kubesift 2.0.0__tar.gz → 2.0.1__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (125) hide show
  1. {kubesift-2.0.0 → kubesift-2.0.1}/PKG-INFO +1 -1
  2. {kubesift-2.0.0 → kubesift-2.0.1}/deploy/helm/kubesift/Chart.yaml +2 -2
  3. {kubesift-2.0.0 → kubesift-2.0.1}/deploy/helm/kubesift/templates/deployment.yaml +4 -0
  4. {kubesift-2.0.0 → kubesift-2.0.1}/deploy/helm/kubesift/values.yaml +2 -1
  5. {kubesift-2.0.0 → kubesift-2.0.1}/docs/configuration.md +10 -1
  6. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/__init__.py +1 -1
  7. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/analysis/llm.py +104 -6
  8. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/cli.py +46 -13
  9. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/config.py +9 -0
  10. kubesift-2.0.1/tests/test_llm.py +192 -0
  11. kubesift-2.0.0/tests/test_llm.py +0 -103
  12. {kubesift-2.0.0 → kubesift-2.0.1}/.dockerignore +0 -0
  13. {kubesift-2.0.0 → kubesift-2.0.1}/.gitattributes +0 -0
  14. {kubesift-2.0.0 → kubesift-2.0.1}/.github/workflows/ci.yml +0 -0
  15. {kubesift-2.0.0 → kubesift-2.0.1}/.github/workflows/pages.yml +0 -0
  16. {kubesift-2.0.0 → kubesift-2.0.1}/.github/workflows/release.yml +0 -0
  17. {kubesift-2.0.0 → kubesift-2.0.1}/.gitignore +0 -0
  18. {kubesift-2.0.0 → kubesift-2.0.1}/.pre-commit-hooks.yaml +0 -0
  19. {kubesift-2.0.0 → kubesift-2.0.1}/CONTRIBUTING.md +0 -0
  20. {kubesift-2.0.0 → kubesift-2.0.1}/Dockerfile +0 -0
  21. {kubesift-2.0.0 → kubesift-2.0.1}/LICENSE +0 -0
  22. {kubesift-2.0.0 → kubesift-2.0.1}/README.md +0 -0
  23. {kubesift-2.0.0 → kubesift-2.0.1}/SECURITY.md +0 -0
  24. {kubesift-2.0.0 → kubesift-2.0.1}/action.yml +0 -0
  25. {kubesift-2.0.0 → kubesift-2.0.1}/deploy/helm/kubesift/templates/NOTES.txt +0 -0
  26. {kubesift-2.0.0 → kubesift-2.0.1}/deploy/helm/kubesift/templates/_helpers.tpl +0 -0
  27. {kubesift-2.0.0 → kubesift-2.0.1}/deploy/helm/kubesift/templates/rbac.yaml +0 -0
  28. {kubesift-2.0.0 → kubesift-2.0.1}/deploy/helm/kubesift/templates/secret.yaml +0 -0
  29. {kubesift-2.0.0 → kubesift-2.0.1}/deploy/helm/kubesift/templates/service.yaml +0 -0
  30. {kubesift-2.0.0 → kubesift-2.0.1}/docs/architecture.md +0 -0
  31. {kubesift-2.0.0 → kubesift-2.0.1}/docs/checks.md +0 -0
  32. {kubesift-2.0.0 → kubesift-2.0.1}/docs/mcp.md +0 -0
  33. {kubesift-2.0.0 → kubesift-2.0.1}/docs/preflight.md +0 -0
  34. {kubesift-2.0.0 → kubesift-2.0.1}/docs/production.md +0 -0
  35. {kubesift-2.0.0 → kubesift-2.0.1}/docs/releasing.md +0 -0
  36. {kubesift-2.0.0 → kubesift-2.0.1}/docs/slack.md +0 -0
  37. {kubesift-2.0.0 → kubesift-2.0.1}/pyproject.toml +0 -0
  38. {kubesift-2.0.0 → kubesift-2.0.1}/scripts/e2e.sh +0 -0
  39. {kubesift-2.0.0 → kubesift-2.0.1}/scripts/e2e_check.py +0 -0
  40. {kubesift-2.0.0 → kubesift-2.0.1}/scripts/gen_check_docs.py +0 -0
  41. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/analysis/__init__.py +0 -0
  42. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/analysis/changes.py +0 -0
  43. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/analysis/correlate.py +0 -0
  44. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/analysis/describe.py +0 -0
  45. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/analysis/engine.py +0 -0
  46. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/checks/__init__.py +0 -0
  47. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/checks/base.py +0 -0
  48. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/checks/config_refs.py +0 -0
  49. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/checks/events.py +0 -0
  50. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/checks/hygiene.py +0 -0
  51. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/checks/lifecycle.py +0 -0
  52. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/checks/network.py +0 -0
  53. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/checks/nodes.py +0 -0
  54. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/checks/operations.py +0 -0
  55. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/checks/pods.py +0 -0
  56. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/checks/storage.py +0 -0
  57. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/checks/workloads.py +0 -0
  58. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/cluster/__init__.py +0 -0
  59. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/cluster/client.py +0 -0
  60. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/cluster/snapshot.py +0 -0
  61. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/cost.py +0 -0
  62. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drill_scenarios/01-crashloop-missing-env.yaml +0 -0
  63. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drill_scenarios/02-imagepull-wrong-tag.yaml +0 -0
  64. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drill_scenarios/03-oomkilled-low-memory.yaml +0 -0
  65. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drill_scenarios/04-service-selector-mismatch.yaml +0 -0
  66. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drill_scenarios/05-missing-probe-crashloop.yaml +0 -0
  67. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drill_scenarios/06-pending-insufficient-resources.yaml +0 -0
  68. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drill_scenarios/07-missing-secret.yaml +0 -0
  69. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drill_scenarios/08-pvc-missing-storageclass.yaml +0 -0
  70. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drill_scenarios/09-netpol-blocks-dns.yaml +0 -0
  71. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drill_scenarios/10a-rollout-good.yaml +0 -0
  72. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drill_scenarios/10b-rollout-bad.yaml +0 -0
  73. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drill_scenarios/__init__.py +0 -0
  74. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/drills.py +0 -0
  75. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/knowledge.py +0 -0
  76. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/leader.py +0 -0
  77. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/lint.py +0 -0
  78. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/mcp_server.py +0 -0
  79. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/models.py +0 -0
  80. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/notify.py +0 -0
  81. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/postmortem.py +0 -0
  82. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/preflight.py +0 -0
  83. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/redact.py +0 -0
  84. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/remediate.py +0 -0
  85. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/report.py +0 -0
  86. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/runbooks.py +0 -0
  87. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/server/__init__.py +0 -0
  88. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/server/app.py +0 -0
  89. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/server/static/index.html +0 -0
  90. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/slack.py +0 -0
  91. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/store.py +0 -0
  92. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/timetravel.py +0 -0
  93. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/upgrade.py +0 -0
  94. {kubesift-2.0.0 → kubesift-2.0.1}/src/kubesift/util.py +0 -0
  95. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/01-crashloop-missing-env.yaml +0 -0
  96. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/02-imagepull-wrong-tag.yaml +0 -0
  97. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/03-oomkilled-low-memory.yaml +0 -0
  98. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/04-service-selector-mismatch.yaml +0 -0
  99. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/05-missing-probe-crashloop.yaml +0 -0
  100. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/06-pending-insufficient-resources.yaml +0 -0
  101. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/07-missing-secret.yaml +0 -0
  102. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/08-pvc-missing-storageclass.yaml +0 -0
  103. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/09-netpol-blocks-dns.yaml +0 -0
  104. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/10a-rollout-good.yaml +0 -0
  105. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/10b-rollout-bad.yaml +0 -0
  106. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/apply-scenario.sh +0 -0
  107. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/cleanup.sh +0 -0
  108. {kubesift-2.0.0 → kubesift-2.0.1}/test-scenarios/expected.json +0 -0
  109. {kubesift-2.0.0 → kubesift-2.0.1}/tests/conftest.py +0 -0
  110. {kubesift-2.0.0 → kubesift-2.0.1}/tests/fixtures.py +0 -0
  111. {kubesift-2.0.0 → kubesift-2.0.1}/tests/test_analysis.py +0 -0
  112. {kubesift-2.0.0 → kubesift-2.0.1}/tests/test_checks.py +0 -0
  113. {kubesift-2.0.0 → kubesift-2.0.1}/tests/test_client.py +0 -0
  114. {kubesift-2.0.0 → kubesift-2.0.1}/tests/test_features.py +0 -0
  115. {kubesift-2.0.0 → kubesift-2.0.1}/tests/test_ha.py +0 -0
  116. {kubesift-2.0.0 → kubesift-2.0.1}/tests/test_hardening.py +0 -0
  117. {kubesift-2.0.0 → kubesift-2.0.1}/tests/test_interfaces.py +0 -0
  118. {kubesift-2.0.0 → kubesift-2.0.1}/tests/test_lint_redact.py +0 -0
  119. {kubesift-2.0.0 → kubesift-2.0.1}/tests/test_more_checks.py +0 -0
  120. {kubesift-2.0.0 → kubesift-2.0.1}/tests/test_production.py +0 -0
  121. {kubesift-2.0.0 → kubesift-2.0.1}/website/README.md +0 -0
  122. {kubesift-2.0.0 → kubesift-2.0.1}/website/favicon.svg +0 -0
  123. {kubesift-2.0.0 → kubesift-2.0.1}/website/index.html +0 -0
  124. {kubesift-2.0.0 → kubesift-2.0.1}/website/main.js +0 -0
  125. {kubesift-2.0.0 → kubesift-2.0.1}/website/styles.css +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.5
2
2
  Name: kubesift
3
- Version: 2.0.0
3
+ Version: 2.0.1
4
4
  Summary: Kubernetes troubleshooting that explains itself: evidence-based diagnosis, change correlation, cost and upgrade checks. CLI, kubectl plugin, MCP server and in-cluster service.
5
5
  Project-URL: Homepage, https://github.com/OpeyemiAdeniji/KubeSift
6
6
  Project-URL: Issues, https://github.com/OpeyemiAdeniji/KubeSift/issues
@@ -2,7 +2,7 @@ apiVersion: v2
2
2
  name: kubesift
3
3
  description: Evidence-based Kubernetes troubleshooting service (dashboard, API, Alertmanager receiver, background scans)
4
4
  type: application
5
- version: 2.0.0
6
- appVersion: "2.0.0"
5
+ version: 2.0.1
6
+ appVersion: "2.0.1"
7
7
  home: https://github.com/OpeyemiAdeniji/KubeSift
8
8
  keywords: [kubernetes, troubleshooting, sre, observability]
@@ -60,6 +60,10 @@ spec:
60
60
  - name: KUBESIFT_LLM_BASE_URL
61
61
  value: {{ . | quote }}
62
62
  {{- end }}
63
+ {{- with .Values.llm.timeout }}
64
+ - name: KUBESIFT_LLM_TIMEOUT
65
+ value: {{ . | quote }}
66
+ {{- end }}
63
67
  - name: KUBESIFT_SEND_LOGS_TO_LLM
64
68
  value: {{ .Values.llm.sendLogs | quote }}
65
69
  {{- with .Values.prometheus.url }}
@@ -23,7 +23,8 @@ existingSecret: ""
23
23
  llm:
24
24
  provider: none # anthropic | openai | openrouter | ollama | none
25
25
  model: ""
26
- baseUrl: "" # e.g. http://ollama.ollama:11434/v1
26
+ baseUrl: "" # e.g. http://ollama.ollama:11434 for Ollama
27
+ timeout: 0 # seconds to wait for one answer; 0 = default (180, Ollama 900)
27
28
  anthropicApiKey: ""
28
29
  openaiApiKey: ""
29
30
  openrouterApiKey: ""
@@ -35,9 +35,11 @@ KubeSift never needs an LLM: every finding comes from deterministic checks. An L
35
35
  |---|---|---|
36
36
  | `llm_provider` | `auto` | `auto` picks the first of `ANTHROPIC_API_KEY`, `OPENROUTER_API_KEY`, `OPENAI_API_KEY` that is set, else `none` |
37
37
  | `llm_model` | per provider | Anthropic: `claude-opus-5-5` |
38
- | `llm_base_url` | per provider | Any OpenAI-compatible endpoint (vLLM, LiteLLM, Azure gateway) |
38
+ | `llm_base_url` | per provider | Any OpenAI-compatible endpoint (vLLM, LiteLLM, Azure gateway), or your Ollama server |
39
39
  | `llm_api_key` | from provider env var | |
40
40
  | `llm_effort` | `medium` | Anthropic effort: `low`, `medium`, `high` |
41
+ | `llm_timeout` | `180` (`900` for Ollama) | Seconds to wait for one answer. A slow answer is not retried |
42
+ | `llm_context` | sized to the prompt | Ollama only: context window in tokens (8192–32768 by default) |
41
43
  | `send_logs_to_llm` | `true` | `false` keeps log lines out of prompts |
42
44
 
43
45
  The Anthropic provider sends requests with server-side refusal fallback enabled (`fallbacks: "default"`), so a
@@ -51,6 +53,13 @@ addresses in log lines are redacted before anything is stored or sent. To keep e
51
53
  export KUBESIFT_LLM_PROVIDER=ollama KUBESIFT_LLM_MODEL=llama3.1
52
54
  ```
53
55
 
56
+ KubeSift talks to Ollama's native API: it turns off hidden reasoning ("thinking") so answers come faster, sizes the
57
+ context window to the prompt (Ollama's default window is small and silently drops the start of a long prompt), and
58
+ constrains the output to the expected JSON. A diagnosis prompt is a few thousand tokens. On a machine with a GPU a
59
+ 7–8B model answers in seconds; on a laptop CPU it can take many minutes, so raise `llm_timeout`, pick a smaller
60
+ model, use one of Ollama's cloud models, or run with `--no-llm`. If the model does not answer in time KubeSift
61
+ shows the rule-based diagnosis instead.
62
+
54
63
  ## Usage data and cost
55
64
 
56
65
  | Name | Default | Notes |
@@ -1,3 +1,3 @@
1
1
  """KubeSift: evidence-based Kubernetes troubleshooting."""
2
2
 
3
- __version__ = "2.0.0"
3
+ __version__ = "2.0.1"
@@ -2,7 +2,9 @@
2
2
 
3
3
  Providers:
4
4
  anthropic Claude through the official Anthropic SDK (pip install 'kubesift[anthropic]')
5
- openai any OpenAI-compatible endpoint: OpenAI, OpenRouter, Ollama, vLLM, LiteLLM, Azure gateways
5
+ openai any OpenAI-compatible endpoint: OpenAI, OpenRouter, vLLM, LiteLLM, Azure gateways
6
+ openrouter OpenRouter (OpenAI-compatible)
7
+ ollama local models through Ollama's native API
6
8
  none rules only
7
9
  """
8
10
 
@@ -40,7 +42,8 @@ class AnthropicProvider(Provider):
40
42
  except ImportError as exc:
41
43
  raise LLMError("Install the Anthropic SDK: pip install 'kubesift[anthropic]'") from exc
42
44
  self._anthropic = anthropic
43
- kwargs: dict[str, Any] = {"timeout": 180.0}
45
+ self.timeout = settings.resolved_llm_timeout()
46
+ kwargs: dict[str, Any] = {"timeout": self.timeout}
44
47
  if settings.llm_api_key:
45
48
  kwargs["api_key"] = settings.llm_api_key
46
49
  if settings.llm_base_url:
@@ -68,8 +71,12 @@ class AnthropicProvider(Provider):
68
71
  response = self.client.messages.create(**request)
69
72
  except a.APIStatusError as exc:
70
73
  raise LLMError(f"Anthropic API error {exc.status_code}: {exc.message}") from exc
74
+ except a.APITimeoutError as exc:
75
+ raise LLMError(_timeout_message(self.model, self.timeout)) from exc
71
76
  except a.APIConnectionError as exc:
72
77
  raise LLMError(f"Cannot reach the Anthropic API: {exc}") from exc
78
+ except a.APITimeoutError as exc:
79
+ raise LLMError(_timeout_message(self.model, self.timeout)) from exc
73
80
  except a.AuthenticationError as exc:
74
81
  raise LLMError("Anthropic rejected the API key (set ANTHROPIC_API_KEY).") from exc
75
82
  except a.RateLimitError as exc:
@@ -84,6 +91,15 @@ class AnthropicProvider(Provider):
84
91
  return parse_json(text)
85
92
 
86
93
 
94
+ def _timeout_message(model: str, seconds: float) -> str:
95
+ return (f"{model} did not answer within {seconds:.0f}s. Local models on a CPU can take many minutes: "
96
+ "raise llm_timeout (KUBESIFT_LLM_TIMEOUT), use a smaller or hosted model, or run with --no-llm.")
97
+
98
+
99
+ def _with_schema(system: str, schema: dict[str, Any]) -> str:
100
+ return system + "\n\nReply with one JSON object matching this schema:\n" + json.dumps(schema)
101
+
102
+
87
103
  class OpenAICompatibleProvider(Provider):
88
104
  name = "openai"
89
105
 
@@ -92,6 +108,7 @@ class OpenAICompatibleProvider(Provider):
92
108
  self.model = settings.resolved_model()
93
109
  self.api_key = settings.resolved_api_key()
94
110
  self.provider = settings.resolved_provider()
111
+ self.timeout = settings.resolved_llm_timeout()
95
112
  if self.provider != "ollama" and not self.api_key and "localhost" not in self.base_url:
96
113
  raise LLMError(f"No API key for {self.provider}; set {self.provider.upper()}_API_KEY or KUBESIFT_LLM_API_KEY.")
