Roadmap

Roadmap

PromptPilot's roadmap holds two principles:

1. The SLM stays the harness brain — it does not become the coder.
2. When the SLM is uncertain, the correct action is passthrough.

Shipped

The pieces below are in main and covered by the test suite.

• v2 control plane — JSON ExecutionSpec with route ∈ {clarify, answer, passthrough, act} plus intent, scope, target_files, risk, memory_record. v1 prose envelope kept as fallback parser. (prpt/core/spec.py, prpt/normalizers/slm_openai_v2.py)
• Four interchangeable judges — MaxHaikuJudge, CodexCliJudge, AnthropicApiJudge, OpenAiJudge, auto-detected in priority order. (prpt/judges/judge.py)
• Hybrid auth pattern — API-key SLM + subscription LLM, ~4,500× subsidy ratio measured on chain5. (Hybrid Mode, Benchmarks)
• Handoff / restart workflow — prpt checkpoint, prpt bootstrap, prpt restart for collapsing heavy sessions to handoff.md and resuming fresh. (prpt/handoff.py)
• Session memory with referential gate — recent turns prepended; --gate-session classifier skips history when prompt is self-contained. Default-on after N=5 chain1 retest showed +60% success / −28.7% cps for WITH-session.
• Onboarding subcommands — prpt setup (one-time, with smoke test), prpt doctor (re-check, no install), prpt install-hook for both Claude Code and Codex.
• Tool-output compression hook — regex-based PostToolUse compressor for pytest / grep / git diff / installer logs. (.codex/hooks/compress_tool_output.py)
• Compliance posture documentation — side-by-side breakdown vs the OpenClaw / OpenCode pattern enforced April 2026. (Comparison)
• Wiki publishing automation — scripts/publish_wiki.sh keeps the wiki mirrored from docs/. (Wiki Publishing)

In progress

• Honest measured benchmarks — chain1, chain4, chain5 numbers are landed; broader workload coverage (Django / Rails / JS repos) not yet measured.
• Route-accuracy fixture set — labelled prompts to score clarify / answer / passthrough / act decisions, not just per-call cost.
• Preservation recall metrics — programmatic checks for file paths, test names, commands, flags, symbols, stack frames, and explicit constraints.

Planned

• Cross-provider / model-size A/B for the SLM harness (Haiku vs gpt-5.4-nano vs gpt-5.4-mini).
• Surface high-risk transformations to the user before invoking the coding agent.
• More worked passthrough and compression examples.
• Easier telemetry inspection during replay (current prpt stats is a flat JSONL summary).

Non-goals

• Replacing frontier coding models with the SLM.
• Maximizing token reduction when context may be lost.
• Letting the SLM make irreversible code changes.
• Token handling for subscription auth (we invoke the official binary; the credential never enters our process).

---

See also: Benchmarks · Telemetry and Replay · Wiki Publishing