Tlamatini – Local-first AI dev assistant with 68 agents and hybrid RAG

"one who knows" — a locally-deployed AI developer assistant

Tlamatini (Nahuatl for "one who knows") is a locally-deployed AI developer assistant that pairs a hybrid RAG pipeline (FAISS + BM25, metadata extraction, context budgeting) with a Multi-Turn tool-orchestration layer, ACPX delegation to external coding-agent CLIs (Claude Code, Cursor, Codex, Gemini, Qwen, …), and a visual workflow designer with 68 drag-and-drop agents.

Local-first by default: the full RAG pipeline, the Multi-Turn execution loop, and every workflow agent run on your machine — embeddings and chat are driven by your local[Ollama]install. Cloud LLMs (Claude API, Ollama Pro/Max) and ACPX delegation to cloud CLIs are opt-in per-request, never the default. Sensitive code never leaves the box unless you explicitly route it out.

Latest — v1.9.0 (2026-05-26): STM32er, zero-config firmware bridge.A newSTM32eragent (canvas node + Multi-Turn toolchat_agent_stm32er

) brings the catalog to68 agents. It bridges the[STM32 Template Project MCP]to scaffold, build, flash, observe (serial / SWD), and reset STM32 firmware.Zero-config auto-bootstrapmeans the user only installs STM32CubeIDE + Tlamatini — STM32er downloads, installs, and validates the MCP server on first use. Acritical-mission safety preflightvalidates the toolchain and a positively-confirmed connected ST-LINK probe before flashing, andrefusesrather than producing or flashing mis-targeted firmware. Three new catalog demos ship in migration0103

(STM32 GENESIS / BLINKY / HIL OBSERVATORY). See[§3.15].

🌐 Website ·

·

▶️ One-minute teaser·

📖 Long-form docs·

🏷️ Versioning

🎬 More demos1. Overview2. Quickstart (source mode)3. Using the Chat (/agent/

)3.1. Chat layout in 30 seconds3.2. Setting code as context3.3. Tutorial: a one-shot question (no toggles)3.4. Tutorial: theinternet toggle3.5. Tutorial: theMulti-Turn toggle3.6. Tutorial: theExec Report toggle3.7. Tutorial: theACPX toggle3.8. From chat to flow: theCreate Flow button3.9. Why Chat-created flows are safer now3.10. TheDB menu — Backup, Set DB, and the start-up swap-in3.11. TheACPX-Skills menu — Browse, Configure, Diagnostics, Reload3.12. Tutorial: command a window from chat (chat_agent_windower

)3.13. Tutorial: drive a browser from chat (chat_agent_playwrighter

)3.14. Tutorial: run Kali Linux tools from chat (chat_agent_kalier

)3.15. Tutorial: build and flash STM32 firmware from chat (chat_agent_stm32er

)

4. Visual Workflow Designer (/agentic_control_panel/

)4.1. Canvas anatomy4.2. Tutorial: your first flow (3 agents)4.3. Saving and .flw

files4.4. Validate and Start now compile the live canvas4.5. , Resume, Stop4.6. FlowHypervisor (watchdog)4.7. FlowCreator (let an LLM design the flow)4.8. Parametrizer (chain outputs into the next agent's config)4.9. Gatewayer (external triggers)

5. ACPX — External Coding-Agent CLIs as Tools6. Unreal MCP — Driving Unreal Engine 5 from Tlamatini6.1. What Unreal MCP is6.2. The MCP plugin source (theMCP git location)6.3. Installing and enabling the plugin inside your UE5 project6.4. The command catalog (up to 53 commands across 9 categories)6.5. Using Unreal MCP from the chat (chat_agent_unrealer

)6.6. Using Unreal MCP on the canvas (the visualUnrealer node)6.7. What the agent actually does, end-to-end6.8. Exec Report integration6.9. Bullet-proof checklist for Unreal Engine users6.10. Troubleshooting Unreal MCP

7. Building a Frozen Distribution8. Configuration (Tlamatini/agent/config.json

)9. Architecture at a Glance10. Embedding-Memory Pre-Flight Guard (GPU hosts)11. Orphan-Process Cleanup (conhost.exe

reaper)12. Troubleshooting13. Versioning14. Contributing & License

Tlamatini (Nahuatl for "one who knows") is a Django/Channels app you run on your own machine. It packages a hybrid RAG pipeline, a Multi-Turn tool-calling LLM loop, an ACPX runtime that spawns external coding-agent CLIs as child processes, an Unreal MCP client that drives Unreal Engine 5 from chat or canvas, and a drag-and-drop workflow designer with 68 agent types — into one local install. Backends: Ollama (local), Anthropic Claude (cloud), Qwen vision (Ollama).

License: GPL-3.0 · Repo: https://github.com/XAIHT/Tlamatini.git · Platform tested: Windows 11 (cross-platform for source mode).

Real RAG over your code— FAISS + BM25 hybrid retrieval, code-aware metadata extraction, Reciprocal Rank Fusion, context budgeting, OOM fallback.Multi-Turn mode— the LLM becomes anoperator: shell, Python, APIs, SQL, file ops, screenshots, keyboard/mouse automation, email, Telegram, WhatsApp, STM32 firmware build/flash — chained in one conversation.ACPX— delegate sub-tasks to external CLIs (claude

,cursor-agent

,codex

,gemini

,qwen-code

, plus 8 more) and relay output between them.Visual workflow designer— design.flw

flows once, run them unattended, schedule with Croner, watch them with FlowHypervisor.Self-aware— a first-person self-knowledge map (Tlamatini.md

) is injected into the LLM's prompt on every chain, so Tlamatini can answer accurately about her own architecture, runtime modes, ports, and pages. Builds packaged with--self-modify

ship her own source tree (TlamatiniSourceCode/

) so she can read, inspect, and modify herself.

Everything runs locally. The whole app packages into a one-click Windows .exe

distribution (Part §7).

First system-usage walkthrough a complete project and summarizing its source codeInstalling OpenCV end-to-end in Multi-TurnUninstalling Poco — Exec Report and matching flowImplementing a FlowCreator-aided agentic flowA complete Cybersec enhancement with Tlamatini!!!

This is the fastest way to be productive: clone, install, run. No installer, no admin, no frozen build. Five minutes.

Requirement	Recommended	Notes
Python	3.12.10
The only version Tlamatini is fully tested on.
OS	Windows 11	Linux/macOS work for chat + designer; Mouser/Keyboarder are Windows-leaning.
RAM	16 GB+	32 GB comfortable for bigger embedding models.
Disk	~10 GB	Most is local LLM models.
LLM server	Ollama
Default. Cloud Claude/Gemini also supported.

You do not need administrator rights for any of the steps below.

Open PowerShell normally (do not Run as Administrator), then:

$env:OLLAMA_INSTALL_DIR = "$env:LOCALAPPDATA\Programs\Ollama"
irm https://ollama.com/install.ps1 | iex

Close the window, open a fresh PowerShell, and verify:

ollama --version
ollama serve     # leave running in its own window if it's not already up
Invoke-WebRequest http://127.0.0.1:11434/api/tags -UseBasicParsing

Tlamatini expects Ollama at http://127.0.0.1:11434

.

ollama pull Nomic-Embed-Text:latest
ollama pull glm-5:cloud
ollama pull qwen3.5:cloud
ollama pull gpt-oss:120b-cloud
ollama pull qwen3.5:397b-cloud
ollama pull llama3.2-vision:11b

Tag	Used for
Nomic-Embed-Text:latest
RAG embeddings (default — small VRAM footprint, ~600 MB resident)
glm-5:cloud
Default chat + Multi-Turn unified-agent + MCP file-search
qwen3.5:cloud
Default vision (Image-Interpreter)
gpt-oss:120b-cloud
Several workflow-agent templates (Monitor-Log, Notifier, Prompter, Summarizer, …)
qwen3.5:397b-cloud
Default FlowCreator
llama3.2-vision:11b
Local vision fallback

You can substitute any tag — just edit Tlamatini/agent/config.json

(see §8.1) or the relevant agent's config.yaml

.

Optional: swap to a higher-detail embedding model.If your retrieval quality on dense, technical corpora is not good enough with the default, you can switch toqwen3-embedding:8b

from theConfig → Modelsmenu inside the app (or by editingembeding-model

inconfig.json

and reconnecting).Use with caution:qwen3-embedding:8b

is roughly10× heavier in VRAMthanNomic-Embed-Text:latest

(~6.24 GB resident vs ~600 MB on a Q4_K_M quant) and will trip the embedding-memory pre-flight guard (see[§10]) on 8 GB consumer GPUs. Pull it first withollama pull qwen3-embedding:8b

.

Four of the six default model tags in §2.3 carry the :cloud

suffix — glm-5:cloud

, qwen3.5:cloud

, gpt-oss:120b-cloud

, and qwen3.5:397b-cloud

. Those are Ollama Cloud models: they live on Ollama's servers, not on your machine, and ollama pull

only registers a stub that proxies inference to the cloud. Reaching that cloud requires a logged-in Ollama account and a subscription tier that allows the workload you intend to run.

The plan structure (prices are deliberately omitted from this README because they change — check ** https://ollama.com/pricing** for the current numbers):

Plan	Cloud-model access	Why it matters for Tlamatini
Free
1 cloud model concurrently, light usage. Local open-weights models are unlimited.	Enough to try a single cloud model for a one-shot chat. Not enough for Tlamatini's default config, which pins different cloud models for chat (glm-5:cloud ), FlowCreator (qwen3.5:397b-cloud ), several workflow agents (gpt-oss:120b-cloud ), and vision (qwen3.5:cloud ) — so a real Multi-Turn run typically needs 2–3 cloud models loaded at once.
Pro
3 concurrent cloud models, ~50× the Free monthly quota, access to the larger cloud-only models, ability to upload / share private models.	The realistic minimum for running Tlamatini out-of-the-box with its shipped cloud-model defaults — Multi-Turn + Exec Report + occasional Image-Interpreter calls.
Max
10 concurrent cloud models, ~5× the Pro quota, designed for sustained heavy agentic workloads.	Recommended for long-running ACPX relays, FlowHypervisor-supervised flows, and Croner-driven unattended runs that chain many cloud calls per hour.

If you do not want to subscribe, you can run Tlamatini entirely on local open-weights models. Edit Tlamatini/agent/config.json

(chained-model

, unified_agent_model

, mcp_file_search_model

, flow_creator_model

, image_interpreter_model

) and every agent config.yaml

that names a :cloud

tag, and swap them for a model you have pulled locally (for example, llama3.1:8b

, qwen2.5-coder:14b

, mistral-nemo:12b

). Performance and quality will scale with your GPU/CPU — Multi-Turn and ACPX both work fine on a sufficiently large local model.

API keys are separate. This subscription only governs *:cloud

Ollama models. The ACPX runtime can additionally spawn external coding-agent CLIs that bring their own credentials (Anthropic API key for claude

, OpenAI key for codex

, Google key for gemini

, etc.) — those are configured in Tlamatini/agent/config.json

under acpx.agents.<id>.env

and are unaffected by your Ollama plan. See §5.6 for the easy-button setup. (Unreal MCP is not part of ACPX — it's its own MCP surface, documented in §6.)

git clone https://github.com/XAIHT/Tlamatini.git
cd Tlamatini

python -m venv venv
venv\Scripts\activate
source venv/bin/activate

pip install -r requirements.txt

python Tlamatini/manage.py migrate
python Tlamatini/manage.py createsuperuser
python Tlamatini/manage.py collectstatic --noinput
python Tlamatini/manage.py runserver --noreload

--noreload

is important: Daphne's auto-re does not coexist well with the wrapped-runtime subprocess pool.

The console title becomes Tlamatini

and stdout/stderr are tee'd into Tlamatini/tlamatini.log

(truncated on every start). When debugging, tlamatini.log

is the first thing to read.

Open http://127.0.0.1:8000/

and log in with the superuser you just created. Then:

/agent/

— the chat (Part§3)/agentic_control_panel/

— the visual designer (Part§4)/admin/

— Django admin (change passwords, manage users)

If you used the

installer(Part[§7]) instead of cloning, the default credentials areuser

/changeme

. Change them at first login via/admin/

.

┌───────────────────────────────────────────────────────────────────────────────┐
│ Tlamatini  [Context ▼] [Open in… ▼] [MCPs ▼] [Tools ▼] [Agents ▼] [Config ▼] [DB ▼] │ ← top nav
├───────────────────────────────────────────────────────────────────────────────┤
│  Multi-Turn ☐   Exec Report ☐   ACPX ☐   internet ☐    Clear ⌫               │ ← four toggles
├───────────────────────────────────────────────────────────────────────────────┤
│  ┌──── chat ────────────────┐   ┌──── code canvas ────────────────┐          │
│  │  conversation history    │   │  syntax-highlighted, with copy  │          │
│  └──────────────────────────┘   └─────────────────────────────────┘          │
├───────────────────────────────────────────────────────────────────────────────┤
│  Type your prompt here…                                              [Send]   │
└───────────────────────────────────────────────────────────────────────────────┘

The four toolbar toggles are independent. Tick whatever combination fits the task — each one is its own tutorial section below.

Newer builds also expose a Config dropdown in the same navbar. Config -> Models

edits the most common model-name fields, and Config -> URLs

edits the Ollama / unified-agent / MCP endpoint values through validated dialogs instead of hand-editing JSON. The chat/canvas divider was also tightened so resizing the right-hand canvas feels more predictable during long editing sessions.

The newest entry in that navbar is the DB dropdown: DB -> Backup database

snapshots the live SQLite file to a directory you pick, and DB -> Set DB

stages a db.sqlite3

file of your choice for the next session — Tlamatini swaps it in before Django opens the database, archives the previous one under DB/Older/<timestamp>/

, then continues normal start-up. Full walkthrough in §3.10.

Click Context in the top nav:

Menu entry	What it does
Set directory as context
Loads a folder. Tlamatini reads every text file, splits, embeds, builds FAISS+BM25, grounds answers in your code.
Set file as context
Single-file scope.
Set canvas as context
Use the code currently shown in the canvas (handy for iterative editing).
Clear context
Drops the loaded context.

Set directory as context now loads a project at any depth under the app root. The old browser showDirectoryPicker()

only exposed the leaf folder name, so deeply-nested projects could not be reached; it was replaced by a backend native Win32 folder picker (views.pick_context_directory_view

, route pick_context_directory/

) that returns the real absolute path. path_guard.is_within_application_root()

then accepts the application root or any descendant of it, and agent_page_init.js

falls back to manual path entry on non-Windows hosts.

A green banner at the top shows the current context path. If embedding runs out of memory, Tlamatini packs the source files as a fallback context — retrieval quality drops, access to your code does not.

If you refresh the browser and Tlamatini restores a saved context automatically, the input now stays disabled until the contextual RAG chain has actually finished rebuilding. That closes the old "restored banner arrived before the context was really ready" race on the first load stage.

Leave every checkbox unticked. Type:

"Write a Python function that validates an email address with a regex. Just the function."

The bot answers in one shot. Code lands in the right-hand canvas with copy/save buttons. This is the legacy chat path — fast, no tools, no internet.

Tick internet when the question genuinely needs fresh web data:

"What is the latest stable version of FastAPI right now?"

Tlamatini classifies the prompt with a small LLM call ("does this need the web?"), then DuckDuckGo-searches, summarizes the top results, and inlines the summary into the LLM's context. Leave it unticked for everything else (the round-trip adds latency).

This is the big one. Multi-Turn turns Tlamatini from answerer into operator:

• The planner picks the relevant subset of Tlamatini's 75 Multi-Turn tools— 20 core Python tools (execute_command

,agent_starter

,googler

, the image-analysis pair, thechat_agent_run_*

lifecycle helpers, …), 43 wrapped chat-agent tools, and 12 ACPX/Skill tools — binding at mostmax_selected_tools

per request (default cap:20). - The unified-agent loop runs up to 4096 iterations(theunified_agent_max_iterations

default) — call tool, see result, decide next, chain. - Wrapped sub-agents run in headless background runtimes (no console pop-ups).

Try this: tick Multi-Turn, send

"Take a screenshot of my desktop and save it to

C:\Tlamatini-test\shot.png

."

Watch the chat. The LLM picks chat_agent_shoter

, calls it with the right args, reads the JSON result, and replies "Done — saved to C:\Tlamatini-test\shot.png." Open the file. The screenshot is there.

Symptom	Fix
LLM says "Tool X is not available"	The planner did not bind it. Check [Planner._select] console lines; add matching keywords to your prompt or raise max_selected_tools .
Same tool fired twice with identical args	Suppressed by the dedup guard — the second call returns "skipped — duplicate".
4096 iterations exhausted	You probably hit a polling loop. Use chat_agent_sleeper instead of busy-polling.

Multi-Turn stacks with Set-Context: the LLM reasons over your code and runs tools on the result.

Below the prose answer, Tlamatini appends per-agent execution tables — one HTML table per kind of state-changing agent that fired. Each row = one real tool call + ✓/✗.

Tick Multi-Turn + Exec Report and send:

"Create

C:\test\hello.txt

withHi from Tlamatini

, then read it back and tell me its size."

After the prose, you see:

─── List of File Creator Operations ───
 1  │ filepath='C:\test\hello.txt' content='Hi …'    │  ✓
─── List of Executer Operations ───
 1  │ type C:\test\hello.txt                         │  ✓

What gets a table: state-changing tools only (execute_command

, execute_file

, unzip_file

, decompile_java

, every chat_agent_*

that touches the system, all five acp_*

lifecycle tools — merged into one "List of ACPx Operations" — and invoke_skill

). Read-only tools (Crawler, Googler, Prompter, Summarizer, File-Interpreter/Extractor, Image-Interpreter, Shoter, Sleeper, monitor_, run_, window_present

) are intentionally absent. Tables persist into chat history — reload the page and they are still there.

ACPX lets the chat delegate to external coding-agent CLIs running on your box. Picture it:

You ─► Tlamatini chat ─► acp_doctor → acp_spawn(claude) → acp_send_and_wait
                                  │
                                  ▼ subprocess.Popen
                                claude CLI / gemini / cursor / codex / qwen / …

When ACPX is ticked, the planner sees the 12 ACPX/Skill tools. When unticked, those tools are filtered out — the chat behaves like legacy Multi-Turn. (Implemented in agent/acpx/__init__.py::filter_acpx_tools()

.)

Prereq: at least one external CLI on PATH

. The simplest:

npm install -g @anthropic-ai/claude-code
claude --version

Then drop your key in Tlamatini/agent/config.json

(or use the setup-new-acpx-key skill — much easier).

Tick Multi-Turn + ACPX + Exec Report and send:

"Use ACPX to spawn the claude CLI in

C:/Development/Tlamatini

, ask it to summarize CLAUDE.md in 5 bullet points, harvest the answer, and kill the session."

You see: acp_doctor

(always first) → acp_spawn(agent_id="claude", task=…)

→ acp_send_and_wait

→ acp_kill

. The 5 bullets land in the prose, and the Exec Report shows a "List of ACPx Operations" table with all four rows.

ACPX deep dive in Part §5.

When a Multi-Turn run succeeds and used at least one state-changing tool, Tlamatini renders a Create Flow button on the message header. Click → download a .flw

JSON file mirroring the exact tool sequence, laid out left-to-right, ready to load in the visual designer:

Starter ─► Crawler ─► File Creator ─► Ender

You can re-open it in /agentic_control_panel/

and run it as an unattended workflow. The LLM is no longer in the loop.

The button gates on four conditions: Multi-Turn was on, ≥1 mappable tool succeeded, an LLM-based classifier marked the answer SUCCESS (fails open on internal error), and the user is logged in.

Older Chat-created .flw

files were generated almost entirely in the browser. That worked for simple chains, but it meant the browser had to remember many backend facts:

• what each agent is called on disk;
• which config field means "my input";
• which config field means "my output";
• which agents are special, like Ender or Parametrizer;
• which values are safe to save into a portable .flw

file.

That is a lot of responsibility for a button.

Now the browser still builds the first draft, but the backend normalizes it through the Agent Contract Registry before the file is downloaded. In plain English: Tlamatini checks the flow against the same agent rules that ACP uses to run it.

What this means for you:

• Repeated tools stay repeated. If Multi-Turn ran Executer five times, the flow contains five Executer nodes, not one overwritten node.
• Agent names are normalized. Names like Kyber-KeyGen

,kyber_keygen

, andKyber Keygen

are resolved to the right template. - Secrets are protected where the contract knows about them. Remote chat super-agents such as TeleTlamatini and WhatsTlamatini have credential-like fields redacted on export.

• The .flw

file stays portable. It does not storeC:/Development/...

or the installed app path besideTlamatini.exe

.

If backend normalization is temporarily unavailable, the old browser generator remains as a fallback so the button does not become useless.

The whole of Tlamatini — chat history, agents, Tool/MCP toggles, sessions, your user — lives in a single SQLite file. The DB dropdown gives you a safe, GUI-first way to handle that file: a read-only Backup path, a destructive-but-deferred Set DB path, and a built-in audit trail under DB/Older/

.

Opens a dialog with one input — the target directory. The path is live-validated (350 ms debounce): the page asks GET /agent/check_backup_directory/?path=…

as you type and colors the status line green / amber / red:

State	Status	Meaning
🟢	Directory exists. db.sqlite3 will be saved here.
Ready to back up.
🟠	A filename was specified — please specify the directory only.
You typed a file path; the output is always named db.sqlite3 so it can be loaded back later.
🔴	Directory does not exist.
Missing on disk.

Click Backup → Tlamatini calls POST /agent/backup_db/

, resolves the live database path via settings.DATABASES['default']['NAME']

(so source / frozen behave identically), and shutil.copy2

s it to <your-dir>/db.sqlite3

. The live database stays open and unchanged.

The opposite direction: replace the database on the next start-up. Same dialog idiom, stricter validation. The input is the full path to a db.sqlite3 file; the page asks

GET /agent/check_set_db_file/?path=…

as you type:| State | Status | Meaning | |---|---|---| | 🟢 | File exists. It will be loaded on the next start-up. | Real db.sqlite3 with a valid SQLite header. | | 🟠 | File found, but its name is not "db.sqlite3". Tlamatini will still stage it as db.sqlite3. | Snapshot-style names (db_2026-05-14.sqlite3 ) work — the staging step renames. | | 🟠 | Specify the full path to a db.sqlite3 file, not a directory. | You typed a directory; Set DB needs a file. | | 🔴 | The selected file does not look like a SQLite database. | First 16 bytes don't match the SQLite format 3\x00 magic. | | 🔴 | File does not exist. | Missing on disk. |

Click Set → POST /agent/set_db/

copies your file into <base>/DB/ToLoad/db.sqlite3

(where <base>

is the executable directory in frozen mode, the inner Django project directory in source mode). The live database is not touched — SQLite is held open by Django, so the actual replacement must wait for a process restart.

Immediately after staging succeeds, the dialog is replaced by a yellow ⚠ warning panel with a single OK button:

The selected database will be loaded the next time Tlamatini starts. If you want it loaded immediately, you must restart Tlamatini completely so the swap-in can run BEFORE Django opens the live database.

If you click Cancel instead of Set, the staging dialog closes and nothing is written.

The actual replacement lives at the very top of Tlamatini/manage.py

, in _apply_pending_db_swap()

. It runs before any Django import so Django's SQLite connection pool is never holding a stale file descriptor at the moment of the swap:

manage.py main()
    │
    ▼
_apply_pending_db_swap()
    │
    ▼
[ DB/ToLoad/db.sqlite3 exists? ]
    │
    ├─ NO  ──► return (no-op, normal start-up continues)
    │
    └─ YES ──► [1] mkdir DB/Older/<YYYY-MM-DD_HHMMSS>/
               [2] shutil.move(live db.sqlite3 → Older/<timestamp>/db.sqlite3)
               [3] shutil.move(DB/ToLoad/db.sqlite3 → live db.sqlite3)
               [4] return
    │
    ▼
from django.core.management import execute_from_command_line   ← only NOW Django wakes up

Three guarantees:

A Reconnect from the navbar is NOT enough. The swap window is only openbeforethe Django process opens its SQLite pool. You mustfully restart Tlamatini(close the console / kill the exe, then launch again).** Atomic moves, no copies.**Both legs useshutil.move

(filesystem rename when possible, copy+delete across mounts). A second launch withDB/ToLoad/

empty is automatically a no-op — no "stuck flag" to clear.Mode-correct path resolution. Frozen mode reads<exe_dir>/DB/ToLoad/db.sqlite3

(where you can browse to it in Explorer); source mode reads<repo>/Tlamatini/DB/ToLoad/db.sqlite3

(next tomanage.py

). The livedb.sqlite3

path is computed the same way Django does —_MEIPASS/db.sqlite3

under PyInstaller,<manage.py dir>/db.sqlite3

in source — so the swap-in always writes to exactly the path Django will open.

If anything fails inside the swap-in (locked file on Windows, corrupt source, permission error), the function catches the exception, prints --- [DB SWAP] Skipped due to error: …

to tlamatini.log

, and lets Tlamatini start normally with the previous database. A bad ToLoad file must never lock you out of your own database.

Every successful swap-in leaves a complete record under <base>/DB/Older/<YYYY-MM-DD_HHMMSS>/db.sqlite3

. Because Set DB moves (not copies) the prior live database, this archive is the only built-in recovery path:

DB/
├─ ToLoad/                 ← empty most of the time; momentary home of next-session pick
│   └─ README.md
└─ Older/
    ├─ 2026-05-14_153022/db.sqlite3   ← was live before swap #1
    ├─ 2026-05-14_164410/db.sqlite3   ← was live before swap #2
    └─ README.md

To roll back, drop the archived db.sqlite3

back into ToLoad/

and restart — the swap-in will archive the current live database under a fresh timestamp and promote your roll-back pick. Tlamatini never auto-deletes anything from Older/

; prune by hand when the tree gets noisy, but remember each file is a full snapshot of chat history + agents + sessions + your user.

Both directories must exist on day one (the swap-in opens them with os.makedirs(exist_ok=True)

, but having them pre-seeded with docs prevents user confusion):

Source / dev mode:Tlamatini/Tlamatini/DB/{ToLoad,Older}/README.md

are checked into the repo. The README files are the "git keepers" — without them, git would silently drop the empty directories.Frozen mode:build.py

extends itsempty_dirs

tuple with"DB/ToLoad"

and"DB/Older"

. The PyInstaller post-build step creates both underdist/manage/

, thepkg.zip

packager preserves them via explicit zip entries, and end-users get the tree from the very first launch.