97
114
 
@@ -104,16 +121,20 @@ class OpenAICompatibleProvider(Provider):
104
121
  body: dict[str, Any] = {
105
122
  "model": self.model,
106
123
  "messages": [
107
- {"role": "system", "content": system + "\n\nReply with one JSON object matching this schema:\n"
108
- + json.dumps(schema)},
124
+ {"role": "system", "content": _with_schema(system, schema)},
109
125
  {"role": "user", "content": user},
110
126
  ],
111
127
  "response_format": {"type": "json_object"},
112
128
  }
113
- with httpx.Client(timeout=180.0) as http:
129
+ with httpx.Client(timeout=httpx.Timeout(self.timeout, connect=15.0)) as http:
114
130
  for attempt in range(3):
115
131
  try:
116
132
  r = http.post(f"{self.base_url}/chat/completions", headers=headers, json=body)
133
+ except httpx.TimeoutException as exc:
134
+ if isinstance(exc, httpx.ConnectTimeout) and attempt < 2:
135
+ continue
136
+ # a slow answer is not retried: the next attempt would redo the same slow work
137
+ raise LLMError(_timeout_message(self.model, self.timeout)) from exc
117
138
  except httpx.HTTPError as exc:
118
139
  if attempt == 2:
119
140
  raise LLMError(f"Cannot reach {self.base_url}: {exc}") from exc
@@ -130,6 +151,81 @@ class OpenAICompatibleProvider(Provider):
130
151
  raise LLMError("LLM request failed after retries")
131
152
 
132
153
 
154
+ class OllamaProvider(Provider):
155
+ """Ollama's native API: turns off hidden reasoning, sizes the context window to the prompt (Ollama's
156
+ default window is small and silently drops the start of longer prompts) and constrains output to the schema."""
157
+
158
+ name = "ollama"
159
+ MAX_OUTPUT = 4096
160
+
161
+ def __init__(self, settings: Settings):
162
+ base = (settings.resolved_base_url() or "http://localhost:11434").rstrip("/")
163
+ self.base_url = base[:-3] if base.endswith("/v1") else base
164
+ self.model = settings.resolved_model()
165
+ self.timeout = settings.resolved_llm_timeout()
166
+ self.context = settings.llm_context
167
+ self.api_key = settings.resolved_api_key()
168
+
169
+ def status(self) -> tuple[bool, str]:
170
+ """Is the server up and the model pulled? Used by `kubesift doctor`."""
171
+ headers = {"Authorization": f"Bearer {self.api_key}"} if self.api_key else {}
172
+ try:
173
+ r = httpx.get(f"{self.base_url}/api/tags", headers=headers, timeout=5.0)
174
+ r.raise_for_status()
175
+ names = {m.get("name") for m in r.json().get("models") or []}
176
+ except (httpx.HTTPError, ValueError) as exc:
177
+ return False, f"cannot reach Ollama at {self.base_url} ({exc.__class__.__name__}); is it running? (ollama serve)"
178
+ wanted = self.model if ":" in self.model else f"{self.model}:latest"
179
+ if wanted not in names and self.model not in names:
180
+ return False, f"Ollama is running but has no model {self.model!r}; run: ollama pull {self.model}"
181
+ return True, f"Ollama at {self.base_url} has {self.model}"
182
+
183
+ def context_size(self, text_chars: int) -> int:
184
+ if self.context:
185
+ return self.context
186
+ needed = text_chars // 3 + self.MAX_OUTPUT # ~3 characters per token for JSON-heavy text
187
+ return min(max(8192, -(-needed // 2048) * 2048), 32768)
188
+
189
+ def complete_json(self, system: str, user: str, schema: dict[str, Any]) -> dict[str, Any]:
190
+ system = _with_schema(system, schema)
191
+ body: dict[str, Any] = {
192
+ "model": self.model,
193
+ "stream": False,
194
+ "think": False,
195
+ "format": schema,
196
+ "options": {"num_ctx": self.context_size(len(system) + len(user)), "num_predict": self.MAX_OUTPUT,
197
+ "temperature": 0.2},
198
+ "messages": [{"role": "system", "content": system}, {"role": "user", "content": user}],
199
+ }
200
+ url = f"{self.base_url}/api/chat"
201
+ headers = {"Authorization": f"Bearer {self.api_key}"} if self.api_key else {}
202
+ with httpx.Client(timeout=httpx.Timeout(self.timeout, connect=10.0)) as http:
203
+ for _ in range(3):
204
+ try:
205
+ r = http.post(url, json=body, headers=headers)
206
+ except httpx.ConnectError as exc:
207
+ raise LLMError(f"Cannot reach Ollama at {self.base_url}; is it running? (ollama serve)") from exc
208
+ except httpx.TimeoutException as exc:
209
+ raise LLMError(_timeout_message(self.model, self.timeout)) from exc
210
+ except httpx.HTTPError as exc:
211
+ raise LLMError(f"Ollama request failed: {exc}") from exc
212
+ if r.status_code == 404 and "not found" in r.text.lower():
213
+ raise LLMError(f"Ollama has no model {self.model!r}; run: ollama pull {self.model}")
214
+ if r.status_code == 400 and "think" in body and "think" in r.text.lower():
215
+ body.pop("think") # older Ollama or a model without a thinking switch
216
+ continue
217
+ if r.status_code == 400 and isinstance(body.get("format"), dict):
218
+ body["format"] = "json" # older Ollama without schema-constrained output
219
+ continue
220
+ if r.status_code >= 400:
221
+ raise LLMError(f"Ollama returned HTTP {r.status_code}: {r.text[:300]}")
222
+ data = r.json()
223
+ if data.get("done_reason") == "length":
224
+ raise LLMError("The model ran out of output tokens before finishing its answer.")
225
+ return parse_json((data.get("message") or {}).get("content") or "")
226
+ raise LLMError("Ollama rejected the request.")
227
+
228
+
133
229
  def parse_json(text: str) -> dict[str, Any]:
134
230
  text = re.sub(r"```(?:json)?", "", text).strip()
135
231
  start, end = text.find("{"), text.rfind("}")
@@ -147,6 +243,8 @@ def get_provider(settings: Settings) -> Provider | None:
147
243
  return None
148
244
  if name == "anthropic":
149
245
  return AnthropicProvider(settings)
150
- if name in ("openai", "openrouter", "ollama"):
246
+ if name == "ollama":
247
+ return OllamaProvider(settings)
248
+ if name in ("openai", "openrouter"):
151
249
  return OpenAICompatibleProvider(settings)
152
250
  raise LLMError(f"Unknown LLM provider {name!r}; use anthropic, openai, openrouter, ollama or none.")
