docs/modules/console.md

sAGI/console.py — the bankML console: bankML as itself, and what it measures

Summary

The console (0.3.7) is a small page with four tabs over bankml serve: Interaction (ask; the streamed answer and its receipt), Admin (CPU, RAM and GPU sliders and D3 charts of measured use), Logging (every exchange) and Infotags (iNFT publication metadata). It talks to bankML as itself — sAGI/personas/bankml.persona — and gives that persona a SELF block measured at each question, so what bankML says about its own tokens, speed, load and power is measurement, not invention.

Technical usage

python3 sAGI/console.py [--host 127.0.0.1] [--port 7875]   # loopback hosts only; bankml serve at $BANKML_SERVE_LISTEN (127.0.0.1:18093)

route	what
GET /, /app.js, /style.css, /vendor/d3.v7.min.js, /vendor/d3.LICENSE	the page and its assets (D3 v7.9.0, ISC, vendored)
GET /api/state	/bankml, /bankml/usage, /bankml/metrics, the saved resources, the persona and its doctrine root, the SELF block
POST /api/ask {"message", "history"?, "max_tokens"?, "temperature"?, "top_k"?, "top_p"?, "min_p"?, "repeat_penalty"?, "seed"?}	NDJSON: {"piece"} lines, then {"done", "error", "receipt", "metrics", "answer_sha256_ok"}; the exchange (question, answer, receipt, the engine's metrics record, the SELF block it was asked with) is appended to $BANKML_UI_STATE/console.jsonl
POST /api/resources {"threads", "ram_gb", "gpu_limit"}	one verified restart of the native engine with rollback (models.apply_resources)
GET /api/log	the last 200 exchanges and the engine log's tail
GET /api/infotags	ERC-721 metadata: name, description, attributes, and the RFC 6962 root and CIDs

/api/ask sends the persona's system prompt with the SELF block appended, the last 12 user and assistant turns of history, and the question to /v1/chat/completions (streamed, max_tokens 256 unless set). The SELF block is rendered for the model as one sentence per measurement with its unit (self_text), "not measured" for a null — a bare JSON key (last_eval_tps) was read as seconds in testing.

How it is verified

testing/test_console.py (18 checks, in the gate, no engine needed): every asset served; the CSP; unknown paths 404; a foreign Host 403; a cross-origin POST 403; a non-JSON POST 400; an empty question 400; SELF says "not measured" without an engine and names units with one; the Infotags root equals savante.merkle_root over the log's exact lines. Live on Bonsai-1.7B: "I am bankML…", its token totals and generation speed read correctly from SELF, and every receipt's sha256 matches the streamed text.

Advantages and efficiency

• No framework, no network. Python's standard library and one vendored D3; CSP default-src 'none' with 'self'.
• Safe controls. The sliders restart the engine through the importer's verified switch with rollback; the console refuses any host but loopback, because it can restart the engine.
• Proof without disclosure. The exchanges stay on disk; Infotags publishes their Merkle root and CID, as Savante's .history commitments do.
• The persona's prompt is cached. The persona comes first and the SELF block last, so after the first question the engine reuses the persona's prefix (first answer 37 s, then 16 s to the first token on Bonsai-1.7B).

Limitations

• The persona's system prompt is about 550 tokens: the first question of a session pays for it on the CPU.
• It forwards only temperature, top_k, top_p, min_p, repeat_penalty and seed; the rest of the 0.3.7 sampler chain and logprobs are reachable through /v1 directly, not from the page. One conversation at a time (bankML's single slot; requests sent together are answered in turn).
• The page has no DeltaVerse substrate background yet.

See also

metrics.md · gpu.md · sys.md · ../usage.md §6c · ../install.md