mirror of
https://github.com/pewdiepie-archdaemon/odysseus.git
synced 2026-10-08 16:02:20 +02:00
bind artifact completion to required output path
This commit is contained in:
@@ -1455,6 +1455,36 @@ def execution_targets_required_artifact(tool_name, arguments, required_artifacts
|
|||||||
return any(str(path or '').rstrip('/') in serialized for path in required_artifacts)
|
return any(str(path or '').rstrip('/') in serialized for path in required_artifacts)
|
||||||
|
|
||||||
|
|
||||||
|
def artifact_completion_tool_schemas(offered_schemas, required_artifacts):
|
||||||
|
"""Bind the sole artifact writer to the runner-declared output file.
|
||||||
|
|
||||||
|
This applies only after native execution enters its reserved artifact-write
|
||||||
|
phase. A JSON-Schema ``const`` gives the provider the exact destination
|
||||||
|
instead of relying on it to recover the path from a long conversation.
|
||||||
|
Multiple outputs and directory targets stay unconstrained because choosing
|
||||||
|
one of those paths requires model intent.
|
||||||
|
"""
|
||||||
|
targets = [str(path or '').strip().rstrip('/') for path in required_artifacts]
|
||||||
|
targets = [path for path in targets if path and Path(path).suffix]
|
||||||
|
if len(targets) != 1 or len(tuple(required_artifacts or ())) != 1:
|
||||||
|
return offered_schemas
|
||||||
|
target = targets[0]
|
||||||
|
bound = copy.deepcopy(offered_schemas)
|
||||||
|
for schema in bound:
|
||||||
|
function = schema.get('function') or {}
|
||||||
|
if canonical(function.get('name')) != 'write_file':
|
||||||
|
continue
|
||||||
|
properties = (function.get('parameters') or {}).get('properties') or {}
|
||||||
|
path_schema = properties.get('path')
|
||||||
|
if not isinstance(path_schema, dict):
|
||||||
|
continue
|
||||||
|
path_schema['const'] = target
|
||||||
|
path_schema['description'] = (
|
||||||
|
f'Write this exact required artifact path: {target}'
|
||||||
|
)
|
||||||
|
return bound
|
||||||
|
|
||||||
|
|
||||||
def protocol_safe_tool_calls(calls):
|
def protocol_safe_tool_calls(calls):
|
||||||
"""Keep malformed model calls out of the next provider request."""
|
"""Keep malformed model calls out of the next provider request."""
|
||||||
safe_calls = copy.deepcopy(calls)
|
safe_calls = copy.deepcopy(calls)
|
||||||
@@ -4334,6 +4364,10 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
|
|||||||
canonical(schema['function']['name']), 0
|
canonical(schema['function']['name']), 0
|
||||||
) < round_number
|
) < round_number
|
||||||
]
|
]
|
||||||
|
if artifact_write_phase and not successful_artifact_write:
|
||||||
|
round_offered = artifact_completion_tool_schemas(
|
||||||
|
round_offered, required_artifacts,
|
||||||
|
)
|
||||||
research_choice = None
|
research_choice = None
|
||||||
if not required_artifacts:
|
if not required_artifacts:
|
||||||
round_offered, research_choice, _ = bounded_research_tool_policy(
|
round_offered, research_choice, _ = bounded_research_tool_policy(
|
||||||
|
|||||||
@@ -5,7 +5,7 @@ import jsonschema
|
|||||||
import pytest
|
import pytest
|
||||||
import re
|
import re
|
||||||
|
|
||||||
from src.clean_agent_preview import conversation, readonly_call, preview_call_allowed, evaluate_preview_call, authorized_write_families, compact_schemas, normalize_preview_function_args, normalize_preview_call_args, private_browser_dom_batch, private_browser_state_transition, private_browser_success_repeat_limit, stream_preview, denied_response, execution_has_write_effect, requests_mutation, claims_completion, recent_successful_write_families, scope_preview_contract, multimodal_image_count, attachment_reference_count, active_document_context_message, active_email_context_message, targets_active_editor, active_editor_whole_draft_request, active_editor_suggestion_request, scope_active_editor_contract, native_execution_limits, interactive_execution_limit, runtime_required_artifacts, execution_targets_required_artifact, document_suggestions_event, document_suggestion_quality_error, required_read_tool_choice, required_active_editor_tool_choice, sealed_read_arguments, email_identifier_error, requested_item_limit, contract_item_limit, notes_terminal_response, documents_terminal_response, shell_listing_terminal_response, shell_output_terminal_response, direct_shell_output_request, ui_panel_terminal_response, ui_toggle_state_result, calendar_terminal_response, memory_terminal_response, tasks_terminal_response, task_list_requires_synthesis, skills_terminal_response, cookbook_servers_terminal_response, prior_short_answer_for_no_tool_summary, prior_collection_repeat_answer, prior_failed_operation_answer, prior_cookbook_server_answer, prior_workspace_path_answer, prior_web_source_answer, bounded_web_evidence_answer, inherit_referential_read_arguments, normalized_search_intent, requested_web_source_links, web_source_links, requested_web_link_limit, preserve_requested_web_recency, ground_referenced_note_content, note_search_result_empty, note_referent_error, research_referent_error, private_browser_open_url, private_browser_effective_url, web_fetch_observation_is_boilerplate, broad_current_web_request, record_tool_execution, align_structured_tool_history, provider_request_messages, provider_compatible_tool_choice_request, offered_tool_alias, dependent_write_prerequisite_error, bounded_research_tool_policy, retrieved_source_urls, serialize_required_email_attachment_chain
|
from src.clean_agent_preview import conversation, readonly_call, preview_call_allowed, evaluate_preview_call, authorized_write_families, compact_schemas, normalize_preview_function_args, normalize_preview_call_args, private_browser_dom_batch, private_browser_state_transition, private_browser_success_repeat_limit, stream_preview, denied_response, execution_has_write_effect, requests_mutation, claims_completion, recent_successful_write_families, scope_preview_contract, multimodal_image_count, attachment_reference_count, active_document_context_message, active_email_context_message, targets_active_editor, active_editor_whole_draft_request, active_editor_suggestion_request, scope_active_editor_contract, native_execution_limits, interactive_execution_limit, runtime_required_artifacts, execution_targets_required_artifact, artifact_completion_tool_schemas, document_suggestions_event, document_suggestion_quality_error, required_read_tool_choice, required_active_editor_tool_choice, sealed_read_arguments, email_identifier_error, requested_item_limit, contract_item_limit, notes_terminal_response, documents_terminal_response, shell_listing_terminal_response, shell_output_terminal_response, direct_shell_output_request, ui_panel_terminal_response, ui_toggle_state_result, calendar_terminal_response, memory_terminal_response, tasks_terminal_response, task_list_requires_synthesis, skills_terminal_response, cookbook_servers_terminal_response, prior_short_answer_for_no_tool_summary, prior_collection_repeat_answer, prior_failed_operation_answer, prior_cookbook_server_answer, prior_workspace_path_answer, prior_web_source_answer, bounded_web_evidence_answer, inherit_referential_read_arguments, normalized_search_intent, requested_web_source_links, web_source_links, requested_web_link_limit, preserve_requested_web_recency, ground_referenced_note_content, note_search_result_empty, note_referent_error, research_referent_error, private_browser_open_url, private_browser_effective_url, web_fetch_observation_is_boilerplate, broad_current_web_request, record_tool_execution, align_structured_tool_history, provider_request_messages, provider_compatible_tool_choice_request, offered_tool_alias, dependent_write_prerequisite_error, bounded_research_tool_policy, retrieved_source_urls, serialize_required_email_attachment_chain
|
||||||
from src.tool_capabilities import capabilities_for_tool
|
from src.tool_capabilities import capabilities_for_tool
|
||||||
|
|
||||||
|
|
||||||
@@ -454,6 +454,53 @@ def test_runtime_required_artifacts_includes_runner_declared_directory():
|
|||||||
) == ('/tmp_workspace/results',)
|
) == ('/tmp_workspace/results',)
|
||||||
|
|
||||||
|
|
||||||
|
def test_artifact_completion_schema_binds_single_required_file_without_mutating_source():
|
||||||
|
source = [{
|
||||||
|
'type': 'function',
|
||||||
|
'function': {
|
||||||
|
'name': 'write_file',
|
||||||
|
'description': 'Write a file.',
|
||||||
|
'parameters': {
|
||||||
|
'type': 'object',
|
||||||
|
'properties': {
|
||||||
|
'path': {'type': 'string'},
|
||||||
|
'content': {'type': 'string'},
|
||||||
|
},
|
||||||
|
'required': ['path', 'content'],
|
||||||
|
},
|
||||||
|
},
|
||||||
|
}]
|
||||||
|
|
||||||
|
bound = artifact_completion_tool_schemas(
|
||||||
|
source, ('/workspace/output.html',),
|
||||||
|
)
|
||||||
|
|
||||||
|
path_schema = bound[0]['function']['parameters']['properties']['path']
|
||||||
|
assert path_schema['const'] == '/workspace/output.html'
|
||||||
|
assert '/workspace/output.html' in path_schema['description']
|
||||||
|
assert 'const' not in source[0]['function']['parameters']['properties']['path']
|
||||||
|
|
||||||
|
|
||||||
|
def test_artifact_completion_schema_does_not_bind_directory_or_multiple_outputs():
|
||||||
|
source = [{
|
||||||
|
'type': 'function',
|
||||||
|
'function': {
|
||||||
|
'name': 'write_file',
|
||||||
|
'parameters': {
|
||||||
|
'type': 'object',
|
||||||
|
'properties': {'path': {'type': 'string'}},
|
||||||
|
},
|
||||||
|
},
|
||||||
|
}]
|
||||||
|
|
||||||
|
assert artifact_completion_tool_schemas(
|
||||||
|
source, ('/workspace/results/',),
|
||||||
|
) == source
|
||||||
|
assert artifact_completion_tool_schemas(
|
||||||
|
source, ('/workspace/a.txt', '/workspace/b.txt'),
|
||||||
|
) == source
|
||||||
|
|
||||||
|
|
||||||
def test_runtime_required_artifacts_does_not_promote_inputs_to_outputs():
|
def test_runtime_required_artifacts_does_not_promote_inputs_to_outputs():
|
||||||
assert runtime_required_artifacts(
|
assert runtime_required_artifacts(
|
||||||
'Read /workspace/input/data.json and write /workspace/results/report.json.',
|
'Read /workspace/input/data.json and write /workspace/results/report.json.',
|
||||||
|
|||||||
Reference in New Issue
Block a user