@@ -39,6 +39,14 @@ class State:
39
39
  state = State()
40
40
 
41
41
 
42
+ def _print_version(value: bool) -> None:
43
+ if value:
44
+ from kubesift import __version__
45
+
46
+ print(f"kubesift {__version__}")
47
+ raise typer.Exit()
48
+
49
+
42
50
  @app.callback()
43
51
  def main_options(
44
52
  kubeconfig: str | None = typer.Option(None, "--kubeconfig", envvar="KUBECONFIG_PATH", help="Path to kubeconfig."),
@@ -46,6 +54,8 @@ def main_options(
46
54
  llm: str | None = typer.Option(None, "--llm-provider", help="anthropic | openai | openrouter | ollama | none"),
47
55
  model: str | None = typer.Option(None, "--model", help="LLM model name."),
48
56
  prometheus: str | None = typer.Option(None, "--prometheus", help="Prometheus URL for usage history."),
57
+ version: bool = typer.Option(False, "--version", "-V", is_eager=True, callback=_print_version,
58
+ help="Show the version and exit."),
49
59
  ) -> None:
50
60
  state.settings = load_settings(kubeconfig=kubeconfig, context=context, llm_provider=llm, llm_model=model,
51
61
  prometheus_url=prometheus)
@@ -56,12 +66,32 @@ def _fail(msg: str, code: int = 1) -> NoReturn:
56
66
  raise typer.Exit(code)
57
67
 
58
68
 
59
- def _snapshot(namespace: str | None, from_file: str | None = None, with_logs: bool = True):
69
+ def _load_snapshot_file(path: str):
60
70
  from kubesift.cluster.snapshot import Snapshot
61
71
 
72
+ try:
73
+ data = json.loads(Path(path).read_text(encoding="utf-8"))
74
+ except FileNotFoundError:
75
+ _fail(f"snapshot file not found: {path}")
76
+ except (OSError, UnicodeDecodeError) as exc:
77
+ _fail(f"cannot read {path}: {exc}")
78
+ except json.JSONDecodeError as exc:
79
+ _fail(f"{path} is not a KubeSift snapshot (invalid JSON: {exc})")
80
+ if not isinstance(data, dict) or "objects" not in data:
81
+ _fail(f"{path} is not a KubeSift snapshot (create one with: kubesift snapshot save)")
82
+ return Snapshot.from_dict(data)
83
+
84
+
85
+ def _llm_status(text: str, use_llm: bool) -> str:
86
+ if use_llm and state.settings.resolved_provider() == "ollama":
87
+ return f"{text} (local model: this can take several minutes without a GPU)"
88
+ return text
89
+
90
+
91
+ def _snapshot(namespace: str | None, from_file: str | None = None, with_logs: bool = True):
92
+
62
93
  if from_file:
63
- data = json.loads(Path(from_file).read_text(encoding="utf-8"))
64
- snap = Snapshot.from_dict(data)
94
+ snap = _load_snapshot_file(from_file)
65
95
  if namespace:
66
96
  snap.namespace = namespace
67
97
  return snap
@@ -126,7 +156,7 @@ def diagnose(
126
156
 
127
157
  snap = _snapshot(namespace, from_file)
128
158
  store = _store()
129
- with err.status("Analysing…"):
159
+ with err.status(_llm_status("Analysing…", llm)):
130
160
  report = investigate(state.settings, snap, use_llm=llm, window=parse_duration(since), store=store)
131
161
  if store:
132
162
  store.save_report(report, retention_days=state.settings.history_days)
@@ -175,7 +205,7 @@ def ask(
175
205
  ns = mentioned_namespace(q, [n.get("metadata", {}).get("name", "") for n in snap.all("namespaces")])
176
206
  if ns:
177
207
  snap.namespace = ns
178
- with err.status("Thinking…"):
208
+ with err.status(_llm_status("Thinking…", llm)):
179
209
  report, answer = do_ask(state.settings, snap, q, use_llm=llm, store=_store())
180
210
  if output == "json":
181
211
  print(json.dumps({"answer": answer, "report": report.model_dump(mode="json")}, indent=2, default=str))
@@ -537,6 +567,11 @@ def doctor() -> None:
537
567
  if provider == "none":
538
568
  line(None, "no LLM configured: rule-based diagnosis only (set ANTHROPIC_API_KEY, OPENAI_API_KEY, "
539
569
  "OPENROUTER_API_KEY or KUBESIFT_LLM_PROVIDER=ollama)")
570
+ elif provider == "ollama":
571
+ from kubesift.analysis.llm import OllamaProvider
572
+
573
+ ready, detail = OllamaProvider(s).status()
574
+ line(True if ready else None, f"LLM: {detail}")
540
575
  else:
541
576
  line(True, f"LLM: {provider} / {s.resolved_model()}")
542
577
  line(True if s.prometheus_url else None, f"Prometheus: {s.prometheus_url or 'not configured (optional)'}")
@@ -665,17 +700,13 @@ def diff(
665
700
  output: str = OUT,
666
701
  ) -> None:
667
702
  """Time travel: what is different between two moments (stored history or snapshot files)."""
668
- from kubesift.cluster.snapshot import Snapshot
669
703
  from kubesift.timetravel import diff as do_diff
670
704
  from kubesift.timetravel import digest
671
705
  from kubesift.util import now, parse_duration
672
706
 
673
- if to_file:
674
- later = Snapshot.from_dict(json.loads(Path(to_file).read_text(encoding="utf-8")))
675
- else:
676
- later = _snapshot(namespace, with_logs=False)
707
+ later = _load_snapshot_file(to_file) if to_file else _snapshot(namespace, with_logs=False)
677
708
  if from_file:
678
- earlier = digest(Snapshot.from_dict(json.loads(Path(from_file).read_text(encoding="utf-8"))))
709
+ earlier = digest(_load_snapshot_file(from_file))
679
710
  elif since:
680
711
  store = _store()
681
712
  earlier = store.digest_before(later.context, now() - parse_duration(since)) if store else None
@@ -853,8 +884,10 @@ def knowledge_import(pack_file: str) -> None:
853
884
  try:
854
885
  pack = yaml.safe_load(Path(pack_file).read_text(encoding="utf-8")) or {}
855
886
  result = import_pack(store, pack, state.settings.runbooks_dir)
856
- except ValueError as exc:
857
- _fail(str(exc))
887
+ except FileNotFoundError:
888
+ _fail(f"knowledge pack not found: {pack_file}")
889
+ except (OSError, ValueError, yaml.YAMLError) as exc:
890
+ _fail(f"cannot import {pack_file}: {exc}")
858
891
  console.print(f"Added {result['notes_added']} notes, skipped {result['notes_skipped']}, "
859
892
  f"wrote {result['runbooks_written']} runbooks.")
