prjct-cli 2.43.2 → 2.43.4

package/CHANGELOG.md

@@ -2,8 +2,31 @@
 
 ## [Unreleased]
 
+### Performance
+- **Hot recall queries now use the type index.** Every `type LIKE 'memory.%'` filter (recall, vault index, embeddings backfill anti-join, memory cap, transcript dedup) was a full `SCAN events` — SQLite can't derive scan bounds from a parameterized LIKE. Rewritten as indexed range predicates with exact-parity tests; cost stays flat as the events table grows.
+- **Prompt hook git snapshot cached for 3s.** The UserPromptSubmit hook forked 3 git processes per turn (~40ms); agentic bursts forked hundreds. The daemon now reuses the snapshot within a 3s TTL per cwd.
+- **Upgrade migration no longer opens every project DB.** A `.cli-version` marker per project dir reduces the post-upgrade scan from a SQLite open per directory (~3ms × tens of thousands of dirs ≈ up to a minute) to one tiny file read each.
+- Vault writes skip redundant per-file `mkdir` syscalls (dir-ensured cache with delete-retry fallback).
+
 ### Fixed
-- **CRITICAL: the memory cap no longer deletes knowledge.** `capEntries` (runs on every `prjct sync`) counted ALL `memory.%` events against the 500-row cap — high-churn telemetry (`memory.post_edit` fires on every file edit) inflated the total and the age-ordered delete silently destroyed the OLDEST remembered decisions/gotchas/learnings while keeping hundreds of newer telemetry rows. `memory.remember.*` is now invisible to the cap (count and delete); knowledge leaves the log only via `prjct forget`. The delete is also exact-id based instead of an id-range sweep. If a past sync capped your project: the `memories` mirror table still holds the rows with their original ids — restorable.
+- **Vault can no longer go stale across upgrades.** The regen fingerprint now includes the CLI version: every upgrade forces one cheap full regen per project, so a builder/format change always renders (previously the old fingerprint blocked new output until unrelated inputs changed).
+- **Codex detected in non-interactive shells.** `detectCodex` accepts `~/.codex/auth.json` (a logged-in install) as evidence alongside binary-on-PATH — hooks/daemon run without the interactive PATH, which silently skipped Codex MCP wiring and skill repair.
+- Removed dead `memoryService.search()` (1,000-row JS substring scan, zero callers — `prjct search` uses the FTS5 + semantic blend).
+- Archive-storage tests no longer leak a real-path SQLite connection across the pathManager patch (stray async sync-publish could cache it and hijack the next test's isolation).
+
+## [2.43.4] - 2026-06-11
+
+### Performance
+
+- indexed range queries, git TTL cache, version-aware vault fingerprint, upgrade-scan guard (#428)
+
+
+## [2.43.3] - 2026-06-11
+
+### Bug Fixes
+
+- Codex first-class — skill under 1KB cap, MCP wiring, real AGENTS.md, doctor checks (#427)
+
 
 ## [2.43.2] - 2026-06-11
 

package/README.md

@@ -509,6 +509,14 @@ Codex is detected by the `codex` CLI on PATH (context file `AGENTS.md`). The
 sandbox is non-interactive/non-TTY, so prjct-cli emits the same static, prompt-free
 status line as any agent; add `--md` for fully markdown-structured output.
 
+**What does Codex get from prjct?**
+Three surfaces, all installed/healed automatically: a compact skill at
+`~/.codex/skills/prjct/SKILL.md` (kept under Codex's ~1KB skill cap), the
+prjct MCP server wired into `~/.codex/config.toml` (`prjct_*` tools), and a
+vendor-neutral routing block in the project's `AGENTS.md` written by
+`prjct init`. Codex has no lifecycle hooks, so AGENTS.md + MCP are its
+session-start context and live tool surface.
+
 **How do I quickly find the local `.prjct/` directory?**
 It's in your **project repo root** (created by `prjct init` / first `prjct`
 command) and is `.gitignore`d — that's why `git status` never shows it. Find it:

package/bin/prjct

@@ -91,15 +91,19 @@ ensure_setup() {
   # postinstall is unreliable (--ignore-scripts, npm policies), so the
   # bin shim self-heals on every invocation. mtime-based: source newer
   # than dest → overwrite. Common case (skill up-to-date) = two stats.
-  SKILL_SRC="$ROOT_DIR/templates/skills/prjct/SKILL.md"
-  if [ -f "$SKILL_SRC" ]; then
-    for SKILL_DEST in "$HOME/.claude/skills/prjct/SKILL.md" "$HOME/.codex/skills/prjct/SKILL.md"; do
-      SKILL_DIR=$(dirname "$SKILL_DEST")
-      if [ ! -f "$SKILL_DEST" ] || [ "$SKILL_SRC" -nt "$SKILL_DEST" ]; then
-        mkdir -p "$SKILL_DIR" 2>/dev/null
-        cp "$SKILL_SRC" "$SKILL_DEST" 2>/dev/null
-      fi
-    done
+  # Each agent gets ITS OWN template: Codex enforces a hard ~1KB limit
+  # on the whole SKILL.md and silently rejects the entire skill when
+  # the (much larger) Claude baseline lands there.
+  install_skill() {
+    if [ -f "$1" ] && { [ ! -f "$2" ] || [ "$1" -nt "$2" ]; }; then
+      mkdir -p "$(dirname "$2")" 2>/dev/null
+      cp "$1" "$2" 2>/dev/null
+    fi
+  }
+  install_skill "$ROOT_DIR/templates/skills/prjct/SKILL.md" "$HOME/.claude/skills/prjct/SKILL.md"
+  # Codex only if present — don't create ~/.codex on machines without it
+  if [ -d "$HOME/.codex" ] || command -v codex >/dev/null 2>&1; then
+    install_skill "$ROOT_DIR/templates/codex/SKILL.md" "$HOME/.codex/skills/prjct/SKILL.md"
   fi
 }