Tlamatini ships with 24 skills — markdown SKILL.md packages under agent/skills_pkg/

that the LLM can invoke through invoke_skill('<name>', '{...args...}')

. They cover everything from the canonical acp-router

(pick the right external CLI for an intent) and summarize

(compress text faithfully) to setup-new-acpx-key

, skill-creator

, code-review

(senior-engineer git-diff review with an APPROVE/REQUEST_CHANGES verdict), security-audit

(multi-scanner SAST/secret/dependency sweep) and kali-pentest

(an authorized Kali Linux assessment runbook that drives the Kalier agent / MCP-Kali-Server), the tlamatini_*

audit/lint/refactor helpers, and integration stubs for GitHub / Notion / Slack / Gmail / Jira / Todoist / Trello / Weather.

Before 2026-05-17 the only way to interact with them was through the LLM (list_skills

to enumerate, invoke_skill

to run). The ACPX-Skills navbar dropdown — added next to Agents and Config in the chat toolbar — gives you an operator-grade admin surface that does NOT require the LLM. Four entries:

Opens a two-pane modal: a left-side list of all 24 skills (with a green/red dot for enabled / disabled and a runtime tag) and a right-side detail pane that shows the selected skill's full identity — description, runtime (in-process / acpx), acpx_agent

if any, budgets (max_iterations · max_seconds · max_tokens), trigger keywords, requires_tools

and requires_mcps

, inputs and outputs (with required-field markers), and the full markdown body. A search box at the top filters by name or description as you type. Pure read — nothing is written back.

Backed by GET /agent/skills/

(list payload) and GET /agent/skills/<name>/

(deep detail). Use it when you want to know what a skill actually does before you ask the LLM to call it, or when you've just authored a new SKILL.md and want to confirm it parsed correctly.

A checkbox grid — one row per skill — that mirrors the existing MCPs and Agents dialogs exactly. Toggle a checkbox off, click Continue, and the row's Skill.enabled

flips to false

. Two consequences immediately:

list_skills

(the LLM's enumeration tool) filters that skill out of its returned array.invoke_skill('<name>', ...)

returns{"ok": false, "code": "SKILL_DISABLED"}

instead of running.

Toggling back to enabled restores the skill. This is the right knob when (a) you want to hide an unfinished skill from the planner, (b) you don't have the API key for an integration skill (e.g. notion

without NOTION_TOKEN

) and don't want the LLM to keep trying, or (c) you're running a demo and want a minimal tool surface.

The toggle goes over the same WebSocket channel as set-mcps

/ set-tools

/ set-agents

— payload encoding name=description=true/false,name=description=true/false,...

. Backend handler is in consumers.py::receive

and calls save_skill(name, enabled)

which touches only the enabled

column.

A cross-check report that catches drift between the skill catalog and the rest of the system. Sections:

Missing tool dependencies— for each skill whoserequires_tools

lists a tool that's currentlydisabled in the Tools dialog, lists the skill + the unmet tools. (A disabled tool means the skillwouldfail at runtime — Diagnostics surfaces it before the LLM tries.)Missing MCP dependencies— same idea against disabledMcp

rows.Unknown ACPX agents— for skills withruntime: acpx

, flags anyacpx_agent

value that isn't in theAcpAgent

table (typo, removed CLI, etc.).Orphan DB rows—Skill

rows whose SKILL.md file no longer exists on disk. Usually a sign that someone deleted a skill directory without running Reload.

Each section is collapsed when clean (✓ green) and expanded with red ⚠ counts when something's wrong. Run it after editing SKILL.md files or after toggling tools/MCPs to confirm nothing is silently broken. Backed by GET /agent/skills/_/diagnostics/

— pure read, no writes.

A single-click button that re-runs the registry boot pipeline: rescan agent/skills_pkg/

, refresh every Skill

DB row's metadata (description, runtime, frontmatter_json, body_sha256), prune any DB row whose SKILL.md is gone. The user-toggled enabled

field is preserved across reload.

Use this after you've authored or edited a SKILL.md on disk — no server restart needed. The success toast tells you the new skill count.

By design, the Skill

DB table stays at "enumeration + enable/disable" only, exactly the way the Tool

and Mcp

tables work. Per-skill permissions (filesystem read/write globs, allowed shell commands, network deny/allow), budgets (max_iterations / max_seconds / max_tokens), and the skill's body all live in the SKILL.md frontmatter on disk and are the only source of truth. The admin UI deliberately does NOT let you override them from the browser — if you want to change a permission, edit the SKILL.md and click Reload. This keeps git diff

honest: every behavioural change to a skill shows up in a file, not in a database row that the next backup would silently archive.

• HTTP endpoints: agent/views.py

(list_skills_view

,skill_detail_view

,reload_skills_view

,skills_diagnostics_view

) — wired inagent/urls.py

. - WebSocket toggle: agent/consumers.py::receive

(set-skills

branch) →save_skill(name, enabled)

. The connect path also callsskill_establishment()

for every row so the frontend'sskills = []

cache hydrates on session start, mirroring howtools[]

andagents[]

hydrate. - Tool-surface gating: agent/acpx/tools.py::_disabled_skill_names()

— fails open on DB exception so a broken admin layer never silently hides skills. - Frontend dialogs: agent/static/agent/js/skills_dialog.js

(the Configure / Browse / Diagnostics / Reload dialogs) +agent/static/agent/css/skills_dialog.css

. - Coverage: 14 tests in agent/tests.py

(SkillsAdminEndpointTests

,SkillsToolSurfaceGatingTests

,SkillsNavbarTemplateContractTests

).

Windower is the desktop window manager of the chat surface — the third member of the desktop-UI trio: where Mouser clicks inside a window and Keyboarder types into one, Windower commands the window itself. It is implemented self-contained on the Win32 API (pywin32 win32gui

/win32con

/win32process

• ctypes

), porting the window-management subset of Microsoft's Windows-MCP — including the cross-process AttachThreadInput

focus-transfer dance that lets a background process reliably raise a foreground window. It is Windows-only and state-changing, so it appears in the Exec Report.

Tick only the Multi-Turn checkbox (Windower is a normal Multi-Turn tool — it is not behind the ACPX/Skill surface, so the ACPX checkbox is not required). Then ask, for example:

"Open Notepad, bring it to the front and maximize it, then tell me its size."

Tlamatini will launch the app (chat_agent_executer

), confirm it is up (chat_agent_window_present

), then call ** chat_agent_windower** with

action='maximize'

and window_title='Notepad'

— and you will watch the window come to the foreground and fill the screen. The tool promotes its result fields (

action

, window_title

, matched

, match_count

, state

, left

, top

, width

, height

) to the top level of its JSON return, so the answer can report the live geometry without parsing logs.action

can be any of: list

(enumerate every open window with its geometry + state), focus

, minimize

, maximize

, restore

, move

, resize

, move_resize

, close

(by title), topmost

/ untopmost

(always-on-top), or arrange

(snap/tile to left/right/top/bottom halves, the four quadrants, center, or full). Matching is by match_mode

∈ substring

(default) / exact

/ regex

, with match_index

to pick among same-titled windows. Use Windower — not Mouser — whenever the goal is the window as a whole (bring to front, tile, resize, close by title).

Two ready-made showcases live in the chat Prompts dropdown (the Catalog of Prompts): WINDOW SPOTLIGHT (basic — maximize + list + close) and WINDOW CHOREOGRAPHY (medium — restore → tile left → tile right → top-left quadrant → move/resize → list → close, so a single window visibly dances around the screen). Pick one, send, and watch.

Playwrighter drives a real browser (Playwright — Chromium / Firefox / WebKit) through a scripted, interactive, stateful flow. Where Crawler does a one-shot static fetch and the googler

tool only searches, Playwrighter clicks, fills forms, waits for elements, extracts text/attributes, screenshots, asserts, and downloads — so it can log into a site, submit a multi-step form, click through a wizard, or scrape a JavaScript-rendered single-page-app behind a login. It needs Playwright installed (pip install playwright && playwright install

); set ** headless=false** to

watchit drive, and

(alias

hold_open_seconds=N

hold_open_ms

) to keep the browser visible for N seconds afterthe last step

beforeit closes — that's the "wait a few seconds before closing so I can see it" knob; just ask Tlamatini to wait and it passes it for you.

Tick only the Multi-Turn checkbox (Playwrighter, like Windower, is a normal Multi-Turn tool — ACPX is not required). Then ask, for example:

"Open Wikipedia in a visible browser, search for ‘Nahuatl’, and tell me the first paragraph of the article."

Tlamatini calls ** chat_agent_playwrighter** with

start_url

, headless='false'

(and hold_open_seconds

if you asked it to keep the browser open), and the whole script as a single JSON string in (the flat

steps_json

key=value

request grammar cannot express a list-of-dicts, so the script is passed as JSON and the agent json.loads

it). Each step is {"action": <verb>, ...}

; supported verbs are goto

, click

, dblclick

, fill

, type

, press

, select

, check

/uncheck

, wait_for

, wait

, extract_text

, extract_attr

, screenshot

, assert_visible

, assert_text

, and download

. The run reports status

, final_url

, steps_run

, assert_result

, and the extracted values, so a downstream step (or a Forker, on the canvas) can branch on the verdict.Two ready-made showcases live in the Prompts dropdown: BROWSER SPOTLIGHT (basic — open example.com

with a visible browser, extract the heading, assert the link, screenshot) and BROWSER WIZARD (medium — a visible multi-step Wikipedia search: fill → click → wait → extract → assert → screenshot). The canvas counterpart is the visual Playwrighter node (see §4 and §9.5); the YAML steps

list is its authoring form.

Kalier bridges Tlamatini to Kali Linux offensive-security tooling through the MCP-Kali-Server. That project runs a small Flask API server (server.py

) on the Kali box exposing /api/command

, /api/tools/<tool>

and /health

; Kalier talks to it directly over HTTP (Python-stdlib urllib

, no extra packages in the agent pool), so it is the canonical tool for AI-assisted penetration testing, recon, and CTF solving. It is state-changing, so it appears in the Exec Report.

First, get the MCP-Kali-Server (server.py

) running on your Kali machine and reachable from Tlamatini. Tlamatini is the embedded client — you no longer need Claude Desktop's client.py

; instead you set the Kali box URL once in ** Config ▸ URLs → Kali server (Kalier)** (the

kali_server_url

key in config.json

, default http://127.0.0.1:5000

). That default already works when Kali runs in WSL2 with localhost forwarding or when you SSH-tunnel the port (ssh -L 5000:localhost:5000 user@KALI_IP

); for a LAN Kali box set it to http://<KALI_LAN_IP>:5000

. (See for the full zero-client walkthrough.)

Tlamatini-Kali-Setup.md

Authorized targets only— run Kalier solely against systems you own or are explicitly authorized to test (engagement, lab, CTF).

Tick only the Multi-Turn checkbox (Kalier is a normal Multi-Turn tool — not behind the ACPX/Skill surface). Then ask, for example:

"Scan 10.0.0.5 with an nmap -sCV on ports 1-1000 and summarize the open services."

Tlamatini calls ** chat_agent_kalier** with

action='nmap'

, target='10.0.0.5'

, scan_type='-sCV'

, ports='1-1000'

— and auto-injects your configured from

server_url

kali_server_url

, so you never repeat the Kali box address in a prompt (the LLM only passes server_url=

explicitly to hit a different one-off box). The action

field selects the capability: command

(any shell command on the Kali box), nmap

, gobuster

, dirb

, nikto

, sqlmap

, metasploit

, hydra

, john

, wpscan

, enum4linux

, or health

(probe the server and which tools are installed — a good first call when you are unsure the API is reachable). The tool returns the Kali tool's stdout/stderr verbatim and captures an INI_SECTION_KALIER

block (action

, endpoint

, subject

, return_code

, success

, timed_out

, server_url

) for the Exec Report and Parametrizer.On the canvas the same capability is the visual Kalier node (see §4 and §9.5): chain Starter → Kalier (nmap) → Parametrizer → Kalier (gobuster) → Forker → Ender

to build a fully unattended, branch-on-result assessment pipeline. The visual node and the chat tool share the same MCP-Kali-Server contract.

STM32er bridges Tlamatini to STM32 microcontroller firmware development through the STM32 Template Project MCP — a FastMCP server that exposes project scaffolding, build, flash, serial / SWD observation, and reset. It is state-changing (it compiles firmware and writes to hardware), so it appears in the Exec Report. The visual canvas counterpart is the STM32er node (see §4 and §9.5); both surfaces share the same MCP contract.

Zero-config auto-bootstrap — you only install STM32CubeIDE + Tlamatini.With no on-diskserver_script

configured (the new default), STM32erdownloads the MCPfrom its git repo (a shallowgit clone

, or a GitHub-zip fallback whengit

is absent) into a per-user cache (%LOCALAPPDATA%/Tlamatini/STM32TemplateProjectMCP

),pip-installs its deps(mcp

+pyserial

) if they are missing, andvalidatesthe server — all on first use. Nothing to clone by hand, no path to set. Abootstrap

action triggers this explicitly;auto_bootstrap

(defaulttrue

) does it lazily before the first real action.

⚠️ Critical-mission safety preflight.Before it compiles or flashes anything, STM32er runs avalidate

preflight that checks the environment:arm-none-eabi-gcc

, STM32CubeIDE,make

/cmake

,STM32_Programmer_CLI

, the ST-LINK USB driver,a positively-confirmed connected ST-LINK probe, and the target device family. If the environment is wrong — or the request would target the wrong STM32 family — STM32errefusesrather than producing or flashing mis-targeted firmware. Hardware isconditional: compile-only actions (build

,list_artifacts

,clean

,create_project

,write_source

) needno board; hardware actions (flash

,erase

,reset

,serial_*

, SWD reads,live_*

)requirea connected ST-LINK. The bundled MCP template isSTM32F407VG-specific, so a cross-STM32F

-family device mismatch is refused (a multi-family MCP fork is future work).

