lex-llm-bedrock 0.5.8 → 0.5.10
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.
- checksums.yaml +4 -4
- data/CHANGELOG.md +13 -0
- data/lib/legion/extensions/llm/bedrock/helpers/callable.rb +6 -5
- data/lib/legion/extensions/llm/bedrock/provider/converse_helpers.rb +1 -18
- data/lib/legion/extensions/llm/bedrock/provider/invoke_model_helpers.rb +3 -0
- data/lib/legion/extensions/llm/bedrock/thinking_modes.rb +23 -7
- data/lib/legion/extensions/llm/bedrock/translator/chunk_parsing.rb +4 -2
- data/lib/legion/extensions/llm/bedrock/translator/read_helpers.rb +11 -18
- data/lib/legion/extensions/llm/bedrock/version.rb +1 -1
- metadata +1 -1
checksums.yaml
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
---
|
|
2
2
|
SHA256:
|
|
3
|
-
metadata.gz:
|
|
4
|
-
data.tar.gz:
|
|
3
|
+
metadata.gz: f71988b85a83937a63f4f746b47ced891eab1020bf8257530cacf1530c9dc221
|
|
4
|
+
data.tar.gz: 567c99945ef05ab78f7fcc9c6f2b2cddbfa439381c24db3c9b0d395c8067a127
|
|
5
5
|
SHA512:
|
|
6
|
-
metadata.gz:
|
|
7
|
-
data.tar.gz:
|
|
6
|
+
metadata.gz: 05bded0be0925b28bc33e504a29b112d2740e9601c27549c826071dd2ad5eb70da6b7b55bfe7cc1192ee23b919f23844f191d52dc61abe978126c2aa809479ed
|
|
7
|
+
data.tar.gz: 2b9ef635ea825e718397850cdfc0fa195ebf391c26a8a95087893d364ca82526aea78d08a10359f1c556dc94f5198d107f4c173427fab5729407fa8f466d3528
|
data/CHANGELOG.md
CHANGED
|
@@ -1,5 +1,18 @@
|
|
|
1
1
|
# Changelog
|
|
2
2
|
|
|
3
|
+
## [0.5.10] - 2026-11-18
|
|
4
|
+
|
|
5
|
+
### Fixed
|
|
6
|
+
- Preserve explicit thinking budgets on Claude Sonnet 4.6 and Opus 4.6, which support both manual and adaptive thinking. A budget-only request no longer becomes adaptive low effort and silently permits thinking to be skipped. Both provider dispatch paths and canonical render targets use the shared thinking-mode decision; effort-only requests and adaptive-only models retain adaptive rendering.
|
|
7
|
+
- Preserve provider-reported `output_tokens_details.thinking_tokens` in canonical usage. Streaming completion shares the translator's usage parser and merges final cumulative output counts with initial input/cache counts, including when thinking text is omitted.
|
|
8
|
+
|
|
9
|
+
## [0.5.9] - 2026-11-17
|
|
10
|
+
|
|
11
|
+
### Fixed
|
|
12
|
+
- Classify Bedrock `ServiceUnavailableException` as request-local overload. Temporary AWS capacity failures no longer mark every offering on the instance unavailable and cause unrelated requests to fail routing.
|
|
13
|
+
- Classify an explicitly disposed callable as exact-instance unavailability; readiness remains responsible for recovery.
|
|
14
|
+
- Keep system-shape regression coverage valid when the shared canonical constructor rejects a non-String before the provider renderer receives it.
|
|
15
|
+
|
|
3
16
|
## [0.5.8] - 2026-08-25
|
|
4
17
|
|
|
5
18
|
### Added
|
|
@@ -155,6 +155,8 @@ module Legion
|
|
|
155
155
|
|
|
156
156
|
def classify_error(error:)
|
|
157
157
|
case error
|
|
158
|
+
when Legion::Extensions::Llm::Inventory::Errors::CallableDisposedError
|
|
159
|
+
:instance_unavailable
|
|
158
160
|
when Aws::BedrockRuntime::Errors::ThrottlingException,
|
|
159
161
|
Aws::Bedrock::Errors::ThrottlingException
|
|
160
162
|
:rate_limited
|
|
@@ -172,8 +174,9 @@ module Legion
|
|
|
172
174
|
:model_missing
|
|
173
175
|
when Aws::BedrockRuntime::Errors::ServiceUnavailableException,
|
|
174
176
|
Aws::Bedrock::Errors::ServiceUnavailableException
|
|
175
|
-
#
|
|
176
|
-
|
|
177
|
+
# AWS documents this as temporary capacity exhaustion, not
|
|
178
|
+
# authoritative evidence that the exact instance is offline.
|
|
179
|
+
:overloaded
|
|
177
180
|
when Aws::BedrockRuntime::Errors::ServiceError,
|
|
178
181
|
Aws::Bedrock::Errors::ServiceError
|
|
179
182
|
classify_service_error(error: error)
|
|
@@ -189,9 +192,7 @@ module Legion
|
|
|
189
192
|
when 401 then :authentication
|
|
190
193
|
when 403 then :authorization
|
|
191
194
|
when 404 then :model_missing
|
|
192
|
-
#
|
|
193
|
-
# Only the explicit ServiceUnavailableException (handled above) justifies
|
|
194
|
-
# instance_unavailable. Everything else is request-local.
|
|
195
|
+
# Transport overload is request-local, including AWS's named 503.
|
|
195
196
|
when 503, 529 then :overloaded
|
|
196
197
|
else :provider_error
|
|
197
198
|
end
|
|
@@ -227,27 +227,10 @@ module Legion
|
|
|
227
227
|
translator.parse_response(normalize_response(response), model: fallback_model)
|
|
228
228
|
end
|
|
229
229
|
|
|
230
|
-
def cache_read_tokens(usage)
|
|
231
|
-
return nil if usage.nil?
|
|
232
|
-
|
|
233
|
-
value(usage, :cache_read_input_tokens) || value(usage, 'cache_read_input_tokens')
|
|
234
|
-
end
|
|
235
|
-
|
|
236
|
-
def cache_write_tokens(usage)
|
|
237
|
-
return nil if usage.nil?
|
|
238
|
-
|
|
239
|
-
value(usage, :cache_creation_input_tokens) || value(usage, 'cache_creation_input_tokens')
|
|
240
|
-
end
|
|
241
|
-
|
|
242
230
|
# Canonical usage from a wire usage hash/struct (provider spellings
|
|
243
231
|
# translated at this edge, 13 §3).
|
|
244
232
|
def usage_from(usage)
|
|
245
|
-
|
|
246
|
-
input_tokens: value(usage, :input_tokens),
|
|
247
|
-
output_tokens: value(usage, :output_tokens),
|
|
248
|
-
cache_read_tokens: cache_read_tokens(usage),
|
|
249
|
-
cache_write_tokens: cache_write_tokens(usage)
|
|
250
|
-
)
|
|
233
|
+
translator.parse_usage(usage)
|
|
251
234
|
end
|
|
252
235
|
|
|
253
236
|
# One stop-reason edge: the gem's Translator map is the single
|
|
@@ -314,6 +314,9 @@ module Legion
|
|
|
314
314
|
end
|
|
315
315
|
when 'message_delta'
|
|
316
316
|
state[:stop_reason] = (event_json['delta'] || {})['stop_reason']
|
|
317
|
+
# Final output counts are cumulative; retain input/cache counts
|
|
318
|
+
# from message_start without adding output tokens twice.
|
|
319
|
+
state[:final_usage] = (state[:final_usage] || {}).merge(event_json['usage'] || {})
|
|
317
320
|
end
|
|
318
321
|
rescue StandardError => e
|
|
319
322
|
handle_exception(e, level: :error, handled: false,
|
|
@@ -19,16 +19,21 @@ module Legion
|
|
|
19
19
|
# 2. ADAPTIVE (Claude opus-4-6/4-7/4-8/5, sonnet-4-6/5):
|
|
20
20
|
# { type: 'adaptive' } + output_config: { effort: <low|medium|high|xhigh|max> }
|
|
21
21
|
# Gated by beta header 'effort-2025-11-24' in the anthropic_beta list.
|
|
22
|
-
#
|
|
22
|
+
# Adaptive-only models REJECT { type: 'enabled', budget_tokens: N } with
|
|
23
23
|
# `ValidationException: "thinking.type.enabled" is not supported for
|
|
24
24
|
# this model. Use "thinking.type.adaptive" and "output_config.effort"`.
|
|
25
25
|
# The effort value is passed through directly from the canonical
|
|
26
26
|
# Thinking::Config#resolved_effort (the full ladder matches 1:1).
|
|
27
27
|
# 'none' or nil effort → omit output_config (API defaults to high).
|
|
28
28
|
#
|
|
29
|
+
# Opus 4.6 and Sonnet 4.6 support BOTH shapes. Preserve an explicit
|
|
30
|
+
# budget on those models; effort-only requests use adaptive thinking.
|
|
31
|
+
# Converting a supplied budget into adaptive effort changes the request
|
|
32
|
+
# from manual thinking to a mode that may skip thinking altogether.
|
|
33
|
+
#
|
|
29
34
|
# PRECEDENCE: adaptive fragments are checked BEFORE budgeted because
|
|
30
35
|
# `claude-opus-4` (budgeted) is a substring of `claude-opus-4-7`
|
|
31
|
-
# (adaptive).
|
|
36
|
+
# (adaptive). Only the explicit dual-mode list permits both shapes.
|
|
32
37
|
#
|
|
33
38
|
# The match is a substring so it tolerates the many Bedrock model-id
|
|
34
39
|
# decorations (geo prefixes `us.`/`eu.`/`ap.`, `anthropic.` provider
|
|
@@ -60,6 +65,10 @@ module Legion
|
|
|
60
65
|
claude-sonnet-5
|
|
61
66
|
].freeze
|
|
62
67
|
|
|
68
|
+
# AWS still accepts manual budgets on these adaptive-capable models.
|
|
69
|
+
# https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-adaptive-thinking.html
|
|
70
|
+
DUAL_MODE_THINKING_FRAGMENTS = %w[claude-opus-4-6 claude-sonnet-4-6].freeze
|
|
71
|
+
|
|
63
72
|
# Model-id fragments for Claude families that support explicit budgeted
|
|
64
73
|
# extended thinking via { type: 'enabled', budget_tokens: N }.
|
|
65
74
|
BUDGETED_THINKING_FRAGMENTS = %w[
|
|
@@ -83,10 +92,12 @@ module Legion
|
|
|
83
92
|
ADAPTIVE_EFFORT_FRAGMENTS.any? { |fragment| mid.include?(fragment) }
|
|
84
93
|
end
|
|
85
94
|
|
|
86
|
-
# @return [Boolean] true when the model supports { type: 'enabled', budget_tokens: N }
|
|
87
|
-
#
|
|
95
|
+
# @return [Boolean] true when the model supports { type: 'enabled', budget_tokens: N }.
|
|
96
|
+
# Dual-mode 4.6 models support both; adaptive-only models must not
|
|
97
|
+
# accidentally match a broad budgeted-family substring.
|
|
88
98
|
def budgeted_thinking?(model_id)
|
|
89
99
|
return false if model_id.nil? || model_id.to_s.strip.empty?
|
|
100
|
+
return true if DUAL_MODE_THINKING_FRAGMENTS.any? { |fragment| model_id.to_s.include?(fragment) }
|
|
90
101
|
return false if adaptive_thinking?(model_id)
|
|
91
102
|
|
|
92
103
|
mid = model_id.to_s
|
|
@@ -95,6 +106,10 @@ module Legion
|
|
|
95
106
|
BUDGETED_THINKING_FRAGMENTS.any? { |fragment| mid.include?(fragment) }
|
|
96
107
|
end
|
|
97
108
|
|
|
109
|
+
def adaptive_mode?(thinking:, model_id:)
|
|
110
|
+
adaptive_thinking?(model_id) && !(thinking.budget && budgeted_thinking?(model_id))
|
|
111
|
+
end
|
|
112
|
+
|
|
98
113
|
# @return [Boolean] true when the model is KNOWN not to support thinking.
|
|
99
114
|
# Distinct from `!budgeted_thinking?`: a nil/blank/unknown model id is
|
|
100
115
|
# NOT known-unsupported — we then honor an explicit thinking request and
|
|
@@ -156,8 +171,9 @@ module Legion
|
|
|
156
171
|
return nil unless thinking_enabled?(thinking)
|
|
157
172
|
return nil if known_non_thinking?(model_id)
|
|
158
173
|
|
|
159
|
-
#
|
|
160
|
-
|
|
174
|
+
# Preserve explicit budgets where supported; only resolve to an
|
|
175
|
+
# adaptive effort when that is the selected model's usable mode.
|
|
176
|
+
return { type: 'adaptive' } if adaptive_mode?(thinking: thinking, model_id: model_id)
|
|
161
177
|
|
|
162
178
|
budget = thinking.resolved_budget
|
|
163
179
|
if budget.nil?
|
|
@@ -185,7 +201,7 @@ module Legion
|
|
|
185
201
|
# Converse: additionalModelRequestFields + beta array mechanism).
|
|
186
202
|
def adaptive_wire(thinking:, model_id:)
|
|
187
203
|
return nil unless thinking_enabled?(thinking)
|
|
188
|
-
return nil unless
|
|
204
|
+
return nil unless adaptive_mode?(thinking: thinking, model_id: model_id)
|
|
189
205
|
|
|
190
206
|
effort = wire_effort(thinking.resolved_effort)
|
|
191
207
|
result = { thinking: { type: 'adaptive' }, beta_header: EFFORT_BETA_HEADER }
|
|
@@ -102,8 +102,10 @@ module Legion
|
|
|
102
102
|
)
|
|
103
103
|
Canonical::Chunk.tool_call_delta(tool_call: tc, request_id: request_id)
|
|
104
104
|
when 'message_delta'
|
|
105
|
-
stop =
|
|
106
|
-
|
|
105
|
+
stop = safe_key_read(safe_key_read(raw, :delta), :stop_reason)
|
|
106
|
+
usage_raw = safe_key_read(raw, :usage)
|
|
107
|
+
Canonical::Chunk.done(request_id: request_id, stop_reason: map_stop_reason(stop),
|
|
108
|
+
usage: usage_raw ? parse_usage(usage_raw) : nil)
|
|
107
109
|
end
|
|
108
110
|
end
|
|
109
111
|
end
|
|
@@ -124,25 +124,18 @@ module Legion
|
|
|
124
124
|
Canonical::Usage.from_hash(h)
|
|
125
125
|
end
|
|
126
126
|
|
|
127
|
+
# Provider stream completion shares the same wire-to-canonical edge.
|
|
128
|
+
public :parse_usage
|
|
129
|
+
|
|
127
130
|
def build_usage_hash(usage_raw)
|
|
128
|
-
|
|
129
|
-
|
|
130
|
-
|
|
131
|
-
|
|
132
|
-
|
|
133
|
-
|
|
134
|
-
|
|
135
|
-
|
|
136
|
-
}
|
|
137
|
-
else
|
|
138
|
-
{
|
|
139
|
-
input_tokens: safe_key_read(usage_raw, :input_tokens),
|
|
140
|
-
output_tokens: safe_key_read(usage_raw, :output_tokens),
|
|
141
|
-
cache_read_tokens: safe_key_read(usage_raw, :cache_read_input_tokens),
|
|
142
|
-
cache_write_tokens: safe_key_read(usage_raw, :cache_creation_input_tokens),
|
|
143
|
-
thinking_tokens: safe_key_read(usage_raw, :thinking_tokens)
|
|
144
|
-
}
|
|
145
|
-
end
|
|
131
|
+
details = safe_key_read(usage_raw, :output_tokens_details)
|
|
132
|
+
{
|
|
133
|
+
input_tokens: safe_key_read(usage_raw, :input_tokens),
|
|
134
|
+
output_tokens: safe_key_read(usage_raw, :output_tokens),
|
|
135
|
+
cache_read_tokens: safe_key_read(usage_raw, :cache_read_input_tokens),
|
|
136
|
+
cache_write_tokens: safe_key_read(usage_raw, :cache_creation_input_tokens),
|
|
137
|
+
thinking_tokens: safe_key_read(details, :thinking_tokens) || safe_key_read(usage_raw, :thinking_tokens)
|
|
138
|
+
}
|
|
146
139
|
end
|
|
147
140
|
end
|
|
148
141
|
end
|