v0.8.41 tracker: hostability, long-session hardening, and orientation cache

[Hmbown]

Release theme

v0.8.41 should pick up what is valuable after v0.8.40 without reopening the release train: hostability, long-session reliability, native context-management improvements, cleaner tool-call ergonomics, and a deliberate maintainer pass over open PRs/issues.

Track 1: Tencent Lighthouse + Feishu/Lark hostability

• Produce a real VM smoke path for deepseek serve --http plus the Feishu/Lark bridge.
• Document minimum env, secrets, port/process expectations, restart behavior, and how to verify webhook delivery.
• Keep the bridge restart/persisted-turn work from v0.8.40 as the baseline; this milestone should prove it in a small production-like setup.

Track 2: PEEK/Aleph-inspired orientation cache

Build this as a native ds-tui/RLM context-management feature, not as an MCP dependency.

Source inspiration:

• PEEK paper: https://arxiv.org/abs/2605.19932
• PEEK reference implementation: https://github.com/zhuohangu/peek
• Aleph reference project: https://github.com/Hmbown/aleph

Initial shape:

• maintain a small, fixed-size, agent-managed orientation cache for long-running work
• distill useful project/session facts from external files, logs, tool outputs, and prior turns
• expose a bounded peek surface to RLM/subagent workflows without dumping the environment into the prompt
• add eviction/update policy and tests so stale facts do not silently dominate future turns
• make the feature observable enough to debug: where entries came from, when they were updated, and why they were kept

Track 3: Fin/tool-agent and FIM-assisted tool-call ergonomics

Use the v0.8.40 Fin/tool-agent lane as the baseline for tool-bound execution work, then explore whether DeepSeek V4 FIM can make structured tool calls feel less brittle without making execution opaque.

Initial shape:

• evaluate FIM as an intent-to-structured-tool-call completion or repair layer, especially for filling missing arguments and normalizing partial tool plans
• keep actual execution on the existing auditable tool schema and approval/cost pipeline
• compare FIM-assisted repair against the current argument repair and schema-sanitize paths using targeted fixtures
• measure whether Fin/tool-agent delegation reduces main-model context clutter, latency pain, and cost for OCR/search/fetch/probe tasks
• keep the main reasoning model's context clean: prose and synthesis in the parent, machine-bound details in bounded executors

Track 4: PR harvest and issue/PR close-out

Do a maintainer-quality pass over open PRs and stale issues before v0.8.41 ships. The goal is to reduce noise and preserve useful contributor work, not to merge everything.

Initial shape:

• identify small PRs that are still relevant, mergeable, and release-testable
• harvest the bounded valuable part of larger PRs when the whole branch is not safe to merge
• close superseded or stale PRs kindly, with a short explanation and a pointer to the shipped fix or remaining milestone
• preserve contributor credit in changelog/release notes for any harvested or merged work
• use issue evidence plus source/test verification; do not rely on PR descriptions alone
• keep unresolved or high-risk work in v0.8.41 only when it has a concrete reproduction path and owner

Current starting candidates:

• feat: add DEEPSEEK.md as project context file #1852 add DEEPSEEK.md as project context file
• feat: load /project/.deepseek/skills #1850 load /project/.deepseek/skills
• fix(tui): set spillover root in dedup tests for Nix sandbox builds #1844 Nix sandbox spillover-root test fix
• fix(grep): respect cancellation token #1839 cancellation token handling for grep/file search
• fix: 修复 macOS 终端标题悬浮居中，改为贴顶布局 #1837 macOS terminal title layout fix
• feat(update): add version check output to doctor and update #1836 doctor/update version check output
• feat(tui): add "Terminal" theme #1831 Terminal theme
• feat: embed user prompt in snapshot labels for readable /restore listing #1798 readable snapshot labels
• fix(file_search): wrap walker in spawn_blocking + 30s timeout #1790 file_search blocking/timeout containment
• feat: configurable auto-compact threshold with Ctrl+L keybinding #1723 configurable auto-compact threshold with Ctrl+L keybinding
• chore(deps): bump portable-pty to 0.9.0 #1701 portable-pty 0.9 dependency bump
• open Windows/process/input issues already assigned to v0.8.41

Close-out review candidates:

• PRs whose fixes were already shipped or partially harvested in v0.8.40 should get grateful maintainer comments and be closed only after source-level verification
• large architecture PRs that overlap with Fin/tool-agent, PEEK/orientation-cache, plugin/tool-registry, or approval-policy tracks should be split into narrow v0.8.41 work items instead of merged wholesale
• conflicting old UI/theme PRs should be resolved by harvesting still-useful pieces, then closing the stale branch with a clear note

Track 5: remaining hardening after v0.8.40

• Windows process/input containment and any remaining terminal edge cases
• synchronous tool cancellation and stuck-tool behavior
• subagent/RLM output robustness beyond the lossy-stdout fix
• docs/doctor messages that reduce duplicate setup reports

Acceptance criteria

• at least one reproducible Tencent Lighthouse + Feishu/Lark smoke run is documented
• orientation-cache prototype is narrow, tested, observable, and can be disabled
• FIM/tool-call exploration produces either a small tested prototype or a written no-go decision with evidence
• Fin/tool-agent behavior has smoke coverage for model selection, thinking-off routing, cost accounting, and close/eval lifecycle
• PR harvest pass lands or closes the easy valuable open PRs, with evidence and grateful maintainer comments
• any harvested PR work preserves contributor credit in changelog/release notes
• release gates pass before tagging/publishing

Explicit non-goals

• do not block v0.8.40 on the orientation-cache or FIM/tool-call work
• do not introduce a new mandatory MCP dependency for the orientation-cache feature
• do not make tool execution invisible or unauditable in pursuit of elegance
• do not merge broad rewrites unless they close a confirmed, release-tested problem

[Hmbown]

Adding composer text selection to the v0.8.41 must-review path. The current inability to highlight/select text in the composer is a basic interaction gap, not polish. Candidate implementation: PR #1726 (feat(composer): mouse + keyboard text selection with copy/cut). Before merge, verify selection does not regress normal typing, mouse scroll, copy/cut, multiline composer movement, or terminal paste behavior.

[Hmbown]

v0.8.41 check-in: current local cargo check --locked -p deepseek-tui fails after the cockpit sidebar work because SidebarFocus::Now and SidebarFocus::Budget are partial enum additions. The immediate compile errors are in config_ui.rs and tui/sidebar.rs.

More important than the enum plumbing: the cockpit should not duplicate existing Session/Context/Tasks/Agents calculations. Treat the sidebar as an orientation-cache projection over existing sources. Budget should reuse the existing token/cost/currency path; Now should reuse/extract the existing task/tool/subagent line helpers. This should be considered a v0.8.41 acceptance item before local rebuild/install.

[Hmbown]

Closing as part of the v0.8.41 milestone wrap. The hostability / long-session / orientation-cache themes that lived here are now distributed across:

• v0.8.43 (truth surface + publish-out) — picked up the 37 open items (25 issues + 12 PRs) that hadn't shipped; tracked under v0.8.43 tracker: truth surface #1877.
• v0.8.45 (control plane + provider parity for open-weight model hosts) — receives any items that surface as cancel/resume / Windows hostability / sub-agent reliability during v0.8.43 triage.

No work is lost; v0.8.41 milestone is being closed as shipped/superseded.