860
893
 
@@ -27,6 +27,7 @@ _KEY_ENV = {
27
27
  "anthropic": "ANTHROPIC_API_KEY",
28
28
  "openai": "OPENAI_API_KEY",
29
29
  "openrouter": "OPENROUTER_API_KEY",
30
+ "ollama": "OLLAMA_API_KEY", # only for Ollama's hosted API; a local server needs no key
30
31
  }
31
32
 
32
33
 
@@ -43,6 +44,8 @@ class Settings(BaseModel):
43
44
  llm_base_url: str | None = None
44
45
  llm_api_key: str | None = Field(default=None, repr=False)
45
46
  llm_effort: str = "medium"
47
+ llm_timeout: int | None = None # seconds to wait for an answer; default 180 (hosted) or 900 (ollama)
48
+ llm_context: int | None = None # ollama only: context window in tokens; default sized to the prompt
46
49
  send_logs_to_llm: bool = True
47
50
 
48
51
  # Optional data sources
@@ -99,6 +102,12 @@ class Settings(BaseModel):
99
102
  def resolved_base_url(self) -> str | None:
100
103
  return self.llm_base_url or _DEFAULT_BASE_URLS.get(self.resolved_provider())
101
104
 
105
+ def resolved_llm_timeout(self) -> float:
106
+ if self.llm_timeout:
107
+ return float(self.llm_timeout)
108
+ # local models on a laptop CPU can need many minutes for one answer
109
+ return 900.0 if self.resolved_provider() == "ollama" else 180.0
110
+
102
111
  def resolved_api_key(self) -> str | None:
103
112
  if self.llm_api_key:
104
113
  return self.llm_api_key
@@ -0,0 +1,192 @@
1
+ """LLM providers against a local stand-in HTTP server: request shape and response parsing."""
2
+
3
+ import json
4
+ import threading
5
+ import time
6
+ from http.server import BaseHTTPRequestHandler, HTTPServer
7
+
8
+ import pytest
9
+
10
+ from kubesift.analysis.engine import DIAGNOSIS_SCHEMA
11
+ from kubesift.analysis.llm import (
12
+ AnthropicProvider,
13
+ LLMError,
14
+ OllamaProvider,
15
+ OpenAICompatibleProvider,
16
+ get_provider,
17
+ parse_json,
18
+ )
19
+ from kubesift.config import Settings, load_settings
20
+
21
+ ANSWER = {"summary": "s", "root_causes": [], "next_checks": []}
22
+
23
+
24
+ @pytest.fixture
25
+ def llm_server():
26
+ requests: list[dict] = []
27
+ responses: list[tuple[int, dict]] = []
28
+
29
+ class Handler(BaseHTTPRequestHandler):
30
+ def log_message(self, *a):
31
+ pass
32
+
33
+ def do_POST(self):
34
+ body = json.loads(self.rfile.read(int(self.headers["Content-Length"])))
35
+ requests.append({"path": self.path, "headers": dict(self.headers), "body": body})
36
+ code, payload = responses.pop(0) if responses else (200, {})
37
+ if "_sleep" in payload:
38
+ time.sleep(payload.pop("_sleep")) # a model that is slow to answer
39
+ data = json.dumps(payload).encode()
40
+ self.send_response(code)
41
+ self.send_header("Content-Type", "application/json")
42
+ self.send_header("Content-Length", str(len(data)))
43
+ self.end_headers()
44
+ self.wfile.write(data)
45
+
46
+ server = HTTPServer(("127.0.0.1", 0), Handler)
47
+ threading.Thread(target=server.serve_forever, daemon=True).start()
48
+ yield f"http://127.0.0.1:{server.server_port}", requests, responses
49
+ server.shutdown()
50
+
51
+
52
+ def anthropic_message(text, stop="end_turn"):
53
+ return {"id": "msg_1", "type": "message", "role": "assistant", "model": "claude-opus-5-5",
54
+ "content": [{"type": "text", "text": text}], "stop_reason": stop, "stop_sequence": None,
55
+ "usage": {"input_tokens": 10, "output_tokens": 5}}
56
+
57
+
58
+ def test_anthropic_request_and_parse(llm_server):
59
+ pytest.importorskip("anthropic")
60
+ url, requests, responses = llm_server
61
+ responses.append((200, anthropic_message(json.dumps(ANSWER))))
62
+ p = AnthropicProvider(Settings(llm_provider="anthropic", llm_api_key="k", llm_base_url=url))
63
+ assert p.complete_json("sys", "user", DIAGNOSIS_SCHEMA) == ANSWER
64
+ body = requests[0]["body"]
65
+ assert requests[0]["path"].startswith("/v1/messages")
66
+ assert body["model"] == "claude-opus-5-5"
67
+ assert body["output_config"]["format"]["type"] == "json_schema"
68
+ assert body["fallbacks"] == "default"
69
+ assert "server-side-fallback-2026-07-01" in requests[0]["headers"].get("anthropic-beta", "")
70
+
71
+
72
+ def test_anthropic_retries_without_beta_on_400(llm_server):
73
+ pytest.importorskip("anthropic")
74
+ url, requests, responses = llm_server
75
+ responses += [(400, {"type": "error", "error": {"type": "invalid_request_error", "message": "unsupported"}}),
76
+ (200, anthropic_message(json.dumps(ANSWER)))]
77
+ p = AnthropicProvider(Settings(llm_provider="anthropic", llm_api_key="k", llm_base_url=url))
78
+ assert p.complete_json("sys", "user", DIAGNOSIS_SCHEMA) == ANSWER
79
+ assert "fallbacks" not in requests[-1]["body"]
80
+
81
+
82
+ def test_anthropic_refusal(llm_server):
83
+ pytest.importorskip("anthropic")
84
+ url, _, responses = llm_server
85
+ responses.append((200, anthropic_message("", stop="refusal")))
86
+ p = AnthropicProvider(Settings(llm_provider="anthropic", llm_api_key="k", llm_base_url=url))
87
+ with pytest.raises(LLMError):
88
+ p.complete_json("sys", "user", DIAGNOSIS_SCHEMA)
89
+
90
+
91
+ def test_openai_compatible(llm_server):
92
+ url, requests, responses = llm_server
93
+ responses += [(400, {"error": "no json mode"}),
94
+ (200, {"choices": [{"message": {"content": "```json\n" + json.dumps(ANSWER) + "\n```"}}]})]
95
+ p = OpenAICompatibleProvider(Settings(llm_provider="openrouter", llm_api_key="k", llm_base_url=url, llm_model="m"))
96
+ assert p.complete_json("sys", "user", DIAGNOSIS_SCHEMA) == ANSWER
97
