coding-agent

coding-agent

A Claude Code plugin that turns a Jira ticket into a reviewed pull request — autonomously.

coding-agent is a multi-agent development pipeline for go-stablenet (a geth fork with WBFT consensus). You point it at a ticket; it analyzes, plans, designs, implements, tests, opens a PR, folds in review feedback, and merges — pausing for your confirmation on anything irreversible.

It is built on two ideas:

• Orchestration over a document-backed state machine. Every stage writes its artifact (analysis.md, plan.md, design-v{N}.md, test-report.md) to disk, so a truncated context or a new session resumes exactly where it left off.
• Retrieval-grounded decisions. Instead of guessing about a large unfamiliar codebase, the planner queries a knowledge service (cks) with RAG (semantic search) and graph-RAG (call graphs, impact and concurrency analysis), and grounds every design choice in real code.

---

How it works

Jira ticket (STABLE-xxxx)
   │  jira-gateway MCP  ── sensitive-info filter (secrets blocked before they reach the LLM)
   ▼
TICKET_INTAKE → ANALYSIS → PLANNING → DESIGN → IMPLEMENTATION → EVALUATION → COMPLETION
                   │                                                │
              cks retrieval                                   4-stage gate
            (RAG + graph-RAG)                       (unit+race · lint · security · chainbench)
                                                                    │
                                                   PASS → PR + Jira update
                                                   FAIL → bugfix cycle (≤3) or BLOCKED

Four isolated agents do the work; the orchestrator is the only one that sees the whole flow:

Agent	Role
orchestrator	Drives state transitions, MCP pre-flight, PR/Jira completion, bug-cycle re-entry
planner	ANALYSIS / PLANNING / DESIGN. The sole cks consumer — RAG + graph-RAG retrieval
implementer	Branch isolation, one commit per atomic step, build handoff
evaluator	4-stage verification: unit (+-race), lint/format, security scan, chainbench integration

It talks to three MCP servers: jira-gateway (in this repo, a sensitive-info proxy in front of Jira), cks (code-knowledge-system, a sibling repo that composes semantic + graph retrieval), and chainbench (a sibling repo, the deterministic test runner). The agent-facing tool surface is frozen in contract/agent-mcp.schema.json and enforced by contract/lint-tool-names.sh.

Security model. Sensitive data is blocked before it reaches the model, not after. All inbound Jira content passes through the jira-gateway filter (regex + entropy + allowlist → REDACTED/BLOCKED); all outbound text (PR bodies, commit bodies, Jira comments) passes through the pr-sanitize skill using the same shared/patterns.json.

---

Install

The plugin is distributed as a Claude Code marketplace plugin from this GitHub repo.

/plugin marketplace add 0xmhha/coding-agent
/plugin install coding-agent@coding-agent

Restart Claude Code, then run /help — you should see /coding-agent:work, /coding-agent:review, /coding-agent:status, and /coding-agent:merge.

Local development install. To run from a clone instead, point your user config at the plugin directory:

{ "plugins": { "coding-agent": { "path": "/abs/path/to/coding-agent/plugin" } } }

---

Configure

The plugin reads secrets and server locations from environment variables that plugin/.mcp.json forwards into the MCP servers. Export them in your shell profile so Claude Code's child processes inherit them.

# Jira (required) — token: https://id.atlassian.com/manage-profile/security/api-tokens
export JIRA_BASE_URL="https://your-domain.atlassian.net"
export JIRA_USER_EMAIL="[email protected]"
export JIRA_API_TOKEN="atlassian_api_token_here"

# cks — the code-knowledge service (sibling repo; see SETUP.md to build)
export CKS_MCP_BIN="$HOME/Work/code-knowledge-system/bin/cks-mcp"
export CKS_CONFIG="$HOME/Work/code-knowledge-system/cks.yaml"

# chainbench — the deterministic test runner (sibling repo)
export CHAINBENCH_DIR="$HOME/Work/chainbench"

Requirement	Why
Claude Code	Hosts the plugin
Atlassian (Jira) Cloud	Source of tickets
gh CLI ≥ 2.50	PR create / comment / merge
code-knowledge-system (cks) + Ollama + bge-m3	Code retrieval (RAG + graph-RAG). Without it, cks runs degraded and the pipeline still works at lower retrieval quality
chainbench	Evaluator Stage 4 (integration). Skippable; Stage 4 fails loudly if absent

code-knowledge-system and chainbench are sibling repositories resolved by path at runtime. Building and indexing them (the slow bge-m3 embed of go-stablenet) is covered step by step in docs/SETUP.md.

After install, verify the contract is intact: