@psiclawops/hypermem 0.9.6 → 0.9.9

package/docs/INTEGRATION_VALIDATION.md

@@ -11,6 +11,7 @@ This page is the operator validation contract for HyperMem releases. It describe
 | Memory plugin | `~/.openclaw/plugins/hypermem/memory-plugin` | OpenClaw memory slot integration for `memory_search` and retrieval surfaces | `openclaw plugins list` shows `hypermem` loaded |
 | Runtime staging tool | `hypermem-install` | Copies package runtime into `~/.openclaw/plugins/hypermem` | staged directory contains `dist`, `plugin`, `memory-plugin`, `bin` |
 | Status CLI | `hypermem-status` | Health, database, vector, and runtime summary | `hypermem-status --health` exits cleanly or reports only healthy-empty state |
+| Runtime validator | `hypermem-validate-runtime` | Seeds a deterministic fixture and validates write, FTS, facts, vector index/search, warm, and compose | exits 0 with `consistencyScore: 1` |
 | Model audit CLI | `hypermem-model-audit` | Checks model context-window detection and overrides | `hypermem-model-audit --strict` reports no risky unknown models, or known required overrides |
 
 ## Install state machine
@@ -24,8 +25,9 @@ Do not treat installation as complete until all 5 states pass.
 | 3. OpenClaw wired | OpenClaw config points at staged plugins | `plugins.load.paths` includes both staged plugin dirs, slots are set |
 | 4. Runtime loaded | Gateway has loaded both plugins | `openclaw plugins list` shows `hypercompositor` and `hypermem` loaded |
 | 5. Runtime active | HyperMem is composing live turns | logs show `[hypermem] hypermem initialized` and `[hypermem:compose]` |
+| 6. Runtime validated | HyperMem components round-trip deterministic data | `hypermem-validate-runtime` exits 0 |
 
-A successful `hypermem-install` proves only state 2. It does not modify OpenClaw config and does not restart the gateway.
+A successful `hypermem-install` proves only state 2. It does not modify OpenClaw config and does not restart the gateway. A live install is not complete until runtime validation passes or a skipped semantic path is explicitly accepted for lightweight FTS-only mode.
 
 ## Fresh install validation
 
@@ -47,13 +49,13 @@ cat > ~/.openclaw/hypermem/config.json <<'JSON'
 {
   "embedding": { "provider": "none" },
   "compositor": {
-    "budgetFraction": 0.55,
+    "budgetFraction": 0.60,
     "contextWindowReserve": 0.25,
     "targetBudgetFraction": 0.50,
     "warmHistoryBudgetFraction": 0.27,
     "maxFacts": 25,
-    "maxHistoryMessages": 500,
-    "maxCrossSessionContext": 4000,
+    "maxHistoryMessages": 250,
+    "maxCrossSessionContext": 0,
     "maxRecentToolPairs": 3,
     "maxProseToolPairs": 10,
     "keystoneHistoryFraction": 0.15,
@@ -88,6 +90,7 @@ Verify activation:
 openclaw plugins list
 openclaw logs --limit 100 | grep -E 'hypermem|context-engine'
 hypermem-status --health
+hypermem-validate-runtime --allow-no-embedding
 hypermem-model-audit --strict
 ```
 
@@ -116,6 +119,7 @@ Validate after restart:
 openclaw plugins list
 openclaw logs --limit 100 | grep -E 'hypermem|context-engine|falling back'
 hypermem-status --health
+hypermem-validate-runtime --allow-no-embedding
 hypermem-model-audit --strict
 ```
 
@@ -143,6 +147,9 @@ Before publishing a release, run:
 
 ```bash
 npm run validate:version-parity
+npm run validate:sdk-imports
+npm run validate:sdk-latest-canary
+npm run validate:plugin-checkers
 npm run validate:docs
 npm run validate:config
 npm run validate:release-path
@@ -154,6 +161,10 @@ npm pack ./plugin --dry-run
 npm pack ./memory-plugin --dry-run
 ```
 
+The SDK gates are split deliberately: `validate:sdk-imports` enforces reproducible pinned OpenClaw Plugin SDK metadata, while `validate:sdk-latest-canary` rebuilds both plugin packages against the latest published OpenClaw SDK in a temp workspace so a stale pin cannot hide a future import break.
+
+The plugin checker gate is deliberately composite: it runs Plugin Inspector static/runtime reports, isolated dependency-install cold import proof, production/dev dependency audit, Plugin Inspector issue-debt validation for the plugin packages. Raw Plugin Inspector P2 rows are not ignored: they must either disappear or have a proof artifact consumed by `validate:plugin-inspector:debt`.
+
 Runtime validation must install the packed artifact, not the repo working tree. The production-shaped local install path is:
 
 ```bash
@@ -161,8 +172,17 @@ npm run install:runtime:packed
 openclaw gateway restart
 openclaw plugins list
 node ~/.openclaw/plugins/hypermem/bin/hypermem-status.mjs --master
+node ~/.openclaw/plugins/hypermem/bin/hypermem-validate-runtime.mjs --allow-no-embedding
 ```
 
+The final OpenClaw host-runtime plugin checker is:
+
+```bash
+npm run validate:plugins:runtime
+```
+
+That gate packs and installs the release artifact into OpenClaw's managed plugin directory, refreshes the plugin registry, runs `openclaw plugins doctor`, and runs runtime inspection for `hypercompositor` and `hypermem`. Treat HyperMem-owned doctor diagnostics as release-blocking. Unrelated local plugin diagnostics are reported but ignored by the HyperMem gate.
+
 For isolated validation, test the packed artifact from a temp directory:
 
 ```bash

package/docs/TUNING.md

@@ -72,7 +72,7 @@ Estimates on a 200k model (Claude Sonnet). Scale proportionally for smaller wind
 | `standard` (default) | **35–50k** | 55–80k | All layers, default caps |
 | `full` | **40–55k** | 60–85k | All layers, raised caps, cross-session on |
 
-Standard turn 1 is lower than full because `warmHistoryBudgetFraction` is the same, but full enables cross-session context which adds tokens immediately; by turn 5 full overtakes standard as more layers accumulate history.
+Standard turn 1 is intentionally conservative. Full keeps cross-session context opt-in and should be used only when operators accept the extra pressure.
 
 **Turn 1 is where token-conscious users will react.** Light vs full on turn 1 is roughly 25–40k tokens — the number that shows up in provider dashboards. By turn 5 the gap is still real, but the value case is easier to make because the user has already experienced continuity.
 
@@ -145,12 +145,12 @@ The standard semantic configuration. All memory layers are active with 0.9.4 rec
     "budgetFraction": 0.6,
     "contextWindowReserve": 0.25,
     "targetBudgetFraction": 0.5,
-    "warmHistoryBudgetFraction": 0.45,
-    "maxFacts": 28,
+    "warmHistoryBudgetFraction": 0.27,
+    "maxFacts": 25,
     "maxHistoryMessages": 250,
     "maxCrossSessionContext": 0,
-    "keystoneHistoryFraction": 0.20,
-    "keystoneMaxMessages": 15,
+    "keystoneHistoryFraction": 0.15,
+    "keystoneMaxMessages": 12,
     "hyperformProfile": "standard"
   },
   "indexer": {
@@ -187,11 +187,11 @@ For long-running sessions, multi-agent fleets, or any deployment where the agent
   "compositor": {
     "budgetFraction": 0.70,
     "contextWindowReserve": 0.22,
-    "maxFacts": 40,
+    "maxFacts": 35,
     "maxHistoryMessages": 500,
-    "maxCrossSessionContext": 6000,
-    "keystoneHistoryFraction": 0.22,
-    "keystoneMaxMessages": 20,
+    "maxCrossSessionContext": 4000,
+    "keystoneHistoryFraction": 0.18,
+    "keystoneMaxMessages": 18,
     "wikiTokenCap": 600,
     "hyperformProfile": "full"
   },
@@ -234,7 +234,7 @@ Three pre-built profiles ship with hypermem. Each configures every setting to a
 | Profile | Context window | Budget | Hyperform | Best for |
 |---|---|---|---|---|
 | `light` | 64k | 40k effective | `light` (behavior only) | Single agent, small models, constrained resources |
-| `standard` | 128k | 90k effective | `standard` (behavior + structure) | Normal deployments, small fleets |
+| `standard` | 128k | ~77k effective before reserve | `standard` (behavior + structure) | Normal deployments, safe starter posture |
 | `full` | 200k+ | 160k effective | `full` (behavior + model adaptation) | Multi-agent fleets, large-context models |
 
 ```ts
@@ -432,10 +432,10 @@ effective budget × (1 - targetBudgetFraction) = history budget
 **Worked example (standard profile, 128k model):**
 
 ```
-128,000 × 0.703 = 89,984 (effective budget before reserve)
-89,984 × (1 - 0.25) = 67,488 (effective budget after reserve)
-67,488 × 0.65 = 43,867 (context assembly budget)
-67,488 × 0.35 = 23,621 (history budget)
+128,000 × 0.60 = 76,800 (effective budget before reserve)
+76,800 × (1 - 0.25) = 57,600 (effective budget after reserve)
+57,600 × 0.50 = 28,800 (context assembly budget)
+57,600 × 0.50 = 28,800 (history budget)
 ```
 
 **Model swap resilience:** The budget is computed from the model's actual context window at compose time when OpenClaw passes `tokenBudget`. If runtime metadata is missing, HyperMem falls back to `contextWindowOverrides` and then `contextWindowSize`. Structured tool history is guarded from being overwritten during a budget downshift — the compositor computes the new allocation but doesn't persist a lower-context snapshot to disk, preserving the full history for when the larger model returns.
@@ -542,7 +542,7 @@ This gives only 45% to context assembly and 55% to history. Useful for coding ag
 }
 ```
 
-Raises the fact injection cap from the default 30 to 50, and reduces wiki page space from 600 to 400 tokens.
+Raises the fact injection cap from the default 25 to 50, and reduces wiki page space from 600 to 400 tokens.
 
 **More keystone history (recalled older messages):**
 
@@ -555,7 +555,7 @@ Raises the fact injection cap from the default 30 to 50, and reduces wiki page s
 }
 ```
 
-Reserves 30% of the history budget for keystones (up from default 20%) and allows up to 25 keystone messages (up from 15). Keystones are high-significance older messages that survive pressure trimming ahead of ordinary history.
+Reserves 30% of the history budget for keystones (up from default 15%) and allows up to 25 keystone messages (up from 12). Keystones are high-significance older messages that survive pressure trimming ahead of ordinary history.
 
 ### Adjusting for model context size
 
@@ -880,7 +880,7 @@ Higher `contextWindowReserve` (0.30) gives more headroom for large tool results.
 
 | Knob | Type | Default | What it controls |
 |---|---|---|---|
-| `maxFacts` | number | 30 | Maximum facts surfaced per compose pass. |
+| `maxFacts` | number | 25 | Maximum facts surfaced per compose pass. |
 | `wikiTokenCap` | tokens | 600 | Hard ceiling on wiki/knowledge injection per pass. |
 | `maxTotalTriggerTokens` | tokens | 4000 | Ceiling across all trigger-fired doc chunk collections. |
 
@@ -888,9 +888,9 @@ Higher `contextWindowReserve` (0.30) gives more headroom for large tool results.
 
 | Knob | Type | Default | What it controls |
 |---|---|---|---|
-| `maxHistoryMessages` | number | 500 | Maximum messages in the hot history window. |
-| `keystoneHistoryFraction` | 0.0–0.5 | 0.20 | Fraction of history budget reserved for keystones. 0 disables. |
-| `keystoneMaxMessages` | number | 15 | Max keystone messages injected per pass. |
+| `maxHistoryMessages` | number | 250 | Maximum messages in the hot history window. |
+| `keystoneHistoryFraction` | 0.0-0.5 | 0.15 | Fraction of history budget reserved for keystones. 0 disables. |
+| `keystoneMaxMessages` | number | 12 | Max keystone messages injected per pass. |
 | `keystoneMinSignificance` | 0.0–1.0 | 0.5 | Minimum episode significance for keystone qualification. |
 
 ### Tool history
@@ -899,7 +899,7 @@ Higher `contextWindowReserve` (0.30) gives more headroom for large tool results.
 |---|---|---|---|
 | `maxRecentToolPairs` | number | 3 | Tool call/result pairs kept verbatim. |
 | `maxProseToolPairs` | number | 10 | Older pairs converted to prose stubs. Beyond this, payloads dropped. |
-| `maxCrossSessionContext` | tokens | 4000 | Token ceiling for cross-agent context. 0 disables. |
+| `maxCrossSessionContext` | tokens | 0 | Token ceiling for cross-session context. 0 disables and is the default. |
 
 ### Dynamic reserve
 

package/memory-plugin/dist/index.d.ts

@@ -18,7 +18,7 @@ declare const _default: {
     id: string;
     name: string;
     description: string;
-    configSchema: import("openclaw/plugin-sdk").OpenClawPluginConfigSchema;
-    register: NonNullable<import("openclaw/plugin-sdk/plugin-entry").OpenClawPluginDefinition["register"]>;
-} & Pick<import("openclaw/plugin-sdk/plugin-entry").OpenClawPluginDefinition, "kind" | "reload" | "nodeHostCommands" | "securityAuditCollectors">;
+    configSchema: import("openclaw/plugin-sdk/core").OpenClawPluginConfigSchema;
+    register: NonNullable<import("openclaw/plugin-sdk/core").OpenClawPluginDefinition["register"]>;
+} & Pick<import("openclaw/plugin-sdk/core").OpenClawPluginDefinition, "kind" | "reload" | "nodeHostCommands" | "securityAuditCollectors">;
 export default _default;

package/memory-plugin/dist/index.js

@@ -14,7 +14,7 @@
  *
  * Both plugins share the same HyperMem singleton (loaded from repo dist).
  */
-import { definePluginEntry, emptyPluginConfigSchema } from 'openclaw/plugin-sdk/plugin-entry';
+import { definePluginEntry, emptyPluginConfigSchema } from 'openclaw/plugin-sdk/core';
 import { matchTriggers, TRIGGER_REGISTRY } from '@psiclawops/hypermem';
 import path from 'path';
 import fs from 'fs/promises';
@@ -181,9 +181,11 @@ function createMemorySearchManager(hm, agentId, workspaceDir) {
             try {
                 const vectorStore = hm.getVectorStore();
                 if (vectorStore) {
-                    const vectorResults = await hm.semanticSearch(agentId, query, {
+                    const semanticSearch = hm.semanticSearch;
+                    const vectorResults = await semanticSearch(agentId, query, {
                         limit: maxResults,
                         maxDistance: 1.5,
+                        allowInlineQueryEmbedding: false,
                     });
                     for (const vr of vectorResults) {
                         const score = 1.0 - (vr.distance / 2.0); // normalize distance to 0-1 score

package/memory-plugin/openclaw.plugin.json

@@ -8,6 +8,11 @@
     ],
     "onStartup": false
   },
+  "contracts": {
+    "tools": [
+      "history_query"
+    ]
+  },
   "configSchema": {
     "type": "object",
     "additionalProperties": false,

package/memory-plugin/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@psiclawops/hypermem-memory",
-  "version": "0.9.6",
+  "version": "0.9.9",
   "description": "HyperMem memory plugin for OpenClaw \u2014 bridges HyperMem retrieval into the memory slot",
   "type": "module",
   "main": "dist/index.js",
@@ -29,8 +29,8 @@
       "minGatewayVersion": "2026.4.12"
     },
     "build": {
-      "openclawVersion": "2026.4.9",
-      "pluginSdkVersion": "2026.4.12"
+      "openclawVersion": "2026.5.7",
+      "pluginSdkVersion": "2026.5.7"
     }
   },
   "scripts": {
@@ -40,10 +40,10 @@
     "typecheck": "tsc --noEmit"
   },
   "dependencies": {
-    "@psiclawops/hypermem": "0.9.6"
+    "@psiclawops/hypermem": "0.9.9"
   },
   "devDependencies": {
-    "openclaw": "*",
+    "openclaw": "2026.5.7",
     "typescript": "^5.4.0"
   },
   "peerDependencies": {
@@ -62,5 +62,9 @@
     "openclaw.plugin.json",
     "README.md",
     "LICENSE"
-  ]
+  ],
+  "overrides": {
+    "uuid": "14.0.0",
+    "fast-xml-builder": "1.2.0"
+  }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@psiclawops/hypermem",
-  "version": "0.9.6",
+  "version": "0.9.9",
   "description": "Agent-centric memory and context composition engine for OpenClaw",
   "type": "module",
   "main": "dist/index.js",
@@ -10,7 +10,9 @@
     "hypermem-install": "scripts/install-runtime.mjs",
     "hypermem-model-audit": "bin/hypermem-model-audit.mjs",
     "hypermem-bench": "bin/hypermem-bench.mjs",
-    "hypermem-doctor": "bin/hypermem-doctor.mjs"
+    "hypermem-doctor": "bin/hypermem-doctor.mjs",
+    "hypermem-cleanup": "bin/hypermem-cleanup.mjs",
+    "hypermem-validate-runtime": "bin/hypermem-validate-runtime.mjs"
   },
   "exports": {
     ".": {
@@ -31,13 +33,14 @@
     "install:runtime": "node scripts/install-runtime.mjs",
     "dev": "tsc --watch",
     "typecheck": "tsc --noEmit",
-    "test": "npm run build && node test/smoke.mjs && node test/lightweight-mode.mjs && node test/cross-agent.mjs && node test/vector-search.mjs && node test/embed-existing-coverage.mjs && node test/library.mjs && node test/compositor.mjs && node test/fleet-cache.mjs && node test/fleet-startup-seeding.mjs && node test/rotation.mjs && node test/knowledge-graph.mjs && node test/rate-limiter.mjs && node test/doc-chunker.mjs && node test/compaction-fence-gate.mjs && node test/hybrid-retrieval.mjs && node test/reranker-hotfix.mjs && node test/live-org-registry.mjs && node test/multi-turn-session.mjs && node test/secret-scanner.mjs && node --test test/content-type-classifier.mjs && node --test test/composition-snapshot-integrity.test.mjs && node --test test/composition-snapshot-store.test.mjs && node test/keystone-history.mjs && node test/cross-topic-keystone.mjs && node test/proactive-pass.mjs && node test/topic-synthesis.mjs && node test/session-topic-map.mjs && node test/virtual-sessions.mjs && node test/budget-downshift.mjs && node --test test/image-eviction.mjs && node test/fos-mod.mjs && node test/history-depth-estimator.mjs && node test/unified-pressure-signal.mjs && node test/sprint4-prompt-placement.mjs && node test/model-aware-budgeting-b4.mjs && node test/adaptive-lifecycle.mjs && node test/repair-tool-pairs.mjs && node test/tool-result-guards.mjs && node test/phase-c-fixtures.mjs && node test/oversized-artifact-c2.mjs && node test/oversized-artifact-guards.mjs && node test/doc-chunk-artifact-retrieval.mjs && node test/dreaming-promoter-temporal.mjs",
+    "test": "npm run build && node test/smoke.mjs && node test/lightweight-mode.mjs && node test/cross-agent.mjs && node test/vector-search.mjs && node test/embed-existing-coverage.mjs && node test/library.mjs && node test/compositor.mjs && node test/fleet-cache.mjs && node test/fleet-startup-seeding.mjs && node test/rotation.mjs && node test/knowledge-graph.mjs && node test/rate-limiter.mjs && node test/doc-chunker.mjs && node test/compaction-fence-gate.mjs && node test/hybrid-retrieval.mjs && node --test test/entity-extractor.mjs test/question-shape.mjs test/entity-bridge.mjs && node test/reranker-hotfix.mjs && node test/live-org-registry.mjs && node test/multi-turn-session.mjs && node test/secret-scanner.mjs && node --test test/content-type-classifier.mjs && node --test test/composition-snapshot-integrity.test.mjs && node --test test/composition-snapshot-store.test.mjs && node test/keystone-history.mjs && node test/cross-topic-keystone.mjs && node test/proactive-pass.mjs && node test/topic-synthesis.mjs && node test/session-topic-map.mjs && node test/virtual-sessions.mjs && node test/budget-downshift.mjs && node --test test/image-eviction.mjs && node test/fos-mod.mjs && node test/history-depth-estimator.mjs && node test/unified-pressure-signal.mjs && node test/sprint4-prompt-placement.mjs && node test/model-aware-budgeting-b4.mjs && node test/adaptive-lifecycle.mjs && node test/repair-tool-pairs.mjs && node test/tool-result-guards.mjs && node test/phase-c-fixtures.mjs && node test/oversized-artifact-c2.mjs && node test/oversized-artifact-guards.mjs && node test/doc-chunk-artifact-retrieval.mjs && node test/dreaming-promoter-temporal.mjs",
     "test:quick": "node test/smoke.mjs && node test/lightweight-mode.mjs && node test/library.mjs && node test/compositor.mjs",
     "validate:plugin-pipeline": "npm --prefix plugin run build && node test/plugin-pipeline.mjs && node test/batch-trim-b3.mjs",
     "validate:release-path": "npm run build && npm --prefix plugin run build && node test/release-gateway-path.mjs",
     "test:multi-turn": "node test/multi-turn-session.mjs",
     "test:ci": "npm test && npm --prefix plugin run typecheck",
     "smoke": "node scripts/smoke-test.mjs",
+    "validate:runtime": "node bin/hypermem-validate-runtime.mjs --data-dir /tmp/hypermem-runtime-validation --allow-no-embedding",
     "sync-public": "node scripts/sync-public.mjs",
     "validate:docs": "node scripts/validate-docs.mjs",
     "validate:config": "node scripts/validate-config-surface.mjs",
@@ -45,16 +48,27 @@
     "validate:public-surface": "node scripts/validate-public-surface.mjs",
     "bench:memory": "node bench/data-access-bench.mjs --iterations 1000 --warmup 50",
     "validate:doctor": "node test/doctor-cli.mjs",
-    "build:all": "npm run build && npm --prefix plugin install --prefer-offline && npm --prefix plugin run build && npm --prefix memory-plugin install --prefer-offline && npm --prefix memory-plugin run build",
+    "build:all": "npm run build && npm --prefix plugin run build && npm --prefix memory-plugin run build",
     "validate:history-query": "npm run build && npm --prefix memory-plugin run build && node scripts/validate-history-query.mjs",
     "install:runtime:packed": "node scripts/install-packed-runtime.mjs",
-    "release:install-smoke": "bash release-gate-internal/fresh-install-smoke.sh"
+    "release:install-smoke": "bash release-gate-internal/fresh-install-smoke.sh",
+    "validate:sdk-imports": "node scripts/validate-sdk-imports.mjs",
+    "validate:sdk-latest-canary": "node scripts/validate-sdk-latest-canary.mjs",
+    "validate:plugin-inspector": "node scripts/validate-plugin-inspector.mjs",
+    "validate:plugin-inspector:runtime": "npm run build:all && node scripts/validate-plugin-inspector.mjs --runtime",
+    "validate:plugins:runtime": "node scripts/validate-openclaw-plugin-runtime.mjs",
+    "validate:release-readiness": "npm run validate:version-parity && npm run validate:sdk-imports && npm run validate:sdk-latest-canary && npm run validate:plugin-checkers && npm run validate:docs && npm run validate:config && npm run validate:release-path && npm run validate:public-surface",
+    "validate:plugin-isolated-cold-import": "npm run build:all && node scripts/validate-plugin-isolated-cold-import.mjs",
+    "validate:plugin-inspector:debt": "node scripts/validate-plugin-inspector-issue-debt.mjs",
+    "validate:plugin-checkers": "npm run validate:plugin-inspector && npm run validate:plugin-isolated-cold-import && npm run validate:plugin-inspector:runtime && npm run validate:plugin-inspector:debt && npm run validate:plugin-dev-audit",
+    "validate:plugin-dev-audit": "npm audit && npm --prefix plugin audit && npm --prefix memory-plugin audit"
   },
   "dependencies": {
     "sqlite-vec": "^0.1.9",
     "zod": "^4.0.0"
   },
   "devDependencies": {
+    "@openclaw/plugin-inspector": "0.3.6",
     "@types/node": "^22.0.0",
     "typescript": "^5.7.0"
   },
@@ -66,9 +80,12 @@
     "dist/**/*.d.ts",
     "dist/**/*.d.ts.map",
     "assets/default-config.json",
+    "assets/runtime-validation-fixture.json",
     "bin/hypermem-status.mjs",
     "bin/hypermem-model-audit.mjs",
     "bin/hypermem-doctor.mjs",
+    "bin/hypermem-cleanup.mjs",
+    "bin/hypermem-validate-runtime.mjs",
     "bench/README.md",
     "bench/data-access-bench.mjs",
     "bin/hypermem-bench.mjs",

package/plugin/dist/index.d.ts

@@ -203,8 +203,8 @@ declare const _default: {
     id: string;
     name: string;
     description: string;
-    configSchema: import("openclaw/plugin-sdk").OpenClawPluginConfigSchema;
-    register: NonNullable<import("openclaw/plugin-sdk/plugin-entry").OpenClawPluginDefinition["register"]>;
-} & Pick<import("openclaw/plugin-sdk/plugin-entry").OpenClawPluginDefinition, "kind" | "reload" | "nodeHostCommands" | "securityAuditCollectors">;
+    configSchema: import("openclaw/plugin-sdk/core").OpenClawPluginConfigSchema;
+    register: NonNullable<import("openclaw/plugin-sdk/core").OpenClawPluginDefinition["register"]>;
+} & Pick<import("openclaw/plugin-sdk/core").OpenClawPluginDefinition, "kind" | "reload" | "nodeHostCommands" | "securityAuditCollectors">;
 export default _default;
 //# sourceMappingURL=index.d.ts.map

package/plugin/dist/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAaH,OAAO,KAAK,EACV,cAAc,EACd,eAAe,EACf,iBAAiB,EACjB,cAAc,EACd,aAAa,EAId,MAAM,sBAAsB,CAAC;AAC9B,OAAO,EAUL,kBAAkB,EASnB,MAAM,sBAAsB,CAAC;AAW9B,YAAY,EAAE,cAAc,EAAE,eAAe,EAAE,iBAAiB,EAAE,cAAc,EAAE,aAAa,EAAE,CAAC;AAYlG,KAAK,iBAAiB,GAClB,iBAAiB,GACjB,mBAAmB,GACnB,mBAAmB,GACnB,SAAS,GACT,iBAAiB,GACjB,iBAAiB,GACjB,kBAAkB,GAClB,qBAAqB,GACrB,WAAW,CAAC;AAEhB,KAAK,wBAAwB,GAAG,SAAS,GAAG,UAAU,CAAC;AAEvD,UAAU,0BAA0B;IAClC,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,IAAI,EAAE,wBAAwB,CAAC;IAC/B,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B,yBAAyB,CAAC,EAAE,MAAM,CAAC;IACnC,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B,+BAA+B,CAAC,EAAE,MAAM,CAAC;IACzC,WAAW,CAAC,EAAE,UAAU,GAAG,aAAa,GAAG,QAAQ,CAAC;IACpD,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB;AA0BD,iBAAS,aAAa,CAAC,MAAM,EAAE;IAC7B,IAAI,EAAE,iBAAiB,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;IAChB,gBAAgB,EAAE,OAAO,CAAC;IAC1B,MAAM,EAAE,MAAM,CAAC;CAChB,GAAG,IAAI,CAcP;AAED,iBAAS,aAAa,CAAC,MAAM,EAAE;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,IAAI,EAAE,MAAM,GAAG,QAAQ,GAAG,UAAU,CAAC;IACrC,QAAQ,EAAE,OAAO,CAAC;IAClB,QAAQ,EAAE,MAAM,CAAC;IAEjB,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,gBAAgB,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACjC,SAAS,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE;QAAE,SAAS,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,OAAO,CAAA;KAAE,CAAC,CAAC;IACrF,uBAAuB,CAAC,EAAE,MAAM,CAAC;IACjC,uBAAuB,CAAC,EAAE,MAAM,CAAC;IACjC,wBAAwB,CAAC,EAAE,MAAM,CAAC;IAClC,kBAAkB,CAAC,EAAE,kBAAkB,GAAG,mBAAmB,GAAG,MAAM,CAAC;IACvE,iBAAiB,CAAC,EAAE,iBAAiB,GAAG,oBAAoB,GAAG,sCAAsC,GAAG,kBAAkB,CAAC;IAC3H,wBAAwB,CAAC,EAAE,MAAM,CAAC;IAClC,+BAA+B,CAAC,EAAE,MAAM,CAAC;IACzC,2BAA2B,CAAC,EAAE,SAAS,GAAG,uBAAuB,CAAC;CACnE,GAAG,IAAI,CAcP;AAED,iBAAS,oBAAoB,CAAC,MAAM,EAAE,0BAA0B,GAAG,IAAI,CActE;AAGD,iBAAS,wBAAwB,CAAC,MAAM,EAAE;IACxC,IAAI,EAAE,kBAAkB,GAAG,mBAAmB,GAAG,oBAAoB,CAAC;IACtE,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,6BAA6B,CAAC,EAAE,MAAM,CAAC;IACvC,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,OAAO,CAAC,EAAE,MAAM,EAAE,CAAC;CACpB,GAAG,IAAI,CAcP;AAkBD,iBAAS,UAAU,IAAI,MAAM,CAG5B;AA+CD,QAAA,MAAM,uBAAuB,+LAOnB,CAAC;AACX,KAAK,oBAAoB,GAAG,OAAO,uBAAuB,CAAC,MAAM,CAAC,CAAC;AAqBnE,iBAAS,kBAAkB,CAAC,UAAU,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,IAAI,CAEpE;AAED,iBAAS,gBAAgB,CAAC,UAAU,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,IAAI,CAElE;AAED;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,iBAAS,cAAc,CAAC,UAAU,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,iBAAiB,GAAG,OAAO,CAwB5F;AAED;;;;;;;;;;GAUG;AACH,iBAAS,cAAc,CAAC,MAAM,EAAE;IAC9B,IAAI,EAAE,iBAAiB,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,oBAAoB,CAAC;CAC9B,GAAG,IAAI,CAcP;AAkBD,eAAO,MAAM,mBAAmB;;;;;;;;;;;;;;aAgBrB,IAAI;CASd,CAAC;AAyEF,eAAO,MAAM,iCAAiC,QAAuB,CAAC;AACtE,MAAM,MAAM,qBAAqB,GAAG;IAAE,aAAa,CAAC,EAAE,MAAM,CAAC;IAAC,aAAa,CAAC,EAAE,MAAM,CAAA;CAAE,CAAC;AAwBvF,wBAAgB,8BAA8B,CAAC,GAAG,EAAE,OAAO,GAAG;IAC5D,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,qBAAqB,CAAC,CAAC;IAC7C,QAAQ,EAAE,MAAM,EAAE,CAAC;CACpB,CA4BA;AAED,wBAAgB,sBAAsB,CAAC,IAAI,EAAE;IAC3C,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,iBAAiB,EAAE,MAAM,CAAC;IAC1B,oBAAoB,EAAE,MAAM,CAAC;IAC7B,sBAAsB,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,qBAAqB,CAAC,CAAC;CAChE,GAAG;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,CAoBrC;AAED,MAAM,WAAW,aAAa;IAC5B,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;CACxB;AAED,wBAAgB,oBAAoB,CAAC,KAAK,CAAC,EAAE,MAAM,GAAG,aAAa,CAkBlE;AAED,wBAAgB,cAAc,CAC5B,QAAQ,EAAE;IACR,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,QAAQ,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,QAAQ,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,OAAO,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB,GAAG,IAAI,GAAG,SAAS,EACpB,OAAO,EAAE;IACP,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB,GACA;IACD,gBAAgB,EAAE,aAAa,CAAC;IAChC,eAAe,EAAE,aAAa,CAAC;IAC/B,YAAY,EAAE,OAAO,CAAC;IACtB,eAAe,EAAE,OAAO,CAAC;IACzB,cAAc,EAAE,OAAO,CAAC;IACxB,aAAa,EAAE,OAAO,CAAC;IACvB,eAAe,EAAE,OAAO,CAAC;IACzB,YAAY,EAAE,OAAO,CAAC;CACvB,CAyBA;AAgwGD;;;;GAIG;AACH,wBAAsB,iBAAiB,CAAC,OAAO,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAU1F;;;;;;;;AA2GD,wBAgGG"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAaH,OAAO,KAAK,EACV,cAAc,EACd,eAAe,EACf,iBAAiB,EACjB,cAAc,EACd,aAAa,EAId,MAAM,sBAAsB,CAAC;AAC9B,OAAO,EAUL,kBAAkB,EASnB,MAAM,sBAAsB,CAAC;AAW9B,YAAY,EAAE,cAAc,EAAE,eAAe,EAAE,iBAAiB,EAAE,cAAc,EAAE,aAAa,EAAE,CAAC;AAYlG,KAAK,iBAAiB,GAClB,iBAAiB,GACjB,mBAAmB,GACnB,mBAAmB,GACnB,SAAS,GACT,iBAAiB,GACjB,iBAAiB,GACjB,kBAAkB,GAClB,qBAAqB,GACrB,WAAW,CAAC;AAEhB,KAAK,wBAAwB,GAAG,SAAS,GAAG,UAAU,CAAC;AAEvD,UAAU,0BAA0B;IAClC,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,IAAI,EAAE,wBAAwB,CAAC;IAC/B,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B,yBAAyB,CAAC,EAAE,MAAM,CAAC;IACnC,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B,+BAA+B,CAAC,EAAE,MAAM,CAAC;IACzC,WAAW,CAAC,EAAE,UAAU,GAAG,aAAa,GAAG,QAAQ,CAAC;IACpD,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB;AA0BD,iBAAS,aAAa,CAAC,MAAM,EAAE;IAC7B,IAAI,EAAE,iBAAiB,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;IAChB,gBAAgB,EAAE,OAAO,CAAC;IAC1B,MAAM,EAAE,MAAM,CAAC;CAChB,GAAG,IAAI,CAcP;AAED,iBAAS,aAAa,CAAC,MAAM,EAAE;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,IAAI,EAAE,MAAM,GAAG,QAAQ,GAAG,UAAU,CAAC;IACrC,QAAQ,EAAE,OAAO,CAAC;IAClB,QAAQ,EAAE,MAAM,CAAC;IAEjB,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,gBAAgB,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACjC,SAAS,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE;QAAE,SAAS,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,OAAO,CAAA;KAAE,CAAC,CAAC;IACrF,uBAAuB,CAAC,EAAE,MAAM,CAAC;IACjC,uBAAuB,CAAC,EAAE,MAAM,CAAC;IACjC,wBAAwB,CAAC,EAAE,MAAM,CAAC;IAClC,kBAAkB,CAAC,EAAE,kBAAkB,GAAG,mBAAmB,GAAG,MAAM,CAAC;IACvE,iBAAiB,CAAC,EAAE,iBAAiB,GAAG,oBAAoB,GAAG,sCAAsC,GAAG,kBAAkB,CAAC;IAC3H,wBAAwB,CAAC,EAAE,MAAM,CAAC;IAClC,+BAA+B,CAAC,EAAE,MAAM,CAAC;IACzC,2BAA2B,CAAC,EAAE,SAAS,GAAG,uBAAuB,CAAC;CACnE,GAAG,IAAI,CAcP;AAED,iBAAS,oBAAoB,CAAC,MAAM,EAAE,0BAA0B,GAAG,IAAI,CActE;AAGD,iBAAS,wBAAwB,CAAC,MAAM,EAAE;IACxC,IAAI,EAAE,kBAAkB,GAAG,mBAAmB,GAAG,oBAAoB,CAAC;IACtE,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,6BAA6B,CAAC,EAAE,MAAM,CAAC;IACvC,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,OAAO,CAAC,EAAE,MAAM,EAAE,CAAC;CACpB,GAAG,IAAI,CAcP;AAkBD,iBAAS,UAAU,IAAI,MAAM,CAG5B;AA+CD,QAAA,MAAM,uBAAuB,+LAOnB,CAAC;AACX,KAAK,oBAAoB,GAAG,OAAO,uBAAuB,CAAC,MAAM,CAAC,CAAC;AAqBnE,iBAAS,kBAAkB,CAAC,UAAU,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,IAAI,CAEpE;AAED,iBAAS,gBAAgB,CAAC,UAAU,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,IAAI,CAElE;AAED;;;;;;;;;;;;;;;;;;;;;;;GAuBG;AACH,iBAAS,cAAc,CAAC,UAAU,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,iBAAiB,GAAG,OAAO,CAwB5F;AAED;;;;;;;;;;GAUG;AACH,iBAAS,cAAc,CAAC,MAAM,EAAE;IAC9B,IAAI,EAAE,iBAAiB,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,oBAAoB,CAAC;CAC9B,GAAG,IAAI,CAcP;AAkBD,eAAO,MAAM,mBAAmB;;;;;;;;;;;;;;aAgBrB,IAAI;CASd,CAAC;AAyEF,eAAO,MAAM,iCAAiC,QAAuB,CAAC;AACtE,MAAM,MAAM,qBAAqB,GAAG;IAAE,aAAa,CAAC,EAAE,MAAM,CAAC;IAAC,aAAa,CAAC,EAAE,MAAM,CAAA;CAAE,CAAC;AAwBvF,wBAAgB,8BAA8B,CAAC,GAAG,EAAE,OAAO,GAAG;IAC5D,KAAK,EAAE,MAAM,CAAC,MAAM,EAAE,qBAAqB,CAAC,CAAC;IAC7C,QAAQ,EAAE,MAAM,EAAE,CAAC;CACpB,CA4BA;AAED,wBAAgB,sBAAsB,CAAC,IAAI,EAAE;IAC3C,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,iBAAiB,EAAE,MAAM,CAAC;IAC1B,oBAAoB,EAAE,MAAM,CAAC;IAC7B,sBAAsB,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,qBAAqB,CAAC,CAAC;CAChE,GAAG;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,CAoBrC;AAED,MAAM,WAAW,aAAa;IAC5B,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;CACxB;AAED,wBAAgB,oBAAoB,CAAC,KAAK,CAAC,EAAE,MAAM,GAAG,aAAa,CAkBlE;AAED,wBAAgB,cAAc,CAC5B,QAAQ,EAAE;IACR,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,QAAQ,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,QAAQ,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,OAAO,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB,GAAG,IAAI,GAAG,SAAS,EACpB,OAAO,EAAE;IACP,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB,GACA;IACD,gBAAgB,EAAE,aAAa,CAAC;IAChC,eAAe,EAAE,aAAa,CAAC;IAC/B,YAAY,EAAE,OAAO,CAAC;IACtB,eAAe,EAAE,OAAO,CAAC;IACzB,cAAc,EAAE,OAAO,CAAC;IACxB,aAAa,EAAE,OAAO,CAAC;IACvB,eAAe,EAAE,OAAO,CAAC;IACzB,YAAY,EAAE,OAAO,CAAC;CACvB,CAyBA;AA62GD;;;;GAIG;AACH,wBAAsB,iBAAiB,CAAC,OAAO,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAU1F;;;;;;;;AA2GD,wBAgGG"}

package/plugin/dist/index.js

@@ -19,8 +19,7 @@
  *
  * Session key format expected: "agent:<agentId>:<channel>:<name>"
  */
-import { definePluginEntry } from 'openclaw/plugin-sdk/plugin-entry';
-import { buildPluginConfigSchema } from 'openclaw/plugin-sdk/core';
+import { buildPluginConfigSchema, definePluginEntry } from 'openclaw/plugin-sdk/core';
 import { z } from 'zod';
 import { detectTopicShift, stripMessageMetadata, SessionTopicMap, applyToolGradientToWindow, OPENCLAW_BOOTSTRAP_FILES, rotateSessionContext, TRIM_SOFT_TARGET, TRIM_GROWTH_THRESHOLD, TRIM_HEADROOM_FRACTION, resolveTrimBudgets, resolveAdaptiveLifecyclePolicy, formatToolChainStub, decideReplayRecovery, isReplayState, recordOutputMetrics, 
 // Sprint 3: unified pressure signal
@@ -813,6 +812,67 @@ function extractTextFromInboundContent(content) {
         .map(part => part.text ?? '')
         .join('\n');
 }
+function isToolResultRole(role) {
+    return role === 'tool' || role === 'tool_result' || role === 'toolResult';
+}
+function isPersistableInboundRole(role) {
+    return role === 'user' || role === 'assistant' || role === 'system' || isToolResultRole(role);
+}
+function isSenderMetadataText(text) {
+    const trimmed = text.trim();
+    if (!trimmed)
+        return false;
+    if (/^Sender \(untrusted metadata\):/i.test(trimmed))
+        return true;
+    return stripMessageMetadata(trimmed).trim().length === 0 && /Sender \(untrusted metadata\):/i.test(trimmed);
+}
+function shouldDropInboundMessage(msg) {
+    if (isPersistableInboundRole(msg.role))
+        return false;
+    const text = extractTextFromInboundContent(msg.content);
+    // OpenClaw/webchat may carry adjacent sender envelopes as role='custom'. They
+    // are transport metadata, not conversation turns. Persisting or replaying them
+    // lets the metadata row become the latest assistant-ish history item, which can
+    // mask the real user prompt under load.
+    if (msg.role === 'custom' && isSenderMetadataText(text))
+        return true;
+    // Unknown roles are not part of the provider transcript contract. Dropping is
+    // safer than casting them to assistant in toNeutralMessage().
+    return true;
+}
+function sanitizeInboundMessageForModel(msg) {
+    if (shouldDropInboundMessage(msg))
+        return null;
+    // OpenClaw prepends untrusted inbound metadata directly to user-role text for
+    // the current provider prompt. That context is useful for the live inbound
+    // turn, but if HyperMem replays it from the message window it becomes duplicate
+    // user content and can make the model repeat the previous message. Strip it at
+    // the context-engine boundary while preserving the actual user text.
+    if (msg.role !== 'user')
+        return msg;
+    if (typeof msg.content === 'string') {
+        const stripped = stripMessageMetadata(msg.content);
+        return stripped === msg.content ? msg : { ...msg, content: stripped };
+    }
+    if (Array.isArray(msg.content)) {
+        let changed = false;
+        const content = msg.content.map((part) => {
+            if (!part || part.type !== 'text' || typeof part.text !== 'string')
+                return part;
+            const stripped = stripMessageMetadata(part.text);
+            if (stripped !== part.text)
+                changed = true;
+            return stripped === part.text ? part : { ...part, text: stripped };
+        });
+        return changed ? { ...msg, content } : msg;
+    }
+    return msg;
+}
+function sanitizeInboundMessagesForModel(messages) {
+    return messages
+        .map((msg) => sanitizeInboundMessageForModel(msg))
+        .filter((msg) => Boolean(msg));
+}
 /**
  * Determine whether a user turn is "topic-bearing" (substantive).
  *
@@ -1046,7 +1106,7 @@ function toNeutralMessage(msg) {
         toolCalls = contentBlockToolCalls;
     }
     // OpenClaw uses role 'toolResult' (camelCase). Support all three spellings.
-    const isToolResultMsg = msg.role === 'tool' || msg.role === 'tool_result' || msg.role === 'toolResult';
+    const isToolResultMsg = isToolResultRole(msg.role);
     // Tool results must stay on the result side of the transcript. If we persist them as
     // assistant rows with orphaned toolResults, later replay can retain a tool_result after
     // trimming away the matching assistant tool_use, which Anthropic rejects with a 400.
@@ -1067,6 +1127,27 @@ function toNeutralMessage(msg) {
         toolResults,
     };
 }
+function isPlainUserTurnMessage(msg) {
+    if (msg.role !== 'user' || shouldDropInboundMessage(msg))
+        return false;
+    const neutral = toNeutralMessage(msg);
+    return neutral.role === 'user'
+        && !neutral.toolResults?.length
+        && stripMessageMetadata(neutral.textContent ?? '').trim().length > 0;
+}
+function findBoundaryPlainUserMessage(messages, prePromptMessageCount) {
+    const start = Math.min(prePromptMessageCount - 1, messages.length - 1);
+    for (let i = start; i >= 0; i--) {
+        const msg = messages[i];
+        if (shouldDropInboundMessage(msg))
+            continue;
+        if (isPlainUserTurnMessage(msg))
+            return msg;
+        if (msg.role === 'assistant' || msg.role === 'system' || isToolResultRole(msg.role))
+            return null;
+    }
+    return null;
+}
 // ─── Context Engine Implementation ─────────────────────────────
 /**
  * In-flight warm dedup map.
@@ -1640,7 +1721,7 @@ function createHyperMemEngine() {
             }
             // Skip system messages — they come from the runtime, not the conversation
             const msg = message;
-            if (msg.role === 'system') {
+            if (msg.role === 'system' || shouldDropInboundMessage(msg)) {
                 return { ingested: false };
             }
             try {
@@ -1755,7 +1836,7 @@ function createHyperMemEngine() {
                 const agentId = extractAgentId(sk);
                 for (const message of messages) {
                     const msg = message;
-                    if (msg.role === 'system')
+                    if (msg.role === 'system' || shouldDropInboundMessage(msg))
                         continue;
                     const neutral = toNeutralMessage(msg);
                     if (neutral.role === 'user' && !neutral.toolResults?.length) {
@@ -1785,6 +1866,11 @@ function createHyperMemEngine() {
          *   systemPromptAddition — facts/recall/episodes injected before runtime system prompt
          */
         async assemble({ sessionId, sessionKey, messages, tokenBudget, prompt, model }) {
+            // Drop non-provider transcript metadata before any pass-through, token
+            // estimate, or persistence-adjacent decision. The current user prompt is
+            // carried separately as `prompt`; adjacent role='custom' sender envelopes
+            // must never become the final model message.
+            messages = sanitizeInboundMessagesForModel(messages);
             // ── Tool-loop guard ──────────────────────────────────────────────────────
             // When the last message is a toolResult, the runtime is mid tool-loop:
             // the model already has full context from the initial turn assembly.
@@ -2795,7 +2881,10 @@ ${replayRecovery.emittedText}`
          *
          * IMPORTANT: When afterTurn is defined, the runtime calls ONLY afterTurn —
          * it never calls ingest() or ingestBatch(). So we must ingest the new
-         * messages here, using messages.slice(prePromptMessageCount).
+         * messages here. Some OpenClaw runtimes include the current user + adjacent
+         * transport metadata in prePromptMessageCount, leaving only assistant/tool
+         * outputs in messages.slice(prePromptMessageCount). Reconcile the boundary
+         * user turn explicitly so SQLite preserves the real user→assistant order.
          */
         async afterTurn({ sessionId, sessionKey, messages, prePromptMessageCount, isHeartbeat, runtimeContext }) {
             if (isHeartbeat)
@@ -2804,12 +2893,25 @@ ${replayRecovery.emittedText}`
                 const hm = await getHyperMem();
                 const sk = resolveSessionKey(sessionId, sessionKey);
                 const agentId = extractAgentId(sk);
-                // Ingest only the new messages produced this turn
+                // Ingest only the new messages produced this turn, plus the boundary
+                // user turn when OpenClaw counted it as pre-prompt context. The write
+                // path is idempotent, so if an ingress/runtime variant already persisted
+                // this user message, recordUserMessage() suppresses the duplicate.
                 const newMessages = messages.slice(prePromptMessageCount);
-                for (const msg of newMessages) {
+                const hasNewPlainUserMessage = newMessages
+                    .map(m => m)
+                    .some(m => isPlainUserTurnMessage(m));
+                const boundaryUserMessage = hasNewPlainUserMessage
+                    ? null
+                    : findBoundaryPlainUserMessage(messages, prePromptMessageCount);
+                const currentTurnMessages = boundaryUserMessage
+                    ? [boundaryUserMessage, ...newMessages]
+                    : newMessages;
+                for (const msg of currentTurnMessages) {
                     const m = msg;
-                    // Skip system messages — they come from the runtime, not the conversation
-                    if (m.role === 'system')
+                    // Skip system and non-provider metadata messages — they come from the
+                    // runtime/transport, not the conversation.
+                    if (m.role === 'system' || shouldDropInboundMessage(m))
                         continue;
                     if (m.role === 'toolResult' && extractTextFromInboundContent(m.content).trim() === SYNTHETIC_MISSING_TOOL_RESULT_TEXT) {
                         const toolCallId = typeof m.toolCallId === 'string' ? m.toolCallId : 'unknown';
@@ -2845,7 +2947,7 @@ ${replayRecovery.emittedText}`
                     }
                 }
                 try {
-                    const lastAssistantMessage = [...newMessages].reverse().find(m => m.role === 'assistant');
+                    const lastAssistantMessage = [...currentTurnMessages].reverse().find(m => m.role === 'assistant');
                     if (lastAssistantMessage) {
                         const modelState = await hm.cache.getModelState(agentId, sk).catch(() => null);
                         const promptCacheUsage = runtimeContext?.promptCache?.lastCallUsage;
@@ -2893,7 +2995,7 @@ ${replayRecovery.emittedText}`
                 // Non-fatal: topic detection never blocks afterTurn
                 let adaptiveTopicShiftConfidence;
                 try {
-                    const inboundUserMsg = newMessages
+                    const inboundUserMsg = currentTurnMessages
                         .map(m => m)
                         .find(m => m.role === 'user');
                     if (inboundUserMsg) {
@@ -2954,7 +3056,7 @@ ${replayRecovery.emittedText}`
                         const modelState = await hm.cache.getModelState(agentId, sk);
                         const gradientBudget = modelState?.tokenBudget;
                         const gradientDepth = modelState?.historyDepth;
-                        const inboundUserMsg = newMessages
+                        const inboundUserMsg = currentTurnMessages
                             .map(m => m)
                             .find(m => m.role === 'user');
                         const inboundUserText = inboundUserMsg