3.7 KiB
Investigation working set v8
Historical investigation note: new v10 requests restore full evidence and note/question history; v8/v9 selection is retained only for frozen-request compatibility. Other phase optimizations remain active. See investigation-full-history-v10.md.
New project-backed contexts use investigation_project_v8. Stored v1-v7 contexts and requests keep their original encoding. The canonical case, evidence, full hypothesis state, query provenance, content hashes and final diagnosis inputs remain unchanged. docs/architecture.md is unchanged.
Selection and encoding
The model receives full payloads for the selected hypothesis's previous result, every result whose budget action is still unsettled (including other hypotheses), and up to four explicitly recalled evidence IDs. Payloads use the existing lossless v7 tables/code blocks. No extra LLM summarization is performed.
Other archived evidence is represented in one evidence_index columns/rows table: ID, collector hypothesis, content scope and capture descriptors. Each descriptor retains the exact query/predicate, origin/snapshot, result coverage/truncation/paging/EOF metadata and item count. Missing payloads are explicitly unknown, never empty results or absence evidence. Non-query feedback is discoverable through its evidence ID and content scope, and can be recalled in full.
Other hypotheses carry only title, mechanism, confidence and disposition. The active hypothesis retains its complete working control state. Full alternatives and histories remain archived and available to final diagnosis; this view does not alter the hypotheses themselves.
Recall protocol
next = {"kind":"recall", "evidence_refs":["E000001"]} requests one to four unique current-analysis evidence IDs. Unknown IDs reject the entire decision before assessment settlement, controller updates or execution. Old prompt versions do not accept this action.
The controller applies the normal assessment of the previous result, then saves recall IDs scoped to that hypothesis. Its next context carries exact archived payloads. Scheduling may investigate another hypothesis first; recall selection survives this and process restart. The next non-recall decision clears the selection. A repeat recall replaces it.
Recall creates no evidence, external query, budget reservation, novelty or reward. A previous pending research result is settled once via the usual assessment. Internal recall still consumes an investigation iteration, so existing iteration limits prevent unlimited retrieval loops. The model must retrieve hidden contents before making new claims about them; an inventory entry is not proof of its unseen contents.
Offline comparison
On the interrupted Bug 17 run's 18 stored contexts, user-message size including output schema falls from 1,349,730 characters (v7) to 1,107,519 (v8): 17.9% further reduction. The last context falls from 97,855 to 70,819 characters: 27.6%. This simulates no recall selections and preserves all then-unsettled results. Actual future recall frequency, billed tokens and quality require a live run; these are not monetary savings claims.
No provider request was made for this measurement. Current case/entity source material and active-hypothesis notes are retained; they are separate future optimization candidates.
Evaluation
For the post-fix Bug 17 fixture, an evidence-grounded inconclusive result routed to human review is an acceptable analysis outcome. A technical validation failure, provider block or interruption is not the same result. Neither the fixture nor the controller should force a particular diagnosis or inject the known historical solution.