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

name = "research"
enabled = true
description = "investigate an open-ended question — how something works, why it happens, or what options exist — before making a change"

# 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 = "Research parallelizes well when there are several independent angles to check (different files, different search strategies, different hypotheses) — crew can fan work out across isolated worktrees and report back a synthesis, rather than one session serially reading everything. Prefer this when the question spans a large or unfamiliar area; for a narrow, already-scoped question, reading it yourself directly is faster."
progress_horizon_rounds = 6

[classifier]
# Deliberately excludes diagnose_failure's failure-specific phrasing
# ("investigate the failure", "research a failure") so an open-ended,
# non-failure research question doesn't collide with that workflow.
min_score = 0.3
keywords = [
  "research how",
  "research this codebase",
  "investigate how",
  "look into how",
  "find out how",
  "understand how",
  "explore the codebase for",
  "explore how",
  "what are the options for",
  "survey the",
  "survey how",
  "dig into how",
]
examples = [
  "Research how the caching layer decides what to evict.",
  "Investigate how these two subsystems communicate before we change either one.",
  "Look into what options exist for handling this without a rewrite.",
  "Understand how this feature currently works before proposing a change.",
  "Explore the codebase to find every place this pattern is used.",
]

[[steps]]
id = "scope_question"
title = "Pin down exactly what's being asked"
steer = "State precisely what a satisfying answer would contain before searching — a vague question invites a vague, unfalsifiable answer"

[[steps]]
id = "gather_evidence"
title = "Gather evidence from the real source"
steer = "Read the actual code/docs/history rather than pattern-matching from memory or training-data assumptions; cite file:line for every concrete claim"
tools = ["code_search", "read_file", "list_dir", "find", "git", "web_fetch"]

[[steps]]
id = "consider_parallel_delegation"
title = "Consider delegating breadth to crew"
steer = "If the question spans many files or several independent search angles, delegate to crew rather than reading everything serially yourself — each crew member gets its own isolated worktree and reports back, closer to a promise of specific commits/findings than a shared scratch read"
tools = ["crew"]

[[steps]]
id = "synthesize"
title = "Synthesize findings"
steer = "Separate what you directly confirmed from what you inferred; state which parts of the question remain genuinely open rather than papering over a gap with a plausible-sounding guess"

[[steps]]
id = "verify_uncertain_claims"
title = "Verify any load-bearing but uncertain claim"
steer = "If a conclusion the rest of the answer depends on is still uncertain, check it directly (read the code, run a small probe) rather than presenting an inference as a fact"
tools = ["read_file", "run_command"]