Modernize toward latest Claude Code (prompts, hooks, skills, agents, UI)

.gitattributes

@@ -0,0 +1,22 @@
+# Normalize all text files to LF in the repository.
+# Working trees check out LF on every platform to keep diffs clean.
+* text=auto eol=lf
+
+# Shell scripts must always be LF.
+*.sh text eol=lf
+
+# Binary assets — never normalize or diff as text.
+*.png binary
+*.jpg binary
+*.jpeg binary
+*.gif binary
+*.ico binary
+*.webp binary
+*.woff binary
+*.woff2 binary
+*.ttf binary
+*.otf binary
+*.pdf binary
+*.gz binary
+*.zip binary
+*.wasm binary

CHANGELOG.md

@@ -7,6 +7,89 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### UI
+
+- **TUI restyled toward Claude Code's visual language.** Warm terracotta
+  accent (`#D97757`) replaces DeepSeek blue as the primary accent
+  (composer border + `>` prompt, action bullets, headings, live tool
+  markers), with a darker terracotta substituted in light mode for
+  contrast. Assistant actions and completed tools now lead with the `⏺`
+  bullet; tool results hang under a `⎿` tree connector. The composer is a
+  rounded box with a terracotta `> ` prompt (wrap math and cursor
+  position track the prompt gutter). The startup empty state is a rounded
+  `✻ Welcome to DeepSeek TUI` banner showing model + cwd + quick hints.
+
+### Added
+
+- **`/review` effort tiers.** `/review [--effort low|medium|high|max] <target>`
+  (a bare leading level also works, e.g. `/review high file.rs`) injects a depth
+  directive into the review skill — low/medium report few high-confidence
+  findings, high broadens coverage with flagged uncertainty, max is exhaustive.
+  Mirrors Claude Code's `/code-review` levels. Defaults to medium.
+
+- **Output styles (`concise`, `explanatory`).** The personality overlay
+  system is generalized toward Claude Code's output styles: alongside the
+  tone-shaping Calm/Playful, two behavior-shaping styles ship — `concise`
+  (answer-first, minimal) and `explanatory` (teaches as it works) — plus a
+  `Personality::from_str`/`as_str` API with aliases. Composed into the prompt
+  and tested. (Runtime selection of a non-default style via config/picker is a
+  pre-existing dormant gap — even Playful isn't yet user-selectable — and lands
+  separately.)
+
+- **File-defined sub-agent roles (`~/.deepseek/agents/<name>.md`).** Drop a
+  Markdown file with YAML-ish frontmatter (`name`, `description`, optional
+  `model`/`tools`) plus a body, and that body becomes the sub-agent's system
+  prompt when the matching name is spawned as the `subagent_type` or `role` —
+  overriding the built-in role intro while still appending the shared output
+  contract. Brings Claude Code's file-defined agents to DeepSeek TUI on top of
+  the existing built-in roles (general/explore/plan/review/implementer/
+  verifier). New `agents` module with a loader, lister, and context renderer.
+
+- **`memory-consolidate` bundled skill** — a reflective maintenance pass
+  over the user memory file: classify entries, merge duplicates, fix
+  contradicted facts, prune stale one-offs, and keep the file short
+  enough to stay useful on every turn. Auto-installed alongside
+  `skill-creator` and triggered when the user asks to consolidate/clean
+  up memory. The system-skill installer now bundles a *list* of skills
+  with a recorded known-set marker, so newly-bundled skills install on
+  upgrade while user-deleted skills stay deleted.
+- **Four new lifecycle hook events — `stop`, `subagent_stop`,
+  `pre_compact`, `notification`** — bringing the hook surface in line
+  with Claude Code's most-used lifecycle hooks. `stop` fires when the
+  assistant finishes a turn (with `DEEPSEEK_TURN_STATUS` =
+  `completed`/`interrupted`/`failed`), the natural place to hang
+  auto-format, auto-test, or a custom notifier. `subagent_stop` fires
+  when a sub-agent completes. `pre_compact` fires as context
+  compaction begins (manual `/compact` or auto), before the summary
+  replaces the transcript — use it to snapshot first. `notification`
+  mirrors user-facing desktop notifications to external notifiers.
+  All are configurable via `[[hooks.hooks]]` and listed by `/hooks`.
+
+### Documentation
+
+- **`CLAUDE_CODE_GAP_ANALYSIS.md`** — a grounded comparison of the
+  agent against the latest Claude Code across prompts, tools,
+  skills/subagents, and lifecycle, with a prioritized modernization
+  roadmap.
+
+### Changed
+
+- **System prompt: behavioral alignment with current Claude Code
+  norms** — added Engineering Altitude (solve at the right level; no
+  over/under-engineering), Acting Safely (confirm irreversible &
+  outward-facing actions; look before overwrite), Editing Code: Fit In
+  (match surrounding style, reuse first, no drive-by churn),
+  clickable `path:line`/PR references, scheduling discipline, and
+  sharper plan-mode etiquette. Additions land above the locale closer
+  and keep the prompt byte-stable for unchanged workspaces.
+
+### Repository
+
+- **Enforce LF line endings via `.gitattributes`** — the working tree
+  had drifted to CRLF on every text file, rendering diffs unreviewable
+  (467 files of pure line-ending noise). Diffs now reflect real
+  changes only.
+
 ## [0.8.32] - 2026-05-12
 
