Skip to content

Releases: uniStark/starks_skill

v0.2.0 — memory recall & hardened orchestration

10 Jun 06:23
@uniStark uniStark
v0.2.0
e0c8b4c

Choose a tag to compare

View all tags
v0.2.0 — memory recall & hardened orchestration Latest
Latest

Highlights

  • Memory recall at task start — project memory ($STARKS_MEMORY_DIR/<project>/) is now read before grilling, so settled questions are never re-asked. Deterministic project naming (repo basename, fallback to _index.md) prevents memory fragmentation.
  • Structural anti-recursion on the Codex reviewer side — the cross-review invocation disables the starks skill itself via skills.config, so the reviewer cannot even see it. The STARKS_CROSS_REVIEW env guard remains as a second layer.
  • Structured two-stage review — new prompts/spec-review.md and prompts/code-review.md templates with file:line / blocking-or-suggestion findings.
  • Bounded rework loop — at most 2 rework rounds, then stop and escalate to the user.
  • Safer parallel sub-agents — disjoint write-sets required; worktree isolation mapped for both platforms; per-slice pipelined review.
  • Leaner trigger description — frontmatter no longer summarizes the workflow (it misled models into treating cross-review as mandatory; it is user-opt-in).
  • Memory writer keeps the last 5 dated entries instead of overwriting history; durable preferences go to platform-native memory.
  • Cross-reviewer may browse the repo read-only; timeout pinned to ~10 min.
  • CI: lint runs on every push/PR.

Install

git clone https://github.com/uniStark/starks_skill.git
cd starks_skill
bash scripts/install.sh

Full design rationale in docs/DESIGN.md.

Assets 2
Loading

v0.1.0 — first public release

31 May 18:35
@uniStark uniStark
v0.1.0
ed77721

Choose a tag to compare

View all tags
v0.1.0 — first public release

starks is a task-launcher skill for Claude Code & Codex: it grills the requirements, optionally cross-reviews the plan across models (Claude↔Codex), then runs the work with PM-mode parallel sub-agents — without making simple jobs pay for heavy process.

Highlights

  • Task tiering — trivial / light / full, so simple work stays fast and only real complexity triggers the full flow.
  • Cross-model review (you choose) — at sign-off, optionally send the plan to the other engine for a critical second pass. Never automatic, never silently skipped.
  • PM-mode parallel sub-agents — independent slices fan out; tightly-coupled work stays sequential instead of being force-split.
  • Two-stage review — spec compliance, then code quality.
  • Verification gate — no "done / passing" claim without freshly-run evidence.
  • Optional memory layer — project summaries to a knowledge base (e.g. Obsidian); skipped if unconfigured.
  • Dual-platform — one `SKILL.md`, symlinked into both Claude Code and Codex.
  • Anti-recursion guard — answers once as a cross-reviewer instead of re-entering its own flow.

Install

```bash
git clone https://github.com/uniStark/starks_skill.git
cd starks_skill && bash scripts/install.sh
```

See the README for configuration and the design notes for rationale.

Loading