DeepScientist v1.5.15

Overview

DeepScientist v1.5.15 is a broad maintenance and workflow-hardening release.

Compared with v1.5.14, this release does not only adjust one metrics view. It also updates the prompt and skill contracts, hardens bootstrap and stage routing, improves connector/runtime behavior, expands compatibility, and refreshes several user-facing surfaces.

Major Functional Changes

1. Baseline comparison and metrics surfaces

This release reworks how baseline information is exposed in the quest workspace.

Changes:

• Metrics Overview is now focused on the current active baseline plus main-experiment traces.
• Baseline Compare is separated out as its own comparison surface for confirmed baselines and variants.
• The internal metric builders, quest service, daemon API, and web UI were updated to support this split.
• A dedicated quest-level compare route is now exposed for baseline comparison data.

Why this matters:

• it is clearer which baseline the current experiment line is being judged against
• it is easier to inspect other confirmed baselines and variants without mixing them into the active timeline
• the workspace is less likely to blur "current comparator" and "all known baselines"

2. Baseline confirmation now preserves richer metric contracts

This release strengthens the baseline contract rules so a baseline is less likely to be reduced to only one headline scalar when richer comparable metrics already exist.

Changes:

• artifact.confirm_baseline(...) is now backed by stronger metric-surface handling.
• Prompt and skill rules now explicitly say that primary_metric is only the headline metric.
• Multi-task, multi-split, multi-variant, or multi-metric baseline surfaces should be preserved in metrics_summary, metric_contract, and baseline_variants.
• Existing structured baseline-side contracts should be reused instead of being replaced by a thinner hand-written summary when possible.

Why this matters:

• quests with richer baseline packages keep more of the comparison information needed later
• users can compare variants and task-level performance more accurately
• future runs are less likely to silently optimize toward an over-compressed baseline target

3. Connector-facing interaction messages keep the full answer

This release separates the real user-facing interaction body from the preview/summary layer.

Changes:

• actual connector-facing interaction text is preserved in full
• preview text is generated separately for compact displays and summaries
• artifact interaction journaling records the full message and the preview layer distinctly
• the system prompt now explicitly warns agents not to manually turn the real user message into a shortened preview

Why this matters:

• connector users are less likely to receive milestone or baseline messages that were already truncated with …
• journaling and delivery behavior are easier to reason about
• user-facing updates can stay readable without throwing away decision-critical detail

4. Prompt, SOP, and execution-discipline updates

This release also continues the recent prompt and skill hardening work.

Changes in this version line include:

• stronger SOP and workflow guidance in the system prompt
• clearer bash_exec-first terminal execution rules
• stronger prompt redlines against using the wrong shell path
• clearer baseline-discipline language in the baseline skill and prompt/MCP docs

Why this matters:

• stage behavior is easier to keep aligned with the intended workflow
• future quests are less likely to drift into shell usage or baseline handling that conflicts with system policy
• maintainers have a clearer prompt/skill contract to inspect when workflow quality regresses

5. Bootstrap, routing, and runtime behavior

This release contains several fixes around early quest routing and runtime stability.

Changes:

• bootstrap quest stage routing is more stable
• install path and quest lock import behavior were hardened
• connector/workflow/runtime behavior was refreshed in the broader runtime stack
• baseline lineage and connector delivery behavior received additional hardening

Why this matters:

• new quests are less likely to drift into the wrong stage too early
• route changes around baseline and experiment setup are less fragile
• runtime state is more predictable during startup and resume flows

6. Connector and platform improvements

This version line also includes broader compatibility and connector updates.

Included work in this range:

• Windows compatibility improvements
• connector expansion work for additional connector surfaces
• Weixin delivery retry improvements for the ret=-2 text-send failure class
• connector delivery and workflow behavior updates across the stack

Why this matters:

• more environments can run the product cleanly
• connector delivery is more resilient
• the release is not only about the web UI; it also improves operator-facing delivery paths

7. User-facing UI and content fixes

This release also bundles several user-visible fixes and cleanup items.

Included work in this range:

• improved quest loading behavior
• tutorial locale alignment fixes so tutorial content stays in the chosen UI language more consistently
• landing hero interaction stabilization
• removal of temporary internal maintenance files and image artifacts from the public tree

Why this matters:

• startup and workspace loading are more coherent
• language-switch behavior is less surprising
• the public repository and packaged output are cleaner

User-Visible Impact

If you are using the web workspace, the most visible changes are:

• clearer separation between active baseline tracking and full baseline comparison
• better retention of detail in connector-facing milestone messages
• improved loading and routing behavior for quests
• cleaner language behavior in tutorial-related UI flows

If you are maintaining prompts, skills, or quest workflows, the most important changes are:

• stronger baseline confirmation rules in the system prompt
• stronger bash_exec execution discipline
• more explicit prompt/skill guidance for preserving rich comparison contracts
• more stable bootstrap routing behavior

Upgrade Notes

• Old quests that already stored only one aggregate baseline metric will still show only that aggregate metric until a richer baseline contract is recorded for that quest.
• The npm package for this release is published as @researai/deepscientist@1.5.15.
• This release includes both behavior changes and contract hardening, so prompt/skill consumers should read the updated system prompt and baseline skill expectations rather than assuming only UI changes landed.

Commit Themes Since v1.5.14

Representative changes in this release line include:

• feat: split baseline compare and interaction previews
• fix: stabilize stage routing for bootstrap quests
• fix: harden install path and quest lock import
• feat: front-load bash_exec redlines in system prompt
• feat: enforce bash_exec-only terminal execution
• feat: strengthen prompt SOP and metrics overview surfaces
• Update connector, workflow, and runtime behavior
• Default Codex to yolo and unblock built-in MCP tools
• fix: stabilize landing hero interactions
• fix: keep tutorial language aligned with current ui locale
• feat: improve quest loading and release v1.5.15
• 兼容 windows，修复中英文翻译，新增 connector
• Harden baseline lineage and connector delivery
• fix: retry weixin text sends after ret=-2