Skip to content

roadmap.md

Latest commit

 

History

History
171 lines (147 loc) · 9.01 KB

roadmap.md

File metadata and controls

171 lines (147 loc) · 9.01 KB

Roadmap

Last updated: 2026-05-25

This page is the current product roadmap. Older phase-by-phase execution notes live in docs/superpowers/ and in git history; those historical notes are useful for audit trails, but they are not the source of truth for current status.

Current Position

DeepSeekCode is now usable for Linux/macOS dogfooding and repository work:

  • bare deepseek opens the full-screen TUI in a real terminal;
  • deepseek quickstart provides a side-effect-free first-run readiness check with text and JSON output;
  • deepseek chat remains available as the line-oriented REPL;
  • model-backed tasks can read files, apply patches, run shell checks, inspect diffs, and resume from durable runtime state;
  • local runtime, shell-supervisor, background worktree tasks, MCP/ACP surfaces, GitHub Action bridge, dogfood evidence, and release packaging checks all have repeatable gates;
  • PR #16 full CI passed Linux, macOS, and Windows: https://github.com/willamhou/DeepSeekCode/actions/runs/26334525472
  • verified online multi-file external fixture evidence is tracked for the Python invoice, Rust order, and Node task-report samples under .dscode/dogfood/.
  • v0.1.5 ships GitHub Release assets for Linux x64, Linux arm64, macOS x64, macOS arm64, and Windows x64, with a fully green Release Matrix covering release assets, npm publishing, Homebrew tap publishing, and GHCR: https://github.com/willamhou/DeepSeekCode/actions/runs/26380726246
  • v0.1.5 public release archives are verified on hosted Linux x64, Linux arm64, macOS x64, and macOS arm64 by Release Smoke: https://github.com/willamhou/DeepSeekCode/actions/runs/26380981708
  • v0.1.5 npm/npx packages are published and verified through registry lookup, npx @deepseek-code/cli@0.1.5 version, and clean-directory npm install @deepseek-code/cli@0.1.5.
  • the public Homebrew tap is verified on macOS x64 and macOS arm64: https://github.com/willamhou/DeepSeekCode/actions/runs/26380934049

The Linux/macOS local code-agent CLI milestone is effectively established. The remaining work is mostly external evidence depth, installed service proof, and documentation polish.

Near-Term Priorities

1. Release Hardening For Linux/macOS

  • Keep the release matrix green for Linux x64, Linux arm64, macOS x64, macOS arm64, and Windows x64.
  • Use deepseek update release-smoke --version <version> on Linux x64, Linux arm64, macOS x64, and macOS arm64 to validate downloaded release binaries outside the source checkout.
  • Keep deepseek update publish-status --strict fail-closed on verified online dogfood evidence, release assets, npm package artifacts, Homebrew checksums, and public install readiness.
  • Keep node scripts/check-secrets.js in every release path.

2. Homebrew And npm Publishing

  • Homebrew is published and verified for v0.1.5; keep the tap formula aligned with GitHub Release archives and .sha256 files on each tag.
  • HOMEBREW_TAP_TOKEN and HOMEBREW_TAP_REPOSITORY are configured so future tag workflows can update the tap automatically.
  • v0.1.5 platform npm packages and the root wrapper are published and verified. Keep the Release Matrix npm publish path idempotent and keep the manual NPM Publish retry workflow available for recovery from completed release artifacts.

For the Linux/macOS CLI milestone, Homebrew is already a verified installer path; npm/npx is now the verified distribution path for Node-oriented users.

3. DeepSeek-Native Agent Loop Hardening

  • Use DeepSeek-Native Agent Loop Design as the working design for cache-first prompt layers, tool-call repair, cost-aware model presets, read-only parallel dispatch, and stats/replay surfaces.
  • The initial tool-call repair pipeline has landed: truncated JSON arguments can be repaired, explicit JSON-shaped assistant text can be scavenged into known tool calls, nested object tool schemas can be flattened and re-nested through model.tool_schema_flattening = "auto", repair notices are visible in streams, structured tool_call_repair runtime events are persisted, TUI runtime sessions surface repair evidence, and repeated-call suppression now distinguishes read-only retries from mutating or unknown calls. Deterministic before/after repair evidence is generated by deepseek dogfood repair-cache-evidence --json and runs in the Release Matrix packaging job; keep collecting live model-backed examples as they appear in dogfood runs.
  • Prompt-layer hash diagnostics, deepseek stats, and deepseek events replay|diff MVPs have landed for runtime evidence. Cache, cost, repair, suppression, model preset/route split, tool-call, failed-tool, event-kind deltas, and per-layer prompt token/hash trends can now be inspected without reading raw runtime JSON. deepseek stats --require-prefix-stable provides a CI/dogfood gate for cache-stable prompt-layer hash regressions, and daemon compaction threshold / keep-tail policy can be tuned through runtime config. The release matrix now uploads deterministic repair/cache and prompt-prefix-stability JSON evidence as the deepseek-loop-evidence artifact.
  • Model presets and first-pass budget controls have landed: model.preset = "auto" | "flash" | "pro", deepseek config preset, run/exec --preset, --pro-next, TUI /pro with /pro off and /pro show, visible auto-to-Pro escalation events, model.session_budget_microusd warning/refusal, and cross-process runtime budget metadata. Budget raise/disable flows are now explicit through deepseek config budget raise <MICROUSD>, deepseek config budget +<MICROUSD>, deepseek config budget off, and TUI model budget .... Auto escalation now has explicit local triggers for repeated repair signals, malformed tool calls, tool-call storms, empty read/search loops, validation-after-edit failures, and repeated unproductive steps. The default live dogfood plan/report/evidence gate now also requires MCP dynamic/resource loop-surface coverage, an mcp live gate of at least three runs, and a seven-day live recency gate; the first real online MCP surface batch has passed locally with mcp 27/30, so the remaining work is recurring live dogfood calibration with real online runs. Non-recoverable malformed tool-call parse failures are now loop-recoverable: they become failed model observations with tool_call_parse_failed context for the next model turn instead of hard-ending the run. The repair/cache evidence command now records comparable before/after runtime threads with cache-hit deltas and replay/diff/stats commands.
  • Parallel-safe read dispatch has landed for same-turn batches across local read/search/git/project-map/data-validation tools plus common runtime query tools such as task_list, task_read, agent_list, agent_result, automation_list, and automation_read, plus read-only MCP inventory/prompt/resource bridge tools. They run as ordered read-only chunks when hooks/permission prompts are not active and record meta.parallel_* telemetry on tool results. Writes, shell commands, approvals, arbitrary mcp_call / dynamic MCP tools, and repeated-call recovery remain serial barriers.
  • Treat this as product hardening rather than a blocker for the already established Linux/macOS CLI milestone.

4. More External Model-Backed Samples

  • Keep the Python invoice multi-file fixture as the canonical tracked release evidence sample.
  • Keep the reusable fixture catalog healthy: python-invoice-multifile, rust-order-multifile, and node-task-report are now scaffold-smoked in CI.
  • Refresh Rust and Node online model-backed evidence when release questions need non-Python proof.
  • Require dogfood external-evidence verification with post_validation_passed=true for every sample counted as release evidence.

5. Documentation Compression

  • Keep README focused on install, deepseek quickstart, current gap, demo, and validation.
  • Keep docs/current-status.md focused on current facts and near-term work.
  • Keep docs/release.md as the operator checklist.
  • Treat docs/superpowers/ as historical execution logs, not user-facing status.

6. Broader Product Hardening

  • Record installed systemd/launchd service smoke evidence on clean machines.
  • Record a real VS Code runner/manual GUI fixture for the native agent panel.
  • Continue Windows ConPTY/service validation, while keeping it separate from the Linux/macOS local CLI milestone.
  • Periodically compare the core loop against Claude Code CLI, Codex CLI, and DeepSeek-TUI.

Current Stop Conditions

For the Linux/macOS local CLI milestone, stop treating new work as blocking once these are true:

  • hosted Linux/macOS CI gates remain green;
  • release-binary Linux/macOS smoke evidence exists for the latest release;
  • at least one verified online multi-file external fixture remains tracked;
  • Homebrew public install remains published and verified;
  • README and install docs show deepseek quickstart as the accurate first-run path.

Everything else belongs to the broader product-hardening backlog.