@mindrian_os/cli 2.0.0-beta.11 → 2.0.0-beta.15

package/CHANGELOG.md

@@ -1,3 +1,38 @@
+## [2.0.0-beta.15] - 2026-08-28
+
+### Added
+- `scripts/verify-release` gains gate 10e (Phase 267.3, close-out): the whole-tree
+  `commands/*.md` `interactive_first_reward` audit (`node
+  scripts/check-reward-before-investment.cjs`, the CLI's full-audit default mode, no flag) is
+  now a fail-closed release gate, sited alongside gate 10d's registry audit. Proven against a
+  stripped-fixture A/B test before being wired (a `commands/` copy with one declaration
+  stripped exits 1, the real tree exits 0), matching the Phase 271-05 gate-proof discipline.
+  This closes the debt-ratchet gap `DEFERRED-271-D1` documented: a `--staged`-scoped commit
+  gate cannot measure a repo-wide declaration debt, which is how 67 of 113 commands went
+  undeclared for three months invisibly. The `--staged` commit gate itself is unchanged.
+  67 commands backfilled with a ruled `interactive_first_reward` declaration across this
+  phase (17 blocking Phase 271, plus 25 + 25 in the two backfill waves), and the
+  `data/first-reward-surfaces.json` registry (introduced earlier this phase) now declares
+  the first reward for 4 non-frontmatter surfaces inside `scripts/session-start`.
+
+## [2.0.0-beta.13] - 2026-08-28
+
+### Added
+- Memory and Context Operator MCP (Phase 270): three new atomic tools -- `context_assemble` (exposes the four-leg `getRoomContext` budget model with `estimate_only` cost preview before paying it), `graph_reason` (transitive support-closure and cross-room nearest-decision queries, both graph-native and previously unreachable by a flat directory walk), and `identity_write` (the first writer `~/.mindrian-user.md` has ever had). MCP Resources fixed to resolve the room per-session instead of once at boot (`room://state`), plus `mos://tree` and `mos://room/{slug}/tree` exposing the ICM forest live. **Honest tradeoff, not a win:** measured tool-schema token budget went **up** ~17% (36->39 tools, ~7,167->~8,377 approx tokens), not down -- the three new atomic tools cost more than the consolidation saved. Recorded as such, not spun.
+- `interactive_first_reward` vocabulary gains a ninth member, `live_deliverable` (Phase 267.3): a real, deployed, shareable artifact from an irreversible action (a live URL, a sent file) -- minted rather than forcing `/mos:publish`'s actual reward into an existing, incorrect term. A `data/first-reward-surfaces.json` registry now lets non-frontmatter surfaces (bash hooks, injected prose) declare a first-reward contract the linter can read; `scripts/session-start`'s four surfaces (FIRST_INSTALL, UPDATE, MODE_ROUTING, COLD_START_MENU) are now declared.
+- `/mos:doctor --check-rs-engine --fix` and direct auto-install: `scripts/compute-hsi.py` and `scripts/rs-engine.py` (the entry points for `/mos:act`, `/mos:mos-reason`, `/mos:find-bottlenecks`) now call the plugin's own already-shipped Python dependency auto-installer on missing imports, instead of hard-failing with a manual `pip install` instruction. Six sibling scripts already had this; these two did not.
+
+### Fixed
+- Reference-file citations are now anchored with `${CLAUDE_PLUGIN_ROOT}`, so the files a command actually needs load correctly from a real Data Room instead of only from the plugin dev repo. A bare `references/...` path resolves against your session's current working directory, not the plugin install directory, so these citations worked by pure coincidence in development (where the two happen to be the same folder) and silently failed for every installed user on all three surfaces. **134 of 134 measured sites now anchored** (103 commands/skills/agents/pipelines, plus the final 31 -- 16 methodology commands and `commands/doctor.md` -- unblocked by the `interactive_first_reward` vocabulary fix above and landed clean). Hand-authored skills use the fail-closed `${MINDRIAN_OS_ROOT:-${CLAUDE_PLUGIN_ROOT:?...}}` form, which refuses honestly instead of silently reading the wrong path when a foreign Agent-Skills host loads a skill with no plugin root set. `commands/radar.md`'s five sites are a reasoned exception (a genuine write into a git-tracked source of record, not a read -- anchoring it would redirect writes into the install cache every update wipes) -- allowlisted with a written reason, not silently skipped.
+- Added `scripts/check-plugin-path-anchoring.cjs` to the release gates (`scripts/verify-release`, gate 10c) so this class cannot silently return. A newly authored command, skill, agent, or pipeline file citing a bare plugin-relative path now fails the release gate before it can ship. `data/first-reward-surfaces.json` declarations are checked by a companion gate 10d.
+- Capability Radar Absorption + Routing (Phase 265, 23 plans): a full adversarial code review after all plans passed their own tests caught two real regressions no test suite caught -- `/mos:scout`'s competitor fan-out had Write + Brain access despite promising read-only (fixed by minting a genuinely read-only `agents/competitor-watch-fetcher.md`), and `/mos:vault` had a real human-gate bypass (a file-routing step re-invoked before, not after, the navigator's confirmation).
+- MCP layer correctness fixes (Phase 266): a per-call timeout budget that didn't hold in aggregate (60s vs a 30s host budget), caught by independent re-verification after the phase's own tests passed; fixed and re-measured (15.2s).
+
+### Known limitations, stated plainly
+- The other ~50 commands beyond the 17 fixed this release still lack an `interactive_first_reward` declaration (a repo-wide backfill, not yet started; no release gate currently requires full coverage, so this does not block shipping).
+- A code-review pass this cycle found 5 Critical correctness bugs in the local SQLite graph write chokepoint (`lib/core/navigation/edges.cjs`'s `writeEdge` silently discards some writes while reporting success; a schema mismatch fails writes from one of two database openers; the Brain-supplied edge-type allowlist is bypassed on one write path; the busy-timeout fix is applied at 1 of ~20 database openers; a locked room can misreport as having no database at all) -- found and registered as its own phase, **not fixed in this release**. Full detail: `.planning/ROADMAP.md` Phase 273.
+- Phase 134's original objective (eliminating the Python dependency for HSI/reverse-salient analysis entirely) was tracked as complete but was never actually implemented -- the auto-install fix above is a mitigation, not the real fix. Registered as Phase 272, not done here.
+
 ## [2.0.0-beta.11] - 2026-08-24
 
 ### Added

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mindrian_os/cli",
-  "version": "2.0.0-beta.11",
+  "version": "2.0.0-beta.15",
   "description": "Install MindrianOS into Claude Code with one command -- `npx @mindrian_os/cli`. Ships the MindrianOS plugin (Larry + PWS methodology + Data Room) plus a setup/diagnostics CLI (install/doctor/update).",
   "scripts": {
     "mcp": "node bin/mindrian-mcp-server.cjs",