Skip to main content
Version: 0.6.3

Generated Files Reference

Everything ClaudusBridge writes lives under your project's Saved/ClaudusBridge/ folder. This guide is a complete catalogue: what each file is, when it gets written, what it contains, and how to read or use it.

Files fall into two categories:

  1. Regenerated artifacts — Rewritten on every MCP initialize handshake (idempotent: only rewritten if content actually changed). Safe to delete; the plugin recreates them on the next reconnect.
  2. Accumulating state — Written incrementally as the session progresses. Deleting them loses history.

<ProjectDir> below means the absolute path to the Unreal project hosting ClaudusBridge.

Regenerated artifacts (top-level)​

FileSize (typical)Purpose
CLAUDE.md~150 KBFull plugin architecture + tool reference for any connected AI client to read on connect. Lists every registered tool by group, every Cognition-tier surface, the action contract, EYES MANDATORY rule, memory + checkpoint patterns. The "long" version.
CLAUDUS_AI_SKILL.md~35 KBThe provider-facing operating manual. Specifically tuned for the AI model embedded in the Claudus dashboard (Claude / ChatGPT / Gemini). Teaches identity hygiene, the action handoff protocol, the Neural Cortex workflow, and the verification discipline. A compact version is also injected into the provider's system prompt.
UE5_EDITOR_CONTROLS.md~17 KBUE5 keyboard / mouse / shortcut reference for agents driving the editor through the WebRTC /preview stream. Empirically verified Pixel Streaming gotchas (modifier keys do not survive the data channel; recommended interaction patterns).
cb.py~50 KBThe ergonomic Python helper. Wraps the JSON-RPC envelope + Python unwrap pattern into single-line subcommands (cb spawn, cb tx, cb capture, cb verify, cb remember, …). Pure stdlib, works with any Python 3.8+. See The cb Helper CLI.
watcher_subagent.prompt.md~22 KBSystem prompt for the canonical visual watcher sub-agent. Eyes-on-demand model: idles cheaply, captures only on consequential events or explicit verification requests.
claudus_agent.prompt.md~11 KBPortable execution-agent prompt for Codex, Claude Code, Cursor, Windsurf, and similar runtimes. Used when an external client connects as an execution peer.
claudus_documentation_agent.prompt.md~6 KBDocumentation-specialist prompt. Spawn this sub-agent when you want a specialist that teaches other agents how to use ClaudusBridge from the generated docs.
claudusbridge_watcher.py~16 KBProduction-grade Python watcher daemon. Pure stdlib, JSON logs, exponential-backoff reconnect, signal-handled graceful shutdown, three built-in pattern rules (compile loops, PIE crash-on-start, actor leaks).
claudusbridge_offline.py~15 KBStandalone MCP stdio server for offline tools (build, clean, regenerate project files, launch / kill editor) when the editor itself is not running.
install_claudus_codex_agent.ps1~1.5 KBSingle-command Codex CLI install script that copies the CodexAgentRuntime/ package into Codex's plugin directory.
claudus_bridge.spawn.json~1 KBCodex fallback transport spawn-config for the current build.
rebuild-and-relaunch.ps1~2 KBPowerShell script that runs Build.bat -NoHotReloadFromIDE, closes the running editor, and relaunches the active .uproject. Used by claudus_rebuild_plugin.

Regenerated artifacts (subdirectories)​

CompanionAgents/​

One subfolder per supported external AI host, plus a root README.md index. Each folder contains everything the host needs to install ClaudusBridge as a Claudus-aware peer execution agent.

CompanionAgents/
├── README.md                   # Index of all hosts
├── claude-code/
│   ├── INSTALL.md              # Per-host install instructions
│   ├── PROTOCOL.md             # Shared action-request protocol
│   ├── bootstrap-claude-code.ps1
│   ├── claudus-executor.md     # Agent profile in Claude Code's native format
│   └── mcp-config.json         # MCP server entry to merge into Claude Code config
├── cursor/
│   ├── INSTALL.md
│   ├── PROTOCOL.md
│   ├── bootstrap-cursor.ps1
│   ├── claudus-rules.md        # Cursor agent profile
│   └── mcp-config.json
├── gemini-cli/
│   ├── INSTALL.md
│   ├── PROTOCOL.md
│   ├── bootstrap-gemini-cli.ps1
│   ├── claudus-profile.md      # Gemini CLI agent profile
│   └── mcp-config.json
└── windsurf/
    ├── INSTALL.md
    ├── PROTOCOL.md
    ├── bootstrap-windsurf.ps1
    ├── claudus-memory.md       # Windsurf agent profile
    └── mcp-config.json

See Installing Companion AI Agents for the install flow.

CodexAgentRuntime/​

Codex-specific runtime package. Codex CLI installs the local plugin package; other clients use claudus_agent.prompt.md and the documentation pack instead. Contents include agents/claudus-bridge-executor.md, agents/claudus-provider-peer.md, agents/claudus-bridge.md, agents/claudus-bridge-documentation.md, plus protocol references and the spawn-json fallback.

ClaudusAgents/​

16 native agent manifests representing every role the system understands. Each agent is a <id>.md file plus a manifest.json index (schema claudus.native_agents.v1):

provider_peer                native_tool_executor
planning_orchestrator        multi_agent_orchestrator
cpp_plugin_engineer          blueprint_architect
material_vfx_artist          asset_librarian
build_release_engineer       scope_safety_guard
gallery_artist               documentation_specialist
visual_watcher               memory_context
reviewer                     codex_bridge_executor

Inspect with claudus_list_agents / claudus_get_agent. The runtime profile catalog returned by claudus_list_agent_profiles is a related but distinct surface (runtime-only, doesn't touch disk).

ClaudusSkills/​

19 native skill manifests under skills/<id>/SKILL.md plus a manifest.json (schema claudus.native_skills.v1):

claudus_ai_mastery              claudus_native_agents
skill_creator                   self_programming
native_tool_authoring           task_planning
multi_agent_workflow            workspace_programming
blueprint_material_workbench    asset_workbench
visual_qa                       agentic_security
native_asset_visualization      native_unreal_editing
gallery_image_generation        memory_context
documentation_intelligence      rebuild_validation
project_texture_visual_review

Inspect with claudus_list_installed_skills / claudus_get_installed_skill. The runtime catalogue returned by claudus_list_skills includes additional runtime-only entries.

Accumulating state​

These files grow as you work and are not reset on reconnect. Manage them like any other project state.

claudus_memory.jsonl — the auto-observation ring buffer​

Append-only JSONL, capped at 256 entries (FIFO eviction). Every successful asset creation, every workflow completion, every checkpoint, and every tool failure writes one line here automatically. Read it via claudus_recall_observations(tag, since_hours, contains, limit) instead of grep'ping the file directly.

Sample entry shape:

{
  "id":           "2AED43314ECD5D084BA456848EE33CD7",
  "text":         "Created basic_shape asset `StaticMeshActor_UAID_2CF…` via ai_create.",
  "source":       "asset-created",
  "source_model": "claude-opus-4-7[1m]",
  "created_at":   17491635.348,
  "tags":         ["auto_observation", "asset_created", "kind:basic_shape", "path:/Game/…"]
}

The four source strings are asset-created, workflow-completed, workflow-checkpoint, action-failed. See Auto-Observations for the full tag schema.

memory/ — explicit, agent-curated knowledge​

Per-topic markdown files plus a JSONL log. Where the auto-observation buffer is the flight recorder, this folder is the engineer's notebook — written explicitly by agents (or by you) via memory_remember / cb remember.

memory/
├── _log.jsonl              # Raw append-only log of every memory entry
├── _index.md               # Auto-curated table of contents
├── framing.md              # Camera poses, viewport setups (per-topic)
├── spawning.md             # Actor placement patterns
├── transforms.md           # Location/rotation/scale tricks
├── navigation.md
├── materials.md
├── blueprints.md
├── umg.md
├── niagara.md
├── gotchas.md
├── shortcuts.md
├── process.md
├── ue5_controls.md
└── …                       # Any new topic auto-creates on first remember

Topic files are plain markdown, most-recent-entry-first, editable by hand. See Memory System.

feedback.jsonl — plugin-improvement signal​

Append-only log of every record_feedback call, plus everything the internal watcher catches (compile loops, PIE crash-on-start, actor leaks). The canonical signal for plugin self-improvement. Each line is one JSON object with kind (visual_mismatch, tool_friction, suggest_tool, plugin_bug), description, from_session, optional context.

WorkflowCheckpoints/​

One JSON per checkpoint: ck-<workflow_id>-<timestamp>.json. Contains the snapshot of asset paths + class names + last-modified timestamps captured by claudus_checkpoint_workflow. See Workflow Checkpoints for the schema and the diff-with-jq pattern.

MultiAgentWorkflows/​

One JSON per workflow: maw-<slug>-<timestamp>.json. Schema claudus.multi_agent_workflow.v1. Carries participants, steps, action ids, status, results, notes, evidence. Inspect with claudus_get_multi_agent_workflow and update with claudus_update_multi_agent_workflow. See Multi-Agent Workflows.

TaskPlans/​

One JSON per single-agent multi-step plan: <plan_id>.json. Lighter than multi-agent workflows; appropriate when one agent works alone but wants a durable to-do list with evidence. Manipulated with claudus_create_task_plan / claudus_update_task_plan.

ActionResults/​

Evidence JSON for every action that ran through the Claudus Native Tool Executor. Filename pattern <action_id_prefix>_<tool_name>.json. Used by:

  • The transcript (to attach result_path instead of inlining large payloads)
  • The auto-retry queue (so re-firing a failed action has the original args)
  • claudus_read_action_result(action_id | filename | list) for self-audit

The folder can be cleared whenever you want; the running session keeps its own in-memory state and only re-reads the files when explicitly asked.

Vision/​

PNG evidence images saved by claudus_show_texture_asset, claudus_describe_texture_asset, preview_texture_asset, and the dashboard inference capture path. Filename includes the asset name + timestamp for traceability. Safe to delete; the next vision call recreates as needed.

GallerySourceCache/​

Original encoded bytes of every image generated by claudus_generate_gallery_image. One folder per topic; one PNG per generation. Used as provenance — even after the corresponding Texture2D asset has been re-saved or edited, the cached source bytes show what the model actually produced.

TexturePreviews/ and GalleryImportCache/​

Auxiliary scratch for the texture/gallery pipeline. Safe to delete.

SelfProgrammingBackups/​

Backups taken before every claudus_write_plugin_file / claudus_delete_plugin_file. One backup per edit, named with timestamp and original path. Read-only; the plugin reads from here only when a rollback is explicitly requested.

pending-retries.json​

Action records persisted by claudus_request_native_tool_authoring with params_to_retry or schedule_retry=true. After the subsequent claudus_rebuild_plugin + relaunch, the new editor process re-fires the original call automatically and marks the record as fired_success / fired_error. See claudus_list_pending_retries / claudus_clear_pending_retry.

project-intelligence.json​

Cached project baseline written by claudus_analyze_project. Schema claudus.project_intelligence.v1. Contains project_name, engine_dir, current_map, total_asset_estimate, asset_summary[], naming_patterns[], risks[], recommendations[]. Auto-refreshed when older than 1 h (configurable via claudus_get_project_intelligence(refresh, max_age_seconds)).

rebuild-log.txt​

Plain-text log of the last claudus_rebuild_plugin invocation. Used for post-mortem when a rebuild surprised you.

What gets logged to Unreal's own logs​

Independent of the files above, every chat entry routes through six native UE log categories:

CategoryCarries
LogClaudusChatGPTReplies from the active ChatGPT model
LogClaudusClaudeReplies from the active Claude model
LogClaudusGeminiReplies from the active Gemini model
LogClaudusUserYour own input messages
LogClaudusAgentMessages from connected external execution agents
LogClaudusSystemAction queue events, tool results, audit notes

So <ProjectDir>/Saved/Logs/<Project>.log and Saved/Logs/<Project>-backup-*.log contain the conversation history at the UE-logging level too, in addition to the dashboard-side state described above.

Where to go next​