tickmarkr 1.90.4 → 1.90.6

package/dist/adapters/catalog-remote.js

@@ -89,12 +89,20 @@ function providerModels(modelsDev, preferred) {
         : [];
     // Price is provider-specific. A provider-qualified query must never borrow an identically named
     // model from another provider, where subscription and metered costs can differ materially.
-    const selected = preferred ? preferredEntries : entries;
+    // D-OBS-11 follow-up: that rule presumes the hinted provider EXISTS in the catalog. A hint that
+    // matches nothing (kimi's "moonshot" vs the catalog's "moonshotai"/"kimi-for-coding"; cursor has
+    // no provider at all) used to ZERO the search space and blanket-uncover the whole adapter.
+    // Fail open to the full scan instead — advisory evidence with a visible models.dev id beats none.
+    const selected = preferred && preferredEntries.length > 0 ? preferredEntries : entries;
     return selected
         .map(([, provider]) => record(record(provider)?.models))
         .filter((models) => models !== undefined);
 }
 function findModelsDevModel(catalog, provider, modelId) {
+    // CLI namespaces prefix their catalog ids (kimi-code/k3 vs catalog key k3; omp's openai/gpt-4):
+    // after exact key/id misses, retry with the bare segment after the last "/". Deterministic
+    // provider order; first hit wins — acceptable for advisory evidence, never routing.
+    const bare = modelId.includes("/") ? modelId.slice(modelId.lastIndexOf("/") + 1) : undefined;
     for (const models of providerModels(catalog.modelsDev, provider)) {
         const direct = record(models[modelId]);
         if (direct)
@@ -103,6 +111,13 @@ function findModelsDevModel(catalog, provider, modelId) {
         if (byId)
             return byId;
     }
+    if (bare) {
+        for (const models of providerModels(catalog.modelsDev, provider)) {
+            const direct = record(models[bare]);
+            if (direct)
+                return direct;
+        }
+    }
     return undefined;
 }
 function artificialAnalysisRows(value) {

package/dist/adapters/model-lints.js

@@ -11,8 +11,9 @@ export const MODEL_STALE_DAYS = 30;
 const DAY_MS = 86400000;
 // cursor-agent 2026.07.08 reports 193 mostly-parameterized ids (e.g. gpt-5.3-codex-high-fast); filter the `auto`
 // pseudo-model + effort/speed variant suffixes from the unconfigured-lint aggregation ONLY — doctor.json keeps the
-// raw list (verified 2026-07-10). Data stays raw; lints stay signal.
-const LINT_VARIANT_RE = /^auto$|-(fast|minimal|low|medium|high|xhigh)$/;
+// raw list (verified 2026-07-10). Data stays raw; lints stay signal. -max/-none/-thinking joined the suffix set
+// 2026-08-12 (D-OBS-11: cursor's residual lint list was still mostly effort variants of configured bases).
+const LINT_VARIANT_RE = /^auto$|-(fast|minimal|none|low|medium|high|xhigh|max|thinking)$/;
 const LINT_CAP = 5;
 const TTY_LINT_CAP = 3;
 const DEFAULT_STATE_DIR = ".tickmarkr";

package/dist/gates/baseline.js

@@ -51,9 +51,19 @@ const TRAILING_FAIL_RE = /^\s*(?:test\s+)?\S+(?:\s+\([^)]*\))?\s+(?:\.{3}|-{3,})
 // runner sharing the glyph protocol is read without being enumerated; a status strip drawing ✖
 // mid-line inside chrome matches nothing, the same position rule as the other shapes.
 const GLYPH_FAIL_RE = /^\s*(?:(?:[\w@./-]+:\s*)*)✖\s+\S/;
+// Q137s (verbatim capture, dossier run-2 consult dossier, turbo 2.9.14 + pnpm 10): monorepo
+// drivers prefix child output and summarize failures in their own grammar. Four shapes, all
+// positional — `<pkg>:<task>:` prefixes reuse GLYPH_FAIL_RE's prefix idiom; `Failed:`/`ERROR`
+// require the driver's own `pkg#task` identifier or its literal terminus, so prose or drawn
+// chrome containing the words matches nothing (OBS-278 discipline):
+//   `intake-frontend:lint:  ELIFECYCLE  Command failed with exit code 1.`
+//   `Failed:    intake-frontend#lint`
+//   ` ERROR  intake-frontend#lint: command (…) … exited (1)`
+//   ` ERROR  run failed: command  exited (1)`
+const TURBO_FAIL_RE = /^\s*(?:[\w@./-]+:\s*)*ELIFECYCLE\s+Command failed\b|^\s*Failed:\s+\S+#\S+|^\s*ERROR\s+(?:\S+#\S+:|run failed\b)/;
 // Lines that NAME a failing test — the ones worth headlining to the operator. One list, so recognition
 // and reporting cannot drift apart (a shape that blocks but never gets named cost 3 attempts once).
-const namesFailure = (l) => FAIL_ANCHOR_RE.test(l) || RUNNER_FAIL_RE.test(l) || TRAILING_FAIL_RE.test(l) || GLYPH_FAIL_RE.test(l);
+const namesFailure = (l) => FAIL_ANCHOR_RE.test(l) || RUNNER_FAIL_RE.test(l) || TRAILING_FAIL_RE.test(l) || GLYPH_FAIL_RE.test(l) || TURBO_FAIL_RE.test(l);
 const isFailureShaped = (l) => namesFailure(l) || SUMMARY_FAIL_RE.test(l) || ERROR_ANCHOR_RE.test(l)
     || TSC_ERROR_RE.test(l) || LINTER_ERROR_RE.test(l);
 const VOCAB_RE = /\b(?:error|fail(?:ed|ure|ing)?)\b/i;

package/dist/run/consult.js

@@ -2,7 +2,7 @@ import { mkdirSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 import { getAdapter } from "../adapters/registry.js";
 import { bannerShell, paneDispatchCommand } from "../brand.js";
-import { extractVerdictJson, gateExitTrailer, gatePaneName, generateVerdictNonce, verdictNonceLine } from "../gates/llm.js";
+import { dewrapPaneVerdict, extractVerdictJson, gateExitTrailer, gatePaneName, generateVerdictNonce, verdictNonceLine } from "../gates/llm.js";
 import { classifyVerdictCause } from "../gates/verdict-cause.js";
 import { disallowedBy } from "../route/preference.js";
 import { sh } from "./git.js";
@@ -173,7 +173,22 @@ opts = {}) {
                     await driver.close(slot);
             }
         }
-        return parseConsultVerdict(out, nonce);
+        // D-OBS-12 (dossier runs 1-2, 3/3 valid verdicts destroyed): the pane path is a terminal
+        // scrape — a single-line ~2000-char verdict soft-wraps in rendering and brace balance dies.
+        // Judge/review got dewrapPaneVerdict for exactly this (llm.ts, OBS-209 lineage); the consult
+        // seat never did. Headless output is machine-read stdout and needs no reconstruction.
+        const effective = cfg.visibility.llm === "headless" ? out : dewrapPaneVerdict(out, nonce);
+        const parsed = parseConsultVerdict(effective, nonce);
+        if (!parsed.verdict) {
+            // Q144s / OBS-196 parity: an unparseable verdict persists its raw bytes (pre-dewrap,
+            // rendering verbatim) beside the prompt — three human parks carried zero diagnostic
+            // payload because this artifact did not exist.
+            try {
+                writeFileSync(join(dir, `${d.taskId}-${n}${seatIdx > 0 ? `-s${seatIdx}` : ""}-response.txt`), out);
+            }
+            catch { /* evidence persistence must never fail the seat walk */ }
+        }
+        return parsed;
     };
     // v1.54 T1: ranked seat failover. Walk consult.prefer (adapter:model entries) to the first entry
     // whose adapter is in the live channel set; a failed seat or unparseable verdict falls to the next;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tickmarkr",
-  "version": "1.90.4",
+  "version": "1.90.6",
   "description": "Spec in, verified work out.",
   "type": "module",
   "license": "MIT",