Changelog · HyperKB

Unreleased

S3 protocol checks pause incompatible remote sync while preserving local recording, with persistent warnings across reconnects and explicit, previewable protocol 1 → 2 migration. Pending local edits reconcile after upgrade.
hkb update apply follows newer commits even when release tags are unchanged and can retry installation after a partial update.
Multiple local MCP clients now share coordinated storage without terminating one another; session anchors and provenance remain independent.
Atomic entry/configuration updates, recoverable archives, consistent reindex transactions, and offline hkb doctor / hkb reindex improve recovery.
Sync uses staged three-way merges, independent direction checkpoints, elected workers, conditional leases, and version 2 manifests with immutable blobs and deletion records. Upgrade all clients sharing a remote destination before publication.
Search applies metadata/domain filters consistently, restores full entry context, caches ripgrep epoch maps, and estimates complete context-response size.
Bounded asynchronous MCP dispatch, graceful operation draining, explicit batch truncation, task pagination, and MCP SDK 1.x bounds improve integration.
Refreshed GitHub Pages landing page with an accessible memory walkthrough; canonical Markdown guides generate human documentation and machine-readable references.

@weight metadata: New entry importance field (high, normal, low). High-weight entries get 1.15x score boost and are exempt from staleness penalty.
Staleness-aware scoring: Active entries older than 2x recency half-life are penalized (floor 0.7). Decisions and high-weight entries exempt. New config: recency_half_life_days (default 180).
Expanded status scoring: pending (1.08x boost), in_progress (1.05x boost), completed (0.88x dampen), cancelled (0.65x dampen), blocked (0.95x dampen).
hkb_context: Token-budgeted retrieval — searches and packs entries into a token budget using greedy knapsack with type priority re-ranking.
hkb_context_suggest: Proactive file suggestions for a task, with link-graph expansion to surface connected files.
hkb_anchor: Session-only topic bias — 1.5x soft score boost on matching files across search, briefing, recent, context, and suggest. Not a hard filter.
hkb_narrative: Cross-file chronological story reconstruction via search + wiki-link graph traversal (depth 0–2).
hkb_briefing focus/view: New focus param uses search for topic-ranked briefings with relevance scores. New view param scopes briefings to named file groupings.
hkb_view_set / hkb_view_list: Named context views — lightweight file groupings stored as entries in views.workspaces.

README.md rewritten with full tool reference (30 tools), entry metadata docs, scoring pipeline, config key table, and context-aware retrieval guide.
SKILL.md updated with new tools, workflow progression templates, and @weight guidance.
MCP SERVER_INSTRUCTIONS updated with context retrieval guidance.

New DB column: weight on entries table (schema migration).
New DB method: get_entries_by_keys() for batch entry fetching.
New store methods: build_context(), suggest_context(), build_narrative(), set_view(), get_view(), list_views().
AppContext gains anchors and anchor_files for session state.
84 new tests in test_context_retrieval.py. Total: 735 tests.

Sync worker now performs an immediate sync on startup instead of waiting a full interval
Default sync check interval reduced from 5 minutes to 1 minute
hkb update now compares installed package version against latest git tag, not local vs remote tags