@mindrian_os/cli 1.15.3-beta.8 → 1.15.3

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (3) hide show
  1. package/CHANGELOG.md +836 -0
  2. package/README.md +5 -3
  3. package/package.json +4 -2
package/CHANGELOG.md CHANGED
@@ -1,3 +1,839 @@
1
+ ## [1.15.3] - 2026-07-31
2
+
3
+ ### Added
4
+ - **The room now learns which of its two ranking signals has actually been right for you, not
5
+ just from a fixed prior forever (RCA hedge-fold-has-no-production-trigger).** When Larry picks
6
+ which of several fired reaches to surface, two signals vote: how well an idea seems to fit
7
+ right now, and how far up the fixed list it sits. A Phase 222 layer was supposed to learn,
8
+ from your own past accept/reject choices, how much to trust each of those two signals for
9
+ you specifically, and quietly re-weight them over time. It never once ran, on any install,
10
+ because nothing in the shipped code ever handed it your room's database to learn from, so it
11
+ sat at a permanent, correct-looking cold start. Nothing you saw was wrong: the ranking still
12
+ worked, it was just always computed from the same starting assumption instead of from your
13
+ own outcomes. There is now one deliberate command, `node scripts/hedge-refit-pipeline.cjs
14
+ <room>`, that runs that learning step on purpose, rather than it riding along as a side effect
15
+ of serving a turn. Canon Part 9: it reaches your room's database only through the one
16
+ existing local chokepoint, same as everything else. Canon Part 8: zero network, zero Brain,
17
+ nothing but your own room's past decisions.
18
+
19
+ ### Changed
20
+ - **The stated minimum Node version is now 22.16.0, up from 22.5.0, because that is the version
21
+ where the room.db write-safety setting actually starts working (Phase 236, GRAPHDB-03).** Since
22
+ Phase 218-02 the room's database has been opened with a five second "wait your turn" setting, so
23
+ that when two things try to write at the same moment the second one waits instead of failing
24
+ instantly. That setting is called `timeout`, and it is passed to Node's built-in `node:sqlite`
25
+ module. Two different Node versions matter here and they are easy to confuse. `node:sqlite`
26
+ stopped needing a special startup flag at v22.13.0, but the `timeout` setting itself was not
27
+ added until **v22.16.0**. In between, on 22.13 through 22.15, the module loads fine and the
28
+ code looks correct, but `node:sqlite` accepts settings it does not recognise without
29
+ complaining, so `timeout` is quietly thrown away and contended writes still fail at zero
30
+ milliseconds exactly as before. Nothing warns you. We confirmed this on a live runtime by
31
+ reading `PRAGMA busy_timeout` back after opening: `0` on a version without the option, `5000`
32
+ with it. The old floor of `>=22.5.0` was wrong twice over, because on 22.5 through 22.12
33
+ `require('node:sqlite')` throws outright without the flag. Source for the v22.16.0 figure:
34
+ Context7 against the Node.js v22.x API docs, specifically the `timeout` option's
35
+ version-history entry, not the module's separate unflagging entry.
36
+ **User-visible consequence:** npm will now refuse an install on Node 22.5.x through 22.15.x
37
+ that it previously accepted. That is deliberate. The code genuinely does not run safely on
38
+ those versions. Upgrade Node to 22.16.0 or newer (`nvm install 22`, `fnm install 22`, or your
39
+ package manager) before updating. CI already runs the Node 22 major line, which resolves above
40
+ the new floor, so CI keeps exercising a runtime users actually have.
41
+
42
+ ## [1.15.3-beta.50] - 2026-07-28
43
+
44
+ ### Added
45
+ - **A doctor check that catches a Data Room whose graph never learned how the ideas relate,
46
+ and an automatic repair for the rooms already in that state (Phase 233, RCA items 4c/4d).**
47
+ Two things live inside `room.db`: `BELONGS_TO` edges, which say which artifact sits in which
48
+ section (a filing cabinet), and cascade edges (INFORMS, CONTRADICTS, CONVERGES, INVALIDATES,
49
+ ENABLES, REFINES, ROOT_CAUSES), which say how the ideas actually relate (the part that makes
50
+ the room think). Before Phase 224-02 shipped on 2026-07-23, a failed derivation quietly
51
+ deleted its own retry signal, so a room could end up with the first and never the second and
52
+ nothing anywhere would say so. Phase 224-02 stopped that happening again, but it was
53
+ forward-only: roughly 16 live rooms were already in that state and stayed there. This release
54
+ closes both halves. New `graph-derive-health` doctor class (`/mos:doctor
55
+ --graph-derive-health`, or add `--cascade-rooms` to sweep every room) reports FAIL on exactly
56
+ that shape and WARN on a derive queue stuck past three days or a recorded failure log. New
57
+ `--heal-room` flag re-enqueues every affected room; it is the literal flag the
58
+ v1.13.0-beta.16 rename table has pointed at since before it existed, made real here for the
59
+ first time. And a one-time `graph-derive-heal-retrofit` module repairs already-damaged rooms
60
+ by itself the first time doctor runs after the update, with no flag to discover and nothing
61
+ to opt into, because a user should not have to know their graph was damaged in order to get
62
+ it fixed. Both share ONE detection function, so the check and the repair can never drift
63
+ apart. The heal restores the retry signal; the real edges land the next time an in-session
64
+ derive runs. Tri-Polar: CLI gets the full report, Desktop and Cowork get a one-sentence
65
+ re-derive nudge on the existing SessionStart install-drift slot rather than a raw table.
66
+ Canon Part 8: every new path is a read-only local `room.db` read plus a local JSON queue
67
+ write, zero network and zero Brain. Ground-truth tested (20 scenarios) against real `room.db`
68
+ files and real queue files read back off disk, never a mocked return value.
69
+
70
+ ### Fixed
71
+ - **Your room's discovery engine was ranking its own backup files as its top insights, and
72
+ then throwing every result away (Phase 233, RCA Section 9 Defects #4/#5).** The HSI pass is
73
+ the part that reads your artifacts and says "these two distant pieces are secretly related".
74
+ It walks your room looking for content, and every other walker in the codebase was taught in
75
+ Phase 200 to read one shared list of folders to ignore. This one walker was never migrated,
76
+ so it kept a private copy that had gone stale: it never learned to skip `.snapshots`
77
+ (historical state dumps of your own room) or `sub-rooms` (nested rooms that carry their own
78
+ graph). The result on a real room: 207 files scored, and all twenty of its "top discoveries"
79
+ were near-identical backup copies of each other. Then the edge writer correctly refused every
80
+ one of them, because backup files are not artifacts in your graph, and wrote zero edges. A
81
+ full expensive pass, a confident report, and nothing to show for it. Now `compute-hsi.py`
82
+ reads the same shared list as everything else (so this class of drift cannot recur), and a
83
+ new `--scope-to-nodes` mode scores only artifacts that actually exist in your graph, which is
84
+ both cheaper and the only set that can produce an edge.
85
+ - **Healing a damaged graph now runs its four stages in the one order that works, and the last
86
+ stage stopped deleting the work of the third (Phase 233).** Repairing a room is not one
87
+ action, it is four, and each one eats what the previous one produced: index every artifact as
88
+ a node, score similarity across those nodes, write the semantic edges, then run the
89
+ generative tier. Run them out of order and every stage reports success while producing
90
+ nothing. New `node scripts/graph-heal-pipeline.cjs <room>` runs all four in the mandated
91
+ order, reusing each existing implementation unchanged. Running it live on the room from the
92
+ original investigation exposed one more instance of the same bug class this whole phase is
93
+ about: the fourth stage opened by clearing and rebuilding the graph, which deleted the twenty
94
+ semantic edges the third stage had written seconds earlier. The pipeline printed "wrote 20
95
+ connection edges" into a room that ended up holding none. The rebuild is now suppressed when
96
+ the caller already indexed the room, and that room finished the run with 61 artifact nodes,
97
+ 20 semantic edges and 47 typed relationship edges, up from zero. Every other caller of the
98
+ backfill is byte-unchanged. A missing Python or embedding library degrades exactly the two
99
+ stages that need them and never the other two. Canon Part 8: local only, zero network, zero
100
+ Brain.
101
+ - **The graph derivation engine could quietly dial a dead account instead of scoring your
102
+ work locally (Phase 233, RCA items 4b/4e).** `runDerivation` is the composer that turns a
103
+ pair of your artifacts into a typed relationship edge. It takes the scoring function as an
104
+ argument, and until now, if a caller forgot to pass one, it silently fell back to a hosted
105
+ Anthropic API call. That fallback account has been out of credit for months: a live probe
106
+ returned `400 credit balance is too low`. So "forgot to pass the scorer" did not look like a
107
+ mistake at the point it was made; it looked like working code that failed later, over the
108
+ network, for a reason that had nothing to do with the actual bug. That fallback is now gated.
109
+ Omitting the scorer throws `deriveFn_required_no_hosted_default` immediately, before the room
110
+ database is even opened, and the error names the local scorer to use instead. Nothing shipped
111
+ is affected: every real caller already passes one. An operator who genuinely wants the hosted
112
+ path can still have it by setting `MINDRIAN_ALLOW_HOSTED_DERIVE=1` with a funded
113
+ `ANTHROPIC_API_KEY`, and gets byte-identical behavior to before. Gated, not deleted. Also
114
+ fixed: the background drain's header comment still claimed it "CLEARS the drained entry"
115
+ after every pass, which stopped being true when Phase 224-02 made failed entries survive and
116
+ retry. It now states what actually happens, plus the real division of labor: the background
117
+ pass enqueues, preserves on failure, and scores locally, while in-session `/mos:graph
118
+ --derive` is the wider net that also reaches rooms the background pass structurally cannot
119
+ see. Both default to the same local score-based scorer, and that is no longer a claim in a
120
+ comment: a new regression test swaps the shared scorer for a recorder, runs both paths over
121
+ two identical rooms, and compares the edges that land on disk.
122
+ - **`orchestration rooms-open` reported a confirmation-shaped success while never switching
123
+ the active room.** Live-reproduced 2026-07-22: the call returned a full "Room State" payload
124
+ footed with the correct target room, yet `room-registry get-active` still returned the
125
+ previous room and the next Write was blocked with "Active room is <previous>". Root cause:
126
+ `rooms-open` was a DECLARED-BUT-UNIMPLEMENTED command. It passed Zod validation via its
127
+ membership in `ORCHESTRATION_COMMANDS`, matched no handler, and fell through to a generic
128
+ reference-echo fallback shared by 18 of the 22 orchestration commands. That fallback built
129
+ its response from `commands/rooms.md` read off disk, STATE.md from the boot-frozen `roomDir`
130
+ closure (which is why the payload showed an unrelated room's content), and a verbatim echo of
131
+ the caller's own `room` argument -- then appended "Room operation complete". No byte of the
132
+ response derived from an operation, because none was attempted: before this fix, zero product
133
+ `.cjs` code anywhere called `room-registry set-active`, so the MCP surface advertised
134
+ multi-room management while having no room-switch capability at all (worst on Desktop and
135
+ Cowork, which have no shell fallback). Fixed with a new `lib/core/room-open.cjs` chokepoint
136
+ that wraps the one authoritative writer (Canon Part 7) and gates `ok:true` behind a post-write
137
+ `get-active` read-back, so a success-shaped payload is now structurally impossible unless the
138
+ switch actually landed; it also writes the per-session binding so `write-scope-check.cjs`
139
+ authorizes the session without depending on the raceable global field, and it preserves the
140
+ `commands/rooms.md` Step 2 human gate for reopening an archived room. Structurally, the same
141
+ fallback no longer lets any state-mutating orchestration command claim completion it cannot
142
+ back up: `rooms-new`, `rooms-close`, and `rooms-archive` now carry an explicit NOT EXECUTED
143
+ banner. 10 hermetic regression tests assert registry ground truth, not response shape.
144
+
145
+ ## [1.15.3-beta.48] - 2026-07-26
146
+
147
+ ### Added
148
+ - **Room-graph density read (Phase 232.1).** `/mos:doctor` and `room_state status`/`get-state`
149
+ now self-report node/edge counts per `room.db`, read exclusively through a new read-only
150
+ `navigation.cjs` door (`openRoomDbReadOnlyForCaller`) -- never the mutating door, never a
151
+ direct `room.db` open. Closes SEED-074's own "suggested first move"
152
+ (`.planning/seeds/SEED-074-local-graph-read-layer-lacks-salience-and-query-time-joins.md`):
153
+ the seed's actual PageRank/Louvain/query-time-join target stays gated exactly as written;
154
+ this only makes its own trigger condition (measured room.db density) self-reported instead
155
+ of something a human has to remember to check by hand. No output string anywhere claims a
156
+ room graph is dense, at-risk, or healthy (hard guard, grep-verified). Goal-backward
157
+ verification PASS, 6/6 must-haves (`232.1-VERIFICATION.md`).
158
+ - Local reified-claim `ContradictionEvent` primitive (quick task 260725-9ca), with a hermetic
159
+ acceptance test.
160
+
161
+ ### Fixed
162
+ - **`room_bind` could report `ok:true` while a sibling MCP read tool silently resolved a
163
+ stale, unrelated room.** Root cause: 20 of 21 MCP tool call sites were missing the
164
+ `CLAUDE_CODE_SESSION_ID` stdio fallback that `room_bind` itself already had, so a
165
+ session-id mismatch between the bind call and a later read call could bind one room but
166
+ read another without ever surfacing an error. Fixed via one shared
167
+ `resolveEffectiveSessionId` helper, wired into all 21 sites; live-reproduced, then closed,
168
+ tests green. Full reproduction and root cause also filed in the `rethinking-mindrianos`
169
+ room per this repo's dev-research compositing rule.
170
+ - **Phase 232.1's own room-graph density census could silently drop a room that used the
171
+ `abs_path` registry field instead of `path`,** undermining the one thing the census exists
172
+ to be accurate about. `resolveRoomPath` now checks `abs_path` first, `path` as fallback,
173
+ matching the precedence every other production call site in this codebase already honors;
174
+ pinned by a mutation-verified regression test (reverted, confirmed red, restored, confirmed
175
+ green). The identical gap in the earlier, already-shipped `cascade-rooms-module.cjs` this
176
+ was copied from is left alone -- out of this phase's scope, tracked separately as
177
+ low-severity follow-up debt.
178
+
179
+ ## [1.15.3-beta.46] - 2026-07-23
180
+
181
+ ### Added
182
+ -
183
+
184
+ ### Fixed
185
+ - **The F.8 "bind session to room" Decision Gate (and any PRIMARY-path registry gate)
186
+ force-fired on turns whose ONLY content was an automated background-task-completion
187
+ notification, with zero real user text, blocking continuation with a Stop hook error even
188
+ though there was nothing for the navigator to decide that turn.** Root cause:
189
+ `precedingUserText` resolves to `''` for two very different reasons that
190
+ `scripts/check-card-fire.cjs` could not previously distinguish -- a genuinely terse HUMAN
191
+ turn ("ok", "go on") and a SYNTHETIC preceding transcript record (a `tool_result` envelope
192
+ from a background tool call, or an automated task-notification block) with no
193
+ human-authored text at all. `lib/core/gate-relevance.cjs`'s conservative low-signal branch
194
+ forced (assumed relevant) on both, which is only correct for the first -- there is no human
195
+ turn for a Decision Gate to be relevant or irrelevant to on the second. Fixed by adding a
196
+ new `preceding_user_text_source` signal (`'typed' | 'tool_result' | 'none'`), classified by
197
+ a new `classifyPrecedingUserContentSource` helper in `readTranscriptTurn` and threaded
198
+ through `deriveTurnSignals`; `classifyCardFire`'s PRIMARY-path relevance branch now bypasses
199
+ forcing immediately when the source is confirmed `'tool_result'`
200
+ (`reason: 'preceding-turn-synthetic-no-user-engagement'`), before ever reaching the
201
+ conservative low-signal branch. A genuinely terse human turn against the identical gate
202
+ still force-fires, unweakened (the WR-06/CR-06 floor). Live-reproduced 3 consecutive times
203
+ in one session; the mechanism itself was first diagnosed 2026-07-06 (as a fact, not yet
204
+ treated as a defect) and confirmed recurring 2026-07-11 and 2026-07-22. Full RCA:
205
+ `.planning/debug/resolved/room-bind-gate-fires-on-notification-only-turns.md`.
206
+ - **The semantic-edge derivation queue (`graph-derive-queue.json`) silently cleared itself
207
+ on every failed derivation, in every room, forever -- so no room has ever gotten a real
208
+ INFORMS/CONTRADICTS/CONVERGES cascade edge.** Root cause: `scripts/gsd-graph-derive-drain.cjs`'s
209
+ `drainDerive()` caught a `runDerivation` throw, pushed the failed room to `drained` anyway,
210
+ and rewrote the queue empty with no log -- destroying the retry signal on every SessionStart
211
+ run regardless of outcome. Two of the three originally-stacked causes (a dead standalone
212
+ Anthropic API key; the headless path computing a null single-pair derive with no real
213
+ artifact pairs) had already been resolved upstream by Phase 224-02's switch to local
214
+ embeddings; the silent-clear was the one still live. Fixed via a new `reconcileQueue()`: a
215
+ failed room is now kept and retried (capped at `MAX_DERIVE_ATTEMPTS=5`, then dropped
216
+ `permanent:true` and logged, never silently), and every failure appends to a new
217
+ `<room>/.mindrian/graph-derive-failures.json`. Root-caused by cross-referencing a TDS
218
+ "context rot" research thread against the live debug queue, not filed fresh. Full RCA:
219
+ `.planning/debug/graph-derive-silent-clear-dead-api-derivation.md`. Formalizing the
220
+ derivation transport, healing the ~16 already-damaged rooms, and a doctor health check
221
+ remain open (register items 4b-4e).
222
+ - **`room_search` applied its 50-result cap in raw filesystem-walk order, before any
223
+ relevance ranking -- so a genuinely relevant match in a late-traversed folder could lose to
224
+ 50 incidental early-folder hits, and a same-entity query could return dozens of
225
+ near-duplicate lines from one file.** Root cause: `lib/mcp/tools/room.cjs`'s `searchRoom`
226
+ was a plain `String.includes` grep that pushed matches in directory-entry order and
227
+ returned the instant either `SEARCH_MAX_RESULTS` (50) or `SEARCH_MAX_FILES` (500) was hit --
228
+ capped-then-never-ranked, not ranked-then-capped. Fixed via a rank-then-cap rewrite
229
+ (`collectMatches`/`rankMatches`): match density x0.7 + recency x0.3, with a 5-per-file slice
230
+ cap so one file can't monopolize the result budget. `graph_query` was audited in the same
231
+ pass and found already ranked correctly (a composite relevance score already lives in
232
+ `navigation/neighborhood.cjs`); `whitespace_scan` is unranked but uncapped, deferred pending
233
+ the graph-derive edge-density fix above (little signal to rank while most rooms carry
234
+ near-zero semantic edges). Full RCA: `.planning/debug/graph-query-results-unranked.md`.
235
+ - **The M:OS Canonical Design System v1.1 "bake into all HTML artifacts" mandate was only
236
+ ~30% actually landed, despite its own commit (`a9e1ee88`) and Phase 232-01 claiming it was
237
+ done.** The CSS bundle, loader (`mosStyleTag()`), and `lib/wiki/wiki-layout.cjs`'s
238
+ retokenization were real; `scripts/generate-deck.cjs`, `generate-hub.cjs`,
239
+ `generate-lobby.cjs`, and `generate-snapshot.cjs` had zero reference to `mosStyleTag()`,
240
+ `dashboard/index.html` had zero M:OS tokens, and the mandate's own doc
241
+ (`skills/ui-system/rules/design-system.md`) and `SKILL.md` section 0 did not exist. Wired
242
+ all 4 generators to `mosStyleTag()` (cream default, `data-theme="light"`), injected canonical
243
+ tokens plus a role-based CSS variable alias layer into `dashboard/index.html` and
244
+ `dashboard/export-template.html` (legacy `--mondrian-*`/`--ds-*` names aliased onto canonical
245
+ values in place, not renamed), and authored the missing mandate docs. Code review caught 2
246
+ real regressions before this shipped: (1) 3 of the 4 generators' own pre-existing `<style>`
247
+ blocks redeclared the same token names with old hex values LATER in the document, so the
248
+ mandate rendered in the markup but had zero visual effect by CSS cascade -- fixed by removing
249
+ the colliding redeclarations; (2) `dashboard/index.html`'s dark-to-light polarity flip broke
250
+ hardcoded Cytoscape graph-label colors and hover overlays tuned for the old dark theme --
251
+ fixed by repointing them at the resolved ink values. Both independently re-verified via
252
+ Playwright (`getComputedStyle`) against live-regenerated output, not just diffs.
253
+
254
+ ## [1.15.3-beta.44] - 2026-07-23
255
+
256
+ ### Added
257
+ -
258
+
259
+ ### Fixed
260
+ - **Every Stop hook turn showed a raw Claude Code hook JSON validation error
261
+ ("Hook JSON output validation failed: - : Invalid input") instead of the intended calm
262
+ systemMessage, whenever `scripts/check-card-fire.cjs` force-blocked a turn to demand an
263
+ AskUserQuestion card fire.** Root cause: Claude Code's Stop-hook output schema does not
264
+ define a `hookSpecificOutput` variant for the Stop event at all (the union covers only
265
+ PreToolUse, UserPromptSubmit, and PostToolUse); including the key on a Stop envelope
266
+ rejects the WHOLE envelope (`additionalProperties: false`), not just that key, silently
267
+ replacing the carefully-set `decision`/`reason`/`systemMessage` with a raw schema-error
268
+ dump. This is the 4th live occurrence of the same defect class: fixed once in
269
+ `scripts/on-stop` (v1.10.9 -> v1.10.10, 2026-04-15), then reintroduced in
270
+ `scripts/feynman-minto-guardian.cjs` (under a since-corrected comment that had the rule
271
+ backwards), reintroduced again as a regression inside `scripts/on-stop` itself (the
272
+ Phase 198-09 MCP-first thin-adapter branch), and hit live by a real user via
273
+ `scripts/check-card-fire.cjs` today. Fixed at all 3 sites: `hookSpecificOutput` removed
274
+ outright (and removed from `check-card-fire.cjs`'s own envelope-key allowlist so it can't
275
+ silently slip back in); the calm, human-facing `decision`/`reason`/`systemMessage` fields
276
+ each branch already set are unaffected. New structural regression gate:
277
+ `scripts/check-hook-schema-compatibility.cjs` (previously unwired and, worse, encoding the
278
+ opposite/wrong rule) is corrected and wired into `scripts/verify-release` (section 16) --
279
+ it enumerates every script Claude Code registers as a Stop hook straight off
280
+ `hooks/hooks.json`, follows one level of subprocess invocation, and fails the release if
281
+ any of them would emit a Stop-shaped `hookSpecificOutput` again. Full RCA:
282
+ `.planning/debug/resolved/stop-hook-invalid-hookspecificoutput-schema.md`.
283
+
284
+ ## [1.15.3-beta.42] - 2026-07-23
285
+
286
+ ### Fixed
287
+ - **Windows-only: Python source interpolating shell variables directly (`normwin('$VAR')`) raised a
288
+ `SyntaxError` at Python compile time whenever the interpolated value contained a native Windows
289
+ path with a backslash, before `normwin()` ever ran.** This is the same family of bug as beta.40's
290
+ `os.rename`/`os.replace` fix, but one layer earlier: 36 interpolation sites across
291
+ `scripts/room-registry`, `scripts/resolve-room`, `scripts/update-icm-index`, and
292
+ `scripts/on-cwd-changed` built Python heredoc/`-c` source by quoting a shell variable straight
293
+ into the source string, instead of passing it through `sys.argv` (the safe pattern
294
+ `room-registry`'s own `_write_current_room()` already used). Fixed by converting all 36 sites to
295
+ `sys.argv`-based parameter passing. New regression suite
296
+ (`tests/test-room-registry-windows-python-interp.cjs`, 29/29) includes a load-bearing control
297
+ proving the old interpolation shape fails to compile on a backslash value and the new
298
+ `sys.argv` shape does not. Root-caused and fixed same day as beta.40; independently verified live
299
+ on the reporter's Windows install.
300
+ - **Windows-only: the regression tests' own Python-probe spawn mechanism corrupted probe source
301
+ containing a backslash-next-to-a-quote, at the Windows `CreateProcess` argv-marshalling
302
+ boundary.** Three test files (`test-room-registry-windows-atomic-replace.cjs`,
303
+ `test-room-registry-windows-python-interp.cjs`, `test-room-registry-windows-path.cjs`) each
304
+ routed a Python probe body through `bash -c` as a positional argv element, to survive the two
305
+ bugs above -- safe on Linux/macOS, but Node spawning `bash.exe` on Windows re-quotes argv before
306
+ bash's own `$1` expansion runs, silently mangling the probe. Fixed by writing the probe source to
307
+ a temp `.py` file and executing the file path instead of the source text -- a path has no
308
+ embedded quote/backslash-adjacent-quote sequences, eliminating the defect class rather than
309
+ special-casing which characters are unsafe. Verified green on this dev box (21/21, 29/29, 25/25,
310
+ byte-identical to pre-fix baselines); Windows re-verification of this exact patch is pending.
311
+
312
+ ## [1.15.3-beta.40] - 2026-07-23
313
+
314
+ ### Fixed
315
+ - **Windows-only: the room registry silently wedged after the first write, every write
316
+ after that returning non-zero and never sticking.** Root cause: Python's `os.rename()`
317
+ is not POSIX `rename(2)` on Windows -- it raises `FileExistsError [WinError 183]` when
318
+ the destination already exists, instead of overwriting. Every atomic-write tmp-swap in
319
+ `scripts/room-registry`, `scripts/resolve-room`, `scripts/update-icm-index`, and
320
+ `scripts/on-cwd-changed` used `os.rename(tmp, dst)`, so the first write to a destination
321
+ (cold room creation, first `/mos:rooms list`) always succeeded and looked healthy, while
322
+ every subsequent write (`set-active`, `update`, `archive`, git-config sync) silently
323
+ wedged on Windows, leaving an orphaned `.tmp` and a frozen registry. This repo's own test
324
+ suite runs only under WSL/Linux, where `os.rename` already overwrites happily, so the gap
325
+ was invisible to CI for the test suite's entire lifetime. Found live by a Windows install
326
+ testing v1.15.3-beta.38; fix and both the semantics claim and the end-to-end unwedge were
327
+ independently re-verified live on that same Windows install. Fixed: `os.rename` ->
328
+ `os.replace` at all 9 call sites (byte-identical behavior on Linux/macOS, overwrite-safe
329
+ on Windows). New regression test (`tests/test-room-registry-windows-atomic-replace.cjs`)
330
+ and a new release-time gate (`scripts/verify-release` section 15) now fail the release if
331
+ a bare `os.rename(` reappears anywhere in `scripts/`.
332
+
333
+ ## [1.15.3-beta.38] - 2026-07-23
334
+
335
+ ### Changed
336
+ - **Brain default endpoint flipped from the legacy `mindrian-brain.onrender.com` (Neo4j Aura
337
+ + Pinecone) to the new Memgraph-backed `pws-brain-mcp.onrender.com`** (`lib/core/brain-client.cjs`).
338
+ Step 4 of the approved phased Memgraph migration (step 1, an auth-header double-Bearer-prefix
339
+ fix, shipped dark in beta.36's line). Live-verified against the real production endpoint with
340
+ a real key before flipping: `brain_search`/`brain_schema`/`brain_stats`/`brain_ask`/
341
+ `brain_ask_anything` all return real data (28k+ nodes) through the exact same response shapes
342
+ this client already parses. `brain_query`/`brain_write` (raw Cypher) remain admin-tier gated on
343
+ the new server exactly as they already were on the old one -- regular users never had raw-Cypher
344
+ access on either server, so this introduces zero regression; every caller already degrades
345
+ gracefully to a Tier-0 fallback. Every existing Brain API key works unchanged (same Supabase
346
+ `brain_api_keys` Bearer contract) -- no action required from any user. `MINDRIAN_BRAIN_URL`
347
+ still overrides the default for staging/self-hosted use, unchanged.
348
+
349
+ ## [1.15.3-beta.36] - 2026-07-22
350
+
351
+ ### Fixed
352
+ - **`room_bind` could never write a per-session room binding on stdio, so all CLI sessions
353
+ fell back to one shared, unlocked `registry.json` active-room field.** Root-caused via a
354
+ full RCA (`.planning/debug/resolved/registry-active-room-concurrent-session-collision.md`):
355
+ `writeSessionBinding` -- the only function that sets a session's own room binding -- has
356
+ exactly two call sites, both gated behind `room_bind`'s `effectiveSessionId` check, which
357
+ requires the MCP SDK's `extra.sessionId` (never populated on stdio) or an explicit
358
+ `sessionId` argument (nothing supplied one automatically). So no CLI session could ever
359
+ populate its own binding, and every session's write-target resolution fell through to a
360
+ single global field that multiple concurrent `claude` CLI processes on one machine then
361
+ raced to overwrite -- confirmed live with 4 concurrent sessions on this machine, one
362
+ session's active room silently clobbering another's. Fixed: `room_bind` now falls back to
363
+ `process.env.CLAUDE_CODE_SESSION_ID` as a third-priority session identifier on stdio
364
+ (precedence: explicit param > SDK `extra.sessionId` > `CLAUDE_CODE_SESSION_ID` >
365
+ `no_session_id`), so a CLI session can finally write a real per-session binding. The
366
+ separate F.8 binding-ambiguity-card logic in the same handler is untouched. New test:
367
+ `tests/test-room-bind-stdio-session-fallback.cjs` (4 assertions); all 21 pre-existing tests
368
+ touching `room_bind`/`tool-router.cjs`/`session-binding.cjs` still pass.
369
+ - **`write-scope-check.cjs`'s own session-identity fallback checked the wrong environment
370
+ variable name.** It read `process.env.CLAUDE_SESSION_ID`, which this runtime never sets;
371
+ the real variable is `CLAUDE_CODE_SESSION_ID`. Fixed with a backward-compatible fallback
372
+ chain (`CLAUDE_CODE_SESSION_ID` first, legacy `CLAUDE_SESSION_ID` second) so existing test
373
+ fixtures that set the old name are unaffected. A smaller contributor to the active-room
374
+ confusion above than first estimated (this hook already preferred the hook payload's own
375
+ `session_id` ahead of the env check), but a real, worthwhile correctness fix on its own.
376
+
377
+ ## [1.15.3-beta.34] - 2026-07-21
378
+
379
+ ### Fixed
380
+ - **Stop hook (`check-card-fire.cjs`) force-fired a stale Decision-Gate card on unrelated
381
+ terse turns.** Fourth live occurrence of the over-enforcement class (dominant reason
382
+ `reached-registry-gate-no-card`, 30 of 41 records in a 24h diagnostic window). Two stacked
383
+ root causes: (1) `lib/core/card-fire-sidechannel.cjs`'s reach-mint record had no
384
+ session/turn scoping and a 10-minute TTL, so one real gate mint anywhere leaked into every
385
+ later turn, every session, for 10 minutes; (2) `lib/core/gate-relevance.cjs`'s
386
+ `gateTopicallyRelevant` defaulted to force-fire whenever the preceding user text carried
387
+ fewer than 2 subject tokens -- true for nearly every terse slash command, so short turns
388
+ were the LEAST protected against a stale gate. Fixed structurally: a turn-scoped freshness
389
+ window (`TURN_FRESH_MS`) replaces the unscoped union, and the relevance floor now checks
390
+ gate staleness (`opts.gateStale`) instead of defaulting to force on low signal. A model
391
+ that has already judged a reach-card gate stale and moved on in prose is no longer
392
+ overridden by the hook. Verified end-to-end against the live incident shape; full
393
+ card-fire/gate-relevance/connector-registry suites green.
394
+
395
+ ## [1.15.3-beta.32] - 2026-07-20
396
+
397
+ ### Added
398
+ - **BlockNote Wiki Convergence (Phase 232): `/mos:wiki` gets a real editing surface.** The wiki
399
+ now opens to a Room Home dashboard (governing thought, Larry's Briefing, gaps, per-section
400
+ progress) instead of the graph, and every article is directly editable in a BlockNote surface
401
+ themed to the M:OS Canonical Design System -- edit, click Save, the change writes straight to
402
+ the room's `.md` file (no confirmation dialog, no conflict check, by design). `[[wikilinks]]`
403
+ render as clickable pills inside the editor; Backlinks and See Also stay wired to the existing
404
+ SQLite graph edges. New: per-article PDF and Word export, and a real `/mos:wiki --export`
405
+ static-share bundle (previously documented, never implemented). The client bundle (React +
406
+ BlockNote) is walled off in its own `lib/wiki/editor-src/` build, so the plugin's own
407
+ dependencies stay CJS-only -- zero React/Next.js/BlockNote added to the install footprint.
408
+ A live browser walkthrough caught and fixed two integration bugs before ship (a save/load URL
409
+ encoding mismatch, and a JSON-vs-plaintext response contract mismatch that would have written
410
+ raw JSON into article files instead of markdown).
411
+
412
+ ## [1.15.3-beta.30] - 2026-07-20
413
+
414
+ ### Added
415
+ -
416
+
417
+ ## [1.15.3-beta.28] - 2026-07-18
418
+
419
+ ### Added
420
+
421
+ - **Eureka: killed two distinct causes of unusable portfolio-scan output, plus a warm-cache
422
+ MCP path.** Live-verified on two independently-chosen real rooms
423
+ (`aion-eureka-synergy`, `iia-deeptech-centers`), not just fixture-green.
424
+
425
+ - **Seam 2 (statement-metadata gap, RESOLVED).** Every entity-entity Opportunity Statement
426
+ was rendering the literal placeholder text "unknown x unknown approach to a unknown x
427
+ unknown cross-domain bridge" instead of a real mechanism. Root cause: Phase 218 wired
428
+ entity nodes (company/technology/market) into the 215 opportunity-statement pipeline but
429
+ patched only the `title` slot for that node class, leaving `section`/`primary_problem`/
430
+ `problems`/`shared_problems` falling through to content-node defaults entity nodes
431
+ structurally can't satisfy. Fixed in `lib/core/eureka/room-native-substrate.cjs`: the
432
+ entity-node branch now inherits `section` from its already-shipped `DESCRIBES` edge to its
433
+ source memory_artifact (55/56 entities have one; pure composition over an existing edge,
434
+ sibling of the prior title fix), with an `entityType` fallback for the remainder, plus a
435
+ relation-edge-typed bridge label (`competes-with`/`uses-component`/`supplies-to`) in
436
+ `scripts/eureka-portfolio-report.cjs` instead of the generic phrase. 22/25 -> 0/25
437
+ "unknown x unknown" statements on both proving rooms.
438
+ - **Seam 3 (candidate-generation gap, RESOLVED).** Real content was getting ranked against
439
+ its own containing section (`problem-definition` x `problem-definition`) because `Section`
440
+ container nodes -- the room's own top-level folder nodes -- were admitted as pairing
441
+ candidates. The critic already had a correctly-firing rejection tag for this
442
+ (`domain_swap_invariant`) but nothing upstream excluded these pairs before they consumed
443
+ ranked-list slots. Fixed with an additive either-endpoint `Section` exclusion at the same
444
+ generation-layer insertion point the 260715-0nj scaffold-pair fix established (Reuse
445
+ Before Build), with an honest `container_pairs_excluded` counter surfaced in provenance.
446
+ 1,575 degenerate pairs (9.8% of the candidate set) excluded on `iia-deeptech-centers`;
447
+ proven a true no-op on `aion-eureka-synergy` (0 Section nodes there) by output diff, not
448
+ by assertion.
449
+ - **New: `eureka-run`/`eureka-status`/`eureka-report` on the `intelligence` MCP router
450
+ tool.** Calls the same governed dispatcher (`scripts/eureka-command.cjs` `main(argv)`)
451
+ in-process instead of spawning a fresh child process per call, so
452
+ `embedding-spine.cjs`'s existing module-level encoder cache stays warm across scans on
453
+ the Phase 198 resident daemon (spiked: same-process call 2 is 0ms vs. call 1's 179ms,
454
+ unmodified cache mechanism). Transport-gated per call: in-process on the http resident
455
+ daemon; detached child on stdio, since `process.stdout` is the JSON-RPC framing channel
456
+ there and the scan writes progress to stdout. Registered on the `intelligence` tool's
457
+ enum only, outside the 65-command CLI/MCP parity array (mirrors the `eureka_critic`
458
+ precedent). `/mos:eureka`'s CLI behavior and output contract are unchanged (verified
459
+ byte-identical).
460
+
461
+ Flagged, deliberately out of scope for this pass: Seam 1 (entity-extraction noise --
462
+ generic-noun and near-duplicate entities still reach the ranker on `aion-eureka-synergy`,
463
+ tracked separately) and a newly-observed WhitespaceZone-dominated pairing pattern on
464
+ `iia-deeptech-centers` (AHP composite doesn't differentiate a room where ~87% of nodes are
465
+ whitespace hypotheses) -- both real, both future work, neither papered over.
466
+
467
+ - **Phase 230: MindrianOS Skill Fleet Optimization -- the harness for testing whether
468
+ MindrianOS's own 124 skills trigger correctly and stay quiet when they shouldn't.**
469
+ Two workstreams. WS1 (trigger-accuracy, all 124 skills): per-family eval-query generation
470
+ exploiting sibling near-misses, a roster-wide judge funnel (one call scores a query against
471
+ all 124 skill descriptions at once, catching competitive collisions isolated per-skill
472
+ grading structurally cannot see), flagged skills escalate to a real live trigger-test loop
473
+ with train/validation-gated description revision. WS2 (code-quality, the ~59 script/
474
+ workflow-backed skills -- the design estimated ~10-20, the real inventory came in ~3x
475
+ higher, disclosed rather than silently re-scoped): adversarially-verified review
476
+ (Refute-or-Promote) with a deterministic evidence-quote anchor so a fabricated finding
477
+ cannot reach the report. Live-smoke-tested end to end on a 13-skill human-approved
478
+ calibration set (`scripts/skillopt-*.cjs`, `lib/core/skillopt-schemas.cjs`,
479
+ `tests/run-all-230.sh`, 9 deterministic legs): the real Skill-fire detector proved correct
480
+ in both directions on fresh live captures (it turned out MindrianOS's own `mos:` skills
481
+ fire via an MCP tool call, not Claude Code's native Skill tool -- caught before anything
482
+ was built on the wrong assumption), and WS2 independently re-discovered the real
483
+ `check-card-fire.cjs` over-enforcement defect (see Fixed, below) with zero false positive
484
+ on a known-clean control. The smoke calibration gate itself came in under tolerance (30%
485
+ agreement vs. an 85% bar) -- accepted as informative, not blocking, since most of the gap
486
+ is real full-roster collisions a human's isolated pre-labels couldn't see plus one disclosed
487
+ query-labeling bug; the reconciliation (fix the labeling bug, re-run smoke) is tracked as
488
+ SEED-061, not silently dropped. **The full 124-skill fleet run and any multi-agent
489
+ Workflow-tool orchestration are explicitly deferred behind a future opt-in -- this release
490
+ ships the harness, not a fleet run.** Nothing was ever written to a real `SKILL.md` or
491
+ script; every proposed change surfaces in a human-approved report only.
492
+
493
+ ### Fixed
494
+
495
+ - **`check-card-fire.cjs` no longer force-fires the Decision-Gate card on plain prose with no
496
+ actual gate.** Two independent over-fire mechanisms, logged three times across 12 days
497
+ (2026-07-05, 2026-07-11, 2026-07-17) before being root-caused against a live 17-record
498
+ intercept-log replay: (1) the backstop's bare numbered-prose detector had a 7/7
499
+ false-positive rate in the logged evidence and zero true catches -- retired outright,
500
+ genuine ASCII-box degrades stay caught by the separate bracket-arm detector, unchanged;
501
+ (2) the primary registry-gated path fired on `ran_entries` alone, which a side-channel
502
+ session-key/TTL union bled into every turn for roughly 10 minutes regardless of relevance
503
+ -- now requires a confirmed, non-empty gate-subject plus relevance against that real
504
+ subject. Verified against all 7 real logged firings (0/7 re-fire) plus the full
505
+ card-fire-specific suite (11/11 + 27 assertions). Trade-off, disclosed not hidden: a lone
506
+ genuine numbered-prose fork no longer force-fires at the hook level and now depends on the
507
+ model's own Phase-210/SEED-021 judgment -- the same trust boundary the existing
508
+ under-firing watch (`feedback_false_success_silent_skip_gates_academy_testers.md`) already
509
+ tracks from the opposite direction.
510
+ - **Per-session room binding no longer re-prompts every turn after a real bind.** The MCP
511
+ `room_bind` tool wrote the session's binding state keyed by the actual Claude session UUID;
512
+ the CLI `UserPromptSubmit` hook read it keyed by `process.env.CLAUDE_SESSION_ID`, which is
513
+ unset in that hook's execution context, so it silently fell back to a
514
+ `sha256(roomDir+day)` hash key that never matched -- confirmed with an exact hash-vs-
515
+ on-disk-filename proof, not inferred. The hook now reads the real session id from its own
516
+ stdin payload first. This also un-breaks Phase 225's zero-score gate (SEED-039), which
517
+ shared the same key-mismatch root and was never separately regressed -- just never covered.
518
+ - **The reach/navigation dial no longer offers a topically-unrelated room or claim with no
519
+ relevance check.** `cross_room` was a permanent member of the reach candidate bank, always
520
+ offered in a cold room's top-3 regardless of what the live conversation was actually about,
521
+ filtered only by advisory instruction text the model had to apply itself. A structural
522
+ relevance gate (`lib/hmi/reach-relevance-gate.cjs`) now suppresses off-topic candidates by
523
+ token overlap against the live turn before they're ever offered; `cross_room` also no
524
+ longer "borrows from itself" (filling its own room-name slot with the current room).
