Feat/deep agent - #62
Draft
Hercilio1 wants to merge 5 commits into
Draft
Conversation
A product-agnostic runtime for autonomous agents built on deepagents, in a new top-level baseapp_ai_langkit/deep_agents package. Nothing else in langkit imports it and it declares no models, so a project that does not install the extra is unaffected. It sits outside base/ deliberately: base/ is imported by every consumer and must stay installable without deepagents. - builder.build_deep_agent: the single construction point. The limit governors are built here and cannot be removed by a caller. Structured output is optional, for a surface whose result is written by its tools rather than returned as one validated object. - node.LangGraphDeepAgent: an LLMNodeInterface implementation, so a deep agent's behavioural prompt is admin-editable through the runner tables like any other node. It carries no tracing backend; run telemetry goes to an on_run_complete hook instead of being written to a run tree the runtime knows how to find. - limits: soft and hard per-run caps on model calls, cumulative tokens and wall clock, plus the structured-output and planning-loop governors. The soft tier takes a finalize_tools set, so a surface with no answer tool can still be forced to finish: emptying the tool list only works when langchain re-appends an answer tool outside it. - skills: conditional composition of SKILL.md playbooks from components, behind a SkillSource protocol. FilesystemSkillSource reads a directory on disk and re-reads it per run, so an edit lands without a restart. - telemetry: which skills a run loaded, and prompt-cache accounting. - testing.ScriptedChatModel: replays a fixed script so a whole graph run is deterministic, and records what it was bound with on each call. 103 tests, all DB-free.
Two independent bugs, either one enough to 500 the prompt editor. Both hit every runner, so no runner's prompts were editable through the admin at all. - LLMRunnerNodeUsagePromptInline.get_fields read `obj.usage_prompt`, but Django builds an inline's fieldsets once per formset rather than once per row, so `obj` is the change view's object: the root LLMRunner, not this inline's parent node. The attribute does not exist there and the `DoesNotExist` the code caught could never fire. It now asks whether a usage-prompt row exists under the object being edited, handling LLMRunner and LLMRunnerNode both, so a nested_admin that starts passing the immediate parent keeps working. - LLMRunnerAdmin mixed unfold's ModelAdmin with nested_admin's. The two do not compose: rendering any nested inline raises `AttributeError: 'InlineAdminFormSet' object has no attribute 'initial_forms'` out of django.contrib.admin.helpers, because unfold hands the nested formset where Django expects a plain one. Verified against unfold and django-nested-admin 4.1.6 on Django 5.2 by removing unfold and nothing else: every runner change page went from 500 to 200. unfold.contrib.inlines does not help, being about nonrelated inlines. The page now renders with stock admin widgets inside unfold's site chrome, which is the trade: a prompt editor that works, styled less well. The regression test lives in the consuming project (red-backend, apps/parser/tests/test_langkit_runner_admin.py), because this repo's own test settings need a database service and an app that is not installed here.
`build_deep_agent` has always accepted `extra_middleware` and appended it after the governors, but the node never passed anything, so a subclass had no way to add its own. That only matters for a surface with no `response_format`, whose result is written by its tools rather than returned as one validated object: middleware is then the only place a run can be inspected before it ends, since there is no answer to intercept. Redstone's parser is the first such consumer, and it uses this to send a run that is about to finish with work outstanding back to the model once. Additive, defaulting to no middleware, so every existing surface builds the same agent.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.