newt-core 0.7.1

Newt-Agent core types, errors, and the NeMoCode-style tier router
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67

name = "tdd_peer_review"
enabled = true
description = "test-driven, multi-agent implementation: red, green, refactor, then an independent peer review before it's done"

# Top-level WorkflowConfig fields — MUST stay before [classifier] opens a
# sub-table, or TOML parses them as unknown (silently ignored) keys of
# [classifier] instead of top-level fields.
delegate_hint = "Peer review is only a real check if it comes from a genuinely independent perspective — the same context that wrote the change is the worst judge of it. Delegate the review step to crew (a fresh sub-agent with no stake in the implementation, working from the diff alone) rather than only self-reviewing; for a design with several independent pieces, crew can also implement in parallel with each member owning one piece plus its own test."
progress_horizon_rounds = 6

[classifier]
min_score = 0.3
keywords = [
  "use tdd",
  "test-driven",
  "test driven",
  "red green refactor",
  "write a failing test first",
  "test-first",
  "build this test-first",
  "peer review this",
  "peer-review this",
  "get a second opinion on this implementation",
  "review this implementation",
]
examples = [
  "Use TDD for this: write a failing test first, then implement.",
  "Build this test-first — red, green, refactor — and then peer review it.",
  "Implement with tests, and get an independent review before calling it done.",
  "Write the test before the implementation, then have it peer reviewed.",
]

[[steps]]
id = "write_failing_test"
title = "Write (or confirm) a failing test"
steer = "Write a test that captures the actual requirement before any implementation exists — a test retrofitted after the code tends to just describe what the code already does, not what it should do"
tools = ["write_file", "edit_file"]

[[steps]]
id = "confirm_red"
title = "Confirm it fails for the RIGHT reason"
steer = "Run it and read the failure — a setup bug or typo failing for the wrong reason isn't red, it's broken; fix the test itself before treating this as a real starting point"
tools = ["run_command"]

[[steps]]
id = "implement_green"
title = "Implement the smallest change that makes it pass"
steer = "No extra scope beyond what the failing test requires — additional generality can come in the refactor step, once it's justified by a second case"
tools = ["edit_file", "write_file", "run_command"]

[[steps]]
id = "refactor"
title = "Refactor now that it's green"
steer = "Clean up naming/structure/duplication; re-run the test after every refactor step so it never goes red silently"
tools = ["edit_file", "run_command"]

[[steps]]
id = "peer_review"
title = "Get an independent peer review before calling it done"
steer = "Delegate to crew for a genuinely independent read of the diff, or if working solo, adversarially critique the change as if it were someone else's PR — the goal is a perspective that didn't write the code and has no reason to be charitable to it"
tools = ["crew", "git"]

[[steps]]
id = "address_findings"
title = "Address review findings, then re-verify"
steer = "Fix what the review surfaced; if the fix touches behavior, go back through confirm_red/implement_green for that change rather than treating the review pass as a rubber stamp"
tools = ["edit_file", "run_command"]