Networking & Ops

HTTPS / LAN access

Access ChatWalaʻau from other devices on your network (phones, tablets, PCs). HTTPS enables the browser Secure Context required for voice input and clipboard on non-localhost origins.

APP_HOST=0.0.0.0
APP_SSL_CERTFILE=.certs/cert.pem
APP_SSL_KEYFILE=.certs/key.pem

Setup with mkcert:

1. Install mkcert and run mkcert -install
2. Issue a certificate: mkcert -cert-file .certs/cert.pem -key-file .certs/key.pem <your-ip> localhost 127.0.0.1
3. Set the env vars above in .env
4. Allow the ports through your firewall (8000 production, 5173 dev)
5. Install the CA certificate (rootCA.pem) on each client device

Access from the LAN at https://<your-ip>:8000. When SSL is not configured, the server runs in HTTP mode as usual (no breaking change).

Corporate networks (TLS-intercepting proxy)

Behind a corporate proxy (Zscaler, Netskope, on-prem SSL middlebox), the first outbound HTTPS call can fail because Python's bundled certifi store does not know your in-house root CA:

httpx.ConnectError: [SSL: CERTIFICATE_VERIFY_FAILED]
certificate verify failed: self-signed certificate in certificate chain

The recommended remedy is the opt-in corp extras, which route TLS validation through the host OS certificate store:

pip install "chatwalaau[corp]"

No source change, no env change, and no behavior change for non-corporate environments. Alternatively, point Python at an explicit CA bundle:

export SSL_CERT_FILE=/path/to/corp-root-ca.pem
export REQUESTS_CA_BUNDLE=/path/to/corp-root-ca.pem

DevUI

Enable the Microsoft Agent Framework DevUI for debugging:

DEVUI_ENABLED=true
DEVUI_PORT=8080
# DEVUI_DISABLE_MCP=true   # excluded by default (cross-loop safety)
# DEVUI_DISABLE_RAG=true   # excluded by default

Access at http://localhost:8080. DevUI receives a dedicated Agent instance that reuses the main agent's function tools, skills, and model client but omits MCP tools and rag_search by default, avoiding cross-loop invocation between the DevUI daemon thread and the main FastAPI event loop.

Prompt dump (what the model actually received)

When an answer is wrong and the configuration looks right, the useful question is what the model was actually sent. Enable the prompt dump:

PROMPT_DUMP_ENABLED=true
# PROMPT_DUMP_DIR=.prompts   # default

Each run writes a timestamped Markdown file containing the assembled system prompt (first turn only -- it is frozen for the session), the flowing conversation, and the tool surface. Prompt content goes only to the file; the logs carry metadata.

In the system prompt, each tool's usage guidance is grouped into its own labeled <tool-guide name="weather">, <tool-guide name="rag">, ... block, so you can see at a glance which instructions belong to which tool.

Turn it off when you are done: the files accumulate and contain your conversations.

Tool surface

The tool surface section lists every built-in function, every MCP server and tool, and every Skill -- including the ones that are not there, with the reason:

### Built-in functions (4 active / 13 excluded)

- file_write                      ACTIVE    approval: always_require
- bash_execute                    EXCLUDED  allowlist
- rag_search                      EXCLUDED  settings (CHROMA_DIR unset or no embeddings offering)
- manage_cron                     EXCLUDED  settings (CRON_ENABLED=false)

### MCP (1 active server / 1 excluded / 1 warning)

- github                          ACTIVE    2/14 tools
    - create_issue                ACTIVE
    - delete_repo                 EXCLUDED  mcp override
- filesystem                      EXCLUDED  allowlist (server not selected)
- slack                           WARNING   not connected; allow-list not applied

Each reason points at the surface that owns the gate:

Reason	Where to fix it
settings	your .env -- the row names the variable or the missing model offering
allowlist	the Declarative Agent editor's tool selection
mcp override	the MCP Tool Manager
skills override	the Skills Manager
not connected	the MCP server is configured but not connected; Reload once it is up

A MISMATCH row means the built agent and your configuration disagree with no explanation -- the tool should have been there and is not, or the reverse. That is a defect, not a setting; capture the dump and report it.

A WARNING not connected row matters for a different reason: while a server is unconnected its tool names are unknown, so a per-tool allow-list cannot be applied and every tool of that server is exposed once it connects.

The section is written on every turn (unlike the system prompt) because tool state can change mid-session -- from the MCP manager, the Skills manager, or by switching Declarative Agent. The digest line shows at a glance which turn it changed on.