defaults to 1. Closes #26.

0.2.5 — 2026-08-25

• Subject title is derived from the --subject slug (or --subject-title), not the first source filename. rkc_spine.py uses the same rule when it creates a missing Subject. Closes #23.

0.2.4 — 2026-08-25

• Intra-run claim merge no longer calls add_link on a node whose path is still None. A second file that repeats a claim from the first no longer crashes extract with 'NoneType' object has no attribute 'read_text'.

0.2.3 — 2026-08-25

• Quote date-like YAML scalars. as_of / timestamp / captured_at stay strings through parse and write. Closes #16 and #17.
• parse_okf wraps yaml.safe_load in ParseError that names the file. rkc_validate.py reports it and continues. Closes #6.
• Default --extractor-version to EXTRACTOR_VERSION. Same bytes under another version are reported, not duplicated, unless --allow-reextract. Idempotent --extract skips the extractor. Closes #18.
• index.md / README.md titles become parent/filename. Closes #19.

0.2.2 — 2026-08-24

• rkc_spine.py / /research-spine repairs Area → Subject → Task on a tree ingested by 0.2.0. --link-tasks, --list-subjects, --area-map. Idempotent.

0.2.1 — 2026-08-24

Bulk-ingest hardening from the first 2,997-file corpus (#13).

Blockers:

• Ingest lookup is O(1) via research/catalogs/ingest-index.json. Rebuild with --rebuild-index (sources + tasks only). #1
• Heuristic extract skips sources over 200 KB (--force-large to override), caps candidates, and matches near-duplicates through a token index instead of all-pairs Jaccard. #2
• Long subject slugs keep a prefix plus an 8-hex digest of the full value so distinct names no longer collapse. #3

Major:

• pyyaml is a required dependency. _mini_yaml unescapes \\n / \\t / \\" / \\\\ and warns when the fallback runs. #4
• _needs_quote treats a trailing colon as quoted. Ingest writes through write_okf. #5
• rkc_validate.py reports the unparsable path and continues. #6
• SourceDocuments store origin_path. index.md / README.md titles use the parent directory. --source-kind is accepted. #7
• Archive-only ingest creates the Subject node. #8
• Ingest writes has_task. --area creates a ResearchArea and writes has_subject. --spine fails validation on a missing ladder. #9

Minor:

• Heuristic extractor drops Markdown/numbered headings and filecite / citeturn export artifacts. #10
• Per-file progress on stderr, ingest-key resume, --errors-file (does not abort the run). Intra-tree --workers is not added: concurrent writers race, and pkill -f orphans multiprocessing children. Shard by Subject instead. #11
• Vendor is free text. Convention now includes perplexity and unknown. #12

0.2.0 — 2026-08-22

Phase 2 extractor.

• rkc_segment.py — global line/char locators, heading/paragraph packing.
• rkc_extract.py — heuristic extract + fail-closed overlay JSON.
• claim_key merge attaches Evidence to existing Claims; near-match same_as; contradicts as a new draft.
• Accepted / reviewed / verified nodes are never auto-superseded.
• Required PR summary under research/catalogs/pr-summaries/ (not an OKF type).
• Ingest key is sha256(bytes|prompt_hash|extractor_version). --extract flag.
• /research-extract skill + command. ADR 007.

0.1.0 — 2026-08-22

Phase 0–3 bootstrap.

• Eight nouns, twelve rels, JSON schemas, registry.
• Five-host packaging (Claude Code, Grok Build, Codex, Cursor, Agent Plugins 1.0).
• /research-ingest, /research-pack, /research-validate.
• Northstar / Lumenfield fiction sample + eval fixtures (spine, claim_key, quote verify, unknown rel).
• ADRs 001–006. Compact PRD under docs/prd/.