Agentic-Graph/graph-large.yaml at main · AdityaPandey172/Agentic-Graph · GitHub

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
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
# The pipeline for changes too large for one pass, on codebases too large to
# read whole.
#
#   context      map the repository and rank files against the goal (no model)
#   plan         break the goal into small steps, as plan.md with a json list
#   approval     a person approves the steps, and may edit or remove them
#   subtasks     each step runs graph-subtask.yaml: code, affected tests, fix
#   test         the full suite, once, over the whole change
#   review       the quality gate over the changed files only
#   advise       a second opinion on the whole diff
#
# The inner loop is fast because it runs only the tests a step can affect; the
# outer checks are exhaustive because they run once. On a legacy codebase the
# gate judges the change rather than years of existing debt.

version: 1
entry: context

defaults:
  agent: claude_code
  model: claude-opus-5
  timeout_seconds: 900

limits:
  fix:
    max_attempts: 3
    on_exceeded: give_up

nodes:
  context:
    impl: aeg.nodes.context
    kind: deterministic
    config:
      max_relevant_files: 25
      max_chars: 12000

  # The decomposing planner, under the name the approval gate and the output
  # nodes already read from.
  plan:
    impl: aeg.nodes.decompose
    kind: agentic
    model: claude-sonnet-5
    skills: [testing-standards]
    tools:
      allowed: [Read, Glob, Grep]
      disallowed: [Write, Edit, MultiEdit, Bash, WebFetch, WebSearch]
    config:
      max_subtasks: 8

  await_approval:
    impl: aeg.nodes.await_approval
    kind: deterministic

  subtasks:
    impl: aeg.nodes.subtasks
    kind: orchestration
    config:
      subtask_graph: graph-subtask.yaml
      max_subtasks: 8

  test:
    impl: aeg.nodes.test
    kind: deterministic
    timeout_seconds: 1800
    config:
      command: ["{python}", "-m", "pytest", "-q"]
      coverage_report: coverage.xml

  review:
    impl: aeg.nodes.review_sonar_cli
    kind: deterministic
    timeout_seconds: 1800
    config:
      scanner: tools/sonar-scanner-7.3.0.5189-windows-x64/bin/sonar-scanner.bat
      host_url: http://localhost:9000
      token_file: infra/.sonar-token
      project_key: "aeg-{run_id}"
      qualitygate_timeout: 600
      scope: changed

  advise:
    impl: aeg.nodes.advise
    kind: agentic
    model: claude-sonnet-5
    skills: [security-review]
    timeout_seconds: 600
    tools:
      allowed: [Read, Glob, Grep]
      disallowed: [Write, Edit, MultiEdit, Bash, WebFetch, WebSearch]
    config:
      advisory_only: true
      max_findings: 20

  # Between a failed test run and fix: rerun the failures once. Flaky tests
  # move on, broken environments park the run, real bugs go to fix.
  diagnose:
    impl: aeg.nodes.diagnose
    kind: deterministic
    timeout_seconds: 600

  fix:
    impl: aeg.nodes.fix
    kind: agentic
    skills: [python-conventions, testing-standards]
    tools:
      allowed: [Read, Write, Edit, MultiEdit, Glob, Grep]
      disallowed: [WebFetch, WebSearch]

  write:
    impl: aeg.nodes.write
    kind: deterministic

  give_up:
    impl: aeg.nodes.give_up
    kind: deterministic

  END:
    kind: terminal

edges:
  context:
    done: plan
  plan:
    done: await_approval
  await_approval:
    approved: subtasks
    rejected: give_up
  subtasks:
    done: test
    fail: give_up
  test:
    pass: review
    fail: diagnose
  diagnose:
    pass: review
    fail: fix
  review:
    pass: advise
    fail: fix
  advise:
    pass: write
    fail: fix
  fix:
    done: test
  write:
    done: END
  give_up:
    done: END