Skip to content

BEHAVIOR_PARITY.md

Latest commit

 

History

History
202 lines (172 loc) · 27.2 KB

BEHAVIOR_PARITY.md

File metadata and controls

202 lines (172 loc) · 27.2 KB
Raw

Behavior Parity — DeepCode vs Claude Code

历史快照（Claude 兼容矩阵）：本文件记录既有兼容目标，不再作为当前完成度或安全保证。当前事实、已知分叉与 Codex 对齐路线见 CODEX_ALIGNMENT_PLAN.md。

This document tracks where DeepCode's behavior aligns with, deviates from, or deliberately enhances Claude Code. It grows alongside the codebase. Last updated reflects what main contains.

Legend: ✅ matches · 🟡 matches with caveats · 🔄 deferred · ⚠️ deliberately differs · 🆕 DeepCode-only addition

2026-06 accuracy pass (audited against code). Several per-row tags below had drifted behind the code. This pass reconciles the Slash-command, CLI-flag, and Tools tables with the actual source. Treat the milestone tags (M3c, M8, …) as historical notes, not current status. Highlights now landed on main: CLI -C / --cd (Codex parity, PR #148); the /diff, /release-notes, and /bug (alias /feedback) slash commands (PR #150); --resume / --continue / --fork-session wired to real session resume (PR #153); the /init 3-phase REPL flow; and the CLI /effort table reading its numbers from EFFORT_PARAMS (PR #147); --permission-mode wired as a true --mode alias (PR #159); and --bare / --no-plugins wired. The Tools table (de-staled in PR #151) was re-verified row-by-row this pass — all markers hold.

Slash commands (30+ in Claude Code, ~32 shipped in DeepCode)

Command Claude Code DeepCode Status
/help ✓ ✓ ✅
/clear ✓ ✓ ✅
/exit / /quit ✓ ✓ ✅
/status / /doctor ✓ ✓ ✅
/model ✓ ✓ ✅ DeepCode constrains to deepseek-* (model picker doesn't show foreign providers)
/mode ✓ ✓ ✅
/effort ✓ ✓ 🟡 — CLI prints the tier table (numbers from EFFORT_PARAMS SSOT); switch via /effort <tier>; arrow-key selector is GUI-only (M6)
/cost / /usage ✓ ✓ ✅
/context ✓ ✓ ✅
/config ✓ ✓ 🟡 — dumps merged settings + /config set <key> <value> (dotted keys, JSON values) writes user settings; no full arrow-key editor
/resume ✓ ✓ ✅ — lists recent sessions; /resume <id|number> switches the live session in-REPL; --resume <id> / -r at launch
/init ✓ ✓ ✅ — interactive 3-phase REPL flow (scan → draft → approve-write AGENTS.md)
/mcp ✓ ✓ ✅
/add-dir ✓ ✓ ✅ — persists to permissions.additionalDirectories; folded into the sandbox's filesystem.allowWrite by every host (CLI REPL/headless + app-server, so desktop/VS Code/LSP too)
/todos ✓ ✓ ✅ — reads <sessionDir>/todos.json written by TodoWrite tool
/plugins ✓ ✓ ✅ — lists wired plugins + contributed hook events + warnings (M5.2)
/compact ✓ ✓ ✅ — manual /compact + automatic threshold trigger in the agent loop
/diff ✓ ✓ ✅ — git diff + untracked files in the working tree (PR #150)
/btw ✓ ✓ 🟡 — queues a "by the way" context note the agent sees with your next message (no turn fired); exact Claude Code behavior may differ
/recap ✓ ✓ ✅ — provider-summarized recap of the session so far
/rewind ✓ ✓ ✅ — 5 ops (code/conversation/both/summarize-from/up-to); Esc Esc bound
/voice ✓ ✓ 🟡 — local whisper.cpp dictation on CLI (/voice: ffmpeg/sox → transcribe → pre-fill) and desktop (🎙 composer button, ffmpeg). Fully on-device; real-mic round-trip needs local verification
/teleport ✓ ✗ 🔄 M8
/desktop ✓ ✗ 🔄 M6
/background ✓ ✓ ✅ — runs a prompt as a background sub-agent via the session TaskManager (alias /bg); agent-started TaskCreate tasks appear too
/batch ✓ ✗ 🔄 — batch-of-prompts not yet wired (use /background per prompt)
/tasks ✓ ✓ ✅ — lists this session's background tasks; /tasks <id> shows one's status + output
/plan ✓ ✗ 🔄 — set via /mode plan in DeepCode
/login / /logout ✓ ✓ ✅ — /logout clears creds + exits; /login stores a new key (next launch)
/export ✓ ✓ ✅ — writes the conversation to a markdown file
/bug (alias /feedback) ✓ ✓ ✅ — prints a prefilled GitHub issue link (model/mode/effort in the body)
/upgrade ✓ ✓ ✅ — prints version + npm i -g deepcode-cli@latest (also the deepcode upgrade subcommand)
/pr_comments ✓ ✓ ✅ — gh pr view comments for the current branch's PR
/review ✓ ✗ (skill avail) 🟡 — via Skill tool
/security-review ✓ ✗ (skill avail) 🟡 — via Skill tool
/schedule ✓ ✗ (skill avail) 🟡
/loop ✓ ✗ (skill avail) 🟡
/terminal-setup ✓ ✗ 🔄
/vim ✓ ✓ ✅ — toggles Vim mode (persists to ~/.deepcode/keybindings.json)
/keybindings ✓ ✓ (read-only) 🟡 — Claude Code opens/creates the keybindings config; ours lists bindings (edit ~/.deepcode/keybindings.json manually)
/agents ✓ ✓ ✅ — lists sub-agents from .deepcode/agents/
/hooks ✓ ✓ ✅ — lists hooks configured in settings.json
/skills ✓ ✓ ✅ — lists built-in + user + project skills
/permissions ✓ ✓ (read-only) 🟡 — shows rules + default mode (interactive editor deferred)
/privacy-settings ✓ ✓ ✅ — summarizes local data locations + what's sent to the DeepSeek API (read-only)
/migrate-installer ✓ ✗ 🔄
/release-notes ✓ ✓ ✅ — prints the latest CHANGELOG.md entry

Settings.json fields

Tracked in packages/core/src/config/types.ts. Roughly 50 fields total; most are stubbed (schema-known but not actively consumed). M2 loads + merges all of them. Subsystems consume as they ship.

Specific deviations:

  • ⚠️ model enum: only deepseek-chat / deepseek-reasoner / deepseek-v4-flash / deepseek-v4-pro (DeepSeek constraint). Aliases align.
  • 🆕 update.* for Mac client auto-update via electron-updater (Claude Code has its own equivalent).
  • 🟡 managed/MDM policy layer: explicit non-goal v1 per §0.2 — schema reserved.

Hook events

Event Claude Code DeepCode Status
PreToolUse ✓ ✓ ✅
PostToolUse ✓ ✓ ✅
Stop ✓ ✓ ✅ — fires when agent loop ends (any reason)
SubagentStop ✓ ✓ ✅ — fires when a Task sub-agent finishes
PreCompact ✓ ✓ ✅ — fires through compaction event bus
PostCompact ✓ ✓ ✅
SessionStart ✓ ✓ ✅
SessionEnd ✓ ✓ ✅
UserPromptSubmit ✓ ✓ ✅
Notification ✓ ✓ ✅ — REPL fires on turn-end (awaiting input)

Hook handler types

Type Claude Code DeepCode Status
command ✓ ✓ ✅ — JSON-on-stdin contract, JSON-on-stdout decoded
http ✓ ✓ ✅ — fetch POST, response.text → stdout; allowedHttpHookUrls enforced
prompt ✓ ✓ ✅ — synthesizes additionalContext (no exec)
mcp_tool ✓ ✓ ✅ — agent loop resolves mcp__<server>__<tool> from the live registry
agent ✓ ✓ ✅ — runs a named sub-agent (re-entrancy-guarded)
if field ✓ ✓ ✅ permission-rule syntax filter

Modes

Mode Claude Code DeepCode Status
default ✓ ✓ ✅
acceptEdits ✓ ✓ ✅
plan ✓ ✓ ✅
auto (LLM classifier) ✓ ✓ ✅ — classifyAutoMode wired in the tool dispatcher
dontAsk ✓ ✓ ✅
bypassPermissions ✓ ✓ ✅ sandbox still enforces

Memory system

  • ✅ CLAUDE.md ↔ DEEPCODE.md (different filename, same semantics)
  • ✅ ~/.claude/CLAUDE.md ↔ ~/.deepcode/DEEPCODE.md
  • ✅ Hierarchical walk cwd → root
  • ✅ @-import recursion (≤ 4 hops, cycle detection)
  • ✅ AGENTS.md auto-import (cross-tool interop)
  • ✅ .deepcode/rules/*.md (path-scoped frontmatter deferred to M4)
  • 🔄 Auto-memory (~/.deepcode/projects/<repo>/memory/) — schema defined, agent-side writes M4+

MCP

  • ✅ stdio transport
  • ✅ http (Streamable HTTP) / sse transports
  • ✅ list_tools + call_tool with mcp__<server>__<tool> qualification
  • ✅ /mcp slash + auto-connect from settings + per-server enabled/disabled
  • ✅ alwaysLoad: false opt-out defers a server's tools behind ToolSearch
  • ✅ static headers + dynamic headersHelper auth
  • ✅ OAuth 2.0 (authorization-code + PKCE, dynamic client registration) via oauth: true; tokens persist under ~/.deepcode/mcp-auth/ + auto-refresh
  • ✅ Elicitation (form mode) — server-initiated structured input → host prompt
  • ✅ deepcode mcp serve — expose DeepCode's stateless tools as an MCP server (stdio)
  • ✅ MCP resources — listed on connect; @server:scheme://path refs expanded in prompts
  • ✅ MCP prompts as slash commands — /mcp__<server>__<prompt> [args]

Tools

Tool Claude Code DeepCode Status
Read ✓ ✓ ✅
Write ✓ ✓ ✅
Edit ✓ ✓ ✅
Bash ✓ ✓ ✅ + M3.5 sandbox wrap
Grep ✓ ✓ ✅ via ripgrep
Glob ✓ ✓ ✅ via fs.glob
Skill ✓ ✓ ✅ M5
Task (subagents) ✓ ✅ TaskTool in BUILTIN_TOOLS — spawns a sub-agent
NotebookEdit ✓ ✅ shipped (tools/notebook.ts)
AskUserQuestion ✓ ✅ shipped; returns null in headless
EnterPlanMode / ExitPlanMode ✓ ✅ shipped; also drivable via /mode plan
EnterWorktree / ExitWorktree ✓ ✅ shipped (tools/worktree-tools.ts)
ToolSearch (deferred load) ✓ ✅ installed when MCP tools opt out of eager load
TaskCreate / Monitor / TaskList / TaskGet / TaskOutput / TaskStop / TaskUpdate ✓ ✅ shipped — background tasks (TASK_TOOLS)
CronCreate / CronList / CronDelete ✓ ✅ shipped — launchd-backed scheduler
ScheduleWakeup ✓ ⚠️ not a tool in DeepCode — use CronCreate / deepcode cron
WebFetch ✓ ✅ shipped M3c-rest — 5 MiB cap + abort
WebSearch ✓ ✅ shipped M3c-rest — DDG default + SearXNG
TodoWrite ✓ ✅ shipped M3c-rest — persists in sessionDir

CLI flags

Flag Status
--help / --version ✅
--mode ✅
--permission-mode ✅ — true --mode alias (sets mode; last of --mode/--permission-mode wins), wired in PR #159
--model / --effort ✅
--max-turns ✅
-C / --cd <dir> ✅ — chdir before running (Codex parity); validated eagerly, bad path exits 2
--system-prompt / --append-system-prompt[-file] ✅
--allowedTools / --disallowedTools ✅
--bare ✅ — suppresses the REPL startup banner (scripting / minimal output)
--settings / --agents / --mcp-config / --plugin-dir / --plugin-url 🟡 — --settings <file> is a trusted highest-precedence override layer; --agents/--mcp-config/--plugin-dir/--plugin-url still parsed-only
--no-plugins / --strict 🟡 — --no-plugins skips plugin discovery + wiring; --strict still parsed-only
-p headless ✅ text/json/stream-json, 5 exit codes
--output-format / --json-schema / --include-partial-messages ✅ output-format + json-schema (lightweight top-level validation) + include-partial-messages all implemented (headless.ts)
--resume <id> / --continue / --fork-session ✅ resume by id (picker if no id, -r), most-recent-in-cwd (-c), fork-into-new

What DeepCode adds that Claude Code doesn't have (yet)

Feature Note
sandbox.* field macOS sandbox-exec + Linux bwrap wrapping for Bash tool, opt-in
update.* field electron-updater integration with GitHub Releases for Mac client (Claude Code's update is upstream-specific)
Cat-shaped icon (...what)

This document will be kept current as each PR lands. M9 release pipeline includes a check that fails CI if a new public behavior isn't documented here.