Tick only the Multi-Turn checkbox (STM32er is a normal Multi-Turn tool — not behind the ACPX/Skill surface). Then ask, for example:

"Scaffold a blinky project for the STM32F407, build it, and flash it to the connected board."

Tlamatini calls ** chat_agent_stm32er** — bootstrapping the MCP if needed, running the safety preflight, then driving the build and flash. The

action

field selects one of 23 MCP tools(project scaffold / build / flash / erase / reset / serial / SWD / observe / …),

2 composites(

serial_session

, live_monitor

), or 2 meta-actions(

bootstrap

, validate

). The tool captures an INI_SECTION_STM32ER

block for the Exec Report and Parametrizer.On the canvas the same capability is the visual STM32er node: chain Starter → STM32er (validate) → Forker → STM32er (build) → STM32er (flash) → STM32er (serial_session) → Ender

to build a fully unattended, validate-gated firmware pipeline. Three ready-made catalog demos ship in migration 0103

: STM32 GENESIS (bootstrap + validate + compile, no board needed), STM32 BLINKY (validate + build + flash), and STM32 HIL OBSERVATORY (a validate-gated real-hardware flash + SWD + serial + reset). The zero-config end-to-end path (download → build → flash → reset) is verified on a real STM32F407G-DISC1, with 122 automated tests in agent/test_stm32er_agent.py

.

📟

Serial output on a Discovery board needs a wire — the VCP is not bridged to the MCU's UART.This matters specifically for the3rd STM32er demo, STM32 HIL OBSERVATORY, whoseserial_session

step reads the firmware'sBOOT tlamatini hil count=…

banner. The on-board ST-LINK on theSTM32F4-Discovery family (including the STM32F407G-DISC1)provides debug (SWD) but —unlike ST— doesNucleoboardsnotinternally route its USBVirtual COM Portto any of the target STM32F407's USART pins. A firmware that prints overUSART2 (PA2 = TX, PA3 = RX)therefore showsnothingon the ST-LINK VCP regardless of baud or timeout: on the PCB those bytes have nowhere to go. To actually read that stream you mustbridge the port yourself with an external USB-to-UART (USB-TTL) adapter— cross-wire adapterRX ← PA2, adapterTX → PA3,GND ↔ GND— and pointserial_session

atthat adapter'sCOM port (not the ST-LINK VCP). No wiring is needed for the demo'sprimaryhardware proof: thelive_monitor

step samples the liveg_blink_count

counter from the running MCU over theSWD debug channel, so it works on a bare, unwired board — which is exactly why STM32er treats anempty VCP read on a Discovery board as, and why the HIL demo proves the firmware is alive over SWD first and treats the serial banner as a bonus.expected, not a failure

The chat is great for one-off tasks. The designer is for jobs you want scheduled, unattended, or identically reproducible.

┌────────────────────────────────────────────────────────────────────────┐
│ ▶ Start  ⏸   ⏹ Stop  ⚠ Hypervisor  💾 Save  📂 Load  ✓ Validate │
├──────────────────┬─────────────────────────────────────────────────────┤
│ Sidebar          │                                                     │
│ ─ Control        │                                                     │
│   Starter, Ender │            CANVAS (#canvas-content)                 │
│ ─ Routing        │       (draggable agents, typed connections,         │
│   Forker, Asker  │        green-running / red-down / yellow-d     │
│ ─ Logic Gates    │        LEDs)                                        │
│   AND OR Barrier │                                                     │
│ ─ Action / etc.  │                                                     │
└──────────────────┴─────────────────────────────────────────────────────┘

• The canvas scrolls: viewport is#submonitor-container

, content layer is#canvas-content

. New canvas-level features should be children of#canvas-content

. Connections are typed: green = "start the target after this finishes" (target_agents

), blue = "monitor this source's log" (source_agents

).Double-click an agent to edit its config.Right-click for description / log / explore-dir / open-cmd / restart.

Goal: run a shell command, take a screenshot, end.

• Drag Starter onto the canvas (top-left). - Drag Executer to its right. - Drag Shoter further right. - Drag Ender at the far right. - Connect: Starter → Executer → Shoter → Ender (drag from the right edge of one to the left edge of the next).
• Double-click Executer, setcommand

todir C:\

(orls /tmp

). - Double-click Shoter, setoutput_dir

to a writable folder. - Leave Ender wiring to Tlamatini. Validate/Start will calculate Ender'starget_agents

kill list from the arrows. - Click ✓ Validate— Tlamatini compiles the visible canvas, then runs structural checks (no orphans, no self-connections, terminal agents reachable). - Click ▶ Start. LEDs go green, then gray. Openoutput_dir

— there's a screenshot.

💾 Save — pick a name. You get a JSON file with positions, configs, and connections. Distribute to colleagues; they 📂 Load the same file and run the same flow. .flw

is also what the chat's Create Flow button emits.

A .flw

file is meant to describe the idea of the flow, not the exact machine it was created on. A good .flw

says:

• "There is a Starter here."
• "There is an Executer there."
• "Starter connects to Executer."
• "Executer uses this script."

It should not say:

• "This flow only works from C:/Development/Tlamatini/...

." - "This flow only works from the install folder on Angel's PC."

• "This Parametrizer mapping exists somewhere in a temporary pool directory, good luck."

Saved flows now carry a small schemaVersion

plus an artifacts

section. The most important artifact today is Parametrizer mappings. When you save a flow with a Parametrizer, Tlamatini keeps the mapping in the .flw

. When you load the flow later, Tlamatini recreates interconnection-scheme.csv

for that Parametrizer in the current session pool.

For a beginner, the practical rule is simple: if you configured Parametrizer with the mapping dialog, Save/Load should remember that mapping.

This is the most important reliability change in the visual designer.

Before, Validate mostly read whatever agent configs already existed in the pool directory. That could become stale:

• You drag nodes around.
• You load a .flw

. - You edit a config.

• You reconnect an edge.
• The pool directory still contains an older config.yaml

. - Validate or Start reads that older file and acts confused.

Now ACP takes a fresh snapshot of the canvas before validation and start. The snapshot includes:

• every visible node;
• each node's position;
• every connection;
• input and output slot numbers;
• current in-browser config;
• Parametrizer mappings.

The backend then compiles that snapshot into real pool config.yaml

files using the Agent Contract Registry. In beginner terms: the picture on the screen becomes the source of truth.

Another important nuance: if you opened an agent dialog and manually edited wiring-sensitive fields such as source_agents

or an Ender kill list, those dialog edits now survive compilation. Canvas edges still contribute their live connections, but a deliberate dialog override is no longer silently discarded by Validate or Start.

What happens when you click ✓ Validate:

• Browser captures the live canvas.
• Backend compiles it in dry-run mode.
• Frontend validates the compiled configs.
• Nothing is written to disk just for validation.

What happens when you click ▶ Start:

• Browser captures the live canvas.
• Backend compiles it in write mode.
• Pool folders/configs are updated.
• Logs are cleared.
• Starter agents launch.

This removes a whole class of "I swear I connected it correctly, why is it running the old thing?" problems.

Button	What happens
⏸
Saves running agents into d_agents.reanim , kills them, leaves logs and reanim* state files intact. LEDs go yellow.
▶ Resume (after )
Reanimates each saved agent with AGENT_REANIMATED=1 . Each agent reads its reanim* files and continues from where it stopped.
⏹ Stop
Hard stop. Ender runs termination logic; reanimation files are cleared.

This is why long-running workflows (Crawler scraping 10k URLs, Parametrizer iterating segments) survive s.

Stop is also safer in mixed flows now: the ACP cleanup path is better at terminating leftover session processes before the next run begins, so partially mixed ACP sessions are less likely to leave orphaned agents behind.

Click ⚠ Hypervisor — a system FlowHypervisor agent starts watching every running agent. It is an LLM that reads each agent's log, builds an NxN connection matrix from the canvas wiring, and emits exactly ** OK** or

. If it raises, the browser pops an alert. Add custom rules to

ATTENTION NEEDED { explanation }

user_instructions

in its config.Drag FlowCreator, double-click, and type a natural-language objective:

"Every hour, crawl our status page; if it shows ERROR, email the on-call engineer."

Click Generate. FlowCreator reads agentic_skill.md

(its design playbook), produces a JSON flow description, and renders agents + connections onto the canvas. Tweak and run.

Tlamatini agents communicate through log files and ** config.yaml**. Parametrizer is the bridge: it reads structured segments from a source agent's log, injects mapped values into a target agent's

config.yaml

, runs the target, restores the config, advances the cursor, repeats.The unified output format every Parametrizer-friendly agent emits:

INI_SECTION_<AGENT_TYPE><<<
key1: value1
key2: value2

multi-line body content (becomes 'response_body')
>>>END_SECTION_<AGENT_TYPE>

25 source agents support this format: Apirer, Gitter, Kuberneter, Crawler, Summarizer, File-Interpreter, Image-Interpreter, File-Extractor, Prompter, FlowCreator, Kyber-KeyGen/Cipher/DeCipher, Gatewayer, Gateway-Relayer, Googler, Playwrighter, ACPXer, Shoter, Mouser, Windower, Unrealer, Reviewer, Analyzer, Kalier.

Canonical example:

Apirer ─► Parametrizer ─► Kyber-Cipher

Apirer hits 3 endpoints → 3 INI_SECTION_APIRER<<<

blocks → Parametrizer maps response_body → buffer

→ Kyber-Cipher runs 3 times, encrypting each body. No manual config editing. -safe. Single-lane queue.

The mapping dialog is now part of normal flow persistence:

• Connect exactly one source into Parametrizer.
• Connect Parametrizer to exactly one target.
• Double-click Parametrizer.
• Click a source field on the left.
• Click a target config field or target marker on the right.
• Save mappings.
• Save the .flw

.

When the .flw

is loaded later, Tlamatini restores the mappings and writes the Parametrizer's interconnection-scheme.csv

again. You do not need to remember which pool directory had the CSV.

One limitation is intentional: one Parametrizer is a single-lane queue from one source to one target. If one API response must feed Emailer and File-Creator, use two Parametrizers.

Two trigger modes:

Mode	When
HTTP webhook
CI server, SaaS callback, cron, curl, internal portal — anything that POSTs. Auth: bearer / hmac / none . Validates → dedups → queues → starts target_agents .
Folder-drop watcher
Industrial / IoT — sensor writes JSON to a shared folder. Gatewayer polls, archives, fires.

Pending events survive crashes via reanim_queue.json

. To accept GitHub-style webhooks (which sign only the body), put the bundled Gateway-Relayer in front.

ACPX = Agent Communication Protocol eXtension. It spawns external coding-agent CLIs as out-of-process child processes, talks to them over stdin/stdout, persists the conversation as NDJSON transcripts, and brokers them to the chat LLM as 12 native tools. It is a Python port of OpenClaw's ACPX plugin — agent_id

mapping, permissionMode

vocabulary, and SKILL.md

frontmatter all match verbatim.

Defined in agent/acpx/agent_registry.py::DEFAULT_ACP_AGENTS

. User overrides go in config.json

under acpx.agents.<id>

.

agent_id | Default command | Transport | Prompt form | |---|---|---|---| claude | claude | oneshot-prompt | claude -p "<task>" | codex | codex | oneshot-prompt | codex exec "<task>" | cursor | cursor-agent | oneshot-prompt | cursor-agent -p "<task>" | gemini | gemini | oneshot-prompt | gemini -p "<task>" | qwen | qwen-code | oneshot-prompt | qwen-code -p "<task>" | tlamatini | python -m agent.acpx.self_acp_server | json-acp | stdin envelope | kiro / kimi / iflow / kilocode / opencode / pi / droid / copilot | (own command) | tui-repl | stdin |

Transport modes:

— fresh process per turn; prompt is a CLI arg; stdin closes; stdout captured to EOF. Theoneshot-prompt

only transport that reliably captures TUI agents on Windows (TUI CLIs detect a piped stdout and refuse to flush in long-lived mode).— child speaks one JSON envelope per turn, ends withjson-acp

{"done": true}

.— long-lived REPL; transport-aware idle rule fires aftertui-repl

startup_grace + idle_seconds

even with zero events (a silent TUI is, by definition, finished).

All return JSON envelopes. Failures: {"ok": false, "reason": "...", "code": "..."}

.

Tool	What it does
acp_doctor()
Health probe + per-agent enumeration with resolvable and cli_version . Always call first.
list_acp_agents()
Cheap enumeration without the version probe.
acp_spawn(agent_id, task, …)
Spawn child. Returns session_id , transport , transcript_path , events . TUI agents return sub-second.
acp_send(session_id, text, …)
Send a follow-up turn.
acp_send_and_wait(session_id, text, until_idle_seconds=10, max_wait_seconds=180)
Send and block until child settles. Prefer this for "wait for the full answer".
acp_kill(session_id)
Terminate. Returns transcript_path so Exec Report can cite it.
acp_transcript(session_id, max_chars, direction)
Read the on-disk NDJSON transcript.
acp_session_status(session_id)
{alive, pid, transcript_size, last_event_at, closed} .
acp_list_sessions()
Enumerate live sessions.
acp_relay(session_id_src, session_id_dst, …)
Single-call hand-off — replaces transcript→string→send.
invoke_skill(skill_name, args_json)
Run a SKILL.md package inside SkillHarness .
list_skills(filter_keywords)
List registered skills.

24 seed skills live under agent/skills_pkg/

(acp-router, summarize, setup-new-acpx-key, skill-creator, code-review, security-audit, kali-pentest, 8× tlamatini-* maintenance helpers, plus OpenClaw-format ports for github / gmail / slack / jira / notion / todoist / trello / weather).