 A "more useful tools" release. v0.8.31 made the tool surface

CLAUDE_CODE_GAP_ANALYSIS.md

@@ -0,0 +1,212 @@
+# DeepSeek TUI vs. Claude Code (latest) — Gap Analysis & Modernization Plan
+
+> Date: 2026-06-02 · Baseline: DeepSeek TUI v0.8.32 · Reference: Claude Code (latest, 2026)
+> Scope: behavioral prompts, tools/capabilities, skills/subagents, lifecycle (hooks/memory).
+
+## TL;DR
+
+DeepSeek TUI is **not** a greenfield port — it is already a feature-rich Claude Code
+analog built on the codex-rs architecture. The bulk of "Claude Code class" capability
+is present: ~50 tools, 28 built-in slash-command modules plus user commands, a
+`SKILL.md`-format skills system, an 8-event hook system, MCP client **and** server,
+sub-agents with `fork_context`, plan/agent/yolo modes, approval policies, statusline,
+checkpoints/undo, scheduling (`automation_*`), goal mode, share, and memory.
+
+So "modernize to match latest Claude Code" is **not** about adding the big primitives —
+they exist. It is about **closing the long tail**: a handful of lifecycle hooks, a more
+structured memory model, named/file-defined agent types, richer output styles, and a set
+of **behavioral-prompt** refinements that bring the agent's *defaults* in line with
+current Claude Code norms (altitude, irreversible-action confirmation, faithful
+reporting, schedule/loop discipline). The prompt work is the highest leverage-to-risk
+ratio and should land first.
+
+Legend — Effort: **S** (hours) · **M** (1–3 days) · **L** (week+). Risk reflects blast
+radius in a ~210k-LOC production codebase.
+
+---
+
+## 1. Behavioral prompts & agent loop  ← highest leverage, lowest risk
+
+`crates/tui/src/prompts/base.md` (217 lines) is mature. It already absorbed the
+`PROMPT_ANALYSIS.md` recommendations (RLM patterns, Sub-Agent Strategy, Parallel-First
+Heuristic, Verification Principle, Composition Pattern, bumped thinking budget). The
+remaining gaps are *current* Claude Code behavioral norms not yet encoded:
+
+| # | Claude Code norm | Present? | Gap | Effort/Risk |
+|---|---|---|---|---|
+| 1.1 | **Altitude** — solve at the right level of abstraction; don't over-fit to the literal ask or over-engineer | partial (decomposition only) | No explicit "right altitude / don't over-engineer / match the request's scope" guidance | S / low |
+| 1.2 | **Confirm hard-to-reverse & outward-facing actions** (deletes, force-push, publishing to external services, sending messages) unless durably authorized | partial (`approvals/auto.md` mentions "pause before destructive") | Not a first-class principle in `base.md`; no "publishing is irreversible / approval doesn't transfer across contexts" framing | S / low |
+| 1.3 | **Match surrounding code style** (comment density, naming, idiom) | no | Not stated; agent may impose its own style | S / low |
+| 1.4 | **Look before you overwrite/delete** — inspect the target; if it contradicts how it was described, surface that instead of proceeding | no | Missing safety reflex | S / low |
+| 1.5 | **Scheduling discipline** — when to *offer* `/automation` (`/schedule`/`/loop` analog) vs. just finish the work now | no | Scheduling tools exist but no guidance on *when* to reach for them; risk of over-offering | S / low |
+| 1.6 | **Clickable references** — render `file:line` and PR/issue refs as links | partial (OSC-8 in TUI, #374) | Prompt doesn't instruct the model to *emit* `path:line` consistently | S / low |
+| 1.7 | **Plan-mode etiquette** — investigate fully, then present a plan for approval before editing; don't ask "is the plan ready?" | partial (`modes/plan.md` is terse) | Could mirror Claude Code's ExitPlanMode discipline more explicitly | S / low |
+
+**Recommendation:** a single surgical pass on `base.md` + `modes/*.md` + `approvals/*.md`.
+All additive, prefix-cache-aware (append below volatile boundary), independently
+testable via the existing snapshot tests (`crates/tui-core/tests/snapshot.rs`,
+`crates/tui/src/prompts.rs`). **Do this first.**
+
+---
+
+## 2. Lifecycle: hooks
+
+`crates/tui/src/hooks.rs` supports 8 events: `SessionStart`, `SessionEnd`,
+`MessageSubmit`, `ToolCallBefore`, `ToolCallAfter`, `ModeChange`, `OnError`, `ShellEnv`.
+
+Mapping to Claude Code's hook set:
+
+| Claude Code hook | DeepSeek equivalent | Status |
+|---|---|---|
+| SessionStart / SessionEnd | same | ✅ |
+| UserPromptSubmit | `MessageSubmit` | ✅ |
+| PreToolUse / PostToolUse | `ToolCallBefore` / `ToolCallAfter` | ✅ |
+| Notification | — | ❌ missing |
+| Stop (agent finished a response) | — | ❌ missing |
+| SubagentStop | — | ❌ missing (sub-agent completion is an internal sentinel, not a user hook) |
+| PreCompact | — | ❌ missing |
+| (n/a) | `ModeChange`, `OnError`, `ShellEnv` | ➕ DeepSeek extras (keep) |
+
+**Gap:** no `PreCompact` (can't snapshot/guard before compaction), no `Stop`/`SubagentStop`
+(can't trigger notifications, auto-format, or auto-test on turn completion), no
+`Notification` event. The `Stop` hook in particular is the most-used Claude Code hook
+(desktop notifications, auto-lint).
+
+**Recommendation:** add `PreCompact`, `Stop`, `SubagentStop`, `Notification` variants to
+`HookEvent` and wire dispatch sites. The enum + config + executor are already factored,
+so this is mostly wiring. **Effort: M / Risk: medium** (touches the agent loop and
+compaction path).
+
+---
+
+## 3. Lifecycle: memory
+
+Today: `note` tool + `crates/tui/src/memory.rs` + `/memory` command + `tools/remember.rs`
++ `tools/recall_archive.rs`. Functional, but free-form.
+
+Claude Code latest uses a **structured** file-based memory: one fact per file under a
+`memory/` dir, each with frontmatter (`name`, `description`, `metadata.type ∈
+{user,feedback,project,reference}`), a `MEMORY.md` index (one line per memory, loaded
+each session), `[[wikilink]]` cross-references, a recall-relevance discipline, and a
+`consolidate-memory` skill to dedupe/prune.
+
+| Aspect | DeepSeek | Claude Code | Gap |
+|---|---|---|---|
+| Storage | free-form notes | one-fact-per-file + frontmatter | structure |
+| Index | — | `MEMORY.md` loaded per session | no always-loaded index |
+| Typing | — | user/feedback/project/reference | no taxonomy |
+| Cross-links | — | `[[name]]` | none |
+| Maintenance | — | `consolidate-memory` skill | none |
+| Recall caveat | — | "memories reflect when written; verify before acting" | not encoded |
+
+**Recommendation:** evolve `memory.rs` toward the structured schema + a session-loaded
+index, and ship a `memory-consolidate` bundled skill. **Effort: M / Risk: low** (additive;
+keep `note` as the write path).
+
+---
+
+## 4. Sub-agents & agent types
+
+Today: `agent_spawn`/`agent_result`/`agent_wait`/… with `fork_context`,
+`subagent_output_format.md`, worktree support already present in
+`tools/subagent/mod.rs` and `snapshot/`. Strong.
+
+Claude Code latest adds **named, file-defined agent types**: `.claude/agents/*.md` (and
+bundled ones) with frontmatter (`name`, `description`, `tools`, `model`), surfaced to the
+model as selectable `subagent_type`s with curated tool/model scoping, plus
+`run_in_background` and "continue an existing agent via SendMessage."
+
+| Aspect | DeepSeek | Claude Code | Gap |
+|---|---|---|---|
+| Spawn ad-hoc child | ✅ | ✅ | — |
+| `fork_context` / prefix-cache reuse | ✅ | ✅ (worktree) | — |
+| **Named reusable agent types** (file-defined, tool/model-scoped) | ❌ | ✅ | library of roles (Explore, Plan, …) |
+| Background agents | partial (`task_shell`) | ✅ (`run_in_background` agents) | distinct background *agents* |
+| Continue an existing agent | `resume_agent` / `agent_send_input` | ✅ SendMessage | ~parity |
+| Worktree isolation | ✅ present | ✅ | verify it's exposed to model |
+
+**Recommendation:** introduce `~/.deepseek/agents/*.md` (+ bundled `explore`, `plan`,
