Deeptask 9.3.2

Deeptask 9.3.2

Subagent transcript memory bound (the last gray-screen gap)

9.3.1 bounded only the MAIN conversation transcript. Subagent sessions (parallelSessionMessages) still accumulated verbatim — a 3-day image-reading run pushed the renderer to 5.4GB and system memory to 92% (host stayed at ~1.1GB).

• Per-session projectors: every subagent session now flows through its own boundTranscript projector (tombstones + batched boundary + reset semantics identical to the main conversation). Projectors are released when their session disappears.
• Image budget inside the verbatim window: even within the 800-row tail, only the newest 24 image-bearing rows keep their images payloads; older image rows are cloned without images (text preserved, clone identity stable per ts+source object so messageUpdated replacements rebuild correctly).
• Subagent panel: tombstoned rows are filtered out and the same collapsed-messages banner appears.

Falsification condition: a long subagent image-reading run whose renderer anon memory still grows linearly after 9.3.2 refutes this fix.

Docs

• README.md / README_EN.md now document the multi-level task-splitting folder system (authoritative checklist + <short-title>/ light files + finished/ archive).
• Marketplace descriptions (en/zh-CN) mention the folder system.

Full Changelog: 9.3.1...9.3.2