Add generalized trace ingestion (loop C) to graph context engine plan#4776
Open
Evanfeenstra wants to merge 2 commits into
Open
Add generalized trace ingestion (loop C) to graph context engine plan#4776Evanfeenstra wants to merge 2 commits into
Evanfeenstra wants to merge 2 commits into
Conversation
Outcome events alone erase the journey: essential read-only context scores as a retrieval miss, and wrong turns/dead ends never become labels. Loop C ingests agent traces through a closed five-verb vocabulary (search/inspect/act/revert/cite) so flows stay emergent — chat→plan→PR→merged and ingest→workflow→benchmark are the same event stream to the engine. Outcome events become plural with optional magnitude (graded outcomes like benchmark scores scale loop A reinforcement), and mandatory retrieval/self provenance breaks the self-fulfilling-label confound while exposing the recall gap. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Test environment is now live. View it at: https://hive-preview-7.sphinx.chat Database expires at: Jul 9, 2026, 9:38 AM UTC |
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Design-doc-only change to
docs/plans/graph-context-engine.md. The existing plan learns only from outcomes (merged PR → touched leaves), which has two label gaps: essential read-only context scores as a retrieval miss, and the agent's wrong turns / dead ends are invisible. This adds loop C — trace ingestion, generalized so the engine never learns what a flow is.What's new in the doc
search,inspect,act,revert,cite). Flow shape is emergent, never declared —chat→plan→task→PR→mergedandingest documents→build workflow→evaluate against legal benchmarklook identical to the engine. Stage names are opaque strings; engine code branching on one is the same bug as a domain node type in engine source.retrieval/selfprovenance on trace records: breaks the self-fulfilling-label confound (agents touch what you show them) and makes the recall gap directly measurable (self-discovered-then-used = what retrieval should have returned).DomainProfilegainstraceIngestandoutcomeEvents[].magnitude; all three instantiations get flow + trace mappings; order-of-work phase 4 ships loop C with loop B (same logging substrate — code-domain tool traces are already persisted); four new success metrics (time-to-context, dead-end rate, self-discovery rate, trace resolution rate).No engine domain-blindness is disturbed: "tool events joined to outcomes" is as domain-generic as the outcome event itself.
🤖 Generated with Claude Code