+ assert requests[0]["headers"]["Authorization"] == "Bearer k"
98
+ assert "response_format" in requests[0]["body"] and "response_format" not in requests[1]["body"]
99
+
100
+
101
+ def test_parse_json_errors():
102
+ with pytest.raises(LLMError):
103
+ parse_json("no json here")
104
+ assert parse_json('Sure! {"a": 1} hope that helps') == {"a": 1}
105
+
106
+
107
+ def test_provider_auto(monkeypatch):
108
+ for k in ("ANTHROPIC_API_KEY", "OPENAI_API_KEY", "OPENROUTER_API_KEY"):
109
+ monkeypatch.delenv(k, raising=False)
110
+ assert Settings().resolved_provider() == "none"
111
+ monkeypatch.setenv("OPENROUTER_API_KEY", "x")
112
+ assert Settings().resolved_provider() == "openrouter"
113
+ assert Settings().resolved_model() == "anthropic/claude-opus-5-5"
114
+
115
+
116
+ def test_llm_timeout_defaults_and_override(monkeypatch):
117
+ assert Settings(llm_provider="ollama").resolved_llm_timeout() == 900
118
+ assert Settings(llm_provider="anthropic").resolved_llm_timeout() == 180
119
+ assert Settings(llm_provider="ollama", llm_timeout=30).resolved_llm_timeout() == 30
120
+ monkeypatch.setenv("KUBESIFT_LLM_TIMEOUT", "1200")
121
+ assert load_settings(llm_provider="openai").resolved_llm_timeout() == 1200
122
+
123
+
124
+ def test_slow_answer_is_not_retried(llm_server):
125
+ url, requests, responses = llm_server
126
+ responses.append((200, {"_sleep": 2.5}))
127
+ p = OpenAICompatibleProvider(Settings(llm_provider="openai", llm_api_key="k", llm_base_url=url, llm_timeout=1))
128
+ with pytest.raises(LLMError, match="did not answer within 1s"):
129
+ p.complete_json("sys", "user", DIAGNOSIS_SCHEMA)
130
+ assert len(requests) == 1
131
+
132
+
133
+ def test_ollama_uses_native_api(llm_server):
134
+ url, requests, responses = llm_server
135
+ responses.append((200, {"message": {"role": "assistant", "content": json.dumps(ANSWER)}, "done": True,
136
+ "done_reason": "stop"}))
137
+ p = get_provider(Settings(llm_provider="ollama", llm_base_url=url + "/v1", llm_model="qwen3.5:4b"))
138
+ assert isinstance(p, OllamaProvider) and p.base_url == url
139
+ assert p.complete_json("sys", "user", DIAGNOSIS_SCHEMA) == ANSWER
140
+ req = requests[0]
141
+ assert req["path"] == "/api/chat"
142
+ body = req["body"]
143
+ assert body["think"] is False and body["stream"] is False and body["format"] == DIAGNOSIS_SCHEMA
144
+ assert body["options"]["num_ctx"] >= 8192 and body["options"]["num_predict"] == OllamaProvider.MAX_OUTPUT
145
+
146
+
147
+ def test_ollama_context_is_sized_to_the_prompt():
148
+ p = OllamaProvider(Settings(llm_provider="ollama"))
149
+ assert p.context_size(1000) == 8192
150
+ assert p.context_size(60000) == 24576 # 20000 prompt tokens + 4096 output, rounded up
151
+ assert p.context_size(10_000_000) == 32768
152
+ assert OllamaProvider(Settings(llm_provider="ollama", llm_context=4096)).context_size(60000) == 4096
153
+
154
+
155
+ def test_ollama_falls_back_for_older_servers(llm_server):
156
+ url, requests, responses = llm_server
157
+ responses += [(400, {"error": 'think value "false" is not supported for this model'}),
158
+ (400, {"error": "invalid format"}),
159
+ (200, {"message": {"content": json.dumps(ANSWER)}, "done": True})]
160
+ p = OllamaProvider(Settings(llm_provider="ollama", llm_base_url=url))
161
+ assert p.complete_json("sys", "user", DIAGNOSIS_SCHEMA) == ANSWER
162
+ assert "think" not in requests[1]["body"] and requests[2]["body"]["format"] == "json"
163
+
164
+
165
+ def test_ollama_explains_common_failures(llm_server):
166
+ url, _, responses = llm_server
167
+ responses.append((404, {"error": 'model "qwen3.5:4b" not found, try pulling it first'}))
168
+ p = OllamaProvider(Settings(llm_provider="ollama", llm_base_url=url, llm_model="qwen3.5:4b"))
169
+ with pytest.raises(LLMError, match="ollama pull qwen3.5:4b"):
170
+ p.complete_json("sys", "user", DIAGNOSIS_SCHEMA)
171
+ responses.append((200, {"message": {"content": '{"summary": "cut'}, "done": True, "done_reason": "length"}))
172
+ with pytest.raises(LLMError, match="ran out of output tokens"):
173
+ p.complete_json("sys", "user", DIAGNOSIS_SCHEMA)
174
+ down = OllamaProvider(Settings(llm_provider="ollama", llm_base_url="http://127.0.0.1:9"))
175
+ with pytest.raises(LLMError, match="is it running"):
176
+ down.complete_json("sys", "user", DIAGNOSIS_SCHEMA)
177
+
178
+
179
+ def test_ollama_status_and_api_key(llm_server, monkeypatch):
180
+ import httpx
181
+
182
+ url, requests, responses = llm_server
183
+ tags = {"models": [{"name": "qwen3.5:4b"}, {"name": "llama3.1:latest"}]}
184
+ monkeypatch.setattr(httpx, "get", lambda *a, **k: httpx.Response(200, json=tags, request=httpx.Request("GET", a[0])))
185
+ assert OllamaProvider(Settings(llm_provider="ollama", llm_model="qwen3.5:4b")).status()[0]
186
+ assert OllamaProvider(Settings(llm_provider="ollama", llm_model="llama3.1")).status()[0]
187
+ ok, detail = OllamaProvider(Settings(llm_provider="ollama", llm_model="mistral")).status()
188
+ assert not ok and "ollama pull mistral" in detail
189
+ monkeypatch.setenv("OLLAMA_API_KEY", "ok-key")
190
+ responses.append((200, {"message": {"content": json.dumps(ANSWER)}, "done": True}))
191
+ OllamaProvider(Settings(llm_provider="ollama", llm_base_url=url)).complete_json("s", "u", DIAGNOSIS_SCHEMA)
192
+ assert requests[-1]["headers"]["Authorization"] == "Bearer ok-key"
@@ -1,103 +0,0 @@
1
- """LLM providers against a local stand-in HTTP server: request shape and response parsing."""