525
+
526
+ ## [1.15.3-beta.26] - 2026-07-16
527
+
528
+ ### Added
529
+ - **Phase 227: Ignite / mode-select timing across turns 1-4 (SEED-060).** The session-start
530
+ mode-selection Decision Gate (Just Talk / Explore+Capture / Build a Room) gets a structural,
531
+ advisory-only backstop: a new `doctor.cjs` check class (`lib/core/mode-select-sidechannel.cjs`
532
+ + `lib/core/doctor/mode-select-checkpoint-module.cjs`) detects a silent skip (the gate neither
533
+ firing a card nor stating a default) and warns, never blocks, never re-fires the gate itself.
534
+ A systemic sweep of methodology skills for the same loose-description auto-fire bypass that
535
+ let `trending-to-absurd` reach for itself on a casual remark before its 2026-06-24 fix found
536
+ and closed 3 trivial instances inline (`MOSDeckEngine`, `client-discovery-interview`,
537
+ `mullins-scaffold`), with the rest reported and explicitly deferred. A scripted regression
538
+ test (`tests/test-227-frontdoor-restraint.cjs`) now proves the front-door restraint fix holds
539
+ without needing a live human tester re-run. `skills/larry-personality/SKILL.md` names ignite
540
+ for the first time and documents the gate's timing with real Hooked-Model (Fogg B=MAP / TARI)
541
+ reasoning: the gate is a Prompt, not an Investment, fired only when the navigator's opener
542
+ does not already signal a lane. `conversation-mode`'s Mode 3 (Build a Room) now routes through
543
+ ignite's Directive/`--express` path instead of calling `/mos:new-project` directly, correctly
544
+ reserving Gate B1's four-door persona pick for sessions that genuinely have not yet
545
+ established a role or venture. Same-day code review (independently re-verified, not
546
+ self-certified) caught and fixed two real defects before this landed: Mode 3's routing text
547
+ originally claimed established context unconditionally even on a cold direct pick with zero
548
+ prior exchange, and the mode-select "card-fired" recorder was wired against text no live code
549
+ path actually renders through `pickShape()`. Both closed at the design level in this same
550
+ pass (independently re-verified: 8/8 + 4/4 tests pass, both diffs re-read after landing),
551
+ not patched around the symptom.
552
+
553
+ - **Phase 229: HUJI Pitch Feedback Module (IN PROGRESS, 8/9 plans -- not yet shipped).**
554
+ MindrianOS's first paying job: turn each student's diarized 5-minute pitch transcript into
555
+ one Minto-structured formative feedback artifact, batch-orchestrated across 200+ submissions
556
+ at a $4-5/unit cost ceiling, local-only scoring (Brain read-only, generic handles per Canon
557
+ Part 8). Built so far: the evidence/feedback zod schemas plus generated JSON Schema
558
+ (`229-01`); a labeled test-inventory harness and `run-all-229.sh` aggregator (`229-02`);
559
+ deterministic code checks covering quote verification, recall, drift, schema validation,
560
+ similarity, cost, and Part-8 hygiene (`229-03`); the `PWS_grading` recipe with a
561
+ score-and-continue rubric (`229-04`); a Stage A intake adapter porting the Claims-Aware
562
+ Fusion Mode A pipeline (`229-05`); an LLM judge spawner with a calibration protocol that
563
+ fails closed below a 0.7 anchor-hygiene bar (`229-06`); and a single-submission runner plus
564
+ batch orchestrator with pool/ledger/resume/retry and G1-G6 per-unit guardrails (`229-07`,
565
+ `229-08`). Not yet shipped: `229-09`, the mandatory demo run and human verdict checkpoint
566
+ this whole pipeline is gated on before any real submission gets scored -- per this project's
567
+ own standing rule, judge accuracy and calibration are a human-verify bar, never
568
+ self-certified.
569
+
570
+ ## [1.15.3-beta.24] - 2026-07-16
571
+
572
+ ### Added
573
+ - No discrete feature completed exactly at this tag. This pre-release snapshot captured
574
+ in-progress work on Phase 227 (the mode-select firing checkpoint, sweep, and Mode 3 routing
575
+ work) and Phase 229 (the HUJI Pitch Feedback Module's Stage A intake adapter) mid-flight,
576
+ neither phase-complete at cut time. Phase 227's full, finished feature set is recorded once
577
+ under `[Unreleased]` above rather than split and duplicated across the interim tags it
578
+ happened to span.
579
+
580
+ ## [1.15.3-beta.22] - 2026-07-16
581
+
582
+ ### Added
583
+ - No discrete feature completed exactly at this tag either, for the same reason as beta.24:
584
+ a pre-release snapshot mid-flight through Phase 227 (the skill-description sweep landed
585
+ here) and Phase 229 (PWS_grading recipe + score-and-continue rubric work).
586
+
587
+ ## [1.15.3-beta.20] - 2026-07-16
588
+
589
+ ### Added
590
+ - **Backfilled here (this changelog under-documented this tag at cut time): five phases that
591
+ actually shipped in this release window.**
592
+ - **Phase 222 (reach-ranking-unification).** The three surfaces that suggest a next move
593
+ (`/mos:suggest-next`, the reach-candidates list, and the auto-fire engine) now always
594
+ agree on the top pick instead of occasionally diverging, and the ranking improves over
595
+ a room's own accept/reject history rather than staying static.
596
+ - **Phase 223 (jtbd-driven-intelligence-pipeline, governed bono).** `/mos:bono` is now an
597
+ 8-phase governed research debate with Six-Thinking-Hats-style scrutiny (the Black hat
598
+ must disconfirm first, the White hat must cite-or-retract) and three explicit navigator
599
+ approval pauses (topic, hypothesis, ruling) instead of one collapsed confirmation. New
600
+ `/mos:intel-pipeline` command runs a staged research pipeline oriented on the room's
601
+ active JTBD (calibrate -> decompose -> fan out -> compute -> synthesize -> close), pausing
602
+ for approval twice and disclosing a thin fan-out pass rather than silently proceeding.
603
+ - **Phase 224 (graph-derivation-harness, SEED-034).** Every markdown write to a room now
604
+ enqueues and background-derives typed graph edges, closing the previously twice-reconfirmed
605
+ 0-typed-edge gap on the write path -- no manual "derive" step required.
606
+ - **Phase 225 (per-session-room-binding, SEED-039).** Session-to-room binding now correctly
607
+ supports multi-room binds. Fixed during code review: a binding answer was silently
608
+ collapsing a multi-room selection down to a single room regardless of what was picked.
609
+ - **Phase 226 (eureka-reasoning-mode-fallback, SEED-058).** When the local embedding encoder
610
+ is unavailable, `/mos:eureka` now degrades to a labeled, lower-confidence REASONING MODE
611
+ result (a real short ranked list with an honest caveat naming the degrade cause) instead of
612
+ a hard `pairs_scored: 0` dead end. `banked` is structurally `false` on every reasoning-mode
613
+ row; a later healthy re-run over the same room surfaces the reasoning-to-embedded delta
614
+ instead of silently replacing the earlier result. Same-day code review found and fixed
615
+ three real data-loss edge cases in the upgrade-delta path before this shipped (a repeated
616
+ degrade could silently overwrite a completed reasoning report with no trace; a stale
617
+ session file could let a later healthy run get clobbered; reseeding could silently orphan
618
+ an in-progress judging session).
619
+ - **Eureka entity extraction gets a two-tier WHAT-vs-WHY classifier.** A free, fully local
620
+ embedding pass (`lib/core/eureka/embedding-classifier.cjs`) now resolves the confident
621
+ majority of candidates at zero API spend, reusing the same encoder Eureka's own ranking
622
+ already depends on. The existing LLM classifier is demoted to an escalation-only path,
623
+ called per artifact only for the genuinely ambiguous residual the embedding tier cannot
624
+ confidently place. Measured on a real room: 61.1% of candidates resolve locally and
625
+ correctly, 14.3% fewer artifact-level LLM calls. Honest degrade throughout: no LLM key
626
+ means a disclosed low-confidence embedding best-guess, never a silent default; `classifier_source`
627
+ now reports `embedding` / `model` / `mixed` / `fallback` so every result states which tier
628
+ produced it. Tunable via `MINDRIAN_WHATWHY_MARGIN` (default 0.10, calibrated against a
629
+ measured holdout set). (Quick task 260714-k44.)
630
+
631
+ ### Fixed
632
+ - **A low-confidence WHY term is no longer structurally indistinguishable from a confident one.**
633
+ When Eureka's two-tier classifier places a candidate as a framework (WHY) term but has no
634
+ working LLM to confirm the low-margin embedding best-guess, the term lands in the artifact's
635
+ `framework_terms` prop. The only trace that the guess was low-confidence lived in the aggregate
636
+ `status.json` counter (`tier2_low_confidence`); once written onto the node, the guess looked
637
+ exactly like a confidently-resolved term. Each term that lands via the no-LLM degrade path is
638
+ now disclosed per-term in an additive `framework_terms_low_confidence` sibling prop (always a
639
+ subset of `framework_terms`); a confident later run removes the marker; and existing readers of
640
+ `framework_terms`, which stays a plain comma-joined scalar, are unaffected. The 219 metadata
641
+ test now pins the disclosure so it cannot silently disappear. Caught live by the run-all-221
642
+ regression chain. (Quick task 260715-cu8.)
643
+ - **Eureka's ranked top-25 no longer refills with scaffold pairs when real entities are thin.**
644
+ Every room stores one `memory_artifact` node per file as document scaffolding. When a room's
645
+ real-entity cohort is thin, those scaffolding nodes were pairing with each other and flooding
646
+ the ranked top-25 with `memory_artifact`-vs-`memory_artifact` pairs that carry no cross-domain
647
+ signal (measured at 72.0 percent of the top-25 on a live room once entity extraction correctly
648
+ thinned the entity population). Scaffold-vs-scaffold pairs are now excluded from the ranked-pair
649
+ candidate set by construction, at the point the candidate list is built, so the structural share
650
+ drops to 0.0 percent on the same live substrate regardless of how sparse the real entities are.
651
+ The exclusions are counted honestly in the report provenance (`scaffold_pairs_excluded`, in both
652
+ the JSON and the markdown table), never silently dropped. Pairs with only ONE scaffolding side
653
+ are unaffected, so a real entity paired with the artifact it came from still ranks. (Quick task
654
+ 260715-0nj.)
655
+ - **Eureka's entity-extraction pre-step no longer fails silently.** `/mos:eureka run`'s
656
+ auto-extraction step (shipped in beta.18) could fail (a thrown error, or the more likely
657
+ internally-caught non-zero return) with zero visible trace: exit 0, status `done`, nothing
658
+ in the report to say extraction never actually populated the graph. This reproduced the
659
+ exact false-success shape found in a live intern QA session. Failures on both paths now
660
+ surface as an additive `extraction_error` field in the eureka status.json plus one stderr
661
+ line; ranking, fallback behavior, and exit codes are unchanged (the degrade-never-throw
662
+ contract stays intact, only the silence is gone). Proven via a RED-then-GREEN reproduction
663
+ test wired into the permanent suite. (Quick task 260714-jjm.)
