claude-dev-env 8.116.5 → 8.117.1
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.
- package/bin/ever-registered-hooks.mjs +2 -0
- package/bin/ever-registered-hooks.test.mjs +4 -0
- package/bin/install-constants.mjs +3 -1
- package/bin/install.cursor-rules.test.mjs +64 -2
- package/bin/install.mjs +15 -2
- package/bin/install.prune.test.mjs +92 -0
- package/bin/install.test.mjs +1 -0
- package/docs/model-choice/escalation.md +3 -3
- package/docs/model-choice/subagents-and-fan-out.md +2 -2
- package/docs/model-choice/which-model.md +1 -1
- package/docs/references/working-style.md +1 -0
- package/docs/rule-guides/question-presentation.md +6 -0
- package/docs/rule-guides/session-title.md +8 -9
- package/docs/rule-guides/subagent-model-effort.md +13 -13
- package/hooks/advisory/owned_pr_review_loop.py +2 -3
- package/hooks/blocking/bash_post_call_dispatcher.py +17 -9
- package/hooks/blocking/bash_pre_tool_use_dispatcher.py +12 -8
- package/hooks/blocking/final_reply_gate.py +5 -9
- package/hooks/blocking/open_tasks_stop_gate.py +9 -20
- package/hooks/blocking/open_tasks_transcript.py +8 -3
- package/hooks/blocking/pii_prevention_blocker_parts/repository_exemption.py +7 -2
- package/hooks/blocking/question_brief_gate.py +382 -0
- package/hooks/blocking/step_note_gate.py +5 -2
- package/hooks/blocking/task_subject_length_gate.py +111 -0
- package/hooks/blocking/test_bash_post_call_dispatcher.py +36 -1
- package/hooks/blocking/test_bash_pre_tool_use_dispatcher.py +57 -0
- package/hooks/blocking/test_final_reply_gate.py +26 -2
- package/hooks/blocking/test_open_tasks_stop_gate.py +21 -11
- package/hooks/blocking/test_question_brief_gate.py +916 -0
- package/hooks/blocking/test_step_note_gate.py +26 -0
- package/hooks/blocking/test_task_subject_length_gate.py +221 -0
- package/hooks/blocking/test_turn_review_child.py +166 -0
- package/hooks/blocking/test_turn_review_claim.py +194 -0
- package/hooks/blocking/test_turn_review_findings.py +124 -1
- package/hooks/blocking/test_turn_review_gate.py +1312 -10
- package/hooks/blocking/test_turn_review_repositories.py +65 -0
- package/hooks/blocking/test_turn_review_snapshot.py +8 -0
- package/hooks/blocking/test_turn_review_state.py +132 -0
- package/hooks/blocking/test_turn_review_transcript.py +125 -0
- package/hooks/blocking/tests/test_repository_exemption.py +11 -0
- package/hooks/blocking/turn_review_child.py +145 -0
- package/hooks/blocking/turn_review_claim.py +306 -0
- package/hooks/blocking/turn_review_findings.py +173 -13
- package/hooks/blocking/turn_review_gate.py +259 -262
- package/hooks/blocking/turn_review_repositories.py +58 -0
- package/hooks/blocking/turn_review_snapshot.py +5 -1
- package/hooks/blocking/turn_review_state.py +260 -0
- package/hooks/blocking/turn_review_transcript.py +158 -0
- package/hooks/features/bash-dispatchers.md +1 -3
- package/hooks/features/conduct-gates.md +19 -9
- package/hooks/features/lifecycle-cleanup.md +1 -1
- package/hooks/features/spawn-routing.md +1 -1
- package/hooks/hooks.json +24 -11
- package/hooks/hooks_constants/final_reply_gate_constants.py +0 -3
- package/hooks/hooks_constants/hook_time_budget.py +61 -0
- package/hooks/hooks_constants/hosted_hook_runner.py +51 -21
- package/hooks/hooks_constants/hosted_hook_runner_constants.py +3 -0
- package/hooks/hooks_constants/nested_project_hooks_constants.py +8 -0
- package/hooks/hooks_constants/once_per_session_constants.py +1 -0
- package/hooks/hooks_constants/open_tasks_stop_gate_constants.py +0 -4
- package/hooks/hooks_constants/owned_pr_review_loop_constants.py +0 -4
- package/hooks/hooks_constants/pre_tool_use_context_runner.py +23 -3
- package/hooks/hooks_constants/pre_tool_use_context_runner_constants.py +5 -0
- package/hooks/hooks_constants/pre_tool_use_stdin.py +1 -1
- package/hooks/hooks_constants/question_brief_gate_constants.py +66 -0
- package/hooks/hooks_constants/session_title_constants.py +8 -29
- package/hooks/hooks_constants/stop_hook_hold.py +54 -0
- package/hooks/hooks_constants/stop_hook_hold_constants.py +9 -0
- package/hooks/hooks_constants/subagent_effort_levels.json +6 -0
- package/hooks/hooks_constants/subagent_effort_route_constants.py +20 -19
- package/hooks/hooks_constants/task_subject_length_gate_constants.py +43 -0
- package/hooks/hooks_constants/test_hook_time_budget.py +104 -0
- package/hooks/hooks_constants/test_hosted_hook_runner.py +12 -0
- package/hooks/hooks_constants/test_pre_tool_use_context_runner.py +16 -0
- package/hooks/hooks_constants/test_pre_tool_use_stdin.py +6 -0
- package/hooks/hooks_constants/test_stop_hook_hold.py +40 -0
- package/hooks/hooks_constants/thread_spawn_pace_hook_constants.py +1 -2