Tick Multi-Turn + ACPX + Exec Report and send:

"Spawn claude in

C:/Development/Tlamatini

, ask it to list the top-level files, harvest the answer, and kill the session."

Expected tool sequence:

acp_doctor
  → acp_spawn(agent_id="claude", task="list top-level files")
    → acp_send_and_wait(session_id, "...")
      → acp_kill(session_id)

"Spawn claude in this dir, ask it to draft a refactor of

worker.py

. Spawn gemini, relay claude's answer to it, ask gemini to critique. Kill both."

Expected sequence:

acp_doctor
  → acp_spawn(claude, draft_task)
    → acp_send_and_wait(session_a, …)
      → acp_spawn(gemini, critique_template)
        → acp_relay(session_a, session_b)     # ONE call — transform=last_assistant_text
          → acp_kill(session_a)
            → acp_kill(session_b)

Without acp_relay

, that hand-off is three calls (acp_transcript

→ string-manipulate → acp_send

). Always prefer the dedicated tool.

Two layers exist in config.json

:

{
  "ANTHROPIC_API_KEY": "sk-ant-...",     // Layer 1: Tlamatini's own cloud calls
  "GEMINI_API_KEY": "AIza...",
  "acpx": {
    "agents": {
      "claude": { "env": { "ANTHROPIC_API_KEY": "sk-ant-..." } },   // Layer 2: spawned child env
      "gemini": { "env": { "GEMINI_API_KEY": "AIza...", "GOOGLE_API_KEY": "AIza..." } },
      "codex":  { "env": { "OPENAI_API_KEY":  "sk-..." } },
      "qwen":   { "env": { "DASHSCOPE_API_KEY": "sk-..." } }
    }
  }
}

Merge order at spawn: {**os.environ, **spec.env}

— explicit per-agent env

wins over an exported shell variable.

Easier path — invoke the setup-new-acpx-key

skill from chat (Multi-Turn + ACPX ticked):

"Use

invoke_skill

withsetup-new-acpx-key

to register my Anthropic key for theclaude

agent_id." (paste the key)

The skill writes data.keys

, patches both config.json

layers, optionally extends regen_secrets.py

, and verifies via acp_doctor

.

Security:config.json

is git-tracked. Usepython regen_secrets.py --mode push-able

to swap real keys for placeholders before commit;--mode keyed

restores fromdata.keys

(gitignored). Never commitdata.keys

.

ACPXer is the canvas-facing version of the 12 LLM-facing tools. One ACPXer node = one full ACPX session lifecycle. It is self-contained — does NOT import agent.acpx

— because pool subprocesses can't import agent.*

. Mirrors the runtime's transport-aware drain inline (~120 lines), writes byte-identical NDJSON transcripts, and emits Parametrizer-compatible INI_SECTION_ACPXER<<<

blocks.

Canonical visual relay flow:

Starter → ACPXer(claude) → Parametrizer → ACPXer(gemini) → Parametrizer → ACPXer(cursor) → File-Creator → Ender

Three different LLMs argue back and forth, fully visual, fully unattended.

The Unrealer agent (#62 in the catalog) lets Tlamatini drive a live Unreal Engine 5 editor through the Unreal MCP plugin's TCP socket protocol. You spawn a chat_agent_unrealer

call from Multi-Turn or drop an Unrealer node on the visual canvas; Tlamatini opens a TCP connection to 127.0.0.1:55557

, sends one JSON command ({"type": <verb>, "params": {...}}

), captures the engine's JSON response into an INI_SECTION_UNREALER<<<

block, and triggers downstream agents. Because the agent forwards whatever command

• params

you give it, the catalog is exactly whatever your connected plugin build exposes — from the base 28-command upstream release up to the 53-command, nine-category extended surface (actor manipulation incl. viewport screenshots, Blueprint creation and graph wiring, input mappings, UMG widget building, in-editor Python/console execution, level I/O, asset import, and material authoring) shipped by Tlamatini's own plugin fork, (the Unreal Engine MCP modified specifically for this system — see §6.2) — without you ever leaving the chat or the canvas.

XAIHT/XaihtUnrealEngineMCP

Unreal MCP is an open-source UE5 plugin (Model Context Protocol over a TCP socket) that runs inside the Unreal Editor process and accepts one JSON command per TCP connection. Each command names a verb (spawn_actor

, create_blueprint

, compile_blueprint

, add_widget_to_viewport

, …) and a dictionary of parameters; the plugin schedules the work onto UE5's game thread, executes it, and writes a JSON response back over the same socket before closing it. The wire shape is small — {"type": <command>, "params": {...}}

going in, {"status": "ok"|"error", "result": {...}}

(or {"success": false, "error": "..."}

) coming back — and is identical across every documented command.

Tlamatini does not embed or compile the plugin. It is a client of whatever UE5 instance the user has already started. The engine must be open, the plugin must be enabled, and its in-engine listener must be bound to 127.0.0.1:55557

(the default — configurable per-call via host

/ port

). Tlamatini contributes the calling side: one wrapped Multi-Turn tool, one visual canvas node, one Agent Contract entry, one Exec Report row family, and one Parametrizer source mapping — all built around the same UnrealConnection

adapter at agent/agents/unrealer/unrealer.py

.

Recommended — Tlamatini's own extended fork. The plugin Tlamatini is built and tested against is the Unreal Engine MCP modified specifically for this system:

Repository:https://github.com/XAIHT/XaihtUnrealEngineMCP.git

What it is: the canonicalchongdashu/unreal-mcp

plugin forked and extended for Tlamatini. It ships the full53-command, nine-category surface this chapter documents — the base editor / blueprint / node / project / umg verbsplus the System / Level / Asset / Material families and the newertake_screenshot

/focus_viewport

/set_pawn_properties

/find_blueprint_nodes

verbs.Plugin folder name:UnrealMCP

Default plugin TCP port:55557

on127.0.0.1

Supported UE versions: Unreal Engine 5.5+

It speaks the identical wire protocol on the identical port as every other build below, so it is a drop-in: Tlamatini's UnrealConnection

adapter needs no client-side changes to use it. Install this one if you want the System / Level / Asset / Material families that the seeded demos idPrompt 60/61/62

(§6.5) exercise.

Upstream base. The XAIHT fork is built on the canonical reference implementation Tlamatini's UnrealConnection

adapter mirrors verbatim:

Repository:https://github.com/chongdashu/unreal-mcp

License: MITSupported UE versions: Unreal Engine 5.5+

If you only ever need the base 28-command surface (editor / blueprint / node / project / umg), the upstream is enough on its own.

Equivalent community forks. Two other forks ship the same wire protocol on the same port; either works with Tlamatini's Unrealer with no client changes:

https://github.com/CrispyW0nton/Unreal-MCP-Ghost

https://github.com/gingerol/vhcilab-unreal-engine-mcp

Pick the build that matches your UE5 version and your team's licensing comfort. If you fork the plugin to add a new command verb, your fork is automatically usable from Tlamatini — there is no client-side allow-list of verbs (the wrapped tool forwards any command

• params

pair verbatim).

The plugin is a per-project install (not engine-wide). Steps:

Clone or download the plugin — the recommendedfork from §6.2, or any compatible build (only theXAIHT/XaihtUnrealEngineMCP

MCPGameProject/Plugins/UnrealMCP

folder matters — different forks may name the folder slightly differently; rename toUnrealMCP

if needed).Drop the folder into your project'sPlugins/

directory so the final path is<YourProject>/Plugins/UnrealMCP/UnrealMCP.uplugin

. Create thePlugins/

folder at the project root if it does not exist.Open the project in UE5. The editor will detect the new plugin and offer to rebuild it for your engine version — accept. If you opened a Blueprint-only project, you will be prompted to install Visual Studio Build Tools / Xcode command-line tools first, since the plugin is C++.Enable the plugin viaEdit → Plugins → search "UnrealMCP" → tick Enabled

. Restart the editor.Confirm the listener is bound. With the editor running, open theOutput Log(Window → Developer Tools → Output Log

) and look for a line such asLogTemp: UnrealMCP listening on 127.0.0.1:55557

. That line is your green light: the plugin is now waiting for JSON commands on the loopback interface.

You do not need to press Play (PIE).The plugin listens ateditorlevel — actor manipulation, Blueprint creation, widget construction, etc. all work against the open project even when PIE is stopped. Some UMG operations (add_widget_to_viewport

) physically render only after the user enters PIE, but the build steps are queued correctly either way.

The Unrealer agent forwards whatever command

• params

you pass it, so the exact catalog is whatever your connected plugin build exposes — there is no client-side allow-list of verbs. The canonical chongdashu/unreal-mcp release ships 28 commands across 5 categories (rows marked base

below); plugin builds that add the System / Level / Asset / Material command handlers — such as Tlamatini's own extended fork XAIHT/XaihtUnrealEngineMCP (§6.2) — bring the total to

53 commands across 9 categories:

Category	Commands	Tier
editor
get_actors_in_level , find_actors_by_name , spawn_actor , create_actor , delete_actor , set_actor_transform , get_actor_properties , set_actor_property , spawn_blueprint_actor , focus_viewport , take_screenshot
base + focus_viewport /take_screenshot
blueprint
create_blueprint , add_component_to_blueprint , set_static_mesh_properties , set_component_property , set_physics_properties , compile_blueprint , set_blueprint_property , set_pawn_properties
base + set_pawn_properties
node
add_blueprint_event_node , add_blueprint_input_action_node , add_blueprint_function_node , connect_blueprint_nodes , add_blueprint_variable , find_blueprint_nodes , add_blueprint_get_self_component_reference , add_blueprint_self_reference
base + find_blueprint_nodes
project
create_input_mapping
base
umg
create_umg_widget_blueprint , add_text_block_to_widget , add_button_to_widget , bind_widget_event , add_widget_to_viewport , set_text_block_binding
base
system
execute_python , execute_console_command , get_class_info , list_assets
extended
level
open_level , save_current_level , save_all , new_level , get_current_level
extended
asset
import_asset , duplicate_asset , rename_asset , delete_asset , save_asset , create_folder
extended
material
create_material , create_material_instance , set_material_parameter , assign_material
extended

execute_python

is the universal escape hatch — it runs an arbitrary Python script inside the editor, so anything in UE5's unreal

Python API (Niagara, Sequencer, landscape, audio, …) is reachable even when no dedicated verb exists. take_screenshot

closes the observe→act loop: spawn or change something, then capture the viewport to verify it. Note that the plugin's headless build/cook/test tools (build_project

, run_automation_tests

, run_macro

) are not part of this catalog — they shell out to UnrealEditor-Cmd

as separate processes and are unreachable over the editor's TCP socket. Chain Unrealer nodes through a Parametrizer for the run_macro

equivalent.

Param shapes vary per command (e.g. spawn_actor

wants name

• type

• location

• rotation

; create_blueprint

wants name

• parent_class

; set_material_parameter

wants material

• parameter

• value

; import_asset

wants source_file

(a disk path) + destination_path

(a /Game

content path)). The Unrealer agent does not validate them — it forwards them as-is, after two defensive fixups: it normalizes /Content/...

content paths to /Game/...

, prunes unset placeholder params, and remaps params.console_command

→ the wire's params.command

for execute_console_command

(so the console line doesn't collide with the top-level command:

selector). The plugin will reply with {"status": "error", "error": "<reason>"}

if a param is missing or malformed, and that error lands verbatim in the INI_SECTION_UNREALER

block so Multi-Turn / Parametrizer can branch on it.

The wrapped Multi-Turn tool chat_agent_unrealer

is the easiest way in. Tick Multi-Turn in the toolbar, leave Exec Report ticked too (the Unreal calls get their own table in the answer), and send a prompt like:

"Run Unreal command with command='spawn_actor' and params.name='MyCube' and params.type='StaticMeshActor' and params.location=[0,0,150]."

The planner picks chat_agent_unrealer

, the wrapped runtime spawns one short-lived unrealer.py

child under agent/agents/pools/_chat_runs_/unrealer_<seq>_<id>/

, the child opens a TCP socket to 127.0.0.1:55557

, sends the JSON command, captures the response, emits the INI_SECTION_UNREALER

block to its log, and exits. The Multi-Turn loop reads the run's log excerpt, parses the section, and returns the full Unreal response JSON to the LLM. The LLM then sees the engine's reply and either reports it to you, branches on it, or fires the next call.

The tool accepts the same overrides documented in config.yaml

:

host='10.0.0.5'

andport=55557

to target a remote UE instance (rare; the plugin binds to loopback by default and you would need to change the bind address inside the plugin or tunnel it).connect_timeout=5

andread_timeout=10

to widen the budgets for slow operations (e.g.compile_blueprint

on a complex graph).

Built-in demo prompts. Migration 0087_add_unrealer_demo_prompt.py

seeds a one-click demo into the Prompts table (idPrompt=25

). Open the chat, click the Prompts dropdown, pick Unreal MCP End-to-End Editor Drive, and Tlamatini will execute ten guided steps spanning the base command categories — sanity-probe (get_actors_in_level

), spawn a StaticMeshActor

(spawn_actor

), verify it (find_actors_by_name

), scaffold a Blueprint (create_blueprint

), add a StaticMeshComponent

(add_component_to_blueprint

), compile (compile_blueprint

), spawn an instance (spawn_blueprint_actor

), build a UMG HUD widget (create_umg_widget_blueprint

→ add_text_block_to_widget

→ add_button_to_widget

→ add_widget_to_viewport

) — and render the result as a per-step HTML report table at the bottom of the answer. Use it as your smoke test the first time you wire the plugin up.

