feat(tools): mission-sandbox — tune retain/observation missions against a validator#2254
Draft
nicoloboschi wants to merge 1 commit into
Draft
feat(tools): mission-sandbox — tune retain/observation missions against a validator#2254nicoloboschi wants to merge 1 commit into
nicoloboschi wants to merge 1 commit into
Conversation
…st a validator
Add hindsight-tools/mission-sandbox, a dev tool for iterating Hindsight's
retain (extraction) and observation (consolidation) missions and verifying
them against an external validator (e.g. the LOCOMO benchmark runner).
A project is `init`-bound to a documents path + API config. You refine a
mission from feedback (+ optional failing examples), then either
`retain apply` (ingest the docs into a new versioned bank `<project>-vN`)
or `observe apply` (clear observations + reconsolidate in place).
`retain check` scores golden-set coverage via chunk-level dry-run
extraction (POST /memories/dry-run-extract) — no re-ingest, nothing stored.
Commands: init / retain {mission,apply,check} / observe {mission,apply} /
inspect / trace / curate / note / status / ui (Next.js viewer).
Draft notes: workspace wiring (root package.json + lockfile) and a CI job
are intentional follow-ups; build artifacts (.next, standalone, dist,
node_modules) and local project data (projects/) are gitignored.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What
Adds
hindsight-tools/mission-sandbox, a dev tool for tuning Hindsight's retain (extraction) and observation (consolidation) missions and verifying them against an external validator (e.g. the LOCOMO benchmark runner). It's a small, opinionated CLI (+ optional Next.js viewer UI).The loop
A project is
init-bound to a documents path + API config. You then iterate:retain mission <proj> --feedback "…" --example "…"(orobserve mission).retain applyingests the docs into a new versioned bank<proj>-vN;observe applyclears + reconsolidates in place.note.retain checkgives a fast inner loop: it scores golden-set coverage of the current mission via chunk-level dry-run extraction (POST /memories/dry-run-extract, the endpoint added in #2205) — no re-ingest, nothing stored.Commands:
init / retain {mission,apply,check} / observe {mission,apply} / inspect / trace / curate / note / status / ui.Notes / follow-ups (why draft)
package.json/lockfile change). The package is self-contained (@vectorize-io/hindsight-mission-sandbox, ownpackage.json);cd hindsight-tools/mission-sandbox && npm installworks. Happy to add the workspace entry + lockfile if we want it installed from root likehindsight-agent-sdk.vitest) following the existing tool pattern..next,standalone,dist,node_modules) and local project data (projects/) are gitignored; only source is committed.Tests
vitestunit tests for the API client (hindsight.test.ts, incl. the dry-run-extract path) and the project store (store.test.ts) — 8 passing. Mission refinement + coverage judging go through Gemini with bounded exponential-backoff retry on transient 5xx (a single 503 used to abort a wholeretain check).