664
+
665
+ ## [1.15.3-beta.18] - 2026-07-13
666
+
667
+ ### Added
668
+ -
669
+
670
+ ## [1.15.3-beta.16] - 2026-07-13
671
+
672
+ ### Added
673
+ - **Opportunity follow-through: surfaced opportunities stop dying as files and one-liners.**
674
+ Every opportunity now flows through the Harvest Formula lifecycle (candidate -> qualified ->
675
+ explored -> promoted | parked | retired) as a real graph node with append-only stage history -
676
+ who advanced it, why, and on what evidence, at every step.
677
+ - **Eureka statements now bank as proposed opportunity nodes.** The portfolio scan's ranked
678
+ statements get a REAL awaited Grounding Guard verdict (a bounded async resolution pass over
679
+ the Phase 212 critic - previously the sync emitter could never await it, so nothing ever
680
+ banked on a live run). Statements the critic passes bank as `opportunity` nodes with
681
+ DERIVED_FROM evidence edges; statements it rejects stay honestly unbanked with the verdict
682
+ named. Tunable via `MINDRIAN_OPPORTUNITY_BANK_PREDICATE` (critic | critic+tail | all).
683
+ - **Harvest sensor (SENS-14): graph events become scored opportunity candidates.** A producer
684
+ on the insight-sensor rail harvests candidates from five lanes (eureka proposals, bridges,
685
+ contradictions, whitespace, meeting filings), classifies each through the Gibson Four-Lens
686
+ (leveraging_resources / challenging_orthodoxies / understanding_needs / harnessing_trends),
687
+ and scores them with HarvestIndex_v1. The bridge lane rides the real extraction edge
688
+ vocabulary (COMPETES_WITH / USES_COMPONENT / SUPPLIES_TO), so it finds genuine cross-entity
689
+ signal on real rooms, not just fixture edges.
690
+ - **Qualification Decision Gate (`/mos:qualify-opportunity`).** Harvested candidates come to
691
+ YOU at a real card showing why each one qualified (Q1..Q8 rubric verdicts + machine-readiness
692
+ components; an unknown is typed `unknown`, never a fabricated zero). Five verbs:
693
+ Qualify+file, Park, Retire, Explore, Skip. A Skip writes a typed REJECTED_BECAUSE edge -
694
+ rejection is data the ranker learns from. Nothing qualifies without your explicit verb.
695
+ - **[Explore]: one explicit action turns a qualified opportunity into deep research**
696
+ (`/mos:explore-opportunity`). Runs the explored-stage chain - deep research, diffusion and
697
+ timing, analogies, web validation - and files a Minto-shaped opportunity artifact (governing
698
+ thought + SCQA + cited sources) into `opportunity-bank/` plus a research corpus artifact into
699
+ `research/`, both through the navigation.cjs gates with typed evidence edges. When the
700
+ engine cannot run, the surface OFFERS an LLM manual fallback at a card - honestly labeled
701
+ `engine_mode: llm_manual_baseline`, never silent, never the default.
702
+ - **Frontmatter metadata extraction slice.** Artifact frontmatter (methodology, status,
703
+ created) now lands as graph properties during extraction, so engines reason over what the
704
+ files already declare.
705
+ - **Web ingestion agent: any URL becomes room knowledge in one governed move.** Paste a link
706
+ in conversation, or run `/mos:research <url>`, and after you approve at the card the page
707
+ is fetched (Tavily Extract, server-side clean markdown), filed as a cited research artifact
708
+ in `research/` (source URL, capture date, content hash, review status: proposed), and its
709
+ entities land in the room graph so every engine can use them. Nothing is ever fetched or
710
+ filed without your explicit verb ([Ingest] [Ingest+Explore] [Skip]).
711
+ - **Pasted-URL sensor (SENS-15).** A bare URL in your turn offers an ingest card
712
+ contextually - it never auto-files, and it stays quiet for URLs inside code blocks,
713
+ quotes, or ones the room already ingested.
714
+ - **Content-hash idempotency + SUPERSEDES versioning.** Re-ingesting an unchanged page is an
715
+ honest no-op; a changed page files a NEW version linked to the prior one - history is
716
+ append-only, nothing is overwritten.
717
+ - **Watched sources: crawl-and-learn on cadence.** Register sources in
718
+ `.mindrian/watched-sources.json` and the scout cadence re-ingests changed pages under a
719
+ per-run cap (default 2), with cadence provenance stamped on every artifact. Findings
720
+ surface as candidates at existing gates - never auto-qualified.
721
+ - **Provider honesty everywhere (research_mode envelope).** Every ingest and research run
722
+ names which provider produced the bytes (tavily-extract / webfetch / manual), which mode
723
+ it ran in (normal / web_degraded_local_fallback / local_only / insufficient_evidence),
724
+ and never reports success with empty results. A failed fetch is a typed refusal, not a
725
+ silent empty.
726
+ - **Part 8 + inbound safety on the new surface.** Outbound carries the URL only through the
727
+ audited egress chokepoint; inbound web content is data end to end (prompt-injection
728
+ inert, size-bounded, path-safe filing, no symlink escape), adversarially test-pinned.
729
+ - **Every research/recovery stage now produces a typed envelope instead of guessing from an
730
+ empty result.** `lib/core/recovery/stage-envelope.cjs` gives all 13 pipeline stages
731
+ (retrieval, discovery, filing, and more) one shared shape: status (ok / empty_valid /
732
+ degraded / failed / blocked), a named failure_class from a frozen 13-class vocabulary,
733
+ retryable, provenance, and timestamps. A zero-result stage and a broken stage used to look
734
+ identical (both "empty"); now they carry different, typed reasons, and a validator enforces
735
+ the pairing rules (a failed/blocked stage MUST name its class; an ok stage MUST NOT).
736
+ - **When a research engine breaks, Mindrian now recovers through a real 6-tier ladder instead
737
+ of just failing.** `dispatchRecovery` reads the typed envelopes and tries, in order: (0)
738
+ nothing wrong, (1) one bounded idempotent retry for a transient failure, (2) a local
739
+ governed substitute (your room's own corpus, or its cache, honestly labeled - never "live"
740
+ when it isn't), (3) an OFFERED high-effort LLM recovery pass at a Decision Gate (never
741
+ silent, never the default), (4) naming the smallest missing thing a human needs to fix (a
742
+ credential, an engine), (5) honest termination when nothing worked - a partial result
743
+ naming exactly which engines are still down, never a complete-looking bundle papering over
744
+ a gap.
745
+ - **Running out of Claude spend mid-recovery is now its own honestly-named failure, not a
746
+ retry loop.** `spend_limit_exceeded` is a structural, first-class failure_class: it forces
747
+ `retryable:false` at the moment the envelope is built (not just checked later), and
748
+ short-circuits straight past every retry/substitute/LLM-recovery tier to a plain human
749
+ message: "raise your limit at claude.ai/settings/usage, or wait for the monthly reset."
750
+ This closes a real gap this exact session hit: four parallel agents stalling out on an
751
+ account spend cap, with no honest way for the system to say so.
752
+ - **The high-effort LLM recovery pass runs through a 7-step, resumable, audited case file,
753
+ never a black box.** A gate-offered recovery run (diagnose -> plan -> execute -> validate
754
+ -> reconcile -> resume -> surface) journals every step to a real case file under
755
+ `.mindrian/recovery/<run_id>/` so a crash mid-run resumes exactly where it left off, never
756
+ re-doing completed steps. Five hard fences, each proven by an adversarial test: the
757
+ Brain-egress boundary can't be weakened from inside a recovery run, an unknown component
758
+ can never be silently upgraded to "supported," every write still goes through the one real
759
+ writer (no raw DB access from a recovery hook), a filing is only ever called "recovered" if
760
+ a readback actually confirms it landed, and hostile text embedded in a source (a fake
761
+ instruction, a fake tool call) is always treated as inert data, never executed.
762
+ - **Recovery outcomes are now honestly composed, never inferred.** `composeRecoveryResult`
763
+ derives one of five outcomes (recovered / partial_recovery / degraded_recovery /
764
+ manual_intervention_required / insufficient_evidence) strictly from what actually happened:
765
+ "recovered" requires every stage envelope to validate AND any attempted filing to be
766
+ readback-confirmed - one unconfirmed filing forces `partial_recovery`, never a false
767
+ "recovered." The result rides as an additive `research_mode` + disclosure field on every
768
+ touched surface (research, opportunity exploration, URL ingestion) without changing any
769
+ existing field's meaning.
770
+ - **A gap in one accessible corpus is never reported as "this doesn't exist."** The
771
+ vantage-error lesson from this exact release wave (an external research pass wrongly
772
+ concluded a shipped phase was "missing" because it only checked one gitignored, unpushed
773
+ corpus) is now a structural, permanent rule: the only gap scope the recovery composer can
774
+ ever emit is `corpus` (a provisional, vantage-scoped gap), never `project` (a claim of
775
+ project-level nonexistence) - enforced by a source scan that fails the build if that ever
776
+ changes, plus a permanent regression fixture that encodes this exact mistake so it can
777
+ never silently return.
778
+ - **14-class recovery matrix, offline and permanent.** Every named failure class (network
779
+ timeout, missing credential, contract violation, policy block, cadence-vs-on-demand,
780
+ multi-engine outage, spend limit, vantage-scoped gap, and more) is asserted end to end
781
+ through the real dispatch and controller seams, with zero network calls - two of the
782
+ fourteen (the vantage rule and spend_limit_exceeded) are locked as PERMANENT fixtures
783
+ precisely because this session discovered both the hard way.
784
+
785
+ ### Fixed
786
+ - **Windows FTS5 crash: eureka degrades bi-modal instead of dying.** On machines whose Node
787
+ SQLite lacks the FTS5 module, the tri-modal index used to crash the whole scan with
788
+ `no such module: fts5`. A capability probe now selects the backend up front: with FTS5 the
789
+ lexical leg runs as before; without it the scan runs honestly on the two remaining legs
790
+ (vector + graph) and stamps `fts_backend: absent (bi-modal degrade)` in provenance. Never a
791
+ crash, never a silent lie. Live-validated on the exact Windows machine that exposed the bug
792
+ (corepower-isolation, 219-VERIFICATION.md Section 4).
