Skip to content

HANDOFF.md

Latest commit

 

History

History
87 lines (70 loc) · 5.14 KB

HANDOFF.md

File metadata and controls

87 lines (70 loc) · 5.14 KB
Raw

LabWeft development handoff

简体中文

Confirmed product choices

Local-first. Existing coding agent is the controller. One user conversation, not several user-operated agents. Model/provider-agnostic tool protocol. Stable research-project identity separate from repo/branch/path. Small core with a Skill entry, persistent evidence and a clickable map. User records stay separate from tool source. No default training, code rewrite, public upload or dataset migration.

Store facts, interpretations, human decisions and scope separately. Historical verification (such as bounded hardware checks) constrains new suggestions. New seed is replication; equivalent-looking config is not proof of duplication. Preserve unknowns, human corrections and old snapshots. A scientific finding may be negative; a full factorial experiment table is not a mandatory goal.

Every research version must visibly state what changed, why, relative to which parent/comparator, and what actually happened. Include before/after values, proposal attribution and source evidence; preserve failed and prepared-only versions. Keep these explanations in version details and study tables after imports and exports. See the per-version annotation contract.

Implemented modules

  • research_tree_demo.py, fixtures/research_tree.json: fast portable replay via demo --scenario research-tree. Includes saved synthetic CPU measurements, candidate sources and evidence snapshots, with no new execution or model calls. The Chinese generation/design handoff is docs/RESEARCH_TREE_DEMO.zh-CN.md.

  • autoresearch.py, research_worker.py, backends/shinka.py: optional pinned Shinka backend, frozen task contracts, durable bounded batches, drained stop and resume, native report import. See docs/AUTORESEARCH.md before extending or executing. Model-backed execution and Luna integration still require deployment validation; offline fixtures do not establish search effectiveness.

  • store.py: project identity, immutable text blobs, event replay, local lock, revision conflicts, schema/reference checks, human-protection convention.

  • scan.py: conservative inventory and static Python symbols; JSON/TOML/CSV ingestion; no source execution or historical commit fabrication.

  • reasoning.py: descriptive seed comparisons, source drift, bounded scope check.

  • server.py: read-only loopback service with token and safe offline export.

  • cli.py, rh.py, bootstrap.py: simple entry points and Skill installation.

  • static/: expandable map with inline study tables, relationship view, grouped model-family comparison, decisions, evidence, bilingual interface, sourced research timeline and separate record revisions.

  • skill/: one-controller procedure and semantic patch contract.

  • tests/: deterministic regression tests; demo.py: synthetic showcase.

What should be next

First run a real repository through ordinary Codex, Codex + EvoSkills, and this Skill on the same model, permissions and source material. Evaluate initial reconstruction, then a fresh session with changed results. Do not claim one approach wins until measured. Investigate false duplicate proposals AND false exemptions of necessary experiments.

Before broadening scope, fix problems surfaced by these cases: user-confirmed hardware result with missing original files; renamed experiment; independent new seed; changed effective batch/precision/protocol; stale result folder; shared baseline; custom YAML configs; multi-repository research identity; repeated scans and retained human corrections. Add lightweight future-run capture to reduce historical gaps, without requiring retroactive training reruns.

Then improve import adapters (starting with common actual user layouts), durable job checkpoints for interrupted long scans, record migrations, source reconciliation, component-level UI editing and accessibility improvements for the multilingual UI. Multi-project offline export and project switching are now available; continue cross-agent compatibility tests before optional remote adapters.

The 2026-10-03 main refresh retains the optional Shinka backend, saved research tree replay, and contributor fixes for organization review and run-level lineage. It adds original branding and paired English/Chinese READMEs. Source publication has explicit owner authorization; public GitHub visibility was verified on 2026-10-03.

Do not start by implementing another chat product, GPU scheduler, model gateway, plugin marketplace, ten messaging platforms, vector database or public hosting.

Release gate

Name/namespace and ownership confirmed. License and any future copied code audited. README demo runs from a clean clone. Recorded OS/browser/agent compatibility matrix. Real-project regression fixture with permission to publish, or a faithful synthetic substitute. No private project snapshots, credentials, or hidden paths in release. Source/decision changes demonstrably traceable. Published benchmark protocol and limitations. Verify the confirmed GitHub repository before any integration; publish only after explicit authorization.