create-tradejs 3.1.29-beta.261 → 3.1.29-beta.263

package/dist/index.js

@@ -64,6 +64,7 @@ var PROJECT_SKILL_NAMES = [
   "strategy-period-revalidate",
   "strategy-forward-start",
   "strategy-forward-status",
+  "tradejs-mcp",
   "strategy-risk-scale",
   "strategy-release"
 ];
@@ -399,6 +400,7 @@ var buildProjectFiles = (targetDir, port, infrastructurePorts = DEFAULT_INFRASTR
         scripts: {
           dev: "tradejs-app dev",
           backtest: "tradejs backtest",
+          "mcp:worker": "tradejs mcp-worker",
           doctor: "tradejs doctor --skip-ml",
           "infra-up": "tradejs infra-up",
           "infra-down": "tradejs infra-down"
@@ -425,6 +427,8 @@ PG_PORT=${infrastructurePorts.postgres}
 REDIS_PORT=${infrastructurePorts.redis}
 REDIS_INSIGHT_PORT=${infrastructurePorts.redisInsight}
 `,
+    "CLAUDE.md": "@AGENTS.md\n\nRead the matching complete skill from .codex/skills before using a TradeJS workflow. OAuth login and consent are performed by the human user.\n",
+    "AGENTS.md": "# TradeJS Project\n\nUse .codex/skills/tradejs-mcp/SKILL.md for remote MCP reads. Verify host, user, deployment/account, revisions, periods and artifact checksums. Missing local data never proves no production activity. Start or cancel jobs only on explicit user request. Never initiate OAuth, change live orders, deploy, send notifications or silently fall back to SSH. Read the complete matching skill under .codex/skills for strategy research; it does not grant production permissions.\n",
     ".gitignore": `node_modules
 .tradejs
 .next
@@ -445,13 +449,29 @@ npm run dev
 Open [http://localhost:${port}/routes/dashboard](http://localhost:${port}/routes/dashboard).
 On the first launch, TradeJS asks you to create the local root password.
 
+## MCP for Codex and Claude Code
+
+Connect your client to the app's \`/mcp\` endpoint using OAuth; the human
+performs login and consent. See https://docs.tradejs.dev/guides/mcp and the
+included \`$tradejs-mcp\` skill. Remote reads must verify the host, user,
+deployment/account and evidence checksums. Do not infer production state from
+local Redis. Start background jobs only on explicit request and never silently
+fall back to SSH. MCP has no live trading, deployment or notification tools.
+Run \`npm run mcp:worker\` separately for bounded cached jobs, after generating
+the exact runtime package manifest. Feedback replay also needs isolated Redis
+and read-only Timescale.
+
 ## Codex TradeJS workflows
 
 The project includes the complete checksum-managed TradeJS skill set under
 \`.codex/skills\`. Use \`$strategy-improvement-research\` to orchestrate a full
 bounded lineage, \`$strategy-backtest-research\` for one preregistered core
-experiment, and \`$ai-train-local-research\` for the frozen gate stage. For
-example:
+experiment, and \`$ai-train-local-research\` for the frozen gate stage.
+AI-gate research uses outer 60/40 by timestamp groups: discover on all first
+60%, diagnose three temporal stability blocks inside development, and keep the
+last 40% sealed until candidate rules are frozen. Internal blocks are not
+independent validation. Already opened history remains retrospective evidence.
+For example:
 
 \`$strategy-candidate-report MarketFlushReversal\`
 

package/dist/skill-bundle/.codex/skills/ai-train-local-research/SKILL.md

@@ -385,23 +385,21 @@ Interpretation:
   windows. Use `both` (default) or `direction_aware` for policies that retain
   both approved sides.
 - `yarn ai-pocket-search` is the default pocket discovery tool for future AI-gate rules. It reconstructs current strategy AI payloads, excludes outcome/current gate-output fields by default, shows progress bars, deduplicates equivalent row-selection pockets, and writes a Markdown report under `data/ai/output`.
-- `ai-pocket-search` uses a time-ordered row holdout by default
-  (`--validationSplit 0.25`) and ranks candidates on that holdout. Treat it as
-  tuning evidence, not an untouched test. For direction-specific discovery use
-  `--direction LONG` or `--direction SHORT`. For release research, reserve a
-  timestamp-grouped tail with `--testSplit ... --sealTest`; the report retains
-  only its timestamp/count bounds and excludes its rows from current-gate and
-  pocket economics. Open it once later with a frozen `ai-gate-ablation.mjs`
-  spec. Plain `--testSplit` still prints test metrics and is therefore an opened
-  historical test, not sealed release evidence. Use `--validationSplit 0` only
-  for legacy full-sample exploration.
-- `ai-pocket-search` uses `--coverageMode auto` by default. It keeps the
-  full-history search for non-provider features and runs separate CMC and
-  Coinalyze cohorts over rows where that context is usable. Each cohort gets
-  its own timestamp-grouped train/tuning/test split, and every reported cohort
-  pocket must contain a predicate from that provider family. Use
-  `--coverageMode full` only when intentionally reproducing the legacy single
-  full-period search.
+- `ai-pocket-search` defaults to outer 60/40 (`--validationSplit 0
+  --testSplit 0.4 --sealTest`). Search and ranking use the whole first 60%.
+  Three consecutive timestamp-grouped blocks inside that development partition
+  diagnose temporal stability; they are not independent holdouts and do not
+  refit pocket discovery. Stable candidates rank before sparse and unstable
+  ones within the bounded shortlist. Sparse support remains research-only.
+  Freeze expressions before opening the last 40% once with native ablation.
+  Never refine thresholds on that tail; previously inspected history remains
+  retrospective evidence. Explicit nonzero validationSplit is a legacy option,
+  not the standard. For direction-specific discovery use LONG or SHORT.
+- Provider coverage cohorts inherit the global development/test boundaries.
+  Filter CMC/Coinalyze coverage within each partition; never resplit a later
+  provider window so that global test rows enter discovery. Every cohort pocket
+  must include a predicate from that provider family. `--coverageMode full`
+  reproduces a single full-history search.
 - `ai-pocket-search` uses `--cadenceMode auto` by default. For a sparse train
   partition it scales discovery-only `minSupport` / `minEvents` down from the
   legacy 20 / 10 defaults using the number of independent timestamp events.

package/dist/skill-bundle/.codex/skills/ai-train-local-research/references/gate-ablation.md

@@ -82,8 +82,8 @@ node .codex/skills/ai-train-local-research/scripts/ai-gate-ablation.mjs \
   --file data/ai/export/ai-dataset-liquiditytails-merged-1784296244106-part1.jsonl \
   --variant 'near-ma-and-zone::filter::additionalIndicators.baseContext.regime.trend.priceDistanceToMaSlowAtr <= 1.2 && additionalIndicators.baseContext.structure.liquidityZones.activeCount >= 1' \
   --featurePattern 'priceDistanceToMaSlowAtr|liquidityZones.activeCount' \
-  --validationSplit 0.2 \
-  --testSplit 0.2 \
+  --validationSplit 0 \
+  --testSplit 0.4 \
   --output data/ai/output/liquiditytails-near-ma-and-zone.md
 ```
 
@@ -206,18 +206,19 @@ Use `--featurePattern '<regex>'` to print matching causal paths, availability,
 ranges, and categories. Do not use `--includeGateContext` for discovery; it is
 only for auditing current gate output fields.
 
-For direction-specific discovery, use `yarn ai-pocket-search --direction LONG`
-or `--direction SHORT`. For release evidence reserve an untouched chronological
-tail with `--testSplit ... --sealTest`; the search ranks pockets using only the
-preceding train and validation rows and reports only the sealed test bounds.
-Plain `--testSplit` exposes test metrics and cannot be called untouched after
-the report is read. Open the sealed tail once with the frozen fixed-rule
-ablation.
+The default is timestamp-grouped outer 60/40: `--validationSplit 0
+--testSplit 0.4`. Discovery seals the last 40% by default. Search uses the whole
+first 60%; three consecutive development blocks diagnose stability. Fixed-rule
+ablation reports the same three blocks and opens the outer test only after the
+candidate expressions are frozen. These internal blocks overlap development;
+they are not independent holdouts. Do not adjust rules on the outer test.
+Explicit nonzero validationSplit remains a labelled legacy option.
 
-When several core candidates must be compared, pass the same exact UTC
-`--tuningSince` and `--testSince` boundaries to every candidate ablation.
-Exact boundaries take precedence over ratio splits and keep sparse candidates
-on one calendar partition contract.
+For direction-specific discovery use LONG or SHORT. For multiple core exports,
+freeze one UTC `--testSince` boundary once and pass it to every ablation.
+With no tuningSince, all earlier rows form development and tuning is empty.
+Exact boundaries override ratios; use the parent development timestamp groups
+to choose the shared 60/40 calendar boundary, rather than moving it per candidate.
 
 Use `--windowStart <UTC> --windowEnd <UTC>` to compare candidates over the same
 calendar window. The start is inclusive, and the end is exclusive. Full-period
@@ -233,8 +234,8 @@ available strategy contains shared LONG or SHORT approval/block pockets:
 ```bash
 node .codex/skills/ai-train-local-research/scripts/ai-gate-ablation.mjs \
   --crossStrategy \
-  --validationSplit 0.2 \
-  --testSplit 0.2 \
+  --validationSplit 0 \
+  --testSplit 0.4 \
   --portfolioCapacity 5 \
   --output data/ai/output/cross-strategy-shared-pockets.md
 ```
@@ -323,8 +324,8 @@ node .codex/skills/ai-train-local-research/scripts/ai-gate-ablation.mjs \
   --strategy LiquidityTails \
   --movingAverageStudy \
   --maPeriods 5,10,15,20,25,30,35,40,45,50,55,60,65,70,75,80,85,90,95,100 \
-  --validationSplit 0.2 \
-  --testSplit 0.2 \
+  --validationSplit 0 \
+  --testSplit 0.4 \
   --json \
   --output data/ai/output/liquiditytails-ma-grid.json
 ```
@@ -346,7 +347,7 @@ or parity is incomplete. Candidate ranking uses train and tuning only; the
 timestamp-grouped test tail is reported after selection and remains exposed
 historical evidence after the first run.
 
-`--crossStrategy` requires positive `--validationSplit` and `--testSplit`.
+`--crossStrategy` requires positive `--testSplit` and non-negative `--validationSplit` (standard: 0 and 0.4).
 Opening the historical test tail makes it exposed evidence. Re-running the tool
 on the same cutoff does not make it untouched again. Every candidate remains
 research-only until the exact frozen rule survives timestamps strictly after
@@ -360,7 +361,7 @@ Every report contains:
 
 - baseline and candidate tables for full history, `180d`, `90d`, `30d`, `7d`;
 - q3+/q4+/q5+ summaries, configurable with `--qualityThresholds`;
-- timestamp-grouped, time-ordered train/tuning/untouched-test splits;
+- timestamp-grouped outer development/test split (60/40 by default), with three temporal stability blocks inside development;
 - direction and monthly stability;
 - matched, removed, and added slices;
 - PnL, winrate, PF, Sharpe, Sortino, Calmar, max drawdown, DD ratios, strict
@@ -396,3 +397,8 @@ Run the tool tests after every change:
 ```bash
 node --test .codex/skills/ai-train-local-research/scripts/ai-gate-ablation.test.mjs
 ```
+
+With the standard 60/40 cross-strategy split, the field `tuning` is the last
+third of development, overlapping the discovery partition. It is a diagnostic
+stability block, not an independent validation set. All three development
+blocks are reported separately; the outer test remains distinct.

package/dist/skill-bundle/.codex/skills/ai-train-local-research/references/reporting.md

@@ -101,16 +101,22 @@ of implying that every approval filled.
 
 ### 7. Validation
 
-Do not rename row-based candidate-selection output to an untouched test.
+Default: outer 60/40 by whole timestamp groups. Search uses the entire first
+60%; report three stability blocks within it. Internal blocks overlap
+development and are not independent validation. Label outer test as historical
+when it has already been inspected. Legacy explicit partitions must name their
+actual boundaries. Do not rename candidate-selection output to an untouched test.
 
 ```md
 ### Validation
 
 | Partition      | Rows | Events | Approved N |  WR |  PF | PnL | MaxDD | Max batch |
 | -------------- | ---: | -----: | ---------: | --: | --: | --: | ----: | --------: |
-| train          |      |        |            |     |     |     |       |           |
-| tuning         |      |        |            |     |     |     |       |           |
-| untouched test |      |        |            |     |     |     |       |           |
+| development 60% |      |        |            |     |     |     |       |           |
+| development block 1 | | | | | | | | |
+| development block 2 | | | | | | | | |
+| development block 3 | | | | | | | | |
+| outer test 40% |      |        |            |     |     |     |       |           |
 ```
 
 ### 8. Acceptance checks

package/dist/skill-bundle/.codex/skills/ai-train-local-research/scripts/ai-gate-ablation.mjs

@@ -32,8 +32,8 @@ Options:
   --minQuality <n>             Main baseline threshold (default: 4)
   --qualityThresholds <list>   qN+ summaries (default: 3,4,5)
   --terminalWindows <list>     Terminal windows in days (default: 180,90,30,7)
-  --validationSplit <ratio>    Trailing timestamp-grouped tuning share (default: 0.25)
-  --testSplit <ratio>          Later timestamp-grouped test share (default: 0)
+  --validationSplit <ratio>    Optional legacy tuning share (default: 0; stability stays in development)
+  --testSplit <ratio>          Outer timestamp-grouped test share (default: 0.4)
   --tuningSince <timestamp>    Exact UTC boundary where tuning starts
   --testSince <timestamp>      Exact UTC boundary where test starts
   --windowStart <timestamp>    Common comparison start, inclusive
@@ -91,8 +91,8 @@ export const parseCliArgs = (argv) => {
     minQuality: 4,
     qualityThresholds: DEFAULT_QUALITY_THRESHOLDS,
     terminalWindows: DEFAULT_WINDOWS,
-    validationSplit: 0.25,
-    testSplit: 0,
+    validationSplit: 0,
+    testSplit: 0.4,
     tuningSince: null,
     testSince: null,
     capacities: DEFAULT_CAPACITIES,
@@ -1193,6 +1193,13 @@ export const splitRowsByTimestamp = (rows, validationSplit, testSplit = 0) => {
 };
 
 export const splitRowsByTimestampBounds = (rows, tuningSince, testSince) => {
+  if (tuningSince == null && Number.isFinite(testSince)) {
+    return {
+      train: rows.filter((row) => row.timestamp < testSince),
+      tuning: [],
+      test: rows.filter((row) => row.timestamp >= testSince),
+    };
+  }
   if (!Number.isFinite(tuningSince) || !Number.isFinite(testSince)) {
     throw new Error(
       'Exact calendar partitions require both tuningSince and testSince',
@@ -1210,6 +1217,21 @@ export const splitRowsByTimestampBounds = (rows, tuningSince, testSince) => {
   };
 };
 
