Deterministic coding workflow for LLM agents
Structures coding tasks as a phase state machine with traceability IDs, artifact templates, and source state verification. Designed for the z.ai platform.
IDLE → SPECIFY → PLAN → IMPLEMENT → VERIFY → DELIVER
↑ │
└──── Error Recovery ◄───────────────────┘
git clone https://github.com/hoshiyomiX/stellar-coding-agent.git /tmp/cap
cd /tmp/cap && bash setup.shInvoke in any session:
Skill(command="stellar-coding-agent")
Look for ☄️ STELLAR · v5.0.0 · ACTIVE — confirms the framework loaded.
rm -rf /home/z/my-project/skills/stellar-coding-agentThe framework provides tools, not rules. Each phase produces an artifact the next phase consumes, creating a chain that prevents skipping straight to code.
| Phase | Output | Why |
|---|---|---|
| SPECIFY | Problem specification | Forces precise thinking before writing code |
| PLAN | Implementation plan with Traceability IDs | Maps requirements to code locations |
| IMPLEMENT | Annotated code | Each block references its Traceability ID |
| VERIFY | Evidence-based report | Automated checks + edge case tracing |
| DELIVER | Summary + compliance report | Traceable record of what was done |
Traceability IDs (IMPL-001, IMPL-002, ...) chain through every phase — requirement → code → verification. If something is dropped, the gap is visible.
Before analyzing git repositories, the framework verifies data freshness:
git fetch → compare HEAD to origin → sync if behind → proceedPrevents stale-checkout analysis (the failure that inspired this feature).
Structured 5-step decision tree: capture → classify → identify actions → fix → re-verify. Git operations have explicit safety rules — git fetch before git pull, no force push without user instruction, stop all git ops if infrastructure blocks.
skill/stellar-coding-agent/
├── SKILL.md # Core framework (phases, SSV, error recovery, PCR)
├── CHANGELOG.md # Version history
├── procedure/
│ ├── phases.md # Phase definitions with entry/exit criteria
│ ├── templates/
│ │ ├── problem-spec.md # SPECIFY artifact
│ │ ├── implementation-plan.md # PLAN artifact (Traceability IDs)
│ │ ├── verification-report.md # VERIFY artifact (evidence capture)
│ │ └── incident-report.md # Error documentation
│ └── decision-trees/
│ └── error-resolution.md # 5-step structured decision tree
├── constraints/
│ ├── code-standards.md # Function, file, import, quality standards
│ └── type-safety.md # Type system constraints with examples
├── knowledge/
│ ├── architecture.md # Runtime environment, directory layout, service topology
│ ├── conventions.md # Coding conventions, state management, import order
│ ├── platform-constraints.md # Sandbox-specific limitations (gateway, routes, SDK)
│ └── error-patterns.md # Common errors with cause → fix mapping
└── memory-template.md # Template for user preference storage
Stop telling the LLM what it MUST do. Start giving it tools it WANTS to use.
v5.0.0 is a philosophical reset based on an honest audit:
- What works: Traceability IDs, templates, SSV, error decision tree — they work because they're useful, not because they're mandatory
- What doesn't work: Compliance enforcement language ("must", "mandatory", "do not skip") — has no measurable effect on LLM behavior regardless of wording
- What's honest: The framework cannot guarantee compliance, force behavior, or persist across sessions. It's text in a skill file. The user is the final judge of quality.
| Version | Summary |
|---|---|
| v5.0.0 | Philosophical reset. Removed compliance theater, kept useful tools. SKILL.md: 181→95 lines. |
| v4.6.0 | Source State Verification (SSV). Evidence tiers in attestation. |
| v4.5.0 | Coexistence mode with fullstack-dev. (Removed in v5.0.0) |
| v4.4.2 | QA Attestation required for all tasks, not just coding. |
| v4.4.1 | Compliance binding after banner. |
| v4.4.0 | Git error classification and safety rules. |
| v4.3.0 | OUTCOME gate, evidence requirement, defect counter. |
| v4.2.0 | Styled activation banner. |
| v4.1.0 | Removed fullstack-dev routing wrapper. Direct invocation. |
| v4.0.0 | Complete redesign: phase state machine, artifact templates, traceability IDs. |
MIT