prjct-cli 2.43.3 → 2.44.0

package/CHANGELOG.md

@@ -3,13 +3,27 @@
 ## [Unreleased]
 
 ### Fixed
-- **CRITICAL: Codex support actually works now.** The bin shim self-healed the ~9.5KB Claude skill baseline into `~/.codex/skills/prjct/SKILL.md` — 9× over Codex's HARD ~1024-byte cap, so Codex silently rejected the entire skill. The shim now installs the Codex-specific template (and only when Codex is present); the skill metadata marker was compacted and a size guard + tests keep the built artifact under the cap with headroom.
-- `prjct init` no longer logs `Generated: AGENTS.md` (and friends) for files it never wrote — it reports only what it actually writes.
+- **Superseded knowledge can no longer surface as live advice.** `searchFts` (the per-prompt trap cue + the FTS leg of every recall) only filtered soft-deletes — an obsolete decision/gotcha could still inject as current. It now prunes author-declared retirements (`supersedes:` / `superseded-by:` / `duplicates:`) using a mirror-wide dead-id set, catching the case where BM25 never co-returns the superseding entry.
+- **Spanish prompts now produce real retrieval signal.** Keyword extraction was ASCII-only ("búsqueda" → mangled, "qué pasó" → nothing); it now tokenizes unicode, deburrs to match FTS5's diacritic-stripped index, and drops Spanish function words. FTS query sanitizing deburrs too.
+
+### Improved
+- **MCP memory tools retrieve like the CLI.** `prjct_mem_list` used plain recency recall and `prjct_mem_similar` a shared-keyword heuristic — subagents (who do the bulk of the editing) got strictly worse retrieval than `prjct context memory`. Both now run the shared `enrichedRecall` pipeline: FTS-first, semantic blend, usefulness rerank, link expansion, ship attribution.
+- **Digest slots are proven-first.** The session-start and subagent digests picked the 3 most RECENT gotchas/decisions; they now overfetch and rerank by the usefulness ledger, so knowledge that keeps paying off wins the slots.
+- **Push-surfaced knowledge earns ship credit.** The pre-edit guard hook, the prompt trap cue, `prjct guard`, and `prjct_guard` now log surfaced entries against the active task — previously only PULL recalls did, so the most effective gotchas decayed in ranking precisely because the push delivery worked.
+
+## [2.44.0] - 2026-06-11
+
+### Features
+
+- retrieval quality — superseded filter, proven-first digests, MCP parity, push-path ship credit, unicode keywords (#429)
+
+
+## [2.43.4] - 2026-06-11
+
+### Performance
+
+- indexed range queries, git TTL cache, version-aware vault fingerprint, upgrade-scan guard (#428)
 
-### Added
-- **Codex MCP wiring.** Setup and sync now ensure `[mcp_servers.prjct]` in `~/.codex/config.toml` (marker-managed TOML block; user-managed entries are never touched) — the `prjct_*` tools finally exist for Codex sessions.
-- **AGENTS.md generation.** `prjct init` writes a vendor-neutral prjct routing block to the project's `AGENTS.md` (marker-merged, user content preserved) when Codex is detected or wizard-selected — Codex has no hooks, so this block is its session-start context.
-- `prjct doctor` now checks the `codex` binary and the installed Claude hooks count.
 
 ## [2.43.3] - 2026-06-11