prjct-cli 2.37.0 → 2.37.2

package/CHANGELOG.md

@@ -2,7 +2,18 @@
 
 ## [Unreleased]
 
-## [2.37.0] - 2026-06-01
+## [2.37.2] - 2026-06-01
+
+### Changed
+- **Quality-workflow methodology rewritten as behavior, not procedure** (in the pulled `workflows.md` reference). The `review` / `qa` / `security` / `investigate` / `ship` / `audit` sections were numbered step-by-step lists (40 procedural steps); they now describe the *observable outcome* — "what good looks like" — and let the agent discover the how. Stop-conditions, the per-role model policy, and the parallel-dispatch mechanics are kept verbatim (those are concrete operational rules, not procedure). Rationale (Deep SWE / Theo): over-specified step-by-step prompts flatter weak models and add noise; behavior-framed prompts force end-to-end exploration and let capable models self-verify. Aligns the methodology with prjct's own anti-harness ethos.
+
+## [2.37.1] - 2026-06-01
+
+Context-efficiency, part 3: trim the recurring per-turn surfaces.
+
+### Changed
+- **Per-prompt topical-memory injection is now selective, not padded.** It fires on every prompt. Previously, when FTS/BM25 returned fewer than 8 matches it padded the set up to 8 with weaker recency/substring matches — spending tokens on low-relevance entries every turn. Now it injects exactly the BM25-ranked matches and only falls back to the recency window when FTS returns *nothing* (fresh DB / no indexed match). Narrow prompts inject few, highly-relevant entries instead of a padded eight. Also condensed the per-block preamble that repeated verbatim each prompt.
+- **MCP tool descriptions condensed** (~820 → ~680 tokens) without losing the "when to call" signal — they load into context whenever the prjct MCP server is connected.
 
 Context-efficiency, part 2: cut the Claude skill's always-on footprint by 73%.