v0.8.41 — hostability, orientation cache, hardening, cockpit UX, tool-calling accuracy

[Hmbown]

v0.8.41 umbrella

Closes #1849.

Tracks

• Hostability — Tencent Lighthouse + Feishu/Lark smoke + docs
• Orientation cache — native PEEK/Aleph-style context, no MCP dep
• Hardening tail — Windows containment, sync tool cancel, sub-agent/RLM output reliability, doctor-noise cleanup
• Cockpit UX — sidebar + presentation of state so the user feels like they're driving
• Tool-calling spike — strict mode on /beta, prefix-as-repair, prompt-rhythm experiment

Status

Work in progress. Individual slices land here as they complete.

[gemini-code-assist]

Code Review

This pull request updates the CHANGELOG.md for version 0.8.41, adding entries for new features like hostability improvements, orientation caching, and UX enhancements. The reviewer identified several inaccuracies in the changelog, including incorrect UI panel names, inconsistent environment variable prefixes, and the need to categorize bug fixes under a 'Fixed' section rather than 'Added'.

[gemini-code-assist]

The changelog entry for v0.8.41 should be refined for accuracy and consistency with the project's standards:

• Section Grouping: The "Hardening" bullet describes bug fixes (macOS file writes, Windows edge cases, etc.). Following the "Keep a Changelog" format used in previous versions (e.g., line 52), these should be moved to a ### Fixed section.
• Panel Name Accuracy: The "Cockpit UX" entry mentions panels named "Now, Recent, Pending, Budget, Levers". However, the implementation in crates/tui/src/tui/sidebar.rs (lines 48-53) uses "Work, Tasks, Agents, Context". The changelog should use the actual UI labels.
• Environment Variable Wiring: The entry mentions DS_TUI_STRICT_TOOLS. However, this variable is not wired up in crates/tui/src/config.rs, and the DS_ prefix is inconsistent with the DEEPSEEK_ prefix used for all other environment variables in the project. Also, the PR description mentions "Windows containment" which is missing from the changelog entry.
• Formatting: Commands and environment variables like DS_TUI_STRICT_TOOLS and deepseek eval should be wrapped in backticks for consistency with previous entries.

Suggested change

	### Added
	- **Hostability.** Tencent Lighthouse deployment flow, Feishu/Lark smoke-tested integration,
	and companion documentation for self-host and cloud install paths.
	- **Orientation cache.** Native PEEK/Aleph-style workspace context injection — no MCP
	dependency — giving the model a persistent, low-latency map of the project.
	- **Hardening.** Fixes for macOS file writes, Windows input/composer edge cases, sync
	tool-cancel reliability, sub-agent/RLM output robustness, and doctor-noise cleanup.
	- **Cockpit UX.** Sidebar and presentation panels (Now, Recent, Pending, Budget, Levers)
	so the user feels like they're driving the agent, not watching from the back seat.
	- **Tool-calling accuracy.** Strict-mode (`/beta`) tool schema classification and
	adaptation behind `DS_TUI_STRICT_TOOLS`, measured on `deepseek eval`.
	### Added
	- **Hostability.** Tencent Lighthouse deployment flow, Feishu/Lark smoke-tested integration,
	and companion documentation for self-host and cloud install paths.
	- **Orientation cache.** Native PEEK/Aleph-style workspace context injection — no MCP
	dependency — giving the model a persistent, low-latency map of the project.
	- **Cockpit UX.** Sidebar and presentation panels (Work, Tasks, Agents, Context)
	so the user feels like they're driving the agent, not watching from the back seat.
	- **Tool-calling accuracy.** Strict-mode (`/beta`) tool schema classification and
	adaptation behind `DS_TUI_STRICT_TOOLS`, measured on `deepseek eval`.
	### Fixed
	- **Hardening.** Fixes for macOS file writes, Windows input/composer edge cases, sync
	tool-cancel reliability, sub-agent/RLM output robustness, and doctor-noise cleanup.

[Hmbown]

Maintainer review note from the v0.8.41 cockpit pass:

Current local check on work/v0.8.41 fails:

cargo check --locked -p deepseek-tui
error[E0004]: non-exhaustive patterns: `SidebarFocus::Now` and `SidebarFocus::Budget` not covered
  - crates/tui/src/config_ui.rs:972
  - crates/tui/src/tui/sidebar.rs:48

The deeper issue is architectural, not just adding two match arms. The cockpit patch currently adds Now/Budget as new sidebar focus variants and separate render functions, but those functions duplicate existing sidebar/session signal sources and recompute data less accurately. Before this lands:

• Budget should reuse the existing context/cost source path from the Session/Context panel: session.total_conversation_tokens, model context window, compact_threshold, auto_compact, displayed_session_cost_for_currency(app.cost_currency), and app.format_cost_amount(...). No hardcoded $, no message-count token estimate.
• Now should reuse or extract from task_panel_lines, active_tool_rows, recent_tool_rows, and subagent_panel_lines; it should not re-derive tool/subagent status from raw history with a weaker model.
• If SidebarFocus::Now / Budget remain, they must be fully threaded through config UI, settings normalization, docs, tests, focus cycling, and render_sidebar. If they are only cockpit projections, prefer not adding new persisted focus variants yet.
• Auto mode still renders Work/Tasks/Agents/Context, so the cockpit implementation needs to decide whether cockpit is the new auto stack or only focused detail panels.

Suggested acceptance before v0.8.41 merges: one source of truth per datum, cockpit as projection/orientation cache, cargo check --locked -p deepseek-tui green, and focused tests for currency formatting plus the /compact hint.

[Hmbown]

v0.8.41 overnight follow-up summary:

Opened 6 review-only PRs from tonight:

• fix(subagent): keep agent_eval recoverable after a child terminates (#1738) #1928 — fix/cherry-1738-agent-eval-recoverable
• fix(tui): unblock TUI on bare-separator @-mention completion (#1921) #1929 — fix/issue-1921-at-slash-freeze
• fix(tui): hide stale completed tasks from Work sidebar (#1913) #1930 — fix/issue-1913-stale-work-sidebar
• fix(tui): move composer history persistence off the UI thread (#1927) #1931 — fix/issue-1927-enter-stall
• docs(readme): note V4 Pro rates are now permanent (#1923) #1932 — fix/issue-1923-pricing-readme
• fix(tui): filter terminal control sequence fragments from composer (#1915) #1933 — fix/issue-1915-composer-escape-leak-rebased

Community PR harvest verdicts verified locally:

• LAND: fix(skills): parse YAML block scalars in SKILL.md frontmatter #1908, fix(tui): copy transcript selections without visual wraps #1906, fix(tui): suppress verbose CLI logging on Windows alt-screen to prevent TUI leak #1910, fix(engine): report loop guard halt as Failed instead of Completed #1859, fix(grep): respect cancellation token #1839
• SUBSET OF fix(tui): suppress verbose CLI logging on Windows alt-screen to prevent TUI leak #1910: fix(tui): set spillover root in dedup tests for Nix sandbox builds #1844
• DO NOT MERGE: fix(tools): replace cross-await RwLock with Semaphore to prevent deadlock #1856

No merges, tags, publishes, force-pushes, or issue closures performed.

[Hmbown]

v0.8.42 triage note: I did not find remaining unique release-blocker scope here for the current local stabilization branch. The actionable v0.8.41 slices were already harvested or routed in later PRs/issues, and v0.8.42 now has its own release-prep branch with changelog/version bumps.

Recommendation after v0.8.42 lands: close this draft as superseded by the harvested PRs and the current v0.8.42/v0.8.43+ milestone split, unless Hunter wants to preserve it as an archival umbrella.