+// Three diagnostics wholly inside the development partition, never test selection.
+export const developmentBlocks = (rows) => {
+  const timestamps = [...new Set(rows.map((row) => row.timestamp))].sort(
+    (a, b) => a - b,
+  );
+  return [0, 1, 2].map((index) => {
+    const from = timestamps[Math.floor((timestamps.length * index) / 3)];
+    const until =
+      index === 2
+        ? Infinity
+        : timestamps[Math.floor((timestamps.length * (index + 1)) / 3)];
+    return rows.filter((row) => row.timestamp >= from && row.timestamp < until);
+  });
+};
+
 const summarizeSplit = (rows, selector, summaryOptions) =>
   summarizeRows(selectRows(rows, selector), getPeriodDays(rows), {
     ...summaryOptions,
@@ -3034,6 +3056,9 @@ export const evaluateCrossPocket = ({
     expectedSign,
   );
   const test = summarizeCrossSlice(split.test, pocket.predicates, expectedSign);
+  const developmentStability = developmentBlocks(split.train).map((block) =>
+    summarizeCrossSlice(block, pocket.predicates, expectedSign),
+  );
   const negativeControl = evaluateNegativeControl(
     split.test,
     pocket.predicates,
@@ -3109,6 +3134,7 @@ export const evaluateCrossPocket = ({
   return {
     condition: pocket.condition,
     predicates: pocket.predicates,
+    developmentStability,
     train,
     tuning,
     test,
@@ -3594,9 +3620,9 @@ export const buildCrossStrategyReport = async ({
   if (groups.length < 2) {
     throw new Error('Cross-strategy research requires at least two exports');
   }
-  if (testSplit <= 0 || validationSplit <= 0) {
+  if (testSplit <= 0 || validationSplit < 0) {
     throw new Error(
-      '--crossStrategy requires positive --validationSplit and --testSplit',
+      '--crossStrategy requires positive --testSplit and non-negative --validationSplit',
     );
   }
   let searchAiPockets = searchAiPocketsOverride;
@@ -3731,6 +3757,10 @@ export const buildCrossStrategyReport = async ({
       left.timestamp - right.timestamp || left.sequence - right.sequence,
   );
   const split = splitRowsByTimestamp(rows, validationSplit, testSplit);
+  if (validationSplit === 0) {
+    // This overlaps development deliberately: diagnostic stability, not holdout tuning.
+    split.tuning = developmentBlocks(split.train).at(-1);
+  }
   const minSharedStrategies = Math.min(
     groups.length,
     Math.max(5, Math.ceil(minFeatureStrategies * 0.6)),
@@ -3968,6 +3998,10 @@ export const buildCrossStrategyReport = async ({
       },
       validationSplit,
       testSplit,
+      tuningRole:
+        validationSplit === 0
+          ? 'last development block, overlapping diagnostic; not independent holdout'
+          : 'legacy separate tuning',
       minFeatureStrategies,
       acceptance: {
         minSharedStrategies,
@@ -4015,8 +4049,8 @@ export const buildAblationReport = ({
   minQuality,
   qualityThresholds,
   terminalWindows,
-  validationSplit,
-  testSplit = 0,
+  validationSplit = 0,
+  testSplit = 0.4,
   tuningSince = null,
   testSince = null,
   windowStart = null,
@@ -4054,12 +4088,12 @@ export const buildAblationReport = ({
   }
   const minTimestamp = windowStart ?? rows[0].timestamp;
   const maxTimestamp = windowEnd ?? rows.at(-1).timestamp;
-  if ((tuningSince == null) !== (testSince == null)) {
+  if (tuningSince != null && testSince == null) {
     throw new Error(
       'Exact calendar partitions require both tuningSince and testSince',
     );
   }
-  const exactCalendarPartitions = tuningSince != null;
+  const exactCalendarPartitions = testSince != null;
   const split = exactCalendarPartitions
     ? splitRowsByTimestampBounds(rows, tuningSince, testSince)
     : splitRowsByTimestamp(rows, validationSplit, testSplit);
@@ -4088,6 +4122,12 @@ export const buildAblationReport = ({
         }
       : {}),
   };
+  const stabilityBlocks = developmentBlocks([...split.train, ...split.tuning]);
+  const stabilitySummary = (selector) =>
+    stabilityBlocks.map((block) => ({
+      ...partitionEvidence(block),
+      metrics: summarizeSplit(block, selector, summaryOptions),
+    }));
   const baselineSelector = (row) => baselineSelectedAt(row, minQuality);
   const approvedSignalTrace = (selector) =>
     selectRows(rows, selector).map((row) => ({
@@ -4121,6 +4161,7 @@ export const buildAblationReport = ({
       maxTimestamp,
       summaryOptions,
     }),
+    developmentStability: stabilitySummary(baselineSelector),
     train: summarizeSplit(split.train, baselineSelector, summaryOptions),
     tuning: summarizeSplit(split.tuning, baselineSelector, summaryOptions),
     test: summarizeSplit(split.test, baselineSelector, summaryOptions),
@@ -4184,6 +4225,7 @@ export const buildAblationReport = ({
         maxTimestamp,
         summaryOptions,
       }),
+      developmentStability: stabilitySummary(candidateSelector),
       train: summarizeSplit(split.train, candidateSelector, summaryOptions),
       tuning: summarizeSplit(split.tuning, candidateSelector, summaryOptions),
       test: summarizeSplit(split.test, candidateSelector, summaryOptions),
@@ -4235,12 +4277,14 @@ export const buildAblationReport = ({
       validationSplit,
       testSplit,
       partitionMode: exactCalendarPartitions ? 'exact-calendar' : 'ratio',
+      developmentStabilityRole:
+        'diagnostic inside development; never outer test selection',
+      developmentStabilityFolds: 3,
       comparisonWindow: explicitWindow
         ? { start: windowStart, end: windowEnd, interval: '[start, end)' }
         : null,
-      tuningSince: exactCalendarPartitions
-        ? new Date(tuningSince).toISOString()
-        : null,
+      tuningSince:
+        tuningSince != null ? new Date(tuningSince).toISOString() : null,
       testSince: exactCalendarPartitions
         ? new Date(testSince).toISOString()
         : null,
@@ -4566,6 +4610,14 @@ export const formatMarkdownReport = (report) => {
           report.run.trainEvents,
           report.baseline.train,
         ),
+        ...(report.baseline.developmentStability ?? []).map((block, index) =>
+          validationSummaryRow(
+            `development block ${index + 1} (diagnostic)`,
+            block.rows,
+            block.events,
+            block.metrics,
+          ),
+        ),
         validationSummaryRow(
           'tuning',
           report.run.tuningRows,
@@ -4573,7 +4625,7 @@ export const formatMarkdownReport = (report) => {
           report.baseline.tuning,
         ),
         validationSummaryRow(
-          'untouched test',
+          'outer test (opened)',
           report.run.testRows,
           report.run.testEvents,
           report.baseline.test,
@@ -4685,13 +4737,20 @@ export const formatMarkdownReport = (report) => {
             report.baseline.train,
             variant.train,
           ),
+          ...(variant.developmentStability ?? []).map((block, index) =>
+            validationComparisonRow(
+              `development block ${index + 1} (diagnostic)`,
+              report.baseline.developmentStability[index].metrics,
+              block.metrics,
+            ),
+          ),
           validationComparisonRow(
             'tuning',
             report.baseline.tuning,
             variant.tuning,
           ),
           validationComparisonRow(
-            'untouched test',
+            'outer test (opened)',
             report.baseline.test,
             variant.test,
           ),

package/dist/skill-bundle/.codex/skills/ai-train-local-research/scripts/ai-gate-ablation.test.mjs

@@ -11,6 +11,7 @@ import {
   applyBenchmarkEventSnapshots,
   balanceCrossStrategyRows,
   buildAblationReport,
+  developmentBlocks,
   buildCrossStrategyReport,
   buildEquitySeries,
   buildMovingAverageVariants,
@@ -1453,3 +1454,34 @@ test('uses inclusive UTC calendar days for terminal active-day ratio', () => {
   assert.equal(report.baseline.periods['7d'].activeDays, 2);
   assert.equal(report.baseline.periods['7d'].activeDayRatio, 0.25);
 });
+
+test('outer 60/40 is the CLI default and stability blocks preserve whole events', () => {
+  const options = parseCliArgs([]);
+  assert.equal(options.validationSplit, 0);
+  assert.equal(options.testSplit, 0.4);
+  const rows = Array.from({ length: 30 }, (_, i) => ({
+    timestamp: Math.floor(i / 2),
+  }));
+  const blocks = developmentBlocks(rows);
+  assert.deepEqual(
+    blocks.map((block) => block.length),
+    [10, 10, 10],
+  );
+  assert.equal(new Set(blocks.flat().map((row) => row.timestamp)).size, 15);
+  for (let i = 0; i < 2; i++)
+    assert.ok(blocks[i].at(-1).timestamp < blocks[i + 1][0].timestamp);
+});
+
+test('shared outer boundary permits development/test without separate tuning', () => {
+  const rows = [1, 2, 2, 3, 4].map((timestamp) => ({ timestamp }));
+  const split = splitRowsByTimestampBounds(rows, null, 3);
+  assert.deepEqual(
+    split.train.map((row) => row.timestamp),
+    [1, 2, 2],
+  );
+  assert.deepEqual(split.tuning, []);
+  assert.deepEqual(
+    split.test.map((row) => row.timestamp),
+    [3, 4],
+  );
+});

package/dist/skill-bundle/.codex/skills/backtest-config-redis/SKILL.md

@@ -15,3 +15,12 @@ description: Fetch a TradeJS backtest or strategy configuration from the local R
 - If the container name differs from `inv-redis`, ask for the correct name.
 - For research lineage, embed the returned JSON and a canonical checksum in the
   note. The mutable Redis key alone is not reproduction evidence.
+
+## Remote data through MCP
+
+When a TradeJS MCP server is configured, follow `$tradejs-mcp` for remote reads.
+Check server/user/deployment identity before interpreting the result. Prefer
+verified diagnostic artifacts and validate their SHA-256 after download;
+keep existing local research evidence as the source for candidate selection.
+Use `backtest_list_configs` for remote named grids. Do not initiate OAuth,
+start a new job from this read-only workflow, or silently fall back to SSH.

package/dist/skill-bundle/.codex/skills/runtime-parity-mismatch-analysis/SKILL.md

@@ -144,3 +144,12 @@ For each case:
 4. Give the next 1-3 checks in the TradeJS codebase.
 If several cases share one cause, group them.
 ```
+
+## Remote data through MCP
+
+When a TradeJS MCP server is configured, follow `$tradejs-mcp` for remote reads.
+Check server/user/deployment identity before interpreting the result. Prefer
+verified diagnostic artifacts and validate their SHA-256 after download;
+keep existing local research evidence as the source for candidate selection.
+Use `backtest_list_configs` for remote named grids. Do not initiate OAuth,
+start a new job from this read-only workflow, or silently fall back to SSH.

package/dist/skill-bundle/.codex/skills/strategy-backtest-research/references/research-notes.md

@@ -245,3 +245,9 @@ After creating or editing notes, run:
 ```bash
 node .codex/skills/strategy-backtest-research/scripts/research-notes-check.mjs
 ```
+
+For AI-gate studies record the standard outer 60/40 timestamp-group split,
+three temporal stability blocks inside the full first 60%, and the frozen
+outer test boundary. The internal blocks are development diagnostics, not
+independent validation. Note every previous exposure of the outer history;
+changing the split cannot make previously inspected data untouched again.

package/dist/skill-bundle/.codex/skills/strategy-candidate-compare/SKILL.md

@@ -52,3 +52,12 @@ Conclude with:
 - whether the candidate remains the selected best candidate under the frozen
   objective;
 - one exact next action. Do not launch it from this skill.
+
+## Remote data through MCP
+
+When a TradeJS MCP server is configured, follow `$tradejs-mcp` for remote reads.
+Check server/user/deployment identity before interpreting the result. Prefer
+verified diagnostic artifacts and validate their SHA-256 after download;
+keep existing local research evidence as the source for candidate selection.
+Use `backtest_list_configs` for remote named grids. Do not initiate OAuth,
+start a new job from this read-only workflow, or silently fall back to SSH.

package/dist/skill-bundle/.codex/skills/strategy-candidate-report/SKILL.md

@@ -49,3 +49,12 @@ sparse 7d/30d row must remain visible, but it is not a standalone veto and does
 not imply that anyone must wait for a calendar deadline.
 
 Finish with one exact next skill: compare, plan, revalidate, or forward-start.
+
+## Remote data through MCP
+
+When a TradeJS MCP server is configured, follow `$tradejs-mcp` for remote reads.
+Check server/user/deployment identity before interpreting the result. Prefer
+verified diagnostic artifacts and validate their SHA-256 after download;
+keep existing local research evidence as the source for candidate selection.
+Use `backtest_list_configs` for remote named grids. Do not initiate OAuth,
+start a new job from this read-only workflow, or silently fall back to SSH.

package/dist/skill-bundle/.codex/skills/strategy-forward-status/SKILL.md

@@ -42,3 +42,12 @@ alone is not success or failure; weight the number and diversity of independent
 events. Finish with one exact recommendation: keep collecting, diagnose via a
 new research task, stop through the deployment runbook, or invoke
 `$strategy-risk-scale`.
+
+## Remote data through MCP
+
+When a TradeJS MCP server is configured, follow `$tradejs-mcp` for remote reads.
+Check server/user/deployment identity before interpreting the result. Prefer
+verified diagnostic artifacts and validate their SHA-256 after download;
+keep existing local research evidence as the source for candidate selection.
+Use `backtest_list_configs` for remote named grids. Do not initiate OAuth,
+start a new job from this read-only workflow, or silently fall back to SSH.

package/dist/skill-bundle/.codex/skills/strategy-improvement-research/SKILL.md

@@ -93,7 +93,7 @@ composition ledger, and the mandatory visual artifacts.
    `production core + current AI-gate` composition is the mandatory baseline;
    the production core with its newly rebuilt gate is a separate research
    candidate and must not replace the current-gate baseline.
-   Use common calendar train/tuning/test boundaries and open every sealed gate
+   Use common calendar development/test boundaries (outer 60/40; three stability blocks inside development) and open every sealed gate
    tail together only after all per-core gate variants are frozen. Run every
    gate board through `$ai-train-local-research`; do not retune or relabel the
    raw-core result inside gate tooling.

package/dist/skill-bundle/.codex/skills/strategy-improvement-research/references/final-composition-board.md

@@ -23,15 +23,14 @@ or a gate tuned on another core's signal population.
    is the baseline; the rebuilt gate is a research candidate. A no-op,
    zero-approval, or rejected gate still receives a checksum-bound report and
    explicit disposition; it is not silently omitted.
-4. Freeze common calendar train, tuning, and test boundaries once in the parent
-   lineage. Keep all rows sharing a timestamp in one partition. Candidate
-   exports may have different row counts, but final comparison windows and test
-   timestamps must not move to make one candidate look better. Discovery may
-   inspect only train and tuning. Open all candidate test tails together after
-   every gate spec is immutable.
-   Use `ai-gate-ablation.mjs --tuningSince <UTC> --testSince <UTC>` with the
-   same values for every candidate; ratio-only splits are not comparable when
-   core event cadence differs.
+4. Freeze common calendar development/test boundaries once in the parent
+   lineage: outer 60/40, preserving whole timestamp groups. Search the whole
+   development partition and diagnose three consecutive blocks inside it.
+   These blocks overlap development, not independent validation. Candidate
+   exports may have different cadence, but the outer boundary must not move.
+   Freeze every gate spec before opening all outer test tails together.
+   Use `ai-gate-ablation.mjs --testSince <UTC>` with the same boundary for every
+   candidate; ratio-only splits are not comparable across different cadence.
 5. Count every inspected core and gate behavior in the final trial ledger. The
    selected strategy is `core identity + candidate-specific gate identity +
 context identity + direction policy + quality threshold`. Reusing economics

package/dist/skill-bundle/.codex/skills/strategy-release/references/evidence-retention.md

@@ -79,7 +79,7 @@ Retain:
   attempt over full/3y/4y/max-covered/365d/180d/90d/30d/7d windows;
 - control/candidate matched, removed, added, changed-outcome, occupancy, regime,
   month, symbol, event, concentration, and capacity evidence;
-- gate train/tuning/test boundaries, feature provenance, threshold rounding,
+- gate development/test boundaries (outer 60/40), three internal stability blocks, feature provenance, threshold rounding,
   sealed-test flag/open timestamp, ablation, support, and one-round selection
   decision; record any accidental early test exposure permanently;
 - the complete `$ai-train-local-research` report sections and structured

package/dist/skill-bundle/.codex/skills/strategy-release/references/release-workflow.md

@@ -427,7 +427,7 @@ with a future unexposed tail.
 ## 7. Use one gate tuning round
 
 Freeze the isolated finalist's raw-core export and the current deterministic
-gate as control. Use one time-grouped, time-ordered train/tuning/test design.
+gate as control. Use one timestamp-grouped outer 60/40 design, with three temporal stability blocks inside the first 60%.
 Audit existing gate rules, run pocket discovery/ablation without outcome or
 execution leakage, and preregister rounded thresholds before opening the test.
 `ai-pocket-search` must reserve the test with `--sealTest`; its discovery report
@@ -710,8 +710,8 @@ yarn ai-train --strategy <Strategy> --file <merged-export-part1.jsonl> \
   --terminalWindows=1460,1095,365,180,90,30,7
 
 yarn ai-pocket-search --strategy <Strategy> \
-  --file <merged-export-part1.jsonl> -n 0 --validationSplit 0.2 \
-  --testSplit 0.2 --sealTest --maxDepth 2 --minSupport 25
+  --file <merged-export-part1.jsonl> -n 0 --validationSplit 0 \
+  --testSplit 0.4 --sealTest --maxDepth 2 --minSupport 25
 
 yarn strategy:release profile --input <trades.jsonl> --variant <finalist-id> \
   --startTime <start-ms> --endTime <end-ms> --days 7,30,90 \

package/dist/skill-bundle/.codex/skills/tradejs-mcp/SKILL.md

@@ -0,0 +1,43 @@
+---
+name: tradejs-mcp
+description: Use an authenticated TradeJS MCP server from Codex or Claude Code to inspect coins, charts, runtime strategies, signals, orders, cached backtests, and verified runtime feedback/parity. Use for remote TradeJS inspection that previously required SSH; never authorize OAuth or trade live on the user's behalf.
+---
+
+# TradeJS through MCP
+
+1. Use the configured TradeJS server. If missing, give the human the connection
+   commands from the knowledge base at `/guides/mcp`. Never start OAuth consent,
+   device authorization, re-login or permission elevation yourself.
+2. Call `tradejs_info`, then `runtime_get_status` for runtime work. Verify the
+   host URL, authenticated user, available scopes, account/deployment and
+   observed times. A local server is not production evidence.
+3. Read `market_list_symbols`, `market_get_snapshot`, `chart_get_context` for
+   coins/charts. Preserve provider, universe, symbol, interval and period.
+   Cached history is the default; report incomplete coverage. Strategy figures
+   come from the recorded signal, not inferred chart annotations.
+4. Use `runtime_list_strategies`, `runtime_list_signals`,
+   `runtime_list_evaluations`, `runtime_list_orders` and
+   `runtime_get_evaluation_summary`. Follow cursors to `0`, including empty
+   pages. Runtime cursors expire after ten minutes and are bound to filters.
+   `runtime_get_signal` takes signalId, deploymentId, timestamp and strategy
+   from the list. Explain the saved decision, gate, order status and evidence.
+   Do not turn missing retained records into proof that nothing happened.
+5. Prefer existing verified `diagnostics_list_reports` / `diagnostics_get_report`
+   before new jobs. For full artifacts call `artifact_get`, concatenate decoded
+   base64 chunks in offset order, verify byte count and SHA-256, then use the
+   relevant local analysis skill. Debug counters never replace sealed evidence.
+6. On an explicit user request, use `backtest_start` or `diagnostics_start`,
+   retain the job id and idempotency key, and poll `job_get`. Backtests snapshot
+   the named grid; limits are 90 days, 20 symbols, eight combinations, cached
+   history, one worker. Evidence capture is limited to seven days. Feedback
+   replay requires a verified evidence id for that deployment and an isolated
+   replay worker. Disconnects do not cancel work; never automatically retry a
+   failed job. Use `job_cancel` only when requested.
+7. Include source host, time bounds, revisions, coverage, checksums and missing
+   evidence in the result. Use local CLI for offline analysis or work exceeding
+   the bounded MCP job contract. SSH needs a separate explicit operational
+   request; do not broaden access to compensate for a missing tool or scope.
+
+MCP has no tools for live order placement/cancellation, runtime config writes,
+release publication, deployment or notifications. OAuth scopes do not grant
+permission to perform those actions. Skills and AGENTS.md do not change that.

package/dist/skill-bundle/.codex/tradejs-skill-bundle.json

@@ -1,7 +1,7 @@
 {
   "schema": "tradejs-skill-bundle/v1",
   "source": "TradeJS-Dev/TradeJS:.codex/skills",
-  "bundleSha256": "f31f221397fa300534800c147d1ea57e549627083723acc7c97b75364d31da11",
+  "bundleSha256": "de2e0a452cd29dc40f183a68c246cf6d4af7710c8fac599a6030b6713accedc6",
   "skills": [
     "ai-train-local-research",
     "backtest-config-redis",
@@ -15,51 +15,52 @@
     "strategy-period-revalidate",
     "strategy-forward-start",
     "strategy-forward-status",
+    "tradejs-mcp",
     "strategy-risk-scale",
     "strategy-release"
   ],
   "files": {
-    ".codex/skills/ai-train-local-research/references/gate-ablation.md": "afc8334de0282fd0fac3886f467a1fca8e67c624a48cb5de66351635a039421e",
+    ".codex/skills/ai-train-local-research/references/gate-ablation.md": "2fff61953de59fc0030004a29a3a9dcee2e5783dd47e1ab6dcfac73c0eaae756",
     ".codex/skills/ai-train-local-research/references/pocket-feature-exclusion.md": "c7af65fcfcd404d5324f0222678126d5e166fb77c4c27afea96369ef427f4bb0",
-    ".codex/skills/ai-train-local-research/references/reporting.md": "a5a5cc6438a8cc228560e302e944f1ff320f4c963988af83e1ba1443728f7149",
-    ".codex/skills/ai-train-local-research/scripts/ai-gate-ablation.mjs": "f68a1a7c898f4c2f46860709a52cabd3abd4cc8cd14443de2b5f827d261be935",
-    ".codex/skills/ai-train-local-research/scripts/ai-gate-ablation.test.mjs": "f13d2284a4c04a5220588652be81d104d2acde9650481cf569fa0efe56118f08",
-    ".codex/skills/ai-train-local-research/SKILL.md": "479e44aad9cb6ab6b5c8931c807c63cf2a1938a59d01f553c20f44645716a23a",
+    ".codex/skills/ai-train-local-research/references/reporting.md": "ae0faaab3b3dd0f4bd27142fa2c9aa889c5e7572a10e9fe1ee79ff57ebbd8653",
+    ".codex/skills/ai-train-local-research/scripts/ai-gate-ablation.mjs": "09336b11919c0d93f5b3916f88a546b2b1a9c7a2e26c30dca90c610a9114d8b0",
+    ".codex/skills/ai-train-local-research/scripts/ai-gate-ablation.test.mjs": "c21f36e88888ad5f444e615b5ddf3419e834d945ff9517318288516d7dd489bf",
+    ".codex/skills/ai-train-local-research/SKILL.md": "b75a1e737a13f71c96a8c45ff11891dc4d1c415c784d4146dab9cdcbe376f49a",
     ".codex/skills/backtest-config-redis/scripts/get_backtest_config.sh": "833e950d6348c5b7a4bd3f00d25af60f6394190bff744bd5ab64db9e43d826d5",
-    ".codex/skills/backtest-config-redis/SKILL.md": "85e7b1fa425aa23dfa7950de05c1d3ce046d044dc3872f3665705e9e2f2a1977",
-    ".codex/skills/runtime-parity-mismatch-analysis/SKILL.md": "289925c0069d1933ec3b09d5d6a020632170716c61aafb7031a0998acf3736f6",
+    ".codex/skills/backtest-config-redis/SKILL.md": "ac4f9d685c12f748a87b1882df1f731dd3ae9ec8bd7c11cb652221fe0463febc",
+    ".codex/skills/runtime-parity-mismatch-analysis/SKILL.md": "f9cbc571baaba7a185550415c3ef31f0f55a9503107c8ace2314fac88ba025bd",
     ".codex/skills/save-strategy-config-from-backtest/agents/openai.yaml": "ba163f9449562241ed954ca2e0f12e22562390740cdcf0d09d7f99f0ef791741",
     ".codex/skills/save-strategy-config-from-backtest/SKILL.md": "a73cddc29fb8fa741fe6226c051e19b0f9a19747babf96d336e7f165fb1fd298",
-    ".codex/skills/strategy-backtest-research/references/research-notes.md": "e65eb2124115a11a5be637964efed06961c7a90eff33d95eaf238b521d1f505a",
+    ".codex/skills/strategy-backtest-research/references/research-notes.md": "e7b85dce73c612ba6821857e511b63754a574cc14d8ba5224466002fe27509aa",
     ".codex/skills/strategy-backtest-research/scripts/backtest-run-metrics.mjs": "dbcc6bb7cab4e0108ed11df503b8a0a337d1a01b77c5b56554ce1264a00f5d57",
     ".codex/skills/strategy-backtest-research/scripts/backtest-run-metrics.test.mjs": "a00d0ae5fb0037cc554282dbe22109a9e34cb64b93d26a6feab69005a2ba637f",
     ".codex/skills/strategy-backtest-research/scripts/fast-ai-export-metrics.mjs": "1d1747d34269ad15a610dcd687334c3d47ed9460d891285ad1dd851c7658d1af",
     ".codex/skills/strategy-backtest-research/scripts/fast-ai-export-metrics.test.mjs": "4caab52646504afd2a6f1e4ba9f43d6b07d714baca89d2a2010545fb4210a9da",
     ".codex/skills/strategy-backtest-research/scripts/research-notes-check.mjs": "435aa6560abcf4e8a60fb86230d8d8de2f2dc9e4c52acc7a272ce30287d40325",
     ".codex/skills/strategy-backtest-research/SKILL.md": "526197e5f196d80dc4f4cc3688d2e0d232eb010267c53d2a71589e4ddddac4e4",
-    ".codex/skills/strategy-candidate-compare/SKILL.md": "34dbca2c77ce74836f5c264131cefb751313147128ab1c3f947002f975be051c",
-    ".codex/skills/strategy-candidate-report/SKILL.md": "3cad1241ce2107d7e30a5d78eafd6429514a10d1a8e9cf3d0e97ffef9d691ab6",
+    ".codex/skills/strategy-candidate-compare/SKILL.md": "62f44b986c879f6247ecce7b98e6286c523e8d690d6f1f9015fddaa56b70a4bf",
+    ".codex/skills/strategy-candidate-report/SKILL.md": "ec4399494dd545268c74dd5b4d84a3fa6f283801a7f95449ce956a9f22db6713",
     ".codex/skills/strategy-forward-start/SKILL.md": "a86311aeb4c5b46e43011806ffa0f665c4ec0fddce8d7f8c1232184e4a14bfa7",
-    ".codex/skills/strategy-forward-status/SKILL.md": "b41f117680259fd7b4d5cbe3c0a1be75a42f08c8cf78f9ad7ccb0d90bfc66269",
+    ".codex/skills/strategy-forward-status/SKILL.md": "7cb021991ff40bb3407eb58e1c9e2bcf597f67b0201059b9e20e4f37fd3320b9",
     ".codex/skills/strategy-improvement-plan/SKILL.md": "f5f0f390941dd3332890a2f0ffae1248abde4925db2cf49a41c0c0c1b47e7864",
-    ".codex/skills/strategy-improvement-research/references/final-composition-board.md": "bc88878e088744ff933f06b54f73af96231d1f22dc162a44c9eb2e81d6cb51aa",
+    ".codex/skills/strategy-improvement-research/references/final-composition-board.md": "d7c4c9575af37fdff60548754da434511d2ef1ab9f0fca8dd211edf45e73e8d3",
     ".codex/skills/strategy-improvement-research/scripts/build-final-composition-spec.mjs": "b51873691fafbcf2dd70d8c9a6da842fec5f8358346c632cfe79211db08b18ba",
     ".codex/skills/strategy-improvement-research/scripts/build-final-composition-spec.test.mjs": "85d06b8fe09ee7611a3fb989173c361ba32f0e9364422925ff221bfe127b2bfa",
     ".codex/skills/strategy-improvement-research/scripts/final-composition-board.mjs": "c7834c9d567d3f02c39fcca2b0116a0097bc0384891a5185edbcafa23eeea1db",
     ".codex/skills/strategy-improvement-research/scripts/final-composition-board.test.mjs": "14572bbb8083f83b7e05b6dad2a9c93a88c212425b09cc31ee15ab7a684ea8c5",
     ".codex/skills/strategy-improvement-research/scripts/freeze-gate-variants.mjs": "17b3c868880ddd5b326d309e6e42eaeef1ea22d625a5db947997ac73b221ad6d",
     ".codex/skills/strategy-improvement-research/scripts/freeze-gate-variants.test.mjs": "963e1c64ab3deef11fd48a96e10f3827af38036dfe2eb28daf7849f19dd414d8",
-    ".codex/skills/strategy-improvement-research/SKILL.md": "6964eb154e40b0b775aa2735b2ebbe998984479f46333ed03eab88d2ee150d48",
+    ".codex/skills/strategy-improvement-research/SKILL.md": "75a0f2420305b269fec0f71f63904b5bf04be00612b015855f0a0f2beface5b7",
     ".codex/skills/strategy-period-revalidate/SKILL.md": "f367ba1c632506a7d9bf92662a40b61df55f7a56ad04f3600a475b3c365bbbcb",
     ".codex/skills/strategy-release/agents/openai.yaml": "46ddb09119b08ae3be3b9509121b57492342012e774789b615767b24270dba17",
     ".codex/skills/strategy-release/references/diagnose-live.md": "dc2effa6eba34e3cfcc1eaf4d2fd1c96d6ccdda971a0edec76c9eed1ec293a3d",
     ".codex/skills/strategy-release/references/direction-policy.md": "e7635e709dfc7301b94f3dbc1421c6a779c7c39ce4f6066b2eece38e4e600bec",
     ".codex/skills/strategy-release/references/directional-parameter-split.md": "c303fbd15183b4b8c5662526d5bc762edc0a21a42c45e141d0dc5efdd3b52d8c",
     ".codex/skills/strategy-release/references/evidence-limitations.md": "523d6541a230b51c2b0b81a6c11c99e3aef69203f22b77e94fda1e3b15321e20",
-    ".codex/skills/strategy-release/references/evidence-retention.md": "8fc5ccbc1003e8dc524f12dd56592eb72839c251245c6f68fd495fd4871e4993",
+    ".codex/skills/strategy-release/references/evidence-retention.md": "ed39682e5d1cf33b5df0c1fb4896a889e0f53c4dc27ef37efe3d35d2f08a205c",
     ".codex/skills/strategy-release/references/historical-hypothesis-audit.md": "3c4d579a0c097e974592b855891446060eb70d8bddf8558d765052a92509b516",
     ".codex/skills/strategy-release/references/professional-research-loop.md": "e440f05dd6b5c803ba20da9f019f745f89477724d46551c53dbb8dab9cffe40f",
-    ".codex/skills/strategy-release/references/release-workflow.md": "aeda374c8b0a5d14629b2a1476a9c6e67e556e7ff7709b566d532334f39863eb",
+    ".codex/skills/strategy-release/references/release-workflow.md": "a103416334b5605addd856f105462b53602d29ace2816f96162b7e2ecd1a0bb6",
     ".codex/skills/strategy-release/references/research-objective.md": "650a29260ff7a2ee5dcac89500f59b4c81c2fbd650d6d247a1c799e45814798a",
     ".codex/skills/strategy-release/references/verdict-contract.md": "7cd2d0ae8c8264da36319a78ff80d76893d5233cbceba4d076973f32b2bb2fe9",
     ".codex/skills/strategy-release/scripts/direction-policy-checkpoint.mjs": "739240707323ec80b16191b447e76dd2295e946564c3a67f8472949d6129b3a1",
@@ -69,6 +70,7 @@
     ".codex/skills/strategy-release/scripts/release-progress-checkpoint.mjs": "28da8715ab51ee7217d633c675d014b8004e5ddda8faf9489ebacacef98f6b9b",
     ".codex/skills/strategy-release/scripts/release-progress-checkpoint.test.mjs": "d378b0ab87cd3ee7bbf384ec5f5810565b480b56533ad3c6700edf5938286a38",
     ".codex/skills/strategy-release/SKILL.md": "590948c2c1ad222c7784b050e0fa0a8fea60efa7b0f23d8d355fd56a3e8a1291",
-    ".codex/skills/strategy-risk-scale/SKILL.md": "2c07846c9da2fd622ea6f5586c42863fc27b35a7d1d00614d5bbe0719da4d39f"
+    ".codex/skills/strategy-risk-scale/SKILL.md": "2c07846c9da2fd622ea6f5586c42863fc27b35a7d1d00614d5bbe0719da4d39f",
+    ".codex/skills/tradejs-mcp/SKILL.md": "a9c5dfde4ccab13e1fa2c3e56f6f3132fb0ed99e492b119a5017588db55b34c4"
   }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "create-tradejs",
-  "version": "3.1.29-beta.261",
+  "version": "3.1.29-beta.263",
   "description": "Create a ready-to-run TradeJS project with local infrastructure and the Web UI.",
   "keywords": [
     "tradejs",