+`code-review` agent defs) and let the model pick a `subagent_type`. **Effort: M / Risk:
+low.**
+
+---
+
+## 5. Skills
+
+Today: `SKILL.md` + frontmatter (ported from codex), `load_skill` tool, `/skills`,
+bundled `skill-creator` + `v4-best-practices`. Format already matches Claude Code's.
+
+**Gaps vs. Claude Code latest:**
+- **Plugin packaging** — Claude Code bundles skills+commands+agents+hooks+MCP as
+  installable *plugins* with a marketplace. DeepSeek has only minimal plugin refs.
+  (**Effort: L / Risk: medium** — defer.)
+- **Skill discovery surfacing** — verify skills are advertised to the model with
+  descriptions for autonomous triggering (Claude Code injects the skill list each
+  session). Appears present via the `## Skills` prompt section; confirm.
+- **Bundled library breadth** — Claude Code ships docx/pdf/pptx/xlsx and workflow skills.
+  DeepSeek ships 2. Optional content expansion. (**S–M, low risk.**)
+
+---
+
+## 6. Output styles vs. personalities
+
+Today: `prompts/personalities/{calm,playful}.md` — *tone* overlays.
+
+Claude Code "output styles" are broader — they can re-shape the agent's whole operating
+posture (e.g. explanatory/teaching modes, terse-vs-verbose, role changes), not just tone.
+
+**Gap:** only two, tone-scoped. **Recommendation:** generalize "personalities" into
+"output styles" with a couple of behavior-shaping presets (e.g. `concise`, `explanatory`)
+and user-definable styles under `~/.deepseek/styles/*.md`. **Effort: S–M / Risk: low.**
+
+---
+
+## 7. Smaller UX/tool gaps
+
+| Item | Claude Code | DeepSeek | Note |
+|---|---|---|---|
+| Structured multi-choice question | `AskUserQuestion` (2–4 options, multi-select) | `request_user_input` (free-form) | add structured option UI — **S** |
+| Out-of-scope task flagging | spawn-task chip | — | nice-to-have — **S** |
+| Session chapters / ToC | mark-chapter + floating ToC | cycle dividers (#395) | partial; could add explicit chapters — **S** |
+| Code-review effort levels | `/code-review low…ultra` | `/review` + `review` tool | add effort tiers — **S** |
+| Deferred-tool search | ToolSearch | `tool_search_tool_regex/bm25` | ✅ parity |
+| Background bash | `run_in_background` | `task_shell_start` | ✅ parity |
+| Checkpoints/rewind | rewind | `/undo`, `revert_turn`, snapshots | ✅ parity |
+| Web search/fetch | WebSearch/WebFetch | `web_search`/`fetch_url`/`web.run` | ✅ parity |
+
+---
+
+## Prioritized roadmap
+
+**Wave 1 — Behavioral prompt alignment (S, low risk) — DO FIRST.**
+Section 1: altitude, irreversible/outward-facing confirmation, match-surrounding-style,
+look-before-overwrite, scheduling discipline, clickable refs, plan-mode etiquette.
+Verifiable via existing prompt snapshot tests. One reviewable commit.
+
+**Wave 2 — Lifecycle hooks (M, medium).** Add `Stop`, `SubagentStop`, `PreCompact`,
+`Notification` to `HookEvent` + dispatch. Biggest single capability gain (auto-test/
+auto-format/notify on completion).
+
+**Wave 3 — Structured memory + consolidate skill (M, low).** Schema + `MEMORY.md` index +
+`memory-consolidate` skill.
+
+**Wave 4 — Named agent types (M, low).** `~/.deepseek/agents/*.md` + bundled roles.
+
+**Wave 5 — Output styles generalization (S–M, low).** Behavior-shaping presets +
+user-defined styles.
+
+**Wave 6 — Smaller UX (S each).** Structured AskUserQuestion, review effort tiers,
+explicit chapters, out-of-scope flagging.
+
+**Deferred — Plugin/marketplace system (L).** Largest, lowest marginal value given skills/
+commands/agents already work individually.
+
+---
+
+## What NOT to change
+
+- The codex-derived crate architecture — sound and well-factored.
+- DeepSeek-specific prompt content (V4 characteristics, prefix-cache economics, RLM,
+  language-mirroring rules) — correct and tailored; not Claude Code's to dictate.
+- DeepSeek-only hook extras (`ModeChange`, `OnError`, `ShellEnv`) — keep.
+- Terminal-output formatting guidance — terminal constraints are real.