@martian-engineering/lossless-claw 0.5.1 → 0.5.3

package/README.md

@@ -59,6 +59,8 @@ openclaw plugins install --link /path/to/lossless-claw
 
 The install command records the plugin, enables it, and applies compatible slot selection (including `contextEngine` when applicable).
 
+> **Note:** If your OpenClaw config uses `plugins.allow`, make sure both `lossless-claw` and any active plugins you rely on remain allowlisted. In some setups, narrowing the allowlist can prevent plugin-backed integrations from loading, even if `lossless-claw` itself is installed correctly. Restart the gateway after plugin config changes.
+
 ### Configure OpenClaw
 
 In most cases, no manual JSON edits are needed after `openclaw plugins install`.
@@ -92,14 +94,16 @@ Add a `lossless-claw` entry under `plugins.entries` in your OpenClaw config:
       "lossless-claw": {
         "enabled": true,
         "config": {
-          "freshTailCount": 32,
+          "freshTailCount": 64,
+          "leafChunkTokens": 80000,
           "contextThreshold": 0.75,
-          "incrementalMaxDepth": -1,
+          "incrementalMaxDepth": 1,
           "ignoreSessionPatterns": [
             "agent:*:cron:**"
           ],
-          "summaryProvider": "anthropic",
-          "summaryModel": "claude-3-5-haiku"
+          "summaryModel": "anthropic/claude-haiku-4-5",
+          "expansionModel": "anthropic/claude-haiku-4-5",
+          "delegationTimeoutMs": 300000
         }
       }
     }
@@ -107,7 +111,7 @@ Add a `lossless-claw` entry under `plugins.entries` in your OpenClaw config:
 }
 ```
 
-`summaryModel` and `summaryProvider` let you pin compaction summarization to a cheaper or faster model than your main OpenClaw session model. When unset, LCM uses OpenClaw's configured default model/provider.
+`leafChunkTokens` controls how many source tokens can accumulate in a leaf compaction chunk before summarization is triggered. The default is `20000`, but quota-limited summary providers may benefit from a larger value to reduce compaction frequency. `summaryModel` and `summaryProvider` let you pin compaction summarization to a cheaper or faster model than your main OpenClaw session model. `expansionModel` does the same for `lcm_expand_query` sub-agent calls (drilling into summaries to recover detail). `delegationTimeoutMs` controls how long `lcm_expand_query` waits for that delegated sub-agent to finish before returning a timeout error; it defaults to `120000` (120s). When unset, the model settings still fall back to OpenClaw's configured default model/provider. See [Expansion model override requirements](#expansion-model-override-requirements) for the required `subagent` trust policy when using `expansionModel`.
 
 ### Environment variables
 
@@ -119,11 +123,11 @@ Add a `lossless-claw` entry under `plugins.entries` in your OpenClaw config:
 | `LCM_STATELESS_SESSION_PATTERNS` | `""` | Comma-separated glob patterns for session keys that may read from LCM but never write to it |
 | `LCM_SKIP_STATELESS_SESSIONS` | `true` | Enable stateless-session write skipping for matching session keys |
 | `LCM_CONTEXT_THRESHOLD` | `0.75` | Fraction of context window that triggers compaction (0.0–1.0) |
-| `LCM_FRESH_TAIL_COUNT` | `32` | Number of recent messages protected from compaction |
+| `LCM_FRESH_TAIL_COUNT` | `64` | Number of recent messages protected from compaction |
 | `LCM_LEAF_MIN_FANOUT` | `8` | Minimum raw messages per leaf summary |
 | `LCM_CONDENSED_MIN_FANOUT` | `4` | Minimum summaries per condensed node |
 | `LCM_CONDENSED_MIN_FANOUT_HARD` | `2` | Relaxed fanout for forced compaction sweeps |
-| `LCM_INCREMENTAL_MAX_DEPTH` | `0` | How deep incremental compaction goes (0 = leaf only, -1 = unlimited) |
+| `LCM_INCREMENTAL_MAX_DEPTH` | `1` | How deep incremental compaction goes (0 = leaf only, 1 = one condensed pass, -1 = unlimited) |
 | `LCM_LEAF_CHUNK_TOKENS` | `20000` | Max source tokens per leaf compaction chunk |
 | `LCM_LEAF_TARGET_TOKENS` | `1200` | Target token count for leaf summaries |
 | `LCM_CONDENSED_TARGET_TOKENS` | `2000` | Target token count for condensed summaries |
@@ -133,8 +137,10 @@ Add a `lossless-claw` entry under `plugins.entries` in your OpenClaw config:
 | `LCM_LARGE_FILE_SUMMARY_MODEL` | `""` | Model override for large-file summarization |
 | `LCM_SUMMARY_MODEL` | `""` | Model override for compaction summarization; falls back to OpenClaw's default model when unset |
 | `LCM_SUMMARY_PROVIDER` | `""` | Provider override for compaction summarization; falls back to `OPENCLAW_PROVIDER` or the provider embedded in the model ref |
+| `LCM_SUMMARY_BASE_URL` | *(from OpenClaw / provider default)* | Base URL override for summarization API calls |
 | `LCM_EXPANSION_MODEL` | *(from OpenClaw)* | Model override for `lcm_expand_query` sub-agent (e.g. `anthropic/claude-haiku-4-5`) |
 | `LCM_EXPANSION_PROVIDER` | *(from OpenClaw)* | Provider override for `lcm_expand_query` sub-agent |
+| `LCM_DELEGATION_TIMEOUT_MS` | `120000` | Max time to wait for delegated `lcm_expand_query` sub-agent completion |
 | `LCM_AUTOCOMPACT_DISABLED` | `false` | Disable automatic compaction after turns |
 | `LCM_PRUNE_HEARTBEAT_OK` | `false` | Retroactively delete `HEARTBEAT_OK` turn cycles from LCM storage |
 
@@ -178,6 +184,7 @@ Plugin config equivalents:
 - `skipStatelessSessions`
 - `summaryModel`
 - `summaryProvider`
+- `delegationTimeoutMs`
 
 Environment variables still win over plugin config when both are set.
 
@@ -195,13 +202,15 @@ If `summaryModel` already includes a provider prefix such as `anthropic/claude-s
 ### Recommended starting configuration
 
 ```
-LCM_FRESH_TAIL_COUNT=32
-LCM_INCREMENTAL_MAX_DEPTH=-1
+LCM_FRESH_TAIL_COUNT=64
+LCM_LEAF_CHUNK_TOKENS=20000
+LCM_INCREMENTAL_MAX_DEPTH=1
 LCM_CONTEXT_THRESHOLD=0.75
 ```
 
-- **freshTailCount=32** protects the last 32 messages from compaction, giving the model enough recent context for continuity.
-- **incrementalMaxDepth=-1** enables unlimited automatic condensation after each compaction pass — the DAG cascades as deep as needed. Set to `0` (default) for leaf-only, or a positive integer for a specific depth cap.
+- **freshTailCount=64** protects the last 64 messages from compaction, giving the model more recent context for continuity.
+- **leafChunkTokens=20000** limits how large each leaf compaction chunk can grow before LCM summarizes it. Increase this when your summary provider is quota-limited and frequent leaf compactions are exhausting that quota.
+- **incrementalMaxDepth=1** runs one condensed pass after each leaf compaction by default. Set to `0` for leaf-only behavior, a larger positive integer for a deeper cap, or `-1` for unlimited cascading.
 - **contextThreshold=0.75** triggers compaction when context reaches 75% of the model's window, leaving headroom for the model's response.
 
 ### Session exclusion patterns

package/docs/configuration.md

@@ -91,6 +91,25 @@ The actual summary size depends on the LLM's output; these values are guidelines
 - Smaller chunks create summaries more frequently from less material.
 - This also affects the condensed minimum input threshold (10% of this value).
 
+### Maximum assembly token budget
+
+`LCM_MAX_ASSEMBLY_TOKEN_BUDGET` (default: none) caps the token budget used for context assembly and compaction threshold evaluation. When set, this takes precedence over both the 128k fallback and runtime-provided budgets.
+
+Set this if you're using a model with a smaller context window:
+
+- **8k models:** `LCM_MAX_ASSEMBLY_TOKEN_BUDGET=7000`
+- **32k models:** `LCM_MAX_ASSEMBLY_TOKEN_BUDGET=30000`
+- **128k+ models:** No need to set (128k fallback is appropriate)
+
+### Summary size cap
+
+`LCM_SUMMARY_MAX_OVERAGE_FACTOR` (default: `3`) controls the hard ceiling on summary sizes relative to the target tokens (`leafTargetTokens` for leaf summaries, `condensedTargetTokens` for condensed summaries).
+
+If a summary exceeds `overage_factor * target_tokens`, it is deterministically truncated. A warning is logged when any summary exceeds `1.5 * target_tokens`.
+
+- **Lower values** (e.g., 2) enforce tighter summaries but may truncate more often with weaker summarizer models.
+- **Higher values** (e.g., 5) allow more LLM flexibility but risk storing oversized summaries.
+
 ## Model selection
 
 LCM uses the same model as the parent OpenClaw session for summarization by default. You can override this:
@@ -99,6 +118,7 @@ LCM uses the same model as the parent OpenClaw session for summarization by defa
 # Use a specific model for summarization
 export LCM_SUMMARY_MODEL=anthropic/claude-sonnet-4-20250514
 export LCM_SUMMARY_PROVIDER=anthropic
+export LCM_SUMMARY_BASE_URL=https://api.anthropic.com
 ```
 
 Using a cheaper/faster model for summarization can reduce costs, but quality matters — poor summaries compound as they're condensed into higher-level nodes.
@@ -112,6 +132,8 @@ When more than one source is present, compaction summarization resolves in this
 
 If `summaryModel` already includes a provider prefix such as `anthropic/claude-sonnet-4-20250514`, `summaryProvider` is ignored for that choice.
 
+For delegated `lcm_expand_query` runs, you can extend the sub-agent wait window with `delegationTimeoutMs` (plugin config) or `LCM_DELEGATION_TIMEOUT_MS` (environment variable). The default is `120000` milliseconds.
+
 ## Session controls
 
 ### Excluding sessions entirely

package/docs/tui.md

@@ -287,6 +287,9 @@ lcm-tui rewrite 44 --all --apply --diff
 # Rewrite with OpenAI Responses API
 lcm-tui rewrite 44 --summary sum_abc123 --provider openai --model gpt-5.3-codex --apply
 
+# Rewrite through a custom OpenAI-compatible proxy
+lcm-tui rewrite 44 --summary sum_abc123 --provider openai --model gpt-5.3-codex --base-url https://proxy.example.com/openai --apply
+
 # Use custom prompt templates
 lcm-tui rewrite 44 --all --apply --prompt-dir ~/.config/lcm-tui/prompts
 ```
@@ -301,6 +304,7 @@ lcm-tui rewrite 44 --all --apply --prompt-dir ~/.config/lcm-tui/prompts
 | `--diff` | Show unified diff |
 | `--provider <id>` | API provider (inferred from `--model` when omitted) |
 | `--model <model>` | API model (default depends on provider) |
+| `--base-url <url>` | Custom API base URL (overrides config and env) |
 | `--prompt-dir <path>` | Custom prompt template directory |
 | `--timestamps` | Inject timestamps into source text (default: true) |
 | `--tz <timezone>` | Timezone for timestamps (default: system local) |
@@ -378,6 +382,9 @@ lcm-tui backfill my-agent session_abc123 --apply --transplant-to 653
 
 # Backfill using OpenAI
 lcm-tui backfill my-agent session_abc123 --apply --provider openai --model gpt-5.3-codex
+
+# Backfill through a custom OpenAI-compatible proxy
+lcm-tui backfill my-agent session_abc123 --apply --provider openai --model gpt-5.3-codex --base-url https://proxy.example.com/openai
 ```
 
 All write paths are transactional:
@@ -404,6 +411,7 @@ An idempotency guard prevents duplicate imports for the same `session_id`.
 | `--fresh-tail <n>` | Preserve freshest N raw messages from leaf compaction |
 | `--provider <id>` | API provider (inferred from model when omitted) |
 | `--model <id>` | API model (default depends on provider) |
+| `--base-url <url>` | Custom API base URL (overrides config and env) |
 | `--prompt-dir <path>` | Custom depth-prompt directory |
 
 ### `lcm-tui prompts`
@@ -479,9 +487,10 @@ If the provider auth profile mode is `oauth` (not `api_key`), set the provider A
 Interactive rewrite (`w`/`W`) can be configured with:
 - `LCM_TUI_SUMMARY_PROVIDER`
 - `LCM_TUI_SUMMARY_MODEL`
+- `LCM_TUI_SUMMARY_BASE_URL`
 - `LCM_TUI_CONVERSATION_WINDOW_SIZE` (default `200`)
 
-It also honors `LCM_SUMMARY_PROVIDER` / `LCM_SUMMARY_MODEL` as fallback.
+It also honors `LCM_SUMMARY_PROVIDER` / `LCM_SUMMARY_MODEL` / `LCM_SUMMARY_BASE_URL` as fallback.
 
 ## Database
 

package/openclaw.plugin.json

@@ -13,6 +13,10 @@
       "label": "Fresh Tail Count",
       "help": "Number of recent messages protected from compaction"
     },
+    "leafChunkTokens": {
+      "label": "Leaf Chunk Tokens",
+      "help": "Maximum source tokens per leaf compaction chunk before summarization"
+    },
     "dbPath": {
       "label": "Database Path",
       "help": "Path to LCM SQLite database (default: ~/.openclaw/lcm.db)"
@@ -44,6 +48,22 @@
     "expansionProvider": {
       "label": "Expansion Provider",
       "help": "Provider override for lcm_expand_query sub-agent (e.g., 'anthropic')"
+    },
+    "delegationTimeoutMs": {
+      "label": "Delegation Timeout (ms)",
+      "help": "Maximum time to wait for delegated lcm_expand_query sub-agent completion before timing out"
+    },
+    "maxAssemblyTokenBudget": {
+      "label": "Max Assembly Token Budget",
+      "help": "Hard ceiling for assembly token budget — caps runtime-provided and fallback budgets. Set for smaller context-window models (e.g., 30000 for 32k models)"
+    },
+    "summaryMaxOverageFactor": {
+      "label": "Summary Max Overage Factor",
+      "help": "Maximum allowed overage factor for summaries relative to target tokens (default 3). Summaries exceeding this are deterministically truncated."
+    },
+    "customInstructions": {
+      "label": "Custom Instructions",
+      "help": "Natural language instructions injected into all summarization prompts (e.g., formatting rules, tone control)"
     }
   },
   "configSchema": {
@@ -66,6 +86,10 @@
         "type": "integer",
         "minimum": 1
       },
+      "leafChunkTokens": {
+        "type": "integer",
+        "minimum": 1
+      },
       "leafMinFanout": {
         "type": "integer",
         "minimum": 2
@@ -111,6 +135,21 @@
       },
       "expansionProvider": {
         "type": "string"
+      },
+      "delegationTimeoutMs": {
+        "type": "integer",
+        "minimum": 1
+      },
+      "maxAssemblyTokenBudget": {
+        "type": "integer",
+        "minimum": 1000
+      },
+      "summaryMaxOverageFactor": {
+        "type": "number",
+        "minimum": 1
+      },
+      "customInstructions": {
+        "type": "string"
       }
     }
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@martian-engineering/lossless-claw",
-  "version": "0.5.1",
+  "version": "0.5.3",
   "description": "Lossless Context Management plugin for OpenClaw — DAG-based conversation summarization with incremental compaction",
   "type": "module",
   "main": "index.ts",

package/src/assembler.ts

@@ -9,6 +9,15 @@ import type { SummaryStore, ContextItemRecord, SummaryRecord } from "./store/sum
 
 type AgentMessage = Parameters<ContextEngine["ingest"]>[0]["message"];
 
+const TOOL_CALL_TYPES = new Set([
+  "toolCall",
+  "toolUse",
+  "tool_use",
+  "tool-use",
+  "functionCall",
+  "function_call",
+]);
+
 // ── Public types ─────────────────────────────────────────────────────────────
 
 export interface AssembleContextInput {
@@ -67,6 +76,10 @@ function buildSystemPromptAddition(summarySignals: SummaryPromptSignal[]): strin
     "",
     "**Recall priority:** Use LCM tools first for compacted conversation history. If LCM does not cover the needed data, prefer any available memory/recall tool before falling back to raw text search.",
     "",
+    "**Conflict handling:** If newer evidence conflicts with an older summary or recollection, prefer the newer evidence. Do not trust a stale summary over fresher contradictory information.",
+    "",
+    "**Contradictions/uncertainty:** If facts seem contradictory or uncertain, verify with LCM tools before answering instead of trusting the summary at face value.",
+    "",
     "**Tool escalation:**",
     "1. `lcm_grep` — search by regex or full-text across messages and summaries",
     "2. `lcm_describe` — inspect a specific summary (cheap, no sub-agent)",
@@ -92,6 +105,7 @@ function buildSystemPromptAddition(summarySignals: SummaryPromptSignal[]): strin
       "3) Answer with citations to summary IDs used",
       "",
       "**Uncertainty checklist (run before answering):**",
+      "- Am I relying on an older summary even though newer evidence disagrees?",
       "- Am I making exact factual claims from a condensed summary?",
       "- Could compaction have omitted a crucial detail?",
       "- Would this answer fail if the user asks for proof?",
@@ -555,6 +569,174 @@ function pickToolIsError(parts: MessagePartRecord[]): boolean | undefined {
   return undefined;
 }
 
+function extractToolCallId(block: { id?: unknown; call_id?: unknown }): string | null {
+  if (typeof block.id === "string" && block.id.length > 0) {
+    return block.id;
+  }
+  if (typeof block.call_id === "string" && block.call_id.length > 0) {
+    return block.call_id;
+  }
+  return null;
+}
+
+function extractToolCallIdsFromAssistant(message: AgentMessage): string[] {
+  if (message?.role !== "assistant" || !Array.isArray(message.content)) {
+    return [];
+  }
+
+  const ids: string[] = [];
+  for (const block of message.content) {
+    if (!block || typeof block !== "object") {
+      continue;
+    }
+    const record = block as { type?: unknown; id?: unknown; call_id?: unknown };
+    if (typeof record.type !== "string" || !TOOL_CALL_TYPES.has(record.type)) {
+      continue;
+    }
+    const id = extractToolCallId(record);
+    if (id) {
+      ids.push(id);
+    }
+  }
+  return ids;
+}
+
+function extractToolResultIdFromMessage(message: AgentMessage): string | null {
+  if (!message || typeof message !== "object") {
+    return null;
+  }
+  if (typeof message.toolCallId === "string" && message.toolCallId.length > 0) {
+    return message.toolCallId;
+  }
+  if (typeof message.toolUseId === "string" && message.toolUseId.length > 0) {
+    return message.toolUseId;
+  }
+  return null;
+}
+
+function collectAssistantToolCallIds(items: ResolvedItem[]): Set<string> {
+  const ids = new Set<string>();
+  for (const item of items) {
+    for (const id of extractToolCallIdsFromAssistant(item.message)) {
+      ids.add(id);
+    }
+  }
+  return ids;
+}
+
+function mergeFreshTailWithMatchingToolResults(
+  freshTail: ResolvedItem[],
+  matchingToolResults: ResolvedItem[],
+): ResolvedItem[] {
+  if (matchingToolResults.length === 0) {
+    return freshTail;
+  }
+
+  const resultsById = new Map<string, ResolvedItem[]>();
+  for (const item of matchingToolResults) {
+    const toolResultId = extractToolResultIdFromMessage(item.message);
+    if (!toolResultId) {
+      continue;
+    }
+    const existing = resultsById.get(toolResultId);
+    if (existing) {
+      existing.push(item);
+    } else {
+      resultsById.set(toolResultId, [item]);
+    }
+  }
+
+  const merged: ResolvedItem[] = [];
+  const usedOrdinals = new Set<number>();
+
+  for (const item of freshTail) {
+    merged.push(item);
+
+    const toolCallIds = extractToolCallIdsFromAssistant(item.message);
+    if (toolCallIds.length === 0) {
+      continue;
+    }
+
+    for (const toolCallId of toolCallIds) {
+      const matches = resultsById.get(toolCallId);
+      if (!matches) {
+        continue;
+      }
+      for (const match of matches) {
+        if (usedOrdinals.has(match.ordinal)) {
+          continue;
+        }
+        merged.push(match);
+        usedOrdinals.add(match.ordinal);
+      }
+    }
+  }
+
+  for (const item of matchingToolResults) {
+    if (!usedOrdinals.has(item.ordinal)) {
+      merged.push(item);
+    }
+  }
+
+  return merged;
+}
+
+function filterNonFreshAssistantToolCalls(
+  items: ResolvedItem[],
+  freshTailOrdinals: Set<number>,
+): AgentMessage[] {
+  const availableToolResultIds = new Set<string>();
+  for (const item of items) {
+    const toolResultId = extractToolResultIdFromMessage(item.message);
+    if (toolResultId) {
+      availableToolResultIds.add(toolResultId);
+    }
+  }
+
+  const filteredMessages: AgentMessage[] = [];
+  for (const item of items) {
+    if (item.message?.role !== "assistant" || freshTailOrdinals.has(item.ordinal)) {
+      filteredMessages.push(item.message);
+      continue;
+    }
+
+    if (!Array.isArray(item.message.content)) {
+      filteredMessages.push(item.message);
+      continue;
+    }
+
+    let removedAny = false;
+    const content = item.message.content.filter((block) => {
+      if (!block || typeof block !== "object") {
+        return true;
+      }
+      const record = block as { type?: unknown; id?: unknown; call_id?: unknown };
+      if (typeof record.type !== "string" || !TOOL_CALL_TYPES.has(record.type)) {
+        return true;
+      }
+      const toolCallId = extractToolCallId(record);
+      if (!toolCallId || availableToolResultIds.has(toolCallId)) {
+        return true;
+      }
+      removedAny = true;
+      return false;
+    });
+
+    if (content.length === 0) {
+      continue;
+    }
+    if (!removedAny) {
+      filteredMessages.push(item.message);
+      continue;
+    }
+    filteredMessages.push({
+      ...item.message,
+      content: content as typeof item.message.content,
+    } as AgentMessage);
+  }
+  return filteredMessages;
+}
+
 /** Format a Date for XML attributes in the agent's timezone. */
 function formatDateForAttribute(date: Date, timezone?: string): string {
   const tz = timezone ?? "UTC";
@@ -692,8 +874,17 @@ export class ContextAssembler {
 
     // Step 3: Split into evictable prefix and protected fresh tail
     const tailStart = Math.max(0, resolved.length - freshTailCount);
-    const freshTail = resolved.slice(tailStart);
-    const evictable = resolved.slice(0, tailStart);
+    const baseFreshTail = resolved.slice(tailStart);
+    const initialEvictable = resolved.slice(0, tailStart);
+    const freshTailOrdinals = new Set(baseFreshTail.map((item) => item.ordinal));
+    const tailToolCallIds = collectAssistantToolCallIds(baseFreshTail);
+    const tailPairToolResults = initialEvictable.filter((item) => {
+      const toolResultId = extractToolResultIdFromMessage(item.message);
+      return toolResultId !== null && tailToolCallIds.has(toolResultId);
+    });
+    const protectedEvictableOrdinals = new Set(tailPairToolResults.map((item) => item.ordinal));
+    const evictable = initialEvictable.filter((item) => !protectedEvictableOrdinals.has(item.ordinal));
+    const freshTail = mergeFreshTailWithMatchingToolResults(baseFreshTail, tailPairToolResults);
 
     // Step 4: Budget-aware selection
     // First, compute the token cost of the fresh tail (always included).
@@ -747,7 +938,7 @@ export class ContextAssembler {
 
     // Normalize assistant string content to array blocks (some providers return
     // content as a plain string; Anthropic expects content block arrays).
-    const rawMessages = selected.map((item) => item.message);
+    const rawMessages = filterNonFreshAssistantToolCalls(selected, freshTailOrdinals);
     for (let i = 0; i < rawMessages.length; i++) {
       const msg = rawMessages[i];
       if (msg?.role === "assistant" && typeof msg.content === "string") {