- package/hooks/hooks_constants/turn_review_gate_constants.py +246 -12
- package/hooks/lifecycle/nested_project_hooks.py +34 -4
- package/hooks/lifecycle/test_nested_project_hooks.py +52 -0
- package/hooks/once_per_session.py +33 -4
- package/hooks/open_task_asks.py +1 -1
- package/hooks/owned_pr_session_record.py +1 -76
- package/hooks/routing/subagent_effort_route.py +107 -13
- package/hooks/routing/subagent_effort_route_hook.py +9 -8
- package/hooks/routing/test_subagent_effort_route.py +201 -14
- package/hooks/routing/test_subagent_effort_route_hook.py +56 -0
- package/hooks/routing/test_thread_spawn_pace_hook.py +32 -0
- package/hooks/routing/thread_spawn_pace_hook.py +3 -3
- package/hooks/session/background_silence_prompt.py +1 -1
- package/hooks/session/test_background_silence_prompt.py +8 -0
- package/hooks/session/test_tracking_nudge.py +125 -2
- package/hooks/session/tracking_nudge.py +28 -18
- package/hooks/session_title_transcript.py +216 -0
- package/hooks/test_once_per_session.py +10 -0
- package/hooks/test_owned_pr_session_record.py +0 -26
- package/hooks/test_session_title_transcript.py +142 -0
- package/hooks/validation/post_tool_use_dispatcher.py +38 -21
- package/hooks/validation/test_post_tool_use_dispatcher.py +43 -0
- package/package.json +1 -1
- package/rules/question-presentation.md +4 -4
- package/rules/subagent-model-effort.md +1 -1
- package/scripts/account_broker_support.py +24 -5
- package/scripts/claude_account_worker.py +4 -3
- package/scripts/dev_env_scripts_constants/account_broker_constants.py +1 -0
- package/scripts/policy_lint/adapter_configuration.py +12 -11
- package/scripts/policy_lint/config/approved_test_pairs.py +27 -0
- package/scripts/policy_lint/config/constants.py +2 -2
- package/scripts/test_account_broker_support.py +43 -1
- package/scripts/test_claude_account_worker.py +1 -0
- package/scripts/tests/test_adapter_configuration.py +10 -2
- package/scripts/tests/test_policy_lint_rules_pairing.py +24 -0
- package/hooks/blocking/owned_pr_merge_stop_gate.py +0 -185
- package/hooks/blocking/session_title_turn_gate.py +0 -417
- package/hooks/blocking/test_owned_pr_merge_stop_gate.py +0 -254
- package/hooks/blocking/test_session_title_turn_gate.py +0 -396
- package/hooks/hooks_constants/owned_pr_merge_stop_gate_constants.py +0 -40
|
@@ -0,0 +1,306 @@
|
|
|
1
|
+
"""Change the session's review state for a prompt and a stop, for the turn review gate.
|
|
2
|
+
|
|
3
|
+
The gate calls these while it holds the lock on the session's state file. A
|
|
4
|
+
stop claims its range before the review starts. It moves the baseline to
|
|
5
|
+
snapshot B, marks the turn closed, and saves both. A second stop during the
|
|
6
|
+
review starts from B, so no range is reviewed twice.
|
|
7
|
+
|
|
8
|
+
A prompt marks the turn open. When the last turn is still open, no stop has
|
|
9
|
+
claimed it yet: its stop is racing this prompt, or the user interrupted it::
|
|
10
|
+
|
|
11
|
+
prompt 1 -> baseline A, turn open
|
|
12
|
+
stop 1 -> starts; the next prompt runs before its claim
|
|
13
|
+
prompt 2 -> turn still open, so A stays the baseline
|
|
14
|
+
stop 1 -> claims A..B and reviews turn 1's edits
|
|
15
|
+
|
|
16
|
+
Edits the user made between those two turns are then reviewed too. When a
|
|
17
|
+
prompt found the turn open less than 10 seconds before the claim, that prompt
|
|
18
|
+
began the next turn, so the stop leaves it open. A later claim closes it,
|
|
19
|
+
which ends the open turn an interrupt left behind.
|
|
20
|
+
|
|
21
|
+
The claim also lists the findings the reply contests. The judge gets the
|
|
22
|
+
checkout path and one JSON object of data::
|
|
23
|
+
|
|
24
|
+
{"requests": [{"id": "TR-1", "commit_range": "A..B",
|
|
25
|
+
"finding": "src/service.py:3 - ...", "reason_to_dismiss": "..."}],
|
|
26
|
+
"user_messages": ["the last five messages the user typed, cut to 2000 characters"]}
|
|
27
|
+
|
|
28
|
+
The judge gets no other part of the
|
|
29
|
+
conversation. A ``Dismiss`` line for an id that is not open, or
|
|
30
|
+
one sent at the hold cap, is logged and not judged.
|
|
31
|
+
|
|
32
|
+
After the review, each new finding gets the next id of the session and is
|
|
33
|
+
saved as open. An approved dismissal marks its finding dismissed. A rejection
|
|
34
|
+
marks it upheld and adds one line to the wake. A stop with a wake adds one to
|
|
35
|
+
the hold count. A stop below the cap with no
|
|
36
|
+
finding, no rejection and no dismissal to judge resets the count. A prompt within 10 seconds
|
|
37
|
+
of the last wake, or one that starts with ``<task-notification``, belongs to
|
|
38
|
+
the wake.