793
+
794
+ ## [1.15.3-beta.14] - 2026-07-12
795
+
796
+ ### Added
797
+ -
798
+
799
+ ### Fixed
800
+ - **The card-fire backstop no longer force-fires on benign numbered lists** (`scripts/check-card-fire.cjs`). The `ASCII_BOX_GLYPH_RE` alternative-4 shape (a bare `1. / 2.` numbered-prose list, added Phase 209-07) matched ANY 2+-item list on shape alone, so an ordinary Action Footer or a step-by-step explanation that shared any incidental vocabulary with the user's turn was misread as an unfired Decision Gate and hard-blocked. New `GATE_FRAMING_RE` co-requirement (CR-05): a bare numbered-prose list counts as a backstop hit ONLY when a choice-framing cue (`?`, or one of `which / would you like / pick / choose / select / type 1`) sits inside the matched span or the ~150 chars before it. Alternatives 1-3 (bracket notation, the `type 1, 2, or 3` literal, the multiline bracket box) stay unconditional, and `ASCII_BOX_GLYPH_RE` itself is byte-identical so the retry-key signature and the Phase 209 regex-matrix tests are untouched. The Phase 209 floor survives: a genuine hand-rolled fork carrying a framing cue still intercepts.
801
+ - **The Stop-hook `reason` slug no longer reaches the user as a fake "Stop hook error"** (`scripts/check-card-fire.cjs`, `buildEnforcementEnvelope`). The 2026-07-05 fix added a `systemMessage` on the premise that Claude Code surfaces `reason` as "Stop hook error: <reason>" ONLY when no `systemMessage` is present; live observation proved that premise FALSE (it renders `reason` regardless). The only lever is the `reason` CONTENT, so it is now a calm, human-safe phrase on BOTH the intercept and degrade branches (CR-06), never the internal slug. The slug is preserved for telemetry, relocated to the new local diagnostic log (below), not deleted. Confirmed `turnContextHash` never reads `reason`, so the bounded-escape retry key is unaffected (asserted by a new non-effect test).
802
+ - **New local-only intercept diagnostic log** (`~/.mindrian/card-fire-intercepts.log`, CR-07). Append-only JSONL written whenever the backstop intercepts or degrades, capturing `{ timestamp, session_id, reason (the ORIGINAL slug), gate_signature, ran_entries, matched_glyph_span, output_text (truncated ~4000 chars) }`. TTL-pruned on every write by the same `RETRY_TTL_MS` the retry side-file uses, so it cannot grow unbounded. Canon Part 8: LOCAL disk only (`~/.mindrian`), never the Brain, never a network wire. This turns the still-open "unexplained backstop trigger" mystery (`live-session-running-stale-plugin-cache-fixes-inert`) into a one-log-read diagnosis on its next occurrence.
803
+ - **`check-card-fire.cjs`'s `gate-is-simple-binary` exemption swallowed genuine two-option forced-choice forks, not just yes/no closers.** The exemption (added 2026-07-05 to stop over-firing on trivial binaries like "Want those?") used a bare `gateLabels.length === 2` cardinality check, which cannot distinguish that from a genuine two-way strategic fork ("run research vs build the plan"). An intern QA session missed 3 such forks in one session because each carried exactly 2 option labels. Now requires the labels to be YES/NO-SHAPED (new `lib/core/gate-relevance.cjs::isYesNoShapedGate`, extracted from `gateAlreadyAnswered`'s existing yes/no answer-matching), not merely 2-in-number, so a real yes/no closer stays exempt while a genuine 2-option fork force-fires like a 3+-way one. Post-merge integration fix: `GATE_FRAMING_RE` widened with a cardinality + choice-noun cue ("two options", "3 paths") so a fork phrased without a literal `?`/`which`/`pick` still force-fires, closing a gap CR-05 would otherwise have silently reopened. `.planning/debug/intern-w1-card-discipline-decay.md`.
804
+ - **MCP `room_state` reads re-resolve the active room per call, same as `room_content` writes** (`lib/mcp/tool-router.cjs`). The `status` / `analyze` / `compute-state` / `get-state` / `suggest-next` branches read the boot-time closure `roomDir` directly, so a mid-session room switch (or a room created after the MCP daemon booted) was invisible to them and `status` falsely reported "No room initialized" against a room with real content. Now reuses the same `resolveWriteTargetDir` resolver the beta.12 `room_content` write fix already proved correct. intern-w1-room-state-false-empty.
805
+ - **`intelligence:research` now actually fetches instead of echoing its own command spec** (`lib/mcp/tool-router.cjs`). The `research` sub-command fell through to the generic `buildContext()` doc+state-echo helper (the same fallback reasoning-only ops like `grade`/`whitespace` correctly use), so it deterministically returned `commands/research.md`'s own frontmatter and spec text for any input, with zero web fetch. Now special-cases `research` to invoke `research-context-extractor.cjs` -> `source-lens-driver.cjs` (Stage 1-4) and return real findings with source/url/evidence-tier; filing (Stage 6-7) stays a human decision per Canon Part 9 role 5, never auto-wired inside a single MCP call. `commands/mva-brief.md` (the only other `reach_id: deep_research` command) checked and confirmed unaffected -- it runs its own Bash script, not this tool. intern-w1-research-reach-broken.
806
+ - **STATE.md is now actually recomputed after a room-section write, not just reported as recomputed** (`lib/core/intelligence-cascade.cjs`, `lib/core/state-ops.cjs`). `scripts/compute-state` only prints the STATE.md body to stdout by design; it never writes the file itself, so every caller owns persistence. The automatic PostToolUse cascade (Step 8) and the MCP `room_state compute-state` command both discarded that stdout while reporting/implying success, so a filed artifact never updated the room's own intelligence layer in the same turn it was filed. Cascade Step 8 now captures and persists the stdout directly; `state-ops.cjs::computeState()` now persists at the single Node chokepoint, mirroring the pattern already correct in `scripts/on-stop` / `on-task-complete` / `on-agent-complete`. intern-w1-state-not-recomputed.
807
+ - **The session-start mode-selection Decision Gate (`skills/conversation-mode/SKILL.md`) can silently skip with zero detectable signal** (RCA `intern-w1-mode-gate-skip`). Root cause: two converging structural gaps. (1) `scripts/build-render-coverage.cjs::buildMdKeyspace()` walked only `commands/*.md`, never `skills/*/SKILL.md`, so a skill-declared `hitl_shape` Decision Gate could never register in `data/render-coverage-registry.json` -- PRIMARY detection was structurally blind to every skill-declared gate. New `buildSkillKeyspace()` (a third, additive registry keyspace, mirroring the existing commands walk) closes this; `skills/conversation-mode/SKILL.md` now registers as `declared_shape: F.1, wired: true`. (2) `scripts/check-shape-declaration.cjs` had no predicate catching a surface that self-declares BOTH a genuine `hitl_shape` fork AND `connector.excluded:true` (the no-fork exemption) at once -- a direct contradiction of this repo's own CLAUDE.md Part 11 text: "a render-only or pure-capability skill is exempt via its existing connector.excluded:true + reason, never via a fork it does not have." A new predicate now WARNs (advisory, non-blocking per the existing Phase 210 policy) on this exact contradiction. Extending PRIMARY detection to skills also surfaced 5 pre-existing, previously-invisible unwired skill declarations (`MOSDeckEngine`, `client-discovery-interview`, `intelligence-orchestrator`, `mullins-scaffold`, `mva-pipeline`) and 54 additional pre-existing hasShape-and-excluded contradictions beyond conversation-mode -- both are real, tracked findings surfaced for the first time by this fix, out of scope to resolve here, and named in `tests/test-209-declared-implies-wired.cjs`. `scripts/check-card-fire.cjs` (the Stop-hook backstop, the third converging gap in the original RCA) is untouched by this fix.
808
+ - **11 SKILL.md files documented `bash scripts/<name>` as if `scripts/` were skill-local** (`skills/rooms`, `publish`, `new-project`, `setup`, `room`, `file-meeting`, `wiki`, `vault`, `ingest-methodology`, `ignite`, `export`, plus `commands/new-project.md`). The scripts only ever existed at the plugin root, so any invocation with cwd != plugin root failed exit 127. Prefixed all 72 call sites with `${CLAUDE_PLUGIN_ROOT}` (quoted), the proven convention already used in ~38 other SKILL.md files, `hooks.json`, and `.mcp.json`. Also removed the co-located `PLUGIN_ROOT="$(dirname "$(dirname "$(readlink -f "$0")")")"` pattern (`skills/rooms/SKILL.md` Step 2.5, `skills/new-project/SKILL.md`, `commands/new-project.md`, referenced by `skills/ignite/SKILL.md`) -- confirmed broken under the Bash tool's actual invocation mechanism (`$0` resolves to the shell binary, computing `/usr` as the plugin root on every call) -- replaced with `${CLAUDE_PLUGIN_ROOT}` throughout. See `.planning/debug/intern-w1-rooms-skill-script-path.md`.
809
+ - **`/mos:rooms new` could silently fail to create a room while narrating success.** `scripts/resolve-room`'s legacy-fallback branch returned the pre-existing `room/` path with exit 0 (success) whether or not a new room was actually registered, indistinguishable from a real registry hit -- the direct mechanism behind a false "Room's live" claim when no `cv-project/` directory or registry entry ever existed. Added a `--strict` mode: a bare legacy fallback (no `--adopt`) now prints a `FALLBACK:` stdout marker and exits 2, never 0 -- fully backward compatible for every existing caller that omits the flag. Also tightened `skills/rooms/SKILL.md` Step 2's legacy-room adoption prompt to the same "FIRE THE CARD -- mandatory" doctrine `/mos:ignite`'s B1/B2 gates carry, added an explicit warning against narrating room creation before `birthRoom()` returns `{ok:true}`, and fixed the routing note that mislabeled Step 2 as "(name/slug capture)" (Step 1 captures the name/slug; Step 2 is the adoption check). See `.planning/debug/resolved/intern-w1-rooms-new-silent-fail.md`.
810
+ - **`/mos:doctor` Class H (`install-incomplete-module.cjs`) no longer false-positives "install incomplete" on a healthy one-command marketplace install.** Class H never received the `resolveActivePluginRoot()` topology-awareness fix Class A/Class I got in Phase 123 -- it only recognized a legacy `.install-receipt.json` or a user-level `statusLine` block as "healthy," neither of which a marketplace-cache install ever produces. Its `--fix` path then wrote a `~/.claude/settings.json` `statusLine` override pointing at the hardcoded legacy install path, which does not exist on a marketplace-cache-only machine -- silently breaking the statusline for the rest of the session (user-level settings override plugin-level; the broken exec never surfaced an error in chat). Class H now checks topology first and reports healthy without touching the legacy signals. `statusline-visibility-module.cjs` (Class G) also now tests the EFFECTIVE resolved statusline command (the user-level override if present, else the plugin's own), not always the plugin's own file, so the self-heal's re-verification can actually catch a broken override instead of reporting "ok" regardless. The SessionStart self-heal's outer timeout (`scripts/check-onboard-statusline.cjs`) is raised from 4000ms to 10000ms to clear its own nested worst-case spawn budget (8000ms) instead of getting killed mid-repair. Fix re-implemented against the post-Phase-217 `lib/core/doctor/*-module.cjs` files (the original worktree's fix predated that migration). RCA: `.planning/debug/resolved/intern-w1-statusline-room-mismatch.md`.