Migration 0100_add_unrealer_extended_demo_prompts.py

adds three more demos that exercise the extended (System / Level / Asset / Material) surface the base demo never touches, at basic → hard complexity:

(basic): the observe→act loop —idPrompt=60

—Unreal Snapshotget_current_level

→spawn_actor

→take_screenshot

(toC:/Temp/unreal_snapshot.png

) →save_current_level

.(medium): content authoring —idPrompt=61

—Unreal Scene Forgelist_assets

→create_folder

→create_material

→create_material_instance

→set_material_parameter

→spawn_actor

→assign_material

→take_screenshot

→save_all

. (It is honest thatset_material_parameter

on a freshly-created blank material may returnstatus: error

— that is expected and recorded, not aborted.)(hard): the System escape hatch —idPrompt=62

—Unreal Python & Introspectionexecute_console_command

(via the agent'sparams.console_command

remap) →get_class_info

→list_assets

→execute_python

(a multi-line script passed as a triple-quotedparams.code

) →take_screenshot

.

All three drive chat_agent_unrealer

exactly like the base demo (tick only Multi-Turn; ACPX not required) and require the same running editor + bound plugin listener.

For unattended .flw

workflows, drop the Unrealer sidebar agent onto the canvas. Each node sends exactly one Unreal command when its turn arrives. The node's config.yaml

is the same one shipped under agent/agents/unrealer/config.yaml

:

host: 127.0.0.1
port: 55557
command: get_actors_in_level
params:
  name: ''
  type: ''
  location: []
connect_timeout: 5
read_timeout: 10
source_agents: []
target_agents: []

The agent emits an INI_SECTION_UNREALER<<<

block to its log, which means Parametrizer can chain Unreal calls together. Registered source fields (agent/services/agent_contracts.py

): host

, port

, command

, status

, error

, response_body

. Canonical multi-step canvas pattern — create a Blueprint, compile it, spawn an instance:

Starter → Unrealer(create_blueprint) → Parametrizer → Unrealer(compile_blueprint)
       → Parametrizer → Unrealer(spawn_blueprint_actor) → Ender

Each Parametrizer copies the previous Unrealer's response_body

(or a specific JSON field within it, via the Parametrizer dialog's interconnection-mapping UI) into the next Unrealer's params

block. Branching on status

(ok

vs error

) via a Raiser between Unrealer and the next Parametrizer gives you per-step exception handling — e.g., abort to a Notifier if compile_blueprint

returns status: error

.

The unrealer.py

script (~120 lines of business logic, plus the standard pool-agent boilerplate) is self-contained: it does NOT import from agent.acpx

, agent.services

, or any other Tlamatini-internal package. Pool subprocesses run as separate Python interpreters with no sys.path

back into the Django app, so the inline UnrealConnection

adapter is a verbatim mirror of the upstream Unreal MCP Python client (with the FastMCP plumbing stripped out). Per execution:

Load, readconfig.yaml

host

,port

,command

,params

, timeouts, andtarget_agents

.Write so the orphan-process reaper (chapter 11) can track the run.agent.pid

Open a fresh TCP socket tohost:port

withTCP_NODELAY

,SO_KEEPALIVE

, and 64 KB send/recv buffers. The Unreal MCP plugin closes the socket after each command, so the agent opens a new one per turn.Sendjson.dumps({"type": command, "params": params})

(no trailing newline) and callrecv()

until a complete JSON document has been assembled (validated by attemptingjson.loads()

on the accumulated bytes after each chunk).Normalize the response shape —{"success": false, ...}

from older plugin builds is rewritten into{"status": "error", "error": ...}

so downstream Parametrizer / Multi-Turn code can rely on a single shape.Emit one atomic with thelogging.info()

callINI_SECTION_UNREALER<<<

block (header:host

,port

,command

,status

,error

; body: pretty-printed Unreal JSON response, capped at 64 KiB). Single-call emission is mandatory for the parser atagent/agents/parametrizer/parametrizer.py

.Trigger, so the canvas can route on thetarget_agents

even on errorstatus

field instead of relying on agent-level fail-stops.Remove and exit.agent.pid

Failure modes the adapter handles gracefully (each turns into {"status": "error", "error": "<reason>"}

plus a non-fatal log line, never an uncaught exception):

Connection refused— the plugin's TCP listener is not bound (editor not running, plugin not enabled, port mismatched).** Socket timeout during receive**— UE5's game thread is busy (e.g.compile_blueprint

on a heavy graph) and exceededread_timeout

. Raiseread_timeout

inconfig.yaml

or in the wrapped tool call.Malformed JSON— the plugin closed mid-write; logged as anerror

status and downstream agents still fire.

chat_agent_unrealer

is registered in _EXEC_REPORT_TOOLS

(agent/mcp_agent.py

) under agent_key="unrealer"

and agent_display="Unrealer"

. When Exec Report is ticked alongside Multi-Turn, every Unreal call shows up as one row in a dedicated List of Unrealer Operations table at the bottom of the answer. Columns: command (left-bordered with the Unrealer caption gradient), success (SUCCESS

/ FAILURE

derived from the underlying tool-call verdict — the same verdict Multi-Turn already uses for dedup and tool_calls_log

). The table styling lives in agent/static/agent/css/agent_page.css

(caption gradient mirrors .canvas-item.unrealer-agent

in agentic_control_panel.css

).

A short pre-flight you can copy into a sticky note before any session:

Check	How
UE5 5.5+ open with a project loaded	File → Project → <YourProject> — and leave the editor focused, not minimized to the system tray
Plugin enabled	Edit → Plugins → UnrealMCP → Enabled = ✓ , restart of editor confirmed
Listener bound	Output Log shows UnrealMCP listening on 127.0.0.1:55557 (or your configured port)
Port not blocked	Test-NetConnection -ComputerName 127.0.0.1 -Port 55557 returns TcpTestSucceeded: True (PowerShell)
Tlamatini server up	python Tlamatini/manage.py runserver --noreload shows the startup banner
Multi-Turn ticked
The toolbar checkbox to the left of Exec Report
Tool enabled	Tools dialog in chat shows Chat-Agent-Unrealer ticked (it ships ticked by default after migration 0086_add_chat_agent_unrealer_tool runs)

Then run the seeded Unreal MCP End-to-End Editor Drive demo prompt (idPrompt 25) as your smoke test. A clean run leaves three artifacts in your project: actor TlamatiniProbe_Cube

, Blueprint BP_TlamatiniProbe

with one spawned instance TlamatiniProbe_Spawned

, and widget /Game/UI/WBP_TlamatiniProbeHUD

. Delete them via right-click in the Content Browser when you are done.

Unreal MCP is aacp_doctor

is not relevant here.workflow-agentsurface, not theACPXsurface — theacp_*

tools talk to external coding-agent CLIs (claude, gemini, …), not to UE5. The corresponding "is the channel alive?" probe for Unrealer is to callchat_agent_unrealer

withcommand='get_actors_in_level'

and check thatstatus == 'ok'

.The plugin is not listening. Check the UE5 Output Log for thestatus: error

/Failed to connect to Unreal at 127.0.0.1:55557

.UnrealMCP listening on …

line. If the line is absent, the plugin is either disabled or failed to build (re-open the project; UE5 will re-prompt to rebuild). If the line is present but the connection still fails, your firewall is blocking loopback (rare on Windows, butRestart-Service mpssvc

and re-test if you have aggressive endpoint security).UE5's game thread is busy. Either widenTimeout receiving Unreal response

.read_timeout

(config.yaml

or the wrapped-tool call), or split the work into smaller commands (e.g. spawn 10 actors with 10 separate calls instead of onespawn_n_actors

macro the plugin does not actually expose).Capitalize thestatus: error

from a Blueprint command, but the verb seems valid.parent_class

exactly as UE5 expects (Actor

,Pawn

,Character

,UserWidget

, …). The plugin does not auto-resolveactor

→Actor

.Widget appears in the Content Browser but is invisible in PIE.add_widget_to_viewport

queues the widget at editor level; you still need to pressPlay in the editor (or calladd_widget_to_viewport

from within a running PIE session) to make it render. This is an Unreal MCP plugin behavior, not a Tlamatini bug.The Output Log shows the plugin received the command but nothing happened in the level. Most often: an actor spawn at a coordinateinside another object's collision volume. UE5 silently fails the spawn. Raiseparams.location

by[0, 0, 150]

and retry.A second instance of UE5 is bound to the same port. Only one UnrealMCP listener can bind to127.0.0.1:55557

per host. Close the second editor instance, or configure each instance to bind to a different port and passport=<n>

per Unrealer call.

For the full debugging trail: pool-agent log lives at <pool>/unrealer_<n>/unrealer_<n>.log

; chat-wrapped runs land under agent/agents/pools/_chat_runs_/unrealer_<seq>_<id>/unrealer_<seq>_<id>.log

. Both contain the outbound JSON command and the inbound Unreal response verbatim.

For shipping a one-click Windows installer to end users.

build.py  ──►  build_uninstaller.py  ──►  build_installer.py
   │                   │                         │
   ▼                   ▼                         ▼
pkg.zip          Uninstaller.exe        dist/Tlamatini_Release/
python build.py

Installs deps, runs collectstatic

, executes PyInstaller, copies required payloads (including README.md

, the self-knowledge map Tlamatini.md

, and bundled jd-cli/

), runs migrations, creates the default user (user

/changeme

), renames the exe to Tlamatini.exe

, copies all 68 agent templates, bundles support scripts (register_flw.ps1

, CreateShortcut.ps1

, Tlamatini.ps1

, Tlamatini.ico

), and zips it all into ** pkg.zip**.

build.py

is strict: missing README.md

, missing jd-cli/

, or missing jd-cli.bat

causes a non-zero exit.

Self-modify builds. Add the --self-modify

flag to ship Tlamatini's own source tree inside the distribution:

python build.py --self-modify

When the flag is present (self_modify = "--self-modify" in sys.argv

), the build copies Tlamatini/agent/TlamatiniSourceCode/

recursively to the install root next to the exe, so it resolves like prompt.pmt

/ config.json

/ Tlamatini.md

, and Tlamatini can read, inspect, and modify her own code at runtime. Without the flag the directory is omitted entirely. The build prints Self-modify build : YES

(or no

) so you can confirm which kind of build you produced. See §9.6 for how the LLM uses it.

python build_uninstaller.py

Builds uninstall.py

into a single --onefile

Tkinter exe. Output: Uninstaller.exe

at the project root.

python build_installer.py

Requires pkg.zip

and Uninstaller.exe

. Builds install.py

with --onedir --windowed

and a splash screen, copies pkg.zip

and Uninstaller.exe

into dist/Installer/

, and assembles dist/Tlamatini_Release/

with SHA-256 verification.

The final distributable is dist/Tlamatini_Release/

— zip the folder, share it.

• Tkinter GUI to choose installation directory (no admin needed).
• Extracts pkg.zip

into<install_path>/Tlamatini/

. - Locks agent venv permissions.

• Writes config.json

. - Copies Uninstaller.exe

. - Creates desktop and Start Menu shortcuts ( Tlamatini.lnk

— falls back to user-scoped paths under restrictive Group Policies). - Registers .flw

to open with Tlamatini. - Cleans the PyInstaller bundle path from helper subprocess environments.

Frozen mode resolves config.json

from the executable directory (or CONFIG_PATH

env var). Template-agent discovery uses <install_dir>/agents

in frozen mode and Tlamatini/agent/agents/

in source mode. _resolve_python_executable()

tries PYTHON_HOME

→ bundled python.exe

→ PATH.

Tlamatini has two operational modes:

Mode	What it means	Where agent templates live
Source / Not-Frozen
You run python Tlamatini/manage.py runserver --noreload from a cloned repo.
Tlamatini/agent/agents/
Frozen / Installed
You run the packaged Tlamatini.exe from the installer.
<install_dir>/agents/

The new Flow Compiler was built to respect both modes. It does not assume your repo is at C:/Development/Tlamatini

, and it does not assume the installed app lives in a specific Program Files folder.

The compiler asks Tlamatini at runtime:

• "Am I frozen?"
• "Where are the agent templates?"
• "Where is this user's session pool?"
• "Which agent contract applies to this node?"

Then it writes only into the current pool:

agents/pools/<session_id>/<agent_name_n>/config.yaml

That path exists in both modes. In source mode it is under the repo's Tlamatini/agent/agents/pools/

. In frozen mode it is under the installed app's agents/pools/

.

For users, the takeaway is simpler: a .flw saved in source mode should load in an installed build, and a .flw saved in an installed build should load back in source mode.

Mode	Resolution order
Source	Tlamatini/agent/config.json
Frozen	<install-dir>/config.json next to the executable
Both	CONFIG_PATH env var wins over both

{
  "embeding-model": "Nomic-Embed-Text:latest",
  "chained-model": "glm-5:cloud",
  "ollama_base_url": "http://127.0.0.1:11434",
  "ollama_token": "",
  "ANTHROPIC_API_KEY": "<ANTHROPIC_API_KEY goes here>",
  "GEMINI_API_KEY": "<GEMINI_API_KEY goes here>",
  "enable_unified_agent": true,
  "unified_agent_model": "glm-5:cloud",
  "unified_agent_base_url": "http://127.0.0.1:11434",
  "unified_agent_temperature": 0.0,
  "unified_agent_max_iterations": 4096,
  "chat_agent_limit_runs": 100
}

unified_agent_max_iterations

caps the Multi-Turn tool loop (default 4096). enable_unified_agent

is the master switch for tool-calling.