|
|
39
|
+
"""
|
|
40
|
+
|
|
41
|
+
from __future__ import annotations
|
|
42
|
+
|
|
43
|
+
import json
|
|
44
|
+
import time
|
|
45
|
+
from dataclasses import dataclass, field
|
|
46
|
+
from pathlib import Path
|
|
47
|
+
|
|
48
|
+
from hooks_constants.turn_review_gate_constants import (
|
|
49
|
+
BLOCK_REASON_TEMPLATE,
|
|
50
|
+
DISMISS_AT_CAP_DIAGNOSTIC_TEMPLATE,
|
|
51
|
+
DISMISS_IGNORED_DIAGNOSTIC_TEMPLATE,
|
|
52
|
+
FINDING_DISMISSED_DIAGNOSTIC_TEMPLATE,
|
|
53
|
+
FINDING_LINE_SEPARATOR,
|
|
54
|
+
FINDING_LINE_WITH_ID_TEMPLATE,
|
|
55
|
+
FINDING_STATUS_DISMISSED,
|
|
56
|
+
FINDING_STATUS_OPEN,
|
|
57
|
+
FINDING_STATUS_UPHELD,
|
|
58
|
+
FINDING_UPHELD_DIAGNOSTIC_TEMPLATE,
|
|
59
|
+
JUDGE_COMMIT_RANGE_TEMPLATE,
|
|
60
|
+
JUDGE_DATA_INDENT,
|
|
61
|
+
JUDGE_DATA_REQUESTS_KEY,
|
|
62
|
+
JUDGE_DATA_USER_MESSAGES_KEY,
|
|
63
|
+
JUDGE_ID_KEY,
|
|
64
|
+
JUDGE_PROMPT_TEMPLATE,
|
|
65
|
+
JUDGE_REQUEST_FINDING_KEY,
|
|
66
|
+
JUDGE_REQUEST_RANGE_KEY,
|
|
67
|
+
JUDGE_REQUEST_REASON_KEY,
|
|
68
|
+
MACHINE_PROMPT_PREFIX,
|
|
69
|
+
MAXIMUM_BLOCKS_IN_A_ROW,
|
|
70
|
+
MAXIMUM_USER_MESSAGE_CHARACTERS,
|
|
71
|
+
OUTSIDE_REPOSITORY_LABEL,
|
|
72
|
+
RACING_STOP_WINDOW_SECONDS,
|
|
73
|
+
REJECTION_LINE_SEPARATOR,
|
|
74
|
+
REJECTION_LINE_TEMPLATE,
|
|
75
|
+
REPOSITORY_FINDINGS_SEPARATOR,
|
|
76
|
+
REPOSITORY_FINDINGS_TEMPLATE,
|
|
77
|
+
SENTENCE_END_CHARACTERS,
|
|
78
|
+
UNKNOWN_FINDING_LABEL,
|
|
79
|
+
WAKE_PART_SEPARATOR,
|
|
80
|
+
WAKE_PROMPT_WINDOW_SECONDS,
|
|
81
|
+
WORD_SEPARATOR,
|
|
82
|
+
)
|
|
83
|
+
from turn_review_findings import JudgeVerdict
|
|
84
|
+
from turn_review_snapshot import working_tree_snapshot
|
|
85
|
+
from turn_review_state import (
|
|
86
|
+
_Finding,
|
|
87
|
+
_ReviewState,
|
|
88
|
+
_finding_number,
|
|
89
|
+
_load_state,
|
|
90
|
+
_log,
|
|
91
|
+
_next_finding_id,
|
|
92
|
+
_save_state,
|
|
93
|
+
)
|
|
94
|
+
|
|
95
|
+
|
|
96
|
+
def _is_wake_prompt(review_state: _ReviewState, prompt_text: object, now: float) -> bool:
|
|
97
|
+
if isinstance(prompt_text, str) and prompt_text.lstrip().startswith(MACHINE_PROMPT_PREFIX):
|
|
98
|
+
return True
|
|
99
|
+
if review_state.last_wake_at is None:
|
|
100
|
+
return False
|
|
101
|
+
return now - review_state.last_wake_at < WAKE_PROMPT_WINDOW_SECONDS
|
|
102
|
+
|
|
103
|
+
|
|
104
|
+
def _prompt_baselines(review_state: _ReviewState, all_roots: list[Path]) -> dict[str, str]:
|
|
105
|
+
all_baselines = dict(review_state.all_baselines) if review_state.is_turn_open else {}
|
|
106
|
+
for each_root in all_roots:
|
|
107
|
+
if str(each_root) in all_baselines:
|
|
108
|
+
continue
|
|
109
|
+
snapshot_commit = working_tree_snapshot(each_root, None)
|
|
110
|
+
if snapshot_commit is not None:
|
|
111
|
+
all_baselines[str(each_root)] = snapshot_commit
|
|
112
|
+
return all_baselines
|
|
113
|
+
|
|
114
|
+
|
|
115
|
+
@dataclass(frozen=True)
|
|
116
|
+
class _JudgeRequest:
|
|
117
|
+
finding_id: str
|
|
118
|
+
finding: _Finding
|
|
119
|
+
reason: str
|
|
120
|
+
|
|
121
|
+
|
|
122
|
+
@dataclass(frozen=True)
|
|
123
|
+
class _StopClaim:
|
|
124
|
+
all_turn_ranges: list[tuple[Path, str, str]] = field(default_factory=list)
|
|
125
|
+
all_judge_requests: list[_JudgeRequest] = field(default_factory=list)
|
|
126
|
+
|
|
127
|
+
|
|
128
|
+
def _judge_requests(
|
|
129
|
+
review_state: _ReviewState,
|
|
130
|
+
all_dismiss_reasons: dict[str, str],
|
|
131
|
+
all_roots: list[Path],
|
|
132
|
+
is_capped: bool,
|
|
133
|
+
) -> list[_JudgeRequest]:
|
|
134
|
+
all_root_names = {str(each_root) for each_root in all_roots}
|
|
135
|
+
all_requests: list[_JudgeRequest] = []
|
|
136
|
+
