Skip to content

PRD.md

Latest commit

 

History

History
101 lines (78 loc) · 4.74 KB

PRD.md

File metadata and controls

101 lines (78 loc) · 4.74 KB
Raw

DevCLI — Product Requirements

1. Problem

Vibe coders live in Claude Code but the terminal fights them:

  • Prompts vanish. Once the initial prompts are lost, projects become nearly unmaintainable. There's no durable, searchable record of what you asked and why.
  • Agent runs are invisible. Subagents, tool calls, and diffs scroll past as raw text. Review fatigue is real (PR volume up 40–60% in AI teams).
  • Session history is ugly. Claude Code writes JSONL to ~/.claude/projects/<proj>/<session>.jsonl. People built 4+ third-party viewers because raw logs are unreadable.
  • AI terminals are agent-blind. Warp/Wave render blocks but don't understand Claude Code specifically. They also push cloud/login (privacy backlash on HN).
  • No prompt sharing for CC users. Prompt libraries exist (PromptHub, Sourcegraph) but none are wired into the Claude Code loop or shareable as a repo.

2. Solution

A local-first terminal that:

  1. Runs a real shell + Claude Code as a subprocess (real PTY, not a wrapper).
  2. Renders Claude Code's session stream as structured blocks (prompts, tool calls, agent spawns, diffs, todos).
  3. Auto-captures every prompt into a git-versioned vault that's searchable and shareable.
  4. Stays light (Tauri, no Electron) and private (no forced login, no telemetry-by-default).

3. Target users

Persona Need
Solo vibe coder Never lose a prompt; see what the agent did; revert bad diffs fast.
Prompt author / creator Curate reusable prompt-packs, publish/share them with a community.
Small team Share prompt conventions; review agent runs without wading through logs.

Non-goal: enterprise SSO, cloud multi-tenant, or replacing the IDE. Terminal-first.

4. Scope

In (v1)

  • Warp-style block terminal (real PTY, real shell).
  • Claude Code launch + JSONL session tailing → block rendering.
  • Prompt vault: auto-capture, git commit, full-text search, tags.
  • devcli share / devcli import prompt-pack format.
  • Session timeline / replay view.
  • Inline git-diff overlay per agent-touched file + revert.
  • Agent/tool blocks (collapsible) + live todo checklist.
  • Zero-login, all-local (SQLite + git).

Out (v1, later)

  • Cloud sync / accounts (optional add-on later).
  • Multi-agent orchestration / worktree kanban (Conductor territory — phase 3).
  • Mobile companion.
  • Non-Claude agents (Codex, Gemini) — architecture allows, not v1.

5. Key user flows

Capture & reuse a prompt

  1. User types a prompt to Claude Code in DevCLI.
  2. DevCLI records it → .devcli/prompts/<slug>.md + git commit + SQLite index.
  3. User later Cmd-K → searches vault → re-runs or edits a saved prompt.

Review an agent run

  1. Claude spawns a subagent / edits files.
  2. DevCLI collapses the run into an agent block with a diff summary.
  3. User expands → sees per-file before/after → clicks Revert on one hunk.

Share a prompt-pack

  1. User tags a set of prompts #nextjs-setup.
  2. devcli share nextjs-setup → exports a self-contained repo (prompts + metadata).
  3. Another dev devcli import <url> → prompts appear in their vault.

6. Requirements

Functional

  • FR1 Real PTY hosting an interactive shell; Claude Code runs inside.
  • FR2 Tail ~/.claude/projects/**/*.jsonl; parse into typed events.
  • FR3 Every user prompt persisted (file + git + index) within 1 s.
  • FR4 Full-text prompt search < 50 ms on 10k prompts (SQLite FTS5).
  • FR5 Block rendering: command, output, tool-call, agent, diff, todo.
  • FR6 Diff overlay with per-hunk revert (git apply -R).
  • FR7 Prompt-pack export/import (portable folder + manifest).
  • FR8 Session timeline scrubber.

Non-functional

  • NFR1 RAM baseline < 40 MB idle (vs Warp ~150–300 MB). Efficiency is a headline promise.
  • NFR2 Cold start < 500 ms.
  • NFR3 No network calls unless the user opts in. No telemetry by default.
  • NFR4 All data local; nothing leaves the machine without explicit share/sync.
  • NFR5 Terminal throughput within 2x of Alacritty (uses its VTE crate).
  • NFR6 macOS first; Linux second; Windows later.

7. Success metrics

  • Prompts captured per active session (target: ~100% of sent prompts).
  • Vault re-use rate (prompts re-run / prompts saved).
  • Prompt-packs shared/imported.
  • Idle RAM stays under budget (NFR1) — this is the differentiator vs Warp.

8. Risks

Risk Mitigation
Claude Code JSONL format changes Version the parser; tolerate unknown event types; pin to schema, fail soft.
Terminal-emulator complexity Reuse alacritty_terminal crate (don't hand-roll VTE).
"Just another CC GUI" perception Lead with terminal + prompt-vault + local-first; those aren't in one tool yet.
Privacy trust Local-only by default; open-source the core; document zero egress.