v4.3.0: remove silent truncation and harden routing

Preserve critical task and memory artifacts without silent clipping, make subtask lifecycle states honest, and restore cache-aware provider failover so degraded routes are visible instead of failing opaquely.

Made-with: Cursor

Author: Anton

.gitignore

@@ -17,6 +17,7 @@ __pycache__/
 .pytest_cache/
 .mypy_cache/
 .ruff_cache/
+.pyinstaller-cache/
 
 # Virtual environments
 venv/

README.md

@@ -6,7 +6,7 @@
 [![macOS 12+](https://img.shields.io/badge/macOS-12%2B-black.svg)](https://github.com/joi-lab/ouroboros-desktop/releases)
 [![Linux](https://img.shields.io/badge/Linux-x86__64-orange.svg)](https://github.com/joi-lab/ouroboros-desktop/releases)
 [![Windows](https://img.shields.io/badge/Windows-x64-blue.svg)](https://github.com/joi-lab/ouroboros-desktop/releases)
-[![Version 4.2.0](https://img.shields.io/badge/version-4.2.0-green.svg)](VERSION)
+[![Version 4.3.0](https://img.shields.io/badge/version-4.3.0-green.svg)](VERSION)
 
 A self-modifying AI agent that writes its own code, rewrites its own mind, and evolves autonomously. Born February 16, 2026.
 
@@ -238,6 +238,7 @@ Full text: [BIBLE.md](BIBLE.md)
 
 | Version | Date | Description |
 |---------|------|-------------|
+| 4.3.0 | 2026-03-19 | Reliability and continuity release: remove silent truncation from critical task/memory paths, persist honest subtask lifecycle states and full task results, restore transient chat wake banner, replace local-model hard prompt slicing with explicit non-core compaction plus fail-fast overflow, route Anthropic/OpenRouter calls without hard provider pinning while keeping parameter guarantees, and align async review calls with shared LLM routing/usage observability. |
 | 4.2.0 | 2026-03-16 | Cross-platform hardening release: replace Unix-only file locking in memory/consolidation with Windows-safe locking, refresh default model tiers (Opus main/code, Sonnet light/fallback, task effort `medium`), improve reconnect recovery with heartbeat/watchdog/history resync, switch local model chat format to auto-detect, and sync public docs with the current codebase and BIBLE structure. |
 | 4.1.0 | 2026-03-16 | Public desktop release: port the v4 architecture and UI into the platform branch, preserve cross-platform packaging and Windows runtime support, and ship signed notarized macOS packaging. |
 | 4.0.9 | 2026-03-15 | Packaging completeness release: bundle `assets/`, restore custom app icon from `assets/icon.icns`, and copy assets into the bootstrapped repo on fresh install so the shipped app and repo are no longer missing the visual asset layer. |

VERSION

@@ -1 +1 @@
-4.2.0
+4.3.0

build.sh

@@ -47,6 +47,9 @@ PY
 
 rm -rf build dist
 
+export PYINSTALLER_CONFIG_DIR="$PWD/.pyinstaller-cache"
+mkdir -p "$PYINSTALLER_CONFIG_DIR"
+
 echo "--- Running PyInstaller ---"
 python3 -m PyInstaller Ouroboros.spec --clean --noconfirm
 

build_linux.sh

@@ -20,6 +20,9 @@ python-standalone/bin/pip3 install -q -r requirements.txt
 
 rm -rf build dist
 
+export PYINSTALLER_CONFIG_DIR="$PWD/.pyinstaller-cache"
+mkdir -p "$PYINSTALLER_CONFIG_DIR"
+
 echo "--- Running PyInstaller ---"
 python -m PyInstaller Ouroboros.spec --clean --noconfirm
 

build_windows.ps1

@@ -23,6 +23,9 @@ Write-Host "--- Installing agent dependencies into python-standalone ---"
 if (Test-Path "build") { Remove-Item -Recurse -Force "build" }
 if (Test-Path "dist") { Remove-Item -Recurse -Force "dist" }
 
+$env:PYINSTALLER_CONFIG_DIR = Join-Path (Get-Location) ".pyinstaller-cache"
+New-Item -ItemType Directory -Force -Path $env:PYINSTALLER_CONFIG_DIR | Out-Null
+
 Write-Host "--- Running PyInstaller ---"
 python -m PyInstaller Ouroboros.spec --clean --noconfirm
 

docs/ARCHITECTURE.md

@@ -1,4 +1,4 @@
-# Ouroboros v4.2.0 — Architecture & Reference
+# Ouroboros v4.3.0 — Architecture & Reference
 
 This document describes every component, page, button, API endpoint, and data flow.
 It is the single source of truth for how the system works. Keep it updated.
@@ -184,6 +184,7 @@ Navigation is a left sidebar with 8 pages.
 - **Progress messages**: background consciousness thinking shown as dimmed bubbles with 💬 prefix.
 - **Typing indicator**: animated "thinking dots" bubble appears when the agent is processing.
 - **Persistence**: chat history loaded from server on page load (`/api/chat/history`), survives app restarts. Fallback to sessionStorage.
+- **Empty-chat init**: if neither server history nor sessionStorage has messages, the UI shows a transient assistant bubble: `Ouroboros has awakened`. This is visual-only and is not written to chat history.
 - Messages sent via WebSocket `{type: "chat", content: text}`.
 - Responses arrive via WebSocket `{type: "chat", role: "assistant", content: text, ts: "ISO"}`.
 - Supports slash commands: `/status`, `/evolve`, `/review`, `/bg`, `/restart`, `/panic`.
@@ -363,7 +364,7 @@ Each iteration (0.5s sleep):
 - Browser tools use thread-sticky executor (Playwright greenlet affinity)
 - All tools have hard timeout (default 360s, per-tool overrides for browser/search/vision)
 - Multi-layer safety: hardcoded sandbox (registry.py) → deterministic whitelist → LLM safety supervisor
-- Tool results truncated per-tool (repo_read/data_read: 80k, run_shell: 40k, default: 15k chars)
+- Tool results use explicit per-tool caps with visible truncation markers (`repo_read`/`data_read`/`knowledge_read`/`run_shell`: 80k, default: 15k chars). Cognitive reads (`memory/*`, prompts, BIBLE/docs, commit/review outputs) are exempt from silent clipping.
 - Context compaction kicks in after round 8 (summarizes old tool results)
 
 ### Git tools (tools/git.py + tools/review.py + supervisor/git_ops.py)
@@ -399,6 +400,7 @@ Multi-layer security:
 3. **LLM Layer 1 (fast)**: Light model checks remaining tool calls for SAFE/SUSPICIOUS/DANGEROUS.
 4. **LLM Layer 2 (deep)**: If flagged, heavy model re-evaluates with "are you sure?" nudge.
 5. **Post-execution revert**: After claude_code_edit, modifications to safety-critical files are automatically reverted.
+- Safety LLM calls now emit standard `llm_usage` events, so safety costs and failures appear in the same audit/health pipeline as other model calls.
 `identity.md` is intentionally mutable (self-creation) and can be rewritten radically;
 the constitutional guard is that the file itself must remain non-deletable.
 
@@ -407,6 +409,7 @@ the constitutional guard is that the file itself must remain non-deletable.
 - Daemon thread, sleeps between wakeups (interval controlled by LLM via `set_next_wakeup`)
 - Loads full agent context: BIBLE, identity, scratchpad, knowledge base, drive state,
   health invariants, recent chat/progress/tools/events (same context as main agent)
+- Owner messages are forwarded to background consciousness in full text (not first-100-char previews).
 - Calls LLM with lightweight introspection prompt
 - Has limited tool access (memory, messaging, scheduling, read-only)
 - **Progress emission**: emits 💬 progress messages to UI via event queue + persists to `progress.jsonl`
@@ -444,6 +447,7 @@ the constitutional guard is that the file itself must remain non-deletable.
 - Stored in `logs/task_reflections.jsonl`; last 20 entries loaded into dynamic context
 - Pattern register: recurring error classes tracked in `memory/knowledge/patterns.md`
   via LLM, loaded into semi-stable context as "Known error patterns"
+- Secondary reflection/pattern prompts use explicit truncation markers when compacted for prompt size; no silent clipping of these helper summaries.
 - Runs synchronously (not in daemon thread) to avoid data loss on shutdown
 
 ### Crash report injection (agent.py)
@@ -453,17 +457,22 @@ the constitutional guard is that the file itself must remain non-deletable.
 - File is NOT deleted — persists so `build_health_invariants()` surfaces
   CRITICAL: RECENT CRASH ROLLBACK on every task until the agent investigates
 
-### Subtask trace summaries
+### Subtask lifecycle and trace summaries
 
-- When a subtask completes, a compact trace summary is included in the result
-- Parent tasks see tool call counts, error counts, and agent notes
-- Trace is truncated to 4000 chars; large traces show first/last 15 calls
+- `schedule_task` now writes durable lifecycle states in `task_results/<id>.json`: `requested` → `scheduled` → `running` → terminal status (`completed`, `rejected_duplicate`, `failed`, etc.)
+- Duplicate rejects are persisted explicitly, so `wait_for_task()` can report honest status instead of pretending the task is still running.
+- Completed subtasks persist the full result text; parent tasks no longer see silently clipped child output.
+- When a subtask completes, a compact trace summary is included alongside the full result.
+- Parent tasks see tool call counts, error counts, and agent notes.
+- Trace compaction remains explicit: max 4000 chars with visible omission markers, plus first/last 15 tool calls for long traces.
 
 ### Context building (context.py)
 
 - As of v3.16.0, the Memory Registry digest (from `memory/registry.md`) is injected into every LLM context to enable source-of-truth awareness.
 - As of v3.20.0, `patterns.md` (Pattern Register) is injected into semi-stable context, and execution reflections from `task_reflections.jsonl` are injected into dynamic context.
 - As of v3.22.0, all docs are always in static context: BIBLE.md (180k), ARCHITECTURE.md (60k), DEVELOPMENT.md (30k), README.md (10k), CHECKLISTS.md (5k).
+- `build_health_invariants()` is split into focused helpers and now also surfaces recent provider/routing errors plus local context overflows.
+- Local-model path no longer silently slices the live system prompt. It compacts non-core sections explicitly and raises an overflow error if core context still cannot fit.
 
 ### Deep review (review.py)
 
@@ -473,6 +482,7 @@ the constitutional guard is that the file itself must remain non-deletable.
 - Fallback to chunked previews if codebase exceeds 600K token budget
 - Security: skips sensitive files (.env, .pem, credentials.json, etc.)
 - Per-file cap: 1MB
+- Multi-model review now uses the shared async `LLMClient` OpenRouter path instead of raw one-off HTTP calls, so provider routing, Anthropic parameter requirements, usage normalization, and cache metadata are aligned with the rest of the runtime.
 
 ---
 

ouroboros/agent.py

@@ -44,6 +44,7 @@
 from ouroboros.agent_task_pipeline import (
     build_trace_summary, emit_task_results, build_review_context,
 )
+from ouroboros.task_results import STATUS_RUNNING, write_task_result
 
 
 _worker_boot_logged = False
@@ -161,6 +162,18 @@ def _prepare_task_context(self, task: Dict[str, Any]) -> Tuple[ToolContext, List
         drive_logs = self.env.drive_path("logs")
         sanitized_task = sanitize_task_for_event(task, drive_logs)
         append_jsonl(drive_logs / "events.jsonl", {"ts": utc_now_iso(), "type": "task_received", "task": sanitized_task})
+        try:
+            write_task_result(
+                self.env.drive_root,
+                str(task.get("id") or ""),
+                STATUS_RUNNING,
+                parent_task_id=task.get("parent_task_id"),
+                description=task.get("description"),
+                context=task.get("context"),
+                result="Task is running.",
+            )
+        except Exception:
+            log.debug("Failed to persist running task status", exc_info=True)
         self._emit_live_log(
             "context_building_started",
             task_id=str(task.get("id") or ""),

ouroboros/agent_task_pipeline.py

@@ -16,11 +16,19 @@
 import time
 from typing import Any, Dict, List
 
+from ouroboros.task_results import STATUS_COMPLETED, write_task_result
 from ouroboros.utils import utc_now_iso, append_jsonl
 
 log = logging.getLogger(__name__)
 
 
+def _truncate_with_notice(text: Any, limit: int) -> str:
+    raw = str(text or "")
+    if len(raw) <= limit:
+        return raw
+    return raw[:limit] + f"\n...[truncated from {len(raw)} chars; omitted {len(raw) - limit}]"
+
+
 def build_trace_summary(llm_trace: dict) -> str:
     """Return a compact human-readable summary of tool calls and agent notes."""
     tool_calls = llm_trace.get("tool_calls", []) or []
@@ -44,6 +52,8 @@ def _fmt_call(idx: int, tc: dict) -> str:
                     if len(v_str) > 60:
                         v_str = v_str[:57] + "..."
                     parts.append(f"{k}={v_str!r}")
+                if len(args) > 2:
+                    parts.append(f"... (+{len(args) - 2} more args)")
                 args_str = ", ".join(parts)
             else:
                 args_str = repr(args)
@@ -148,23 +158,20 @@ def _store_task_result(env: Any, task: Dict[str, Any], text: str,
                        usage: Dict[str, Any], llm_trace: Dict[str, Any]) -> None:
     """Store task result for parent task retrieval."""
     try:
-        results_dir = pathlib.Path(env.drive_root) / "task_results"
-        results_dir.mkdir(parents=True, exist_ok=True)
         trace_summary = build_trace_summary(llm_trace)
-        result_data = {
-            "task_id": task.get("id"),
-            "parent_task_id": task.get("parent_task_id"),
-            "status": "completed",
-            "result": text[:3500] if text else "",
-            "trace_summary": trace_summary,
-            "cost_usd": round(float(usage.get("cost") or 0), 6),
-            "total_rounds": int(usage.get("rounds") or 0),
-            "ts": utc_now_iso(),
-        }
-        result_file = results_dir / f"{task.get('id')}.json"
-        tmp_file = results_dir / f"{task.get('id')}.json.tmp"
-        tmp_file.write_text(json.dumps(result_data, ensure_ascii=False, indent=2))
-        os.rename(tmp_file, result_file)
+        write_task_result(
+            env.drive_root,
+            str(task.get("id") or ""),
+            STATUS_COMPLETED,
+            parent_task_id=task.get("parent_task_id"),
+            description=task.get("description"),
+            context=task.get("context"),
+            result=text or "",
+            trace_summary=trace_summary,
+            cost_usd=round(float(usage.get("cost") or 0), 6),
+            total_rounds=int(usage.get("rounds") or 0),
+            ts=utc_now_iso(),
+        )
     except Exception as e:
         log.warning("Failed to store task result: %s", e)
 
@@ -194,14 +201,14 @@ def _run_task_summary(env, llm, task, usage, llm_trace, drive_logs):
             CONSOLIDATION_REASONING_EFFORT,
         )
         task_id = task.get("id", "unknown")
-        goal = str(task.get("text", ""))[:500]
+        goal = _truncate_with_notice(task.get("text", ""), 500)
         rounds = int(usage.get("rounds") or 0)
         cost = float(usage.get("cost") or 0)
         trace = build_trace_summary(llm_trace)
         prompt = _TASK_SUMMARY_PROMPT.format(
             task_id=task_id, goal=goal or "(no goal text)",
             task_type=task.get("type", "user"), rounds=rounds,
-            cost=cost, trace_summary=trace[:3000],
+            cost=cost, trace_summary=_truncate_with_notice(trace, 3000),
         )
         try:
             msg, _usage = llm.chat(messages=[{"role": "user", "content": prompt}],
@@ -217,7 +224,10 @@ def _run_task_summary(env, llm, task, usage, llm_trace, drive_logs):
                     pass
         except Exception:
             log.warning("Task summary LLM call failed, using fallback", exc_info=True)
-            summary_text = f"Task {task_id} ({task.get('type', 'user')}): {goal[:200]}. {rounds}r, ${cost:.2f}."
+            summary_text = (
+                f"Task {task_id} ({task.get('type', 'user')}): "
+                f"{_truncate_with_notice(goal, 200)}. {rounds}r, ${cost:.2f}."
+            )
         if summary_text:
             append_jsonl(drive_logs / "chat.jsonl", {
                 "ts": utc_now_iso(), "direction": "system",
@@ -351,6 +361,16 @@ def build_review_context(env: Any) -> str:
                 "\nUse repo_read to inspect specific files. "
                 "Use run_shell for tests. Key files below:\n",
             ]
+            if stats.get("truncated"):
+                parts.append(f"\nCompacted files: {stats['truncated']}\n")
+            if stats.get("dropped"):
+                dropped_paths = stats.get("dropped_paths") or []
+                preview = ", ".join(dropped_paths[:5])
+                parts.append(
+                    f"\nDropped files due review budget: {stats['dropped']}"
+                    + (f" ({preview}{' ...' if len(dropped_paths) > 5 else ''})" if preview else "")
+                    + "\n"
+                )
             chunks = chunk_sections(sections)
             parts.append(chunks[0] if chunks else "(No reviewable content found.)")
             return "\n".join(parts)
@@ -380,6 +400,16 @@ def build_review_context(env: Any) -> str:
                 "\nUse repo_read to inspect specific files. "
                 "Use run_shell for tests. Key files below:\n",
             ]
+            if stats.get("truncated"):
+                parts.append(f"\nCompacted files: {stats['truncated']}\n")
+            if stats.get("dropped"):
+                dropped_paths = stats.get("dropped_paths") or []
+                preview = ", ".join(dropped_paths[:5])
+                parts.append(
+                    f"\nDropped files due review budget: {stats['dropped']}"
+                    + (f" ({preview}{' ...' if len(dropped_paths) > 5 else ''})" if preview else "")
+                    + "\n"
+                )
             chunks = chunk_sections(sections)
             parts.append(chunks[0] if chunks else "(No reviewable content found.)")
             return "\n".join(parts)

ouroboros/consciousness.py

@@ -258,10 +258,12 @@ def _think(self) -> None:
 
                 # Report usage to supervisor
                 if self._event_queue is not None:
+                    provider = "local" if _use_local_light else "openrouter"
+                    model_name = f"{model} (local)" if _use_local_light else model
                     self._event_queue.put({
                         "type": "llm_usage",
-                        "provider": "openrouter",
-                        "model": model,
+                        "provider": provider,
+                        "model": model_name,
                         "usage": usage,
                         "cost": cost,
                         "source": "consciousness",