Key knobs: chunk_size

(3000), chunk_overlap

(800), k_vector

/ k_bm25

(100 each), k_fused

(150), enable_bm25

, rrf_k

(60), max_doc_chars

(150000), max_context_chars

(250000), and a context_budget_allocation

map (high_relevance: 0.60, architecture: 0.20, related: 0.15, documentation: 0.05

). See BookOfTlamatini.md

Part VII for the full schema.

{
  "acpx": {
    "cwd": "C:/Development/Tlamatini",
    "stateDir": "C:/Users/<you>/.tlamatini/acpx-state",
    "probeAgent": "gemini",
    "permissionMode": "approve-reads",
    "nonInteractivePermissions": "deny",
    "timeoutSeconds": 180,
    "agents": {
      "claude": { "command": "C:/Users/<you>/AppData/Roaming/npm/claude.cmd",
                  "env": { "ANTHROPIC_API_KEY": "sk-ant-..." } }
    }
  }
}

permissionMode

∈ approve-reads

(default) / approve-all

(DANGEROUS) / deny-all

. The whole acpx

block is optional; on first boot of an upgrade build, boot_acpx()

appends the documented default block atomically.

mcp_system_server_port

(8765),mcp_files_search_server_port

(50051) — MCP daemons.internet_classifier_model

,web_summarizer_model

,web_context_max_chars

— internet toggle.image_interpreter_model

,image_interpreter_base_url

— vision.history_summary_*

,keep_last_turns

— chat-history compression.kali_server_url

(http://127.0.0.1:5000

) — the MCP-Kali-Server address auto-injected intochat_agent_kalier

(see §3.14).stm32_mcp_server_script

(now""

— empty triggers zero-config auto-bootstrap),stm32_mcp_repo_url

,stm32_mcp_install_dir

— the STM32 Template Project MCP for STM32er (see §3.15). Leavestm32_mcp_server_script

empty and STM32er downloads, installs, and validates the server itself on first use.

You no longer need to hand-edit all of those values. On /agent/

, open Config -> Models

or Config -> URLs

to edit the most common runtime knobs in-place. The browser validates model strings / URLs / hosts / ports, the backend validates again, and config_.save_config_updates()

atomically merges only the changed keys into the active config.json

. The same path is used in source mode and frozen builds, so the chat UI and the executable stop drifting onto different config copies.

Browser (Chat / ACP Designer)
    │ WebSocket
    ▼
Django Channels (Daphne ASGI) → AgentConsumer
    │
    ├── RAG Pipeline (FAISS + BM25, RRF, context budgeting, OOM fallback)
    ├── Unified Agent (Multi-Turn loop, planner, wrapped runtimes)
    └── MCP Services (System-Metrics WS, Files-Search gRPC)
    │
    ▼
LLM Backends: Ollama | Claude API | Qwen vision     +     ACPX Runtime → external CLIs

Layer	Responsibility	Where
1. Persisted toggles	DB rows for Mcp / Tool / Agent (UI enable/disable).
agent/models.py
2. Runtime MCP services	System-Metrics (WebSocket) + Files-Search (gRPC) daemons.	agent/mcp_*
3. Context fetcher chains	LCEL sidecars that inject system / files context.	agent/chain_*_lcel.py
4. Main answer chains	Basic / History-aware / Unified. factory.py monkey-patches invoke() .
agent/rag/chains/
5. Unified-agent tools	74 synchronous @tool functions (20 core Python + 42 wrapped chat-agent + 12 ACPX/Skill). Active only in Multi-Turn.
agent/tools.py + agent/chat_agent_registry.py + agent/acpx/

Frontend (toggles) → WebSocket → AgentConsumer → ask_rag() (skips prompt-shape validator)
  → UnifiedAgentChain.invoke() → filter_acpx_tools(tools, acpx_enabled)
    → planner picks ≤20 tools (capability scoring + history-aware boost)
      → MultiTurnToolAgentExecutor: 1..4096 iterations of (LLM call → tool calls → ToolMessage)
        → Exec Report HTML appended (if exec_report_enabled, BEFORE save_message)
          → broadcast → frontend renders, shows Create Flow if all 4 gates pass

Tlamatini has two ways to create flows:

• The chat can infer a flow from Multi-Turn tool calls.
• The ACP canvas can build a flow by dragging agents and drawing arrows.

Those two paths now meet at the same backend contract layer. This is important because flow files are not just pictures. A flow must eventually become a set of real agent folders, and every folder needs a correct config.yaml

.

The contract layer is intentionally small:

File	What it does
agent/services/agent_paths.py
Finds the correct agents/ and agents/pools/ folders in both source mode and frozen mode. It also normalizes names like TeleTlamatini , tele-tlamatini , and teletlamatini into the same agent type.
agent/services/agent_contracts.py
Describes what each agent needs: which config fields hold incoming agents, which fields hold outgoing agents, which agents are singletons, which agents are long-running, which agents should be hidden from validation, which Parametrizer fields can be mapped, and which secrets must be redacted before export.
agent/services/flow_spec.py
Converts old and new .flw shapes into one clean FlowSpec . It accepts legacy sourceIndex / targetIndex links and newer stable sourceId / targetId links.
agent/services/flow_compiler.py
Converts a FlowSpec into the actual pool configs. In dry-run mode it returns the configs for validation. In write mode it updates the current session pool before Start runs.

For beginners, the rule is: the canvas or chat creates a flow idea, then the Flow Compiler turns that idea into executable agent folders.

The compiler does a few quiet but important safety jobs:

• It starts from each agent template's config.yaml

, then merges only the node's custom settings. - It clears and rebuilds managed connection fields, so stale arrows from an old pool do not survive by accident.

• It understands special agent wiring such as AND

,OR

,Asker

,Forker

,Counter

,Ender

,Stopper

, andCleaner

. - It writes interconnection-scheme.csv

for Parametrizer nodes when mappings are saved in the.flw

. - It keeps FlowCreator and FlowHypervisor out of runtime validation because they are helper/control agents, not normal flow workers.

• It redacts known secrets for remote chat ingress agents such as TeleTlamatini andWhatsTlamatini when chat-created flows are exported.

This is the Pareto improvement: a small shared backend layer makes both major features safer. Chat-created flows and ACP-created flows now speak the same format before they touch the runtime.

Family	Members
Control
Starter, Ender, Stopper, Cleaner, Sleeper, Croner
Routing
Raiser, Forker, Asker, Counter
Logic gates
OR, AND, Barrier
Action
Executer, Pythonxer, Prompter, Summarizer, Crawler, Googler, Playwrighter, Apirer, Gitter, Ssher, Scper, Dockerer, Kuberneter, Pser, Jenkinser, Sqler, Mongoxer, Mover, Deleter, Shoter, Mouser, Keyboarder, Windower, File-Creator, File-Interpreter, File-Extractor, Image-Interpreter, J-Decompiler, De-Compresser, Telegramer, TeleTlamatini, WhatsTlamatini, ACPXer, Unrealer, Reviewer, Analyzer, Kalier, STM32er
Cryptography
Kyber-KeyGen, Kyber-Cipher, Kyber-DeCipher (CRYSTALS-Kyber post-quantum)
Utility
Parametrizer, FlowBacker, Gatewayer, Gateway-Relayer, Node-Manager
Terminal / monitoring
Monitor-Log, Monitor-Netstat, Emailer, RecMailer, Notifier, Whatsapper, TelegramRX, FlowHypervisor
AI / design
FlowCreator

Per-agent details (config knobs, lifecycle, naming convention, log markers): see BookOfTlamatini.md

Part IV — The Tlamatini Bestiary. To add a new agent, follow Tlamatini/.agents/workflows/create_new_agent.md

(8-step checklist).

Tlamatini ships with a first-person self-knowledge map — Tlamatini/agent/Tlamatini.md

— that the LLM reads as her own description of who and what she is: her two runtime modes (frozen vs source, and how to tell them apart), the ports she opens (8000

for the web app, 8765

for the System-Metrics MCP, 50051

for the Files-Search MCP), her main pages, her tech stack, her full capability surface, and how she can improve herself. The audience is the LLM alone, so the file deliberately does not follow prompt.pmt

's HTML/contrast styling rules — it is a private self-reference, never rendered to users.

The map is injected into the system prompt at prompt-build time. prompt.pmt

carries a <self_knowledge>{self_knowledge}</self_knowledge>

block, and agent/rag/config.py

fills it in: _load_self_knowledge_block()

reads Tlamatini.md

, brace-escapes it ({

→ {{

, }

→ }}

) so its code snippets cannot collide with the f-string template variables, and fails open — a missing, empty, or unreadable file degrades to a short literal notice instead of raising. The substitution happens at the single prompt-load site in load_config_and_prompt()

, so it covers all four chains (basic, history-aware, unified, prompt-only) without adding a new input variable. Tlamatini.md

is resolved from the application directory exactly like prompt.pmt

and config.json

(the install root next to the .exe

in frozen mode, Tlamatini/agent/

in source mode), and build.py

ships it both via --add-data=…/Tlamatini.md;agent

and by copying it to the install root so frozen resolution next to the exe works. The identity rules in prompt.pmt

point the LLM at Tlamatini.md

whenever a prompt concerns who or what she is, her architecture / modes / ports / pages / internals, or improving herself.

Self-modification is a second, independent capability axis. The optional directory Tlamatini/agent/TlamatiniSourceCode/

, when present, contains Tlamatini's own source code so she can read, inspect, and modify herself — present means a self-able-modify build; absent means a not-self-able-modify build (orthogonal to frozen vs source). The tree is bundled only when build.py

is invoked with the new --self-modify

flag (see §7.2); without the flag it is omitted entirely. Because the directory is optional, prompt.pmt

instructs the LLM to always verify the directory's presence (for example, a Multi-Turn directory listing) before claiming she can read or edit her own code; if it is absent she says so and falls back to the injected self-knowledge plus the docs.

When you load your own project as context (Context ▸ Set directory / Set file as context) and then ask a generic "summarize the project / the source code / the provided context" question, the

loaded context takes priorityover the always-injected self-knowledge — so Tlamatini summarizesyourcode, not herself. This is enforced by aprompt.pmt

loaded-context-priority rule plus a deterministic scope header (agent/rag/utils.py::prepend_loaded_context_scope()

) applied across all four chains.

When you click Set directory as context in the Context menu, Tlamatini walks the directory, splits each file into chunks, and pushes every chunk through Ollama's embedding API to build a FAISS index. On a laptop / consumer GPU a heavy embedding model can occupy 75–95% of total VRAM by itself — and once a chat model is also resident the combined footprint exceeds available memory and the daemon thrashes RAM↔VRAM swap on every embed batch. A 30-second context-load turns into a multi-hour stall.

The embedding-memory pre-flight guard (Tlamatini/agent/embedding_memory_guard.py

) catches this before the embed burst starts. It runs only when an NVIDIA GPU is detected; on CPU-only / AMD / Apple Silicon hosts it is a silent no-op and the legacy load path is unchanged.

The trigger case is the dev box this codebase is calibrated on: an NVIDIA GeForce RTX 4070 Laptop GPU with 8 188 MiB of VRAM. The previously-configured embedding model qwen3-embedding:8b

(7.6 B parameters, Q4_K_M quantization) sits at ~6.24 GB resident — 77.9% of total VRAM. Add the chat model and the daemon evicts something on every embed batch. The fix is either: switch to a smaller model (e.g. nomic-embed-text:v1.5

at ~0.60 GB resident, 7% saturation) or accept the slow path knowingly. The guard surfaces the choice before the heavy work starts, so you can abort, swap the model in config.json

, and restart Ollama — saving an hour of debugging "why is context frozen?".

The guard is wired into agent/consumers.py::setup_contextual_rag_chain

at exactly one point: after the " context" banner is broadcast to the chat, before the heavy asyncio.to_thread(setup_llm_with_context, …)

call that drives the embedding burst. The flow is:

WebSocket "set-directory-as-context"
    ↓
consumers.py:setup_contextual_rag_chain(path_only)
    ↓
broadcast MSG_AGENT__CONTEXT chat bubble
    ↓
► embedding_memory_guard.check_embedding_memory_for_directory(...)
    │
    ├─► returns None (no GPU / under threshold / probe failed)
    │       → proceed silently
    │
    └─► returns warning dict
            → broadcast HTML warning chat bubble
            → proceed anyway (informational, non-blocking)
    ↓
asyncio.to_thread(setup_llm_with_context, ...)
    └─► OllamaEmbeddings + FAISS.from_documents(...)   ← VRAM burst

The check runs inside asyncio.to_thread

so a slow nvidia-smi

or cold /api/show

call never blocks the Channels event loop. The whole block is wrapped in try/except Exception

so any unhandled probe error prints a one-line [EMBED-MEM] Pre-flight check skipped (fail-open)

to tlamatini.log

and the load continues — a diagnostic must never block the user.

The guard reuses the already-cached nvidia-smi -L

probe from agent/gpu_perf.py::_has_nvidia_gpu()

(introduced for the model-pinning hook). The probe runs at most once per process; subsequent calls hit the in-module cache. On CPU-only Linux/Windows, AMD GPUs, and Apple Silicon the probe returns False

once at server start and every subsequent call to the guard returns None immediately — no subprocesses spawned, no HTTP calls made, no overhead.

This is the portability guarantee: a fresh git pull

on a no-GPU box behaves exactly as before the guard existed. The 28 dedicated no-GPU compatibility tests (see 10.7) lock the contract in place.

When the GPU gate passes, the guard predicts the embedding model's resident VRAM in priority order:

Tier	Source	Trigger	Accuracy
A
GET /api/ps size_vram
Model already resident in Ollama	Exact — verbatim daemon value
B
POST /api/show → parameter_count × bits_per_weight × overhead
Model pulled but unloaded	±5% on calibrated models
C
(any probe failure)	Ollama down, model not pulled, cloud model (:cloud suffix)
Returns None → fail-open

Tier B uses a standard llama.cpp / GGUF bits-per-weight table:

Quant	Bits/weight	Quant	Bits/weight
F32
32.0	Q4_K_M
4.83
F16 / BF16
16.0	Q4_K_S
4.58
Q8_0
8.5	Q4_0
4.55
Q6_K
6.56	Q3_K_M
3.91
Q5_K_M
5.69	Q2_K
2.96

Unknown quants fall back to a conservative 5.0

bits/weight. The overhead multiplier accounts for KV cache + activation buffers + GGML allocator slack:

for models with× 1.40

≥ 1 B parameters(large-model regime)for× 2.20

sub-1 B models (proportionally larger KV/buffer overhead)

Calibration against live measurements on the RTX 4070 Laptop:

Model	Params × bits/8 (raw)	Predicted (× overhead)	Measured resident	Error
qwen3-embedding:8b (Q4_K_M)
4.54 GB	6.36 GB (× 1.40)
6.24 GB	+1.9%
Nomic-Embed-Text:latest (F16)
274 MB	603 MB (× 2.20)
600 MB	+0.5%

Tier B also pulls the embedding dimension from /api/show

via any architecture-prefixed *.embedding_length

key (e.g. qwen3.embedding_length=4096

, nomic-bert.embedding_length=768

). Combined with a directory pre-scan that mirrors the exclusion rules of agent/rag/factory.py::CustomText

, it reports a projected FAISS-index RAM size (num_chunks × embedding_dim × 4

bytes, float32). This is RAM, not VRAM, but useful to surface on directories with hundreds of thousands of chunks.

The guard fires when predicted_vram ≥ 0.80 × smallest-GPU total VRAM. Why the smallest GPU (rather than the sum or the largest)? Because Ollama loads each model into a single device — using the max would silently under-report the constraint on heterogeneous multi-GPU rigs.

When the threshold is crossed, the guard returns a structured dict the consumer renders as an HTML chat bubble. A real example from this dev box (artificially threshold-lowered to 70% so qwen3-embed:8b trips):

⚠️ Embedding-memory warning
Embedding model qwen3-embedding:8b needs ~6,378 MiB of VRAM
(currently resident in VRAM), which is 77.9% of the smallest
GPU's total (8,188 MiB) — above the safety threshold of 70%.
Projected FAISS vector store (RAM, not VRAM): ~28 MiB across
1,847 chunks at dim 4096.
Context  will continue, but expect slow embedding
throughput or RAM↔VRAM swap. To eliminate the pressure, switch
embeding-model in config.json to a smaller model
(e.g. nomic-embed-text:v1.5) and restart.

The message is informational and non-blocking — context proceeds. The user picks whether to wait it out, hit Cancel, or change models. The phrasing names the exact config.json

key (embeding-model

, with the spelling preserved from the existing codebase) and a concrete alternative.

Knob	Where	Default	When to change
Trigger threshold	check_embedding_memory_for_directory(..., threshold=)
0.80
Pass 0.70 on smaller GPUs (6 GB cards) where 80% is already too tight.
Large-model overhead	_OVERHEAD_LARGE constant
1.40
If a new model family proves the calibration off by > 10%, recalibrate against /api/ps and bump the constant.
Small-model overhead	_OVERHEAD_SMALL constant
2.20
Same calibration story for sub-1B models.
Bits/weight table	_QUANT_BITS dict
Q4_K_M=4.83 (etc.)
Add new entries when a future GGUF quant ships.

What the guard does NOT do, by deliberate choice:

• It does not abort context . The warning is informational. (If you want abort-on-warning behavior, wire a confirm/cancel WebSocket round-trip — the surface is described inagent_page_init.js

nearset-dir-context

.) - It does not estimate thechat model's VRAM. Only the embedding model is checked, because that is the model the directory-load path forces into VRAM. The chat model is handled bygpu_perf.pin_ollama_model

separately. - It does not persist warnings. Each context-load runs an independent check. - It does not callnvidia-smi

on CPU-only hosts. Both gates (_has_nvidia_gpu_cached

and the_gpu_total_memory_bytes

query) short-circuit before any subprocess spawn. - It does not add new dependencies.subprocess

,urllib.request

, andos.walk

are the only stdlib touchpoints — the same surfaceagent/gpu_perf.py

already uses.

The guard ships with 49 automated tests in Tlamatini/agent/test_embedding_memory_guard.py

, split into seven SimpleTestCase

classes:

Test class	Count	What it pins
QuantTableTests
2	Known quants resolve to standard bits/weight; unknown quants fall back to the conservative default.
PredictFromShowTests
3	Tier-B prediction lands within calibrated bounds for both 7.6 B and 137 M reference models.
EmbeddingDimExtractionTests
2	The dim key is found regardless of architecture prefix (qwen3. , nomic-bert. , future archs).
ChunkEstimatorTests
4	Directory walk honors default + user omissions, respects max_chunks_per_file , and handles single-file mode.
GuardEntryPointTests
8	All entry-point branches: no-GPU, cloud, threshold gate, Tier A /api/ps , Tier B /api/show , probe failure.
FormatMessageTests
2	HTML renders the model name, percent, threshold, and chunk count.
NoGpuCompatibilityTests
28
Every no-GPU failure mode — see breakdown below.

The NoGpuCompatibilityTests

class is the portability proof. Its coverage matrix:

Failure mode	Tests
Module import on no-GPU host has no side effects	test_module_imports_without_side_effects
nvidia-smi binary missing entirely
test_run_cmd_returns_127_for_real_missing_binary , test_total_vram_returns_none_when_nvidia_smi_missing
nvidia-smi exists but driver unloaded
test_total_vram_returns_none_when_driver_unloaded
nvidia-smi times out / crashes
test_run_cmd_absorbs_timeout / ..._permission_error / ..._generic_oserror
nvidia-smi returns empty or garbage output
test_total_vram_returns_none_on_empty_output / ..._on_unparseable_output
Heterogeneous multi-GPU rig	test_total_vram_picks_smallest_gpu_in_heterogeneous_rig
gpu_perf module missing / its probe raises
test_has_nvidia_gpu_falls_back_when_gpu_perf_unimportable , test_has_nvidia_gpu_returns_false_when_gpu_perf_probe_raises
Ollama daemon offline (closed port)	test_ollama_show_returns_none_against_closed_port , test_ollama_ps_returns_none_against_closed_port
Malformed Ollama URLs / empty args	test_ollama_show_returns_none_for_garbage_url
Model not in /api/ps
test_ollama_loaded_vram_returns_none_when_model_not_in_ps , ..._when_ps_fails
Entry on a CPU-only host	test_check_returns_none_on_cpu_only_host
GPU detected but --query-gpu fails
test_check_returns_none_when_nvidia_smi_query_fails
GPU detected but Ollama offline	test_check_returns_none_when_ollama_offline
Pathological 0 MiB GPU reading	test_check_returns_none_when_gpu_zero_total
Empty ollama_base_url in config
test_check_returns_none_for_empty_base_url
Deleted / nonexistent / empty path	3× test_chunk_estimator_* + test_check_with_nonexistent_path_does_not_crash
Unreadable file inside the walked tree	test_chunk_estimator_with_unreadable_file_skips_it
Partial warning dict (missing optional keys)	test_format_warning_message_handles_missing_optional_keys
Live portability proof (real subprocess + real urllib)
test_real_entry_point_call_never_raises

test_real_entry_point_call_never_raises

is the CI gate: it makes the actual subprocess.run(["nvidia-smi", ...])

and urlopen("http://127.0.0.1:11434/...")

calls against whatever the runner offers, and asserts the return is either None

or a well-formed warning dict — never an exception. The same test passes on this RTX 4070 dev box (returns None

because qwen3-embed sits at 77.9%, under the 80% gate) and on a CPU-only CI image (returns None

because the GPU gate fails fast).

Run them yourself:

cd Tlamatini
python manage.py test agent.test_embedding_memory_guard --verbosity=2

Tlamatini now ships a three-tier reaper (Tlamatini/agent/orphan_reaper.py

) that cleans up the console-host children every console subprocess on Windows drags behind it. Without this pass, users were occasionally seeing conhost.exe

processes lingering in Task Manager with the Tlamatini icon — the icon is inherited from the parent EXE that spawned the console — and reasonably concluding that Tlamatini was leaking processes.

On Windows, when a tool (execute_command

, chat_agent_executer

, an ACPX CLI child, an agent-pool Python subprocess, …) spawns a console child, Windows allocates a conhost.exe

companion to host that console. If the immediate parent dies before the OS reaps the console pair, that conhost.exe

outlives Tlamatini. Two compounding causes were closed at once:

The reaper itself sweeps zombies and orphaned console hosts at three lifecycle points (below).Spawn sites were hardened—views.py::execute_starter_agent_view

,execute_ender_agent_view

,restart_agent_view

,execute_flowcreator_view

, every ACPX child inacpx/runtime.py

, and asubprocess.Popen.__init__

guard at the top ofagents/ender/ender.py

(mirrored across every other pool agent) now spawn withCREATE_NO_WINDOW | DETACHED_PROCESS | CREATE_NEW_PROCESS_GROUP

and stdio piped toDEVNULL

. No console is allocated in the first place, so noconhost.exe

companion exists to orphan.

Tier	When it runs	Scope	Visibility
Tier 1
After every Multi-Turn tool call that may have spawned a child (execute_command , execute_file , unzip_file , decompile_java , googler , agent_starter/stopper/parametrizer , every chat_agent_* , every acp_* ). Driven by MultiTurnToolAgentExecutor._reap_after_tool() in agent/mcp_agent.py . Also fires on the tool-exception path so a crashed tool still gets cleaned up.
Dead/zombie descendants of the current PID, plus orphaned conhost.exe whose parent is gone. Pool-cmdline scan is skipped here (cheap path).
Silent. Survivors accumulate on the executor for Tier 2 to surface.
Tier 2
Once, right after the final answer is broadcast to the user. Driven by AgentConsumer._tier2_orphan_sweep() in agent/consumers.py . Runs in a thread (so it doesn't block the WebSocket loop) and merges its survivors with Tier 1's leftovers (de-duped by PID).
Same as Tier 1 plus the agent-pool cmdline scan (kills processes whose cmdline references agents/pools/... but are not tracked by AgentProcess / ChatAgentRun anymore).
If anything survives BOTH tiers, the consumer sends a second chat message listing every surviving name + PID pair so the user can end them manually from Task Manager.
Tier 3
At Tlamatini.exe shutdown — AgentConfig.ready() registers it on the same atexit / SIGINT / SIGBREAK path that already cleans up pools.
Full sweep (self-tree + pool cmdline + console-host orphans).	Logs --- [Tier-3 reaper] killed=… survivors=… errors=… to tlamatini.log . Survivors are listed by name (PID) so a post-mortem reader can audit what refused to die.

A process is considered a "Tlamatini orphan" if any of the following hold:

• It is a descendant of the current Tlamatini PID and its status is ZOMBIE

/DEAD

. - It is a conhost.exe

/openconsole.exe

whose parent PID is in our process tree, OR whose parent PID no longer exists. - Its cmdline

references the agent-pool directory (agents/pools/...

oragents/pools/_chat_runs_/...

) but it is no longer tracked.

Each candidate is escalated terminate → wait 1 s → kill

via psutil

; an "unable-to-kill" outcome surfaces as a survivor, never as an exception. The reaper never raises into the caller — a cleanup that crashes the chat path would be worse than the orphans it tries to kill.

Out of scope on purpose: console hosts spawned by unrelated processes (a different IDE, your shell, another app's child) — the parentage check keeps the sweep narrow.

When Tier 2 detects survivors, the user sees a second chat bubble immediately after the main answer:

⚠ Heads-up: Tlamatini tried to clean up after this request but the following
process(es) refused to terminate. They are most likely harmless leftovers from
a tool you ran, but if you do not recognize them please end them manually from
Task Manager so no Tlamatini-spawned child outlives the app:
  • conhost.exe — PID 18244
  • python.exe — PID 19108

The rendering helper is orphan_reaper.format_survivors_message()

; it returns None

(so no extra message is sent) when the survivor list is empty, which is the common case after the spawn-site hardening landed.

• "connection refused" → ollama serve

in a dedicated terminal. Checkollama_base_url

. - Model not found → ollama list

to see what's pulled. Pull the missing tag. - Remote Ollama → set ollama_token

for bearer auth.

• Set-Context shows no green banner → check file permissions, ensure files are text not binary.
• "Out of memory" during embedding → fallback mode kicks in; retrieval quality drops, files still accessible. Switch to a smaller embedding model. See chapter 10 — the embedding-memory pre-flight guard now warns you about this* before*the embed burst starts on GPU hosts. - Hit max_doc_chars

→ bump it. - Session says it was restored after a refresh, but the input stays disabled briefly → that is expected while the contextual RAG chain rebuilds. Wait for the ready state / spinner to clear before sending the next prompt.

• Did you tick Multi-Turn? Isenable_unified_agent: true

? - "Tool X is not available" → the planner did not bind X. Check [Planner._select]

console lines, add matching keywords to your prompt, or raisemax_selected_tools

. - 4096 iterations exhausted → likely a busy-poll loop. Use chat_agent_sleeper

/chat_agent_run_wait

instead.

source & further reading

github.com — original article