newt-core 0.7.1

Newt-Agent core types, errors, and the NeMoCode-style tier router
Documentation
name = "fix_failing_tests"
enabled = true
description = "run the test suite, fix failures one at a time, and loop until every test passes"

# 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 = "This is a mechanical fix-and-reverify loop. If many tests are failing across unrelated areas, or the fixes are likely to take many iterations, consider calling crew or team to delegate the loop to a sub-agent with its own round budget and an isolated worktree, instead of spending this session's rounds cycling through it yourself."
progress_horizon_rounds = 6

[classifier]
# Higher than the 0.24 convention elsewhere: this workflow's phrasing shares
# common connector words ("fix", "the", "and") with `diagnose_failure`'s own
# test-related examples, and the default threshold let a near-miss steal that
# workflow's match (found while writing this file's own tests). A tighter bar
# means only a genuinely close "run tests in a loop until green" phrasing
# wins here; ambiguous test-adjacent phrasing still falls through to
# `diagnose_failure`'s broader, deliberately looser match.
min_score = 0.4
keywords = [
  "fix all failing tests",
  "fix the failing tests",
  "fix all the failing tests",
  "make the tests pass",
  "make all tests pass",
  "make the test suite pass",
  "get the tests passing",
  "get all tests passing",
  "run the tests until they pass",
  "loop until the tests pass",
  "keep running the tests until",
  "test suite is red",
  "make the test suite green",
  "iterate until all tests pass",
  "fix tests until green",
  "until every test passes",
]
examples = [
  "Run the test suite and fix any failing tests, looping until everything passes.",
  "Keep iterating on the tests until the whole suite is green.",
  "Fix all the failing tests one at a time and re-run until none fail.",
  "The test suite is red, fix it so every test passes.",
  "Make the tests pass: run them, fix what's broken, repeat until green.",
]

# This is a LOOP, not a linear checklist: `loop_or_stop` explicitly sends the
# model back to `fix_one` while any test still fails. Workflow steps have no
# native loop primitive (they are quoted to the model via update_plan as an
# ordered list) — the loop is expressed in `loop_or_stop`'s steer text, which
# is the mechanism this schema actually has for it.
[[steps]]
id = "run_suite"
title = "Run the full test suite for ground truth"
steer = "Run the project's actual test command — find or ask for the real one, never assume — and record EVERY failing test, not just the first"
tools = ["run_command", "read_file"]

[[steps]]
id = "triage"
title = "Order the failures"
steer = "Fix whichever failure is most likely blocking others first (e.g. a compile/import error before an assertion failure); group failures that share one root cause so they're fixed together, not as separate cycles"
tools = ["read_file", "code_search"]

[[steps]]
id = "fix_one"
title = "Fix ONE failure (or one root cause shared by several)"
steer = "Make the smallest coherent change that addresses the root cause; do not touch unrelated failing tests in this step — one cause at a time keeps each cycle verifiable"
tools = ["read_file", "edit_file", "write_file"]

[[steps]]
id = "reverify_suite"
title = "Re-run the FULL suite, not just the test(s) you touched"
steer = "A fix can regress a different test; only the full suite proves nothing else broke — re-running a single test is not enough"
tools = ["run_command"]

[[steps]]
id = "loop_or_stop"
title = "Loop back to fix_one, or stop"
steer = "If any test still fails, return to fix_one for the next failure and repeat run_suite/reverify_suite as needed — do not stop early or hand back a partial summary while failures remain. Stop only when the full suite passes, or when a specific remaining failure is genuinely out of scope or unfixable here — in that case state EXACTLY which test and why, rather than looping on it indefinitely"
tools = ["update_plan"]