2
-
3
- import json
4
- import threading
5
- from http.server import BaseHTTPRequestHandler, HTTPServer
6
-
7
- import pytest
8
-
9
- from kubesift.analysis.engine import DIAGNOSIS_SCHEMA
10
- from kubesift.analysis.llm import AnthropicProvider, LLMError, OpenAICompatibleProvider, parse_json
11
- from kubesift.config import Settings
12
-
13
- ANSWER = {"summary": "s", "root_causes": [], "next_checks": []}
14
-
15
-
16
- @pytest.fixture
17
- def llm_server():
18
- requests: list[dict] = []
19
- responses: list[tuple[int, dict]] = []
20
-
21
- class Handler(BaseHTTPRequestHandler):
22
- def log_message(self, *a):
23
- pass
24
-
25
- def do_POST(self):
26
- body = json.loads(self.rfile.read(int(self.headers["Content-Length"])))
27
- requests.append({"path": self.path, "headers": dict(self.headers), "body": body})
28
- code, payload = responses.pop(0) if responses else (200, {})
29
- data = json.dumps(payload).encode()
30
- self.send_response(code)
31
- self.send_header("Content-Type", "application/json")
32
- self.send_header("Content-Length", str(len(data)))
33
- self.end_headers()
34
- self.wfile.write(data)
35
-
36
- server = HTTPServer(("127.0.0.1", 0), Handler)
37
- threading.Thread(target=server.serve_forever, daemon=True).start()
38
- yield f"http://127.0.0.1:{server.server_port}", requests, responses
39
- server.shutdown()
40
-
41
-
42
- def anthropic_message(text, stop="end_turn"):
43
- return {"id": "msg_1", "type": "message", "role": "assistant", "model": "claude-opus-5-5",
44
- "content": [{"type": "text", "text": text}], "stop_reason": stop, "stop_sequence": None,
45
- "usage": {"input_tokens": 10, "output_tokens": 5}}
46
-
47
-
48
- def test_anthropic_request_and_parse(llm_server):
49
- pytest.importorskip("anthropic")
50
- url, requests, responses = llm_server
51
- responses.append((200, anthropic_message(json.dumps(ANSWER))))
52
- p = AnthropicProvider(Settings(llm_provider="anthropic", llm_api_key="k", llm_base_url=url))
53
- assert p.complete_json("sys", "user", DIAGNOSIS_SCHEMA) == ANSWER
54
- body = requests[0]["body"]
55
- assert requests[0]["path"].startswith("/v1/messages")
56
- assert body["model"] == "claude-opus-5-5"
57
- assert body["output_config"]["format"]["type"] == "json_schema"
58
- assert body["fallbacks"] == "default"
59
- assert "server-side-fallback-2026-07-01" in requests[0]["headers"].get("anthropic-beta", "")
60
-
61
-
62
- def test_anthropic_retries_without_beta_on_400(llm_server):
63
- pytest.importorskip("anthropic")
64
- url, requests, responses = llm_server
65
- responses += [(400, {"type": "error", "error": {"type": "invalid_request_error", "message": "unsupported"}}),
66
- (200, anthropic_message(json.dumps(ANSWER)))]
67
- p = AnthropicProvider(Settings(llm_provider="anthropic", llm_api_key="k", llm_base_url=url))
68
- assert p.complete_json("sys", "user", DIAGNOSIS_SCHEMA) == ANSWER
69
- assert "fallbacks" not in requests[-1]["body"]
70
-
71
-
72
- def test_anthropic_refusal(llm_server):
73
- pytest.importorskip("anthropic")
74
- url, _, responses = llm_server
75
- responses.append((200, anthropic_message("", stop="refusal")))
76
- p = AnthropicProvider(Settings(llm_provider="anthropic", llm_api_key="k", llm_base_url=url))
77
- with pytest.raises(LLMError):
78
- p.complete_json("sys", "user", DIAGNOSIS_SCHEMA)
79
-
80
-
81
- def test_openai_compatible(llm_server):
82
- url, requests, responses = llm_server
83
- responses += [(400, {"error": "no json mode"}),
84
- (200, {"choices": [{"message": {"content": "```json\n" + json.dumps(ANSWER) + "\n```"}}]})]
85
- p = OpenAICompatibleProvider(Settings(llm_provider="openrouter", llm_api_key="k", llm_base_url=url, llm_model="m"))
86
- assert p.complete_json("sys", "user", DIAGNOSIS_SCHEMA) == ANSWER
87
- assert requests[0]["headers"]["Authorization"] == "Bearer k"
88
- assert "response_format" in requests[0]["body"] and "response_format" not in requests[1]["body"]
89
-
90
-
91
- def test_parse_json_errors():
92
- with pytest.raises(LLMError):
93
- parse_json("no json here")
94
- assert parse_json('Sure! {"a": 1} hope that helps') == {"a": 1}
95
-
96
-
97
- def test_provider_auto(monkeypatch):
98
- for k in ("ANTHROPIC_API_KEY", "OPENAI_API_KEY", "OPENROUTER_API_KEY"):
99
- monkeypatch.delenv(k, raising=False)
100
- assert Settings().resolved_provider() == "none"
101
- monkeypatch.setenv("OPENROUTER_API_KEY", "x")
102
- assert Settings().resolved_provider() == "openrouter"
103
- assert Settings().resolved_model() == "anthropic/claude-opus-5-5"
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes