Parity vs Claude Code

Caliban tracks feature parity with Claude Code in a living matrix. This page summarises the current state by theme. The full matrix — including per-row notes and ADR cross-references — lives at docs/parity-gap-matrix.md in the repository.

Legend: ✅ parity · 🟡 partial · 🔴 not yet

Theme summary

A — Permissions & safety ✅

Rule grammar (allow/ask/deny + globs), all six permission modes, the auto-mode classifier, the TUI Ask modal, OS-level sandbox (macOS Seatbelt + Linux bubblewrap), and the full caliban perms CLI with TOML writeback and audit log are all shipped. See ADRs 0020, 0029, 0032, and 0045.

B — Hooks & extensibility ✅

All hook event types (tool, session, compact, config, cwd, file, subagent, permission), hook decision protocol, and plugin packaging are shipped. The mcp/prompt/agent handler types are v1 stubs; per-subagent hook inheritance lands with the fleet spec.

C — Memory & checkpointing ✅

Three-tier prompt prefix, CLAUDE.md ancestor walk + @-imports, auto-memory, claudeMdExcludes, auto-checkpoint per prompt, /rewind, MicroCompact janitor, and tool-result size cap with overflow persistence are all shipped.

D — Configuration / settings ✅

Layered settings (managed > user > project > local), /config interactive editor, live reload, apiKeyHelper pool, and schema validation are shipped (ADR 0026 + 0045). TOML is the primary write format; JSON is accepted on read.

E — TUI ergonomics 🟡

Status bar, mouse scroll, transcript viewer, @file attach, ! shell escape, external editor (Ctrl+G), Ctrl+O transcript dump, background bash (Ctrl+B), image/vision input, permission Ask modal, and reverse history search are shipped. Notable gaps: vim editing mode (🔴), slash-menu typeahead (🟡 partial), multi-line input (🟡 partial), and voice dictation (🔴).

F — Built-in tools ✅

Bash, Edit, Glob, Grep, Read, Write, WebFetch, TodoWrite, Skill, AgentTool, NotebookEdit, MultiEdit, WebSearch, and background-bash are shipped. PowerShell tool and ToolSearch / WaitForMcpServers (relevant once MCP is fully real) are 🔴.

G — Sub-agents ✅

In-process AgentTool, git worktree isolation, background agent fleet (caliband daemon), per-agent memory dir, hook inheritance, and supervisor daemon are all shipped (ADR 0037).

H — MCP ✅

Config validation, real spawn/handshake, stdio + HTTP/SSE + streamable-HTTP transports, per-server permission scoping, /mcp slash, OAuth PKCE flow, elicitation, and resource references are shipped (ADR 0023).

I — Model router & providers ✅

Purpose-keyed routing, fallback chains, hedging, circuit breakers, capability filtering, Anthropic/OpenAI/Ollama/Google/Bedrock/Vertex providers, and effort levels are shipped. Azure Foundry is 🔴; extended-thinking toggle is 🟡 partial.

J — Headless / CI ✅

-p / --print mode, all output formats (text/json/stream-json), input formats, --max-turns, --max-budget-usd, --bare, --json-schema, --include-partial-messages, and --include-hook-events are shipped. GitHub Actions workflow and devcontainer feature are 🔴 (separate sub-projects).

K — Observability / cost ✅

tracing instrumentation, /context, /usage, /compact, proactive autocompact, prompt cache markers, cost tracking, OpenTelemetry export, and the custom status line are shipped. --debug / --debug-file is 🟡 partial. The feedback survey is 🔴.

L — Output styles ✅

All four built-in output styles (Default, Proactive, Explanatory, Learning) and custom output-style files are shipped (ADR 0031).

M — Slash command coverage 🟡

Core commands (/plan, /memory, /skills, /quit, /clear, /help, /init, /context, /usage, /compact, /config, /hooks, /mcp, /model, /effort, /resume, /cost, /export, /rewind, /doctor, /login, /logout, /status) are shipped. Theme customisation and skill-dependent commands (/code-review, /run, /verify, /batch) are 🔴.

N — Long-tail surfaces 🔴

IDE extensions (VS Code / Cursor / JetBrains), GitHub App, claude.ai/code web, iOS app, Slack, Remote Control, Channels, Routines, Deep links, and Teleport are all 🔴. These are parked until terminal/CLI parity is reached.

Notable gaps