for each_id, each_reason in all_dismiss_reasons.items():
|
|
137
|
+
finding = review_state.all_findings.get(each_id)
|
|
138
|
+
skip_diagnostic = _dismiss_skip_diagnostic(each_id, finding, all_root_names, is_capped)
|
|
139
|
+
if skip_diagnostic is not None:
|
|
140
|
+
_log(skip_diagnostic)
|
|
141
|
+
continue
|
|
142
|
+
if finding is not None:
|
|
143
|
+
all_requests.append(_JudgeRequest(each_id, finding, each_reason))
|
|
144
|
+
return all_requests
|
|
145
|
+
|
|
146
|
+
|
|
147
|
+
def _dismiss_skip_diagnostic(
|
|
148
|
+
finding_id: str, finding: _Finding | None, all_root_names: set[str], is_capped: bool
|
|
149
|
+
) -> str | None:
|
|
150
|
+
if finding is None:
|
|
151
|
+
return DISMISS_IGNORED_DIAGNOSTIC_TEMPLATE.format(
|
|
152
|
+
finding_id=finding_id, status=UNKNOWN_FINDING_LABEL
|
|
153
|
+
)
|
|
154
|
+
if finding.status != FINDING_STATUS_OPEN:
|
|
155
|
+
return DISMISS_IGNORED_DIAGNOSTIC_TEMPLATE.format(
|
|
156
|
+
finding_id=finding_id, status=finding.status
|
|
157
|
+
)
|
|
158
|
+
if finding.repository not in all_root_names:
|
|
159
|
+
return DISMISS_IGNORED_DIAGNOSTIC_TEMPLATE.format(
|
|
160
|
+
finding_id=finding_id, status=OUTSIDE_REPOSITORY_LABEL
|
|
161
|
+
)
|
|
162
|
+
if is_capped:
|
|
163
|
+
return DISMISS_AT_CAP_DIAGNOSTIC_TEMPLATE.format(finding_id=finding_id)
|
|
164
|
+
return None
|
|
165
|
+
|
|
166
|
+
|
|
167
|
+
def _claim_locked_stop(
|
|
168
|
+
session_id: object, all_roots: list[Path], all_dismiss_reasons: dict[str, str]
|
|
169
|
+
) -> _StopClaim:
|
|
170
|
+
review_state = _load_state(session_id)
|
|
171
|
+
is_capped = review_state.blocks_in_a_row >= MAXIMUM_BLOCKS_IN_A_ROW
|
|
172
|
+
all_judge_requests = _judge_requests(review_state, all_dismiss_reasons, all_roots, is_capped)
|
|
173
|
+
all_next_baselines: dict[str, str] = {}
|
|
174
|
+
all_turn_ranges: list[tuple[Path, str, str]] = []
|
|
175
|
+
for each_root in all_roots:
|
|
176
|
+
base_commit = review_state.all_baselines.get(str(each_root))
|
|
177
|
+
turn_commit = working_tree_snapshot(each_root, base_commit)
|
|
178
|
+
if turn_commit is None:
|
|
179
|
+
continue
|
|
180
|
+
all_next_baselines[str(each_root)] = turn_commit
|
|
181
|
+
if base_commit is not None and base_commit != turn_commit and not is_capped:
|
|
182
|
+
all_turn_ranges.append((each_root, base_commit, turn_commit))
|
|
183
|
+
if not all_turn_ranges and not all_judge_requests and not is_capped:
|
|
184
|
+
review_state.blocks_in_a_row = 0
|
|
185
|
+
review_state.all_baselines = all_next_baselines
|
|
186
|
+
review_state.is_turn_open = (
|
|
187
|
+
review_state.open_turn_prompt_at is not None
|
|
188
|
+
and time.time() - review_state.open_turn_prompt_at < RACING_STOP_WINDOW_SECONDS
|
|
189
|
+
)
|
|
190
|
+
review_state.open_turn_prompt_at = None
|
|
191
|
+
_save_state(session_id, review_state)
|
|
192
|
+
return _StopClaim(all_turn_ranges, all_judge_requests)
|
|
193
|
+
|
|
194
|
+
|
|
195
|
+
def _judge_prompt(
|
|
196
|
+
repository_root: Path, all_requests: list[_JudgeRequest], all_user_messages: list[str]
|
|
197
|
+
) -> str:
|
|
198
|
+
judge_data = {
|
|
199
|
+
JUDGE_DATA_REQUESTS_KEY: [
|
|
200
|
+
{
|
|
201
|
+
JUDGE_ID_KEY: each_request.finding_id,
|
|
202
|
+
JUDGE_REQUEST_RANGE_KEY: JUDGE_COMMIT_RANGE_TEMPLATE.format(
|
|
203
|
+
base_commit=each_request.finding.base_commit,
|
|
204
|
+
turn_commit=each_request.finding.turn_commit,
|
|
205
|
+
),
|
|
206
|
+
JUDGE_REQUEST_FINDING_KEY: each_request.finding.text,
|
|
207
|
+
JUDGE_REQUEST_REASON_KEY: each_request.reason,
|
|
208
|
+
}
|
|
209
|
+
for each_request in all_requests
|
|
210
|
+
],
|
|
211
|
+
JUDGE_DATA_USER_MESSAGES_KEY: all_user_messages,
|
|