811
+
812
+ ## [1.15.3-beta.12] - 2026-07-06
813
+
814
+ ### Added
815
+ - **Phase 211 Eureka generator now runs at production scale.** The tri-modal room.db retrieval engine (FTS5 lexical + sqlite-vec vector + RRF fusion, `mdbr-leaf-ir` local embedder) completes end-to-end against a real 2117-node room after two blocker fixes (below). This is the GENERATOR half of the "two-in-a-box"; the critic (Phase 212) is planned, not yet shipped.
816
+ - **Generic `csv-to-idea-graph` export capability** (`scripts/csv-to-idea-graph.cjs`). Turns any relationship CSV (a pairs edge-list plus optional node-enrichment CSV, column-mapped via CLI args) into a De Stijl navigable idea-graph through the shipped dashboard template -- Section-clustered by a chosen grouping column, layer-toggled, every node/edge citation-tagged. Extends the `generate-standalone` export family; zero tenant hardcoding. Hermetic `tests/test-csv-to-idea-graph.cjs` 21/21.
817
+
818
+ ### Fixed
819
+ - **MCP `room_content` writes re-resolve the active room per call** (`lib/mcp/tool-router.cjs`). The MCP server froze its write target at boot-time cwd, so a mid-session `room-registry set-active` never reached it and writes (file-opportunity / create-funding / update-funding-stage) misrouted to the spawn-time room. Now each write branch calls `resolve-active-room.cjs` (the canonical resolver -- this was a fifth active-room guesser never migrated onto it, the exact stale-closure class Phase 212 D5 warns against). Also aligned `opportunitySchema` to `fileOpportunity` (title optional with a title-or-program refine, coerced numerics). Commit `7a84d38b`.
820
+ - **Strict-mode no longer fires false room-switch / session-binding gates** (`lib/core/room-classifier-strict-mode.cjs`, `scripts/intent-classifier.cjs`). A bare numeric menu reply (`NUMERIC_PATTERN` made the verb optional) and product-branded paste blocks (brand tokens credited as room-name matches) both triggered spurious "switch rooms" / "bind session" interruptions on nearly every turn. Verb now required; brand/boilerplate stop-set excluded from name-entity credit. Commit `e23060cd`.
821
+ - **`birthRoom` binds the newborn room into the session write scope** (`lib/core/navigation/room-birth.cjs`). It flipped only the registry active pointer; Phase 194 (PSB) made the per-session bound SET the primary write authority, so a just-created room was BLOCKED for writes. Now unions the new slug into the session binding as primary via the shipped `session-binding.cjs`; `ignite` threads the real `CLAUDE_SESSION_ID`. Commit `3ad78e70`.
822
+ - **Frontmatter schema validator reconciled to the actual writers** (`lib/core/frontmatter-schemas.cjs`). The Phase 88.1-07 schema codified an aspirational vocabulary no scaffold/doc/compute-state writer ever emitted, so the plugin's own output failed its own advisory schema (a Canon Part 6 dog-food self-violation) and polluted the offense log. Relaxed ROOM.md/STATE.md/artifact-default required sets to what writers emit, added a USER.md schema, and split violation messages into missing-vs-unexpected. New reconcile test scaffolds a room and asserts zero blocking violations. Commit `2602c65b`.
823
+ - **Embedding OOM on large-N rooms** (`lib/core/eureka/embedding-spine.cjs`): `embedTexts` embedded the whole corpus in one forward pass (~26.7GB ONNX allocation on 2117 nodes). Now batched (`MINDRIAN_EMBED_BATCH`, default 32). Commit `c222ff7d`.
824
+ - **vec0 offline-load failure** (`lib/core/eureka/vector-store.cjs`): the backend was inferred from stale table existence, so a table from a prior run threw `no such module: vec0`. Now a per-process capability probe selects the backend; confirmed sqlite-vec loads on Node 22 via a `better-sqlite3` allowExtension handle (the >=23.5 floor is `node:sqlite`-only). Commit `73698c73`.
825
+ - **Claim-text persistence + read-side fallbacks** (D15): `writeClaimNode` persists claim `text`; tri-modal index read-side fallbacks for claim/WhitespaceZone/Artifact. Commits `3d1b27a4`, `af24b697`.
826
+ - **`.gitignore` room.db patterns** backing the "never commit room.db" comment (Part 8 hygiene). Commit `a4cd48dc`.
827
+
828
+ ### Housekeeping
829
+ - **JHTV tenant data + JHU-specific tooling relocated out of the product** into the `jhtv-oliver-kuntz` room, with a `.gitignore` leak guard (Canon three-layer: tenant data/tooling lives in the Room, never the Plugin). The reusable graph capability was generalized (see Added). Commit `57bad7ed`.
830
+ - **Planning (not shipped code):** Phases 212 (Eureka Grounding Guard critic, 5 plans), 212.5 (graph substrate), 213-215 (15 checked plans total) and SEED-053 (methodology-chain MCP tool) registered for the next arc. 213/214 execution is gated on the curing-track verdict + 212-05 calibration.
831
+
832
+ ## [1.15.3-beta.10] - 2026-07-05
833
+
834
+ ### Added
835
+ - **Ratification-tracked next-actions now surface in the statusline's `Next:` slot.** Quick task 260705-ui4, motivated by the rethinking-mindrianos standing-consultant room's own unconverted research entries (`ratification_status: proposed` frontmatter) going invisible once the routing engine had nothing else to offer. Reuses the existing `next-move-cache` mechanism end to end (Canon Part 7: no second cache) -- new `lib/statusline/ratification-next.cjs` resolves the active room FRESH per call (avoiding the frozen-roomDir staleness class found at `bin/mindrian-mcp-server.cjs:65`), shallow-scans `research/*/` for `ratification_status: proposed`, and returns an enum/count-only cue (`ratify strong (2 open)`, never entry titles or target prose -- Canon Part 8). `persistFromDecision(decision, opts)` gained an opt-in `opts.fallbackProvider` on the case-3 clear leg only; no-opts behavior stays byte-identical, so the existing clear-semantics tests are unmodified. The statusline's context-percentage color contract (50/65/80 thresholds) and risk chip are untouched -- the `Next:` segment itself carries zero color/ANSI treatment, so the new cue needed none. New tests/test-statusline-ratification-next.cjs 12/12; context-aware 19/19 unmodified; live-signals 10/10.
836
+
1
837
  ## [1.15.3-beta.8] - 2026-07-05
2
838
 
3
839
  ### Added
package/README.md CHANGED
@@ -10,7 +10,7 @@
10
10
  Powered by PWS (Problems Worth Solving), an innovation methodology built and tested through 20 years of teaching by Prof. Lawrence Aronhime.
11
11
  Engineered by Jonathan Sagir.
12
12
 
13
- [![Version](https://img.shields.io/badge/version-1.15.1--beta.1-1E3A6E)](CHANGELOG.md)
13
+ [![Version](https://img.shields.io/badge/version-1.15.3--beta.48-1E3A6E)](CHANGELOG.md)
14
14
  [![License](https://img.shields.io/badge/license-BSL_1.1-C8A43C)](LICENSE)
15
15
  [![Works on](https://img.shields.io/badge/CLI_+_Desktop_+_Cowork-2D6B4A)](#three-surfaces)
16
16
 
@@ -50,7 +50,7 @@ You talk; the room writes itself. Every conversation, every meeting, every decis
50
50
 
51
51
  ### The room surfaces what you cannot see
52
52
 
53
- Every time you add something new, the system compares it against everything already there. Larry tells you what just changed, what contradicts what, what connects to what, and what is now missing. You decide: APPROVE, REJECT (with a reason), or DEFER. The reason becomes part of the room. The next scan is smarter.
53
+ Every time you add something new, the system compares it against everything already there. Larry tells you what just changed, what contradicts what, what connects to what, and what is now missing. You decide: APPROVE, REJECT (with a reason), or DEFER. The reason becomes part of the room. The next scan is smarter. When a scan surfaces an opportunity, it does not stop at a headline: you qualify it at a card, and one explicit Explore turns it into cited deep research filed in your opportunity bank. The web works the same way: paste a URL and, once you approve, the page is filed as a cited source in your room and compared against everything already there. When something breaks mid-research, it tells you exactly what happened and what it tried next, never a silent empty result.
54
54
 
55
55
  ---
56
56
 
@@ -121,12 +121,14 @@ The commands below are internals. You never have to memorize them or type them.
121
121
  /mos:bono # a six-hats research-and-debate team on your question
122
122
  /mos:map-unknowns # hunt the claims you are most confident about, and wrong
123
123
  /mos:file-meeting # paste a transcript, Larry files it
124
+ /mos:research <url> # paste a link, approve the card, the page becomes cited room knowledge
124
125
  /mos:opportunities # what grants match this room right now
126
+ /mos:qualify-opportunity # judge a surfaced opportunity at a card; Explore turns it into research
125
127
  /mos:graph "what is the weakest assumption in my financial model?"
126
128
  /mos:grade # honest assessment against real ventures
127
129
  ```
128
130
 
129
- That is a slice of 107 commands across 14 skills and 9 agents. If you do not know which one to run, that is the normal case. Just talk: Larry reaches for the right one.
131
+ That is a slice of 111 commands across 124 skills and 9 agents. If you do not know which one to run, that is the normal case. Just talk: Larry reaches for the right one.
130
132
 
131
133
  ---
132
134
 
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@mindrian_os/cli",
3
- "version": "1.15.3-beta.8",
3
+ "version": "1.15.3",
4
4
  "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).",
5
5
  "scripts": {
6
6
  "mcp": "node bin/mindrian-mcp-server.cjs",
@@ -19,6 +19,7 @@
19
19
  "CHANGELOG.md"
20
20
  ],
21
21
  "dependencies": {
22
+ "@huggingface/transformers": "^4.2.0",
22
23
  "@ig3/markdown-it-wikilinks": "^1.0.2",
23
24
  "@modelcontextprotocol/ext-apps": "^1.5.0",
24
25
  "@modelcontextprotocol/sdk": "^1.29.0",
@@ -30,10 +31,11 @@
30
31
  "gray-matter": "^4.0.3",
31
32
  "markdown-it": "^14.1.0",
32
33
  "semver": "^7.7.4",
34
+ "sqlite-vec": "^0.1.9",
33
35
  "zod": "^3.25.76"
34
36
  },
35
37
  "engines": {
36
- "node": ">=22.5.0"
38
+ "node": ">=22.16.0"
37
39
  },
38
40
  "license": "BSL-1.1"
39
41
  }