eyewright · the third-eye gate
01 · GATE THE LOOP STEP

Accept, reject, or review.

Stop hardcoding the decide step as a magic number. Ask a typed gate.
$ eyewright gate --question "Is the task done?" "Done: all requested changes are implemented, tests pass, PR merged." { "p_done": 0.968, "decision": "consensus-accept", "trace": { "primary": 0.6663, "consults": { "tev1:4b": 0.9938, "laya-multilingual": 0.9421 } } } "Adds a retry loop; tests pass; no test for the retry path." { "p_done": 0.197, "decision": "review-low" }
github.com/eyewright/eyewright
eyewright · the third-eye gate
02 · ANY TYPED DECISION

Yes / no. Choices. Scales.

One call for every decide step — not just a boolean.
$ eyewright ask --question "Which action?" --choices '{"optimize":"…","review":"…","ship":"…"}' { "kind": "choice", "answer": "optimize", "confidence": 0.94 }
github.com/eyewright/eyewright
eyewright · the third-eye gate
03 · ANY HARNESS

MCP tools, in-loop.

Any agent — OpenCode, Cursor, Claude — can call the gate.
eyewright_gate(state, question) → p_done + decision + trace eyewright_ask(state, question, choices?) → answer + confidence
stdio MCP
one line in the harness config
HTTP
/v1/gate · /v1/ask
CLI
gate · ask · serve · mcp
github.com/eyewright/eyewright
eyewright · the third-eye gate
04 · ONE CLICK

Or from the tray.

No terminal needed: click the icon, drop a file, get the same decision.
tray icon
green = "eyewright - ready"
Decide
drop a file · ask a question
result
decision + confidence
github.com/eyewright/eyewright
eyewright · the third-eye gate
05 · LOCAL-FIRST

Nothing leaves the machine.

A local specialist is the main eye; two more consult when unsure. All on localhost.
laya-serve :8000 ← main eye (typed-decisions) Ollama :11434 ← consult (tev1:4b) laya-serve :8000 ← consult (laya-multilingual)
A cloud eye is opt-in — and labeled. ZDR is routing, not encryption.
github.com/eyewright/eyewright
eyewright · the third-eye gate
06 · BENCH

Measured, not vibes.

23.7% acted · 90.8% accurate
GATEACTEDACCURACY
local default23.7%90.8%
+ local System-234.3%94.7%
cloud consult10.0%100%
600-decision bench · github.com/eyewright/eyewright