eng-org-kiro

A Kiro CLI adapter for the eng-org 5-role engineering pipeline. Same governance, same role prompts, same quality gates — driven from Kiro instead of Claude Code.

Built additively: the eng-org Claude Code plugin is read-only and never modified. This adapter is generated from it, so the two can never silently diverge.

Status

0.20.0 — release candidate. Benchmarked at parity with Claude Code across all four modes (A/B/C/L) using the bench's own scorers: detection II over 9 fixtures (Kiro 92.9 / Claude 92.4), autopilot acceptance-criteria pass (100 / 100), plus maker-checker and implementation (100 / 100). See BENCHMARK.md.

Before you rely on it, read KNOWN-LIMITATIONS.md — the honest ledger (headless non-determinism is mitigated, not eliminated; Kiro token/cost is not exposed by the CLI).

Why

• Use either tool. Code in Claude Code when you want; switch to Kiro when you want.
• Cross-tool resume. Because all pipeline state is files + git (spec.md, tl-analysis, trd, dev-reports, tests, reviews, merge-readiness) — not chat memory — a run started in one tool can be continued in the other (bin/resume.sh).
• Same quality. Both adapters carry the same eng-org core version (stamped + bin/doctor.sh verifies parity), the same governance docs, and the same role instructions.

Three install profiles

| You install… | You get | |---|---| | only the Claude plugin (@eztech/eng-org) | full eng-org via /eng-org:* — works standalone | | only this (eng-org-kiro) | full eng-org via Kiro agents + run-pipeline.sh — self-contained, no Claude needed | | both, same repo | shared on-disk state → cross-tool handoff + identical quality |

Layout

bin/generate.mjs      read eng-org agents/governance (read-only) → Kiro agent JSON + vendored core
bin/install.sh        install eng-* agents to ~/.kiro/agents (global) or ./.kiro/agents (--workspace)
bin/init.sh           scaffold governance/ into a project (identical layout to eng-org)
bin/run-pipeline.sh   Kiro orchestrator — 8-stage halt-on-red pipeline (mirrors /eng-org:ship)
bin/resume.sh         cross-tool handoff — show/continue the next incomplete stage
bin/doctor.sh         parity + schema checks (Claude core vs Kiro core, agent validity)
lib/stages.sh         shared 8-stage state machine (the resume brain)
agents/               21 generated Kiro agents (eng-em, eng-dev-*, eng-reviewer-*, eng-test-*, …)
governance-templates/ vendored governance + commands + scripts (self-contained)
bench/                2-tool × 2-version comparison harness for release decisions

Install & quickstart (npm — recommended)

A global install exposes one eng-org-kiro <command> for every step. The 21 agents ship pre-generated in the package, so Kiro-only users never need to generate — just install and go.

npm install -g @eztech/eng-org-kiro      # public install — anyone can use it
kiro-cli login                           # one-time (IAM Identity Center or Builder ID)

eng-org-kiro install                     # copy the 21 agents → ~/.kiro/agents
eng-org-kiro doctor                      # verify agents valid (optional)

cd /path/to/your/project
eng-org-kiro init                        # scaffold governance/ (same as /eng-org:init)

# Talk to it like a teammate (interactive) …
eng-org-kiro chat                        # opens a conversation with the Engineering Manager

# … or drive the whole pipeline headless in one shot
eng-org-kiro run "Add a POST /feedback endpoint with validation and tests"

eng-org-kiro with no arguments prints the full command list.

From a checkout (instead of npm)

Clone the repo and call the bin/ scripts directly — same flow. Only run node bin/generate.mjs if you have a local eng-org core and want to re-generate for parity (npm users skip it — agents are shipped).

node bin/generate.mjs      # optional: regenerate from a local eng-org core
bin/install.sh             # install agents globally (or --workspace .)
bin/doctor.sh              # verify parity + schema
bin/init.sh .              # scaffold governance/
bin/run-pipeline.sh "Add a POST /feedback endpoint with validation and tests"

Cross-tool handoff

# Claude Code hits its usage limit mid-REQ …
eng-org-kiro resume                  # shows stage table + next incomplