v4.5.0: context quality, provenance repair, prompt discipline

Fix provenance: system summaries (task_summary, direction="system") are now
correctly classified across memory.py, consolidator.py, server API, and chat
UI — no longer masquerading as user messages. New amber system bubbles in UI.

Restore execution reflections (task_reflections.jsonl) in live LLM context.
Move Health Invariants to the top of dynamic context block in both the main
task path and background consciousness.

Task-scope recent progress/tools/events when task_id is available, keeping
dialogue continuity broad.

Harden run_shell: detect literal $VAR env-ref misuse in argv-form commands.
Harden claude_code_edit: retry transient first-run Claude CLI auth failure,
structured error classification via _format_claude_code_error.

Full SYSTEM.md editorial rewrite: terminology normalized to 'creator', new
sections (Methodology Check, Anti-Reactivity, Diagnostics Discipline,
Knowledge Retrieval Triggers), stronger Health Invariant reactions, compressed
inventory sections, balanced KB preflight gate.

16 files changed. New test_chat_provenance.py, updated regression tests.

Made-with: Cursor

Author: Anton

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 4.4.0](https://img.shields.io/badge/version-4.4.0-green.svg)](VERSION)
+[![Version 4.5.0](https://img.shields.io/badge/version-4.5.0-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 |
 |---------|------|-------------|
+| 4.5.0 | 2026-03-19 | Context quality and prompt discipline release: fix provenance — system summaries now correctly marked as system, not user, across memory, consolidation, server API, and chat UI (amber system bubbles); restore execution reflections (task_reflections.jsonl) in live LLM context; move Health Invariants to the top of dynamic context block (both task and consciousness paths); task-scope recent progress/tools/events when task_id is available; harden run_shell against literal $VAR env-ref misuse in argv; add Claude CLI first-run retry and structured error classification; full SYSTEM.md editorial rewrite — terminology normalized to 'creator', new Methodology Check / Anti-Reactivity / Diagnostics Discipline / Knowledge Retrieval Triggers sections, stronger Health Invariant reactions, compressed inventory sections. 12 files changed, new regression tests. |
 | 4.4.0 | 2026-03-19 | Safe editing release: `str_replace_editor` tool for surgical edits to existing files, `repo_write` shrink guard blocks accidental truncation of tracked files (>30% shrinkage), full task lifecycle statuses (failed/interrupted/cancelled) with honest status tracking, rescue snapshot discoverability via health invariants, `provider_incomplete_response` classification for OpenRouter glitches, default review enforcement changed to advisory, fix progress bubble opacity and duplicate emoji. |
 | 4.3.1 | 2026-03-19 | Fix: remove semi-transparent dimming from progress chat bubbles and remove duplicate `💬` emoji that appeared in both sender label and message text. |
 | 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. |

VERSION

@@ -1 +1 @@
-4.4.0
+4.5.0

docs/ARCHITECTURE.md

@@ -1,4 +1,4 @@
-# Ouroboros v4.4.0 — Architecture & Reference
+# Ouroboros v4.5.0 — 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.
@@ -179,14 +179,15 @@ Navigation is a left sidebar with 8 pages.
 - **Status badge** (top-right): "Online" (green) / "Thinking..." (amber pulse) / "Reconnecting..." (red).
   Driven by WebSocket connection state and typing events.
 - **Message input**: textarea + send button. Shift+Enter for newline, Enter to send.
-- **Messages**: user bubbles (right, blue-tinted) and assistant bubbles (left, crimson). Assistant messages render markdown.
+- **Messages**: user bubbles (right, blue-tinted), assistant bubbles (left, crimson), and system-summary bubbles (left, amber). Non-user bubbles render markdown.
 - **Timestamps**: smart relative formatting (today: "HH:MM", yesterday: "Yesterday, HH:MM", older: "Mon DD, HH:MM"). Shown on hover.
 - **Progress messages**: background consciousness thinking shown as dimmed bubbles with 💬 prefix.
+- **System summaries**: `direction="system"` entries from `chat.jsonl` are shown in the same timeline with a 📋 label instead of being hidden or treated as user text.
 - **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"}`.
+- Responses arrive via WebSocket `{type: "chat", role: "assistant", content: text, ts: "ISO"}`. On page-load history sync, `/api/chat/history` can also return `role: "system"` entries for internal summaries.
 - Supports slash commands: `/status`, `/evolve`, `/review`, `/bg`, `/restart`, `/panic`.
 
 ### 3.2 Dashboard
@@ -294,7 +295,7 @@ Navigation is a left sidebar with 8 pages.
 | POST | `/api/local-model/stop` | Stop local model server |
 | GET | `/api/local-model/status` | Local model status and readiness |
 | GET | `/api/evolution-data` | Evolution metrics per git tag (LOC, prompt sizes, memory) |
-| GET | `/api/chat/history` | Merged chat + progress messages (chronological, limit param) |
+| GET | `/api/chat/history` | Merged chat + system summaries + progress messages (chronological, limit param) |
 | POST | `/api/local-model/test` | Local model sanity test (chat + tool calling) |
 | WS | `/ws` | WebSocket: chat messages, commands, log streaming |
 | GET | `/static/*` | Static files from `web/` directory (NoCacheStaticFiles wrapper forces revalidation) |
@@ -471,6 +472,8 @@ the constitutional guard is that the file itself must remain non-deletable.
 - 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).
+- `Health Invariants` are placed at the start of the dynamic context block, before drive state/runtime/recent sections, so warnings influence planning before the model reads the noisier tail sections.
+- `build_recent_sections()` keeps recent dialogue broad, but task-scopes recent progress/tools/events when `task_id` is available.
 - `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.
 

ouroboros/consciousness.py

@@ -410,18 +410,18 @@ def _build_context(self) -> str:
             if patterns_text.strip():
                 parts.append("## Pattern Register\n\n" + clip_text(patterns_text, 30000))
 
+        # Health invariants
+        health_section = build_health_invariants(env)
+        if health_section:
+            parts.append(health_section)
+
         # Drive state
         state_json = safe_read(env.drive_path("state/state.json"), fallback="{}")
         parts.append("## Drive state\n\n" + clip_text(state_json, 90000))
 
         # Runtime section (same as main agent)
         parts.append(build_runtime_section(env, bg_task))
 
-        # Health invariants
-        health_section = build_health_invariants(env)
-        if health_section:
-            parts.append(health_section)
-
         # Recent sections — empty task_id so we get ALL tasks' progress/tools/events
         parts.extend(build_recent_sections(memory, env, task_id=""))
 

ouroboros/consolidator.py

@@ -432,6 +432,9 @@ def _format_entries_for_block(entries: List[Dict[str, Any]]) -> str:
         if dir_raw in ("out", "outgoing"):
             direction_prefix = "-> "
             author = "Ouroboros"
+        elif dir_raw == "system":
+            direction_prefix = "[system] "
+            author = "Ouroboros"
         else:
             direction_prefix = ""
             author = e.get("username") or e.get("author") or "User"

ouroboros/context.py

@@ -222,30 +222,75 @@ def build_memory_sections(memory: Memory) -> List[str]:
     return sections
 
 
-def build_recent_sections(memory: Memory, env: Any, task_id: str = "") -> List[str]:
-    """Build recent chat, recent progress, recent tools, recent events sections.
+def _format_recent_reflections(entries: List[Dict[str, Any]], limit: int = 10) -> str:
+    """Format recent execution reflections for dynamic context."""
+    if not entries:
+        return ""
 
-    Legacy note: older process-memory used task_reflections.jsonl and an
-    "Execution reflections" section; task summaries in chat.jsonl are now the
-    primary continuity layer.
-    """
+    blocks: List[str] = []
+    for entry in entries[-limit:]:
+        ts_full = str(entry.get("ts", ""))
+        ts = ts_full[:16] if len(ts_full) >= 16 else ts_full
+        header_bits = [bit for bit in [
+            ts,
+            str(entry.get("task_type", "")).strip(),
+            str(entry.get("task_id", "")).strip(),
+        ] if bit]
+        header = " | ".join(header_bits) or "unknown reflection"
+
+        lines = [f"### {header}"]
+
+        goal = str(entry.get("goal", "")).strip()
+        if goal:
+            lines.append(f"- Goal: {goal}")
+
+        markers = [str(m).strip() for m in (entry.get("key_markers") or []) if str(m).strip()]
+        if markers:
+            lines.append(f"- Markers: {', '.join(markers)}")
+
+        rounds = entry.get("rounds")
+        if rounds not in (None, ""):
+            lines.append(f"- Rounds: {rounds}")
+
+        cost_usd = entry.get("cost_usd")
+        if cost_usd not in (None, ""):
+            lines.append(f"- Cost: ${cost_usd}")
+
+        reflection = str(entry.get("reflection", "")).strip()
+        if reflection:
+            lines.append("")
+            lines.append(reflection)
+
+        blocks.append("\n".join(lines).strip())
+
+    return "\n\n".join(blocks)
+
+
+def build_recent_sections(memory: Memory, env: Any, task_id: str = "") -> List[str]:
+    """Build recent dialogue and process-memory sections."""
     sections = []
 
     chat_summary = memory.summarize_chat(memory.read_jsonl_tail("chat.jsonl", 1000))
     if chat_summary:
         sections.append("## Recent chat\n\n" + chat_summary)
 
     progress_entries = memory.read_jsonl_tail("progress.jsonl", 200)
+    if task_id:
+        progress_entries = [e for e in progress_entries if str(e.get("task_id", "")).strip() == task_id]
     progress_summary = memory.summarize_progress(progress_entries, limit=50)
     if progress_summary:
         sections.append("## Recent progress\n\n" + progress_summary)
 
     tools_entries = memory.read_jsonl_tail("tools.jsonl", 200)
+    if task_id:
+        tools_entries = [e for e in tools_entries if str(e.get("task_id", "")).strip() == task_id]
     tools_summary = memory.summarize_tools(tools_entries)
     if tools_summary:
         sections.append("## Recent tools\n\n" + tools_summary)
 
     events_entries = memory.read_jsonl_tail("events.jsonl", 200)
+    if task_id:
+        events_entries = [e for e in events_entries if str(e.get("task_id", "")).strip() == task_id]
     events_summary = memory.summarize_events(events_entries)
     if events_summary:
         sections.append("## Recent events\n\n" + events_summary)
@@ -254,6 +299,11 @@ def build_recent_sections(memory: Memory, env: Any, task_id: str = "") -> List[s
     if supervisor_summary:
         sections.append("## Supervisor\n\n" + supervisor_summary)
 
+    reflections_entries = memory.read_jsonl_tail("task_reflections.jsonl", 20)
+    reflections_text = _format_recent_reflections(reflections_entries, limit=10)
+    if reflections_text:
+        sections.append("## Execution reflections\n\n" + reflections_text)
+
     return sections
 
 
@@ -732,14 +782,14 @@ def build_llm_messages(
 
     semi_stable_text = "\n\n".join(semi_stable_parts)
 
-    dynamic_parts = [
-        "## Drive state\n\n" + state_json,
-        build_runtime_section(env, task),
-    ]
-
     health_section = build_health_invariants(env)
+    dynamic_parts = []
     if health_section:
         dynamic_parts.append(health_section)
+    dynamic_parts.extend([
+        "## Drive state\n\n" + state_json,
+        build_runtime_section(env, task),
+    ])
 
     dynamic_parts.extend(build_recent_sections(memory, env, task_id=task.get("id", "")))
 

ouroboros/memory.py

@@ -301,6 +301,9 @@ def chat_history(self, count: int = 100, offset: int = 0, search: str = "") -> s
                 raw_text = str(e.get("text", ""))
                 if dir_raw in ("out", "outgoing"):
                     lines.append(f"→ [{ts}] {raw_text}")
+                elif dir_raw == "system":
+                    entry_type = str(e.get("type", "")).strip() or "system"
+                    lines.append(f"📋 [{ts}] [{entry_type}] {raw_text}")
                 else:
                     username = e.get("username") or e.get("author") or "User"
                     lines.append(f"← [{ts}] [{username}] {raw_text}")
@@ -347,6 +350,9 @@ def summarize_chat(self, entries: List[Dict[str, Any]]) -> str:
             raw_text = str(e.get("text", ""))
             if dir_raw in ("out", "outgoing"):
                 lines.append(f"→ {ts_hhmm} {raw_text}")
+            elif dir_raw == "system":
+                entry_type = str(e.get("type", "")).strip() or "system"
+                lines.append(f"📋 {ts_hhmm} [{entry_type}] {raw_text}")
             else:
                 username = e.get("username") or e.get("author") or "User"
                 lines.append(f"← {ts_hhmm} [{username}] {raw_text}")