v0.8.47 tracker: continuity layer

[Hmbown]

Release theme

v0.8.47 should make long work coherent across time and surfaces. The agent should maintain a small, observable orientation layer, and external clients should attach to the same state/evidence contract instead of inventing parallel worlds.

This is the continuity-layer release.

Product thesis

Long tasks need memory, but memory without provenance becomes superstition. Cross-surface work needs protocol, but protocol without shared state becomes drift. Continuity means source-linked memory plus one auditable contract.

In scope

• A bounded orientation cache for project/session facts with source, timestamp, confidence, and invalidation.
• Context-map views that show which facts are guiding the agent.
• Session continuity improvements for non-interactive and resumed flows.
• ACP/app-server/serve protocol alignment around turns, tool calls, response deltas, cancellation, task state, evidence, and output handles.
• Attach/handoff flows that let IDEs, local dashboards, Feishu/Lark, and HTTP clients observe or continue work safely.
• Auth, permission, and workspace trust boundaries made explicit for every external surface.

Candidate issue/PR clusters

• v0.8.41 PEEK/Aleph-inspired orientation-cache prototype follow-up
• context/cache reports: 输入缓存命中率太低了 #1177, token消耗增大了很多 #743, There still seems to be some problems with cache hits缓存命中方面似乎还是有些问题 #1120, Feature: DeepSeek cache-aware prompt diagnostics and wire payload optimization #1253, Improve prefix cache inspection and warmup #1474
• session continuity: Feature: Support session continuity in non-interactive mode (exec --resume / --session-id) #1530
• global/project context: feat: add DEEPSEEK.md as project context for harness integrations like superpowers #1851, PR feat: add DEEPSEEK.md as project context file #1852, [Feature Request] Support loading a global AGENTS.md (~/.deepseek/AGENTS.md) #1157
• project skills/context loading: PR feat: load /project/.deepseek/skills #1850
• exact context tracking: PR fix(compaction): use tiktoken for exact context tracking and enforce … #1376
• ACP/app-server: ACP: expose tool calls (read_file/exec_shell/write_file) via serve --acp #1092, bug: app-server --stdio thread/message does not stream response_delta output #1490, PR feat(acp): add multi-turn tool execution support to ACP server #1693, PR fix: stream app-server stdio thread events from runtime bridge #1141
• IDE/bridge compatibility: ORCA Lab compatibility: connect DeepSeek-TUI as an ORCA agent connector #1801, Add to acp-registry #1447, Can the ACP protocol be accepted to adapt to Paseo? #889, PR feat: serve-centered local API bridge for Playwright/Tauri integration #1692, PR feat(ide-bridge): add MCP-over-WS IDE bridge with cwd-scoped discovery #1820
• typed permissions/workspace trust: PR feat(tui): route shell and file tool approvals through typed execpolicy rules #1413, PR feat(tui): persist permission rules from approval prompts #1509, feat(execpolicy): add typed persistent permission rules #1186

Review lenses

• Security/sandbox reviewer: does memory and cross-surface handoff preserve trust boundaries and resist prompt injection?
• Expert agent operator: does continuity make delegation and resume better, or does it create stale context that is harder to challenge?
• Systems/tooling builder with taste: is the protocol contract small, stable, and shared instead of duplicated per surface?
• Open-source maintainer: can protocol and memory changes be reviewed, documented, and debugged by contributors?

Acceptance criteria

• Orientation entries are bounded, inspectable, and source-linked.
• Users can see why an entry is still present and when it was last refreshed.
• Stale or contradicted facts have an update/eviction path.
• One documented protocol shape covers state, evidence, output deltas, cancellation, and recovery for at least one external surface.
• Permission/workspace trust behavior is explicit and tested for the chosen surface.

[Hmbown]

Updated scope (2026-05-23 integrated roadmap)

v0.8.47 pairs provider abstractions with cross-surface alignment + VS Code beta:

• Provider abstractions: ProviderKind registry becomes configurable; model selection is provider-agnostic. DeepSeek stays first-class.
• Cross-surface alignment: consistent command names, output formats, and error messages across CLI/TUI/API/web.
• VS Code extension beta (EPIC: VS Code extension scaffold #461-VSC: Chat webview panel #463): scaffold, local runtime detection, chat webview. VSIX attached to GitHub Release (not Marketplace published).
• Tool studio + slash command trackers (Slash commands: spatial workbench routing for sessions and work items #1892, Slash commands: tool studio wiring for tools, MCP, skills, providers, and previews #1891, Slash commands: presence, docs, i18n, and contributor-facing help #1890, Slash commands: PEEK-backed command receipts and continuity #1889, Slash commands: control-plane semantics for agents, jobs, hooks, and recovery #1888, Slash commands: cockpit and truth-surface wiring #1887, Slash commands: registry, dispatcher, help, and palette parity audit #1886): registry, dispatcher, palette audit, cockpit wiring.

See .private/codewhale-integrated-roadmap.md.

[carloslindhstrom-code]

4