212
|
+
}
|
|
213
|
+
return JUDGE_PROMPT_TEMPLATE.format(
|
|
214
|
+
repository=repository_root,
|
|
215
|
+
user_message_characters=MAXIMUM_USER_MESSAGE_CHARACTERS,
|
|
216
|
+
judge_data=json.dumps(judge_data, indent=JUDGE_DATA_INDENT, ensure_ascii=True),
|
|
217
|
+
)
|
|
218
|
+
|
|
219
|
+
|
|
220
|
+
def _one_line(judge_reason: str) -> str:
|
|
221
|
+
return WORD_SEPARATOR.join(judge_reason.split())
|
|
222
|
+
|
|
223
|
+
|
|
224
|
+
def _one_sentence(judge_reason: str) -> str:
|
|
225
|
+
return _one_line(judge_reason).rstrip(SENTENCE_END_CHARACTERS)
|
|
226
|
+
|
|
227
|
+
|
|
228
|
+
def _apply_verdicts(
|
|
229
|
+
review_state: _ReviewState, all_verdicts: dict[str, JudgeVerdict]
|
|
230
|
+
) -> list[str]:
|
|
231
|
+
all_rejection_lines: list[str] = []
|
|
232
|
+
for each_id in sorted(all_verdicts, key=_finding_number):
|
|
233
|
+
finding = review_state.all_findings.get(each_id)
|
|
234
|
+
if finding is None or finding.status != FINDING_STATUS_OPEN:
|
|
235
|
+
continue
|
|
236
|
+
verdict = all_verdicts[each_id]
|
|
237
|
+
finding.judge_reason = verdict.reason
|
|
238
|
+
if verdict.is_approved:
|
|
239
|
+
finding.status = FINDING_STATUS_DISMISSED
|
|
240
|
+
_log(
|
|
241
|
+
FINDING_DISMISSED_DIAGNOSTIC_TEMPLATE.format(
|
|
242
|
+
finding_id=each_id, reason=_one_line(verdict.reason)
|
|
243
|
+
)
|
|
244
|
+
)
|
|
245
|
+
continue
|
|
246
|
+
finding.status = FINDING_STATUS_UPHELD
|
|
247
|
+
_log(
|
|
248
|
+
FINDING_UPHELD_DIAGNOSTIC_TEMPLATE.format(
|
|
249
|
+
finding_id=each_id, reason=_one_line(verdict.reason)
|
|
250
|
+
)
|
|
251
|
+
)
|
|
252
|
+
all_rejection_lines.append(
|
|
253
|
+
REJECTION_LINE_TEMPLATE.format(
|
|
254
|
+
finding_id=each_id, reason=_one_sentence(verdict.reason)
|
|
255
|
+
)
|
|
256
|
+
)
|
|
257
|
+
return all_rejection_lines
|
|
258
|
+
|
|
259
|
+
|
|
260
|
+
def _record_new_findings(
|
|
261
|
+
review_state: _ReviewState, all_repository_findings: list[tuple[Path, str, str, list[str]]]
|
|
262
|
+
) -> list[str]:
|
|
263
|
+
all_repository_blocks: list[str] = []
|
|
264
|
+
for each_root, each_base_commit, each_turn_commit, each_finding_lines in all_repository_findings:
|
|
265
|
+
all_lines_with_ids: list[str] = []
|
|
266
|
+
for each_line in each_finding_lines:
|
|
267
|
+
finding_id = _next_finding_id(review_state)
|
|
268
|
+
review_state.all_findings[finding_id] = _Finding(
|
|
269
|
+
each_line, str(each_root), each_base_commit, each_turn_commit
|
|
270
|
+
)
|
|
271
|
+
all_lines_with_ids.append(
|
|
272
|
+
FINDING_LINE_WITH_ID_TEMPLATE.format(finding_id=finding_id, finding=each_line)
|
|
273
|
+
)
|
|
274
|
+
all_repository_blocks.append(
|
|
275
|
+
REPOSITORY_FINDINGS_TEMPLATE.format(
|
|
276
|
+
repository=each_root.name,
|
|
277
|
+
findings=FINDING_LINE_SEPARATOR.join(all_lines_with_ids),
|
|
278
|
+
)
|
|
279
|
+
)
|
|
280
|
+
return all_repository_blocks
|
|
281
|
+
|
|
282
|
+
|
|
283
|
+
def _locked_stop_wake(
|
|
284
|
+
session_id: object,
|
|
285
|
+
all_repository_findings: list[tuple[Path, str, str, list[str]]],
|
|
286
|
+
all_verdicts: dict[str, JudgeVerdict],
|
|
287
|
+
) -> str | None:
|
|
288
|
+
review_state = _load_state(session_id)
|
|
289
|
+
all_wake_parts: list[str] = []
|
|
290
|
+
all_rejection_lines = _apply_verdicts(review_state, all_verdicts)
|
|
291
|
+
if all_rejection_lines:
|
|
292
|
+
all_wake_parts.append(REJECTION_LINE_SEPARATOR.join(all_rejection_lines))
|
|
293
|
+
all_repository_blocks = _record_new_findings(review_state, all_repository_findings)
|
|
294
|
+
if all_repository_blocks:
|
|
295
|
+
all_wake_parts.append(
|
|
296
|
+
BLOCK_REASON_TEMPLATE.format(
|
|
297
|
+
all_findings=REPOSITORY_FINDINGS_SEPARATOR.join(all_repository_blocks)
|
|
298
|
+
)
|
|
299
|
+
)
|
|
300
|
+
if all_wake_parts:
|
|
301
|
+
review_state.blocks_in_a_row += 1
|
|
302
|
+
review_state.last_wake_at = time.time()
|
|
303
|
+
elif review_state.blocks_in_a_row < MAXIMUM_BLOCKS_IN_A_ROW:
|
|
304
|
+
review_state.blocks_in_a_row = 0
|
|
305
|
+
_save_state(session_id, review_state)
|
|
306
|
+
return WAKE_PART_SEPARATOR.join(all_wake_parts) or None
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
"""Read the findings
|
|
1
|
+
"""Read the findings, the dismissal requests and the judge's verdicts for the turn review gate.
|
|
2
2
|
|
|
3
3
|
A review either prints a JSON array of findings or prose that names each
|
|
4
4
|
problem by its changed path and line::
|
|
@@ -14,49 +14,92 @@ The last non-empty array of objects with a ``file`` key wins. Its entries count
|
|
|
14
14
|
when ``file`` is a changed path and ``line`` is a number. With no such array,
|
|
15
15
|
each line that opens with a changed path, a colon and a line number counts,
|
|
16
16
|
after an optional list marker, backtick or emphasis mark.
|
|
17
|
+
|
|
18
|
+
The agent contests a finding with a reply line, and the judge answers in JSON::
|
|
19
|
+
|
|
20
|
+
"Dismiss TR-3: the caller retries" -> {"TR-3": "the caller retries"}
|
|
21
|
+
'[{"id": "TR-3", "verdict": "reject", "reason": "nothing retries"}]'
|
|
22
|
+
-> {"TR-3": JudgeVerdict(is_approved=False, reason="nothing retries")}
|
|
23
|
+
|
|
24
|
+
A dismiss line may open with a list marker, quote mark, backtick or emphasis
|
|
25
|
+
mark. The first line for an id wins, and its reason is cut to 2000 characters.
|
|
26
|
+
The judge's whole answer must be one JSON array, inside at most one code fence.
|
|
27
|
+
Each object holds exactly ``id``, ``verdict`` and ``reason`` once each, and
|
|
28
|
+
each id is one the judge was asked about. Any other answer gives None, such as
|
|
29
|
+
an object that repeats a key or a NaN or Infinity value. So an array or a key
|
|
30
|
+
that the agent planted in its reason and the judge echoed never counts. An entry
|
|
31
|
+
counts when its verdict is ``approve`` or ``reject``, its reason is not empty,
|
|
32
|
+
and no other entry names its id. A character in a reason that UTF-8 cannot
|
|
33
|
+
write, such as a lone surrogate, is kept as its escape text.
|
|
17
34
|
"""
|
|
18
35
|
|
|
19
36
|
from __future__ import annotations
|
|
20
37
|
|
|
21
38
|
import json
|
|
22
39
|
import re
|
|
40
|
+
from collections import Counter
|
|
41
|
+
from dataclasses import dataclass
|
|
42
|
+
from typing import TypeGuard
|
|
23
43
|
|
|
24
44
|
from hooks_constants.turn_review_gate_constants import (
|
|
45
|
+
ALL_JUDGE_ENTRY_KEYS,
|
|
46
|
+
DISMISS_LINE_PREFIX_PATTERN,
|
|
25
47
|
FINDING_LINE_PATTERN_TEMPLATE,
|
|
26
48
|
JSON_ARRAY_START,
|
|
49
|
+
JUDGE_ANSWER_CONSTANT_REASON,
|
|
50
|
+
JUDGE_ANSWER_FENCE_PATTERN,
|
|
51
|
+
JUDGE_ANSWER_REPEATED_KEY_REASON,
|
|
52
|
+
JUDGE_APPROVE_VERDICT,
|
|
53
|
+
JUDGE_ID_KEY,
|
|
54
|
+
JUDGE_REASON_KEY,
|
|
55
|
+
JUDGE_REASON_TEXT_ENCODING,
|
|
56
|
+
JUDGE_REASON_UNENCODABLE_HANDLER,
|
|
57
|
+
JUDGE_REJECT_VERDICT,
|
|
58
|
+
JUDGE_VERDICT_KEY,
|
|
27
59
|
JSON_FINDING_FILE_KEY,
|
|
28
60
|
JSON_FINDING_LINE_KEY,
|
|
29
61
|
JSON_FINDING_LINE_TEMPLATE,
|
|
30
62
|
JSON_FINDING_SUMMARY_KEY,
|
|
63
|
+
MAXIMUM_DISMISS_REASON_CHARACTERS,
|
|
31
64
|
NO_FINDINGS_MARKER,
|
|
65
|
+
REASON_EDGE_CHARACTERS,
|
|
66
|
+
REPLY_LINE_SEPARATOR,
|
|
32
67
|
)
|
|
33
68
|
|
|
34
69
|
|
|
70
|
+
@dataclass(frozen=True)
|
|
71
|
+
class JudgeVerdict:
|
|
72
|
+
"""The judge's answer on one dismissal request."""
|
|
73
|
+
|
|
74
|
+
is_approved: bool
|
|
75
|
+
reason: str
|
|
76
|
+
|
|
77
|
+
|
|
35
78
|
def _json_finding_lines(review_text: str, all_changed_paths: list[str]) -> list[str] | None:
|
|
36
|
-
all_filled_arrays =
|
|
79
|
+
all_filled_arrays = _filled_object_arrays(review_text, JSON_FINDING_FILE_KEY)
|
|
37
80
|
if not all_filled_arrays:
|
|
38
81
|
return None
|
|
39
82
|
return _entries_on_changed_paths(all_filled_arrays[-1], all_changed_paths)
|
|
40
83
|
|
|
41
84
|
|
|
42
|
-
def
|
|
85
|
+
def _filled_object_arrays(text: str, required_key: str) -> list[list[dict[str, object]]]:
|
|
43
86
|
json_decoder = json.JSONDecoder()
|
|
44
|
-
|
|
45
|
-
array_start =
|
|
87
|
+
all_object_arrays: list[list[dict[str, object]]] = []
|
|
88
|
+
array_start = text.find(JSON_ARRAY_START)
|
|
46
89
|
while array_start >= 0:
|
|
47
90
|
try:
|
|
48
|
-
decoded_array, next_search_start = json_decoder.raw_decode(
|
|
49
|
-
except ValueError:
|
|
91
|
+
decoded_array, next_search_start = json_decoder.raw_decode(text, array_start)
|
|
92
|
+
except (ValueError, RecursionError):
|
|
50
93
|
decoded_array, next_search_start = None, array_start + 1
|
|
51
|
-
if decoded_array and
|
|
52
|
-
|
|
53
|
-
array_start =
|
|
54
|
-
return
|
|
94
|
+
if decoded_array and _is_object_array(decoded_array, required_key):
|
|
95
|
+
all_object_arrays.append(decoded_array)
|
|
96
|
+
array_start = text.find(JSON_ARRAY_START, next_search_start)
|
|
97
|
+
return all_object_arrays
|
|
55
98
|
|
|
56
99
|
|
|
57
|
-
def
|
|
100
|
+
def _is_object_array(decoded_array: object, required_key: str) -> bool:
|
|
58
101
|
return isinstance(decoded_array, list) and all(
|
|
59
|
-
isinstance(each_entry, dict) and
|
|
102
|
+
isinstance(each_entry, dict) and required_key in each_entry
|
|
60
103
|
for each_entry in decoded_array
|
|
61
104
|
)
|
|
62
105
|
|
|
@@ -100,3 +143,120 @@ def finding_lines(review_text: str, all_changed_paths: list[str]) -> list[str]:
|
|
|
100
143
|
for each_line in review_text.splitlines()
|
|
101
144
|
if any(each_pattern.search(each_line) for each_pattern in all_patterns)
|
|
102
145
|
]
|
|
146
|
+
|
|
147
|
+
|
|
148
|
+
def dismiss_requests(reply_text: str) -> dict[str, str]:
|
|
149
|
+
"""Return the reason of each ``Dismiss TR-<n>: <reason>`` line in a reply, by finding id.
|
|
150
|
+
|
|
151
|
+
Args:
|
|
152
|
+
reply_text: The text of the agent's last reply.
|
|
153
|
+
|
|
154
|
+
Returns:
|
|
155
|
+
The reason for each finding id, cut to its first
|
|
156
|
+
``MAXIMUM_DISMISS_REASON_CHARACTERS``; the first line for an id wins.
|
|
157
|
+
"""
|
|
158
|
+
prefix_pattern = re.compile(DISMISS_LINE_PREFIX_PATTERN)
|
|
159
|
+
all_requests: dict[str, str] = {}
|
|
160
|
+
for each_line in reply_text.split(REPLY_LINE_SEPARATOR):
|
|
161
|
+
prefix_match = prefix_pattern.match(each_line)
|
|
162
|
+
if prefix_match is None:
|
|
163
|
+
continue
|
|
164
|
+
reason = each_line[prefix_match.end() :].strip(REASON_EDGE_CHARACTERS)
|
|
165
|
+
if reason and not reason[0].isspace():
|
|
166
|
+
all_requests.setdefault(
|
|
167
|
+
prefix_match.group(1), reason[:MAXIMUM_DISMISS_REASON_CHARACTERS]
|
|
168
|
+
)
|
|
169
|
+
return all_requests
|
|
170
|
+
|
|
171
|
+
|
|
172
|
+
def _object_without_repeated_keys(all_pairs: list[tuple[str, object]]) -> dict[str, object]:
|
|
173
|
+
all_fields = dict(all_pairs)
|
|
174
|
+
if len(all_fields) != len(all_pairs):
|
|
175
|
+
raise ValueError(JUDGE_ANSWER_REPEATED_KEY_REASON)
|
|
176
|
+
return all_fields
|
|
177
|
+
|
|
178
|
+
|
|
179
|
+
def _refused_constant(constant_name: str) -> object:
|
|
180
|
+
raise ValueError(JUDGE_ANSWER_CONSTANT_REASON)
|
|
181
|
+
|
|
182
|
+
|
|
183
|
+
def _decoded_answer(result_text: str) -> object | None:
|
|
184
|
+
answer_text = result_text.strip()
|
|
185
|
+
fence_match = re.fullmatch(JUDGE_ANSWER_FENCE_PATTERN, answer_text, re.DOTALL)
|
|
186
|
+
if fence_match is not None:
|
|
187
|
+
answer_text = fence_match.group(1)
|
|
188
|
+
try:
|
|
189
|
+
return json.loads(
|
|
190
|
+
answer_text,
|
|
191
|
+
object_pairs_hook=_object_without_repeated_keys,
|
|
192
|
+
parse_constant=_refused_constant,
|
|
193
|
+
)
|
|
194
|
+
except (ValueError, RecursionError):
|
|
195
|
+
return None
|
|
196
|
+
|
|
197
|
+
|
|
198
|
+
def _is_verdict_array(
|
|
199
|
+
decoded_answer: object, all_requested_ids: list[str]
|
|
200
|
+
) -> TypeGuard[list[dict[str, object]]]:
|
|
201
|
+
return (
|
|
202
|
+
isinstance(decoded_answer, list)
|
|
203
|
+
and bool(decoded_answer)
|
|
204
|
+
and all(
|
|
205
|
+
isinstance(each_entry, dict)
|
|
206
|
+
and set(each_entry) == ALL_JUDGE_ENTRY_KEYS
|
|
207
|
+
and each_entry[JUDGE_ID_KEY] in all_requested_ids
|
|
208
|
+
for each_entry in decoded_answer
|
|
209
|
+
)
|
|
210
|
+
)
|
|
211
|
+
|
|
212
|
+
|
|
213
|
+
def _entry_verdict(all_entry_fields: dict[str, object]) -> JudgeVerdict | None:
|
|
214
|
+
verdict_word = all_entry_fields[JUDGE_VERDICT_KEY]
|
|
215
|
+
reason = all_entry_fields[JUDGE_REASON_KEY]
|
|
216
|
+
if (
|
|
217
|
+
verdict_word in (JUDGE_APPROVE_VERDICT, JUDGE_REJECT_VERDICT)
|
|
218
|
+
and isinstance(reason, str)
|
|
219
|
+
and reason.strip()
|
|
220
|
+
):
|
|
221
|
+
return JudgeVerdict(
|
|
222
|
+
verdict_word == JUDGE_APPROVE_VERDICT, _encodable_text(reason.strip())
|
|
223
|
+
)
|
|
224
|
+
return None
|
|
225
|
+
|
|
226
|
+
|
|
227
|
+
def _encodable_text(reason: str) -> str:
|
|
228
|
+
return reason.encode(
|
|
229
|
+
JUDGE_REASON_TEXT_ENCODING, JUDGE_REASON_UNENCODABLE_HANDLER
|
|
230
|
+
).decode(JUDGE_REASON_TEXT_ENCODING)
|
|
231
|
+
|
|
232
|
+
|
|
233
|
+
def judge_verdicts(
|
|
234
|
+
result_text: str, all_requested_ids: list[str]
|
|
235
|
+
) -> dict[str, JudgeVerdict] | None:
|
|
236
|
+
"""Return the judge's verdict for each requested finding its answer covers.
|
|
237
|
+
|
|
238
|
+
Args:
|
|
239
|
+
result_text: The ``result`` text of the judge run.
|
|
240
|
+
all_requested_ids: The finding ids the judge was asked about.
|
|
241
|
+
|
|
242
|
+
Returns:
|
|
243
|
+
The verdicts by finding id. An id the answer names twice gets none. None when
|
|
244
|
+
the whole answer, inside at most one code fence, is not one strict JSON array
|
|
245
|
+
of objects with exactly an id the judge was asked about, a verdict and a
|
|
246
|
+
reason, each named once.
|
|
247
|
+
"""
|
|
248
|
+
decoded_answer = _decoded_answer(result_text)
|
|
249
|
+
if not _is_verdict_array(decoded_answer, all_requested_ids):
|
|
250
|
+
return None
|
|
251
|
+
all_entry_counts = Counter(each_entry[JUDGE_ID_KEY] for each_entry in decoded_answer)
|
|
252
|
+
all_verdicts: dict[str, JudgeVerdict] = {}
|
|
253
|
+
for each_entry in decoded_answer:
|
|
254
|
+
entry_id = each_entry[JUDGE_ID_KEY]
|
|
255
|
+
entry_verdict = _entry_verdict(each_entry)
|
|
256
|
+
if (
|
|
257
|
+
isinstance(entry_id, str)
|
|
258
|
+
and entry_verdict is not None
|
|
259
|
+
and all_entry_counts[entry_id] == 1
|
|
260
|
+
):
|
|
261
|
+
all_verdicts[entry_id] = entry_verdict
|
|
262
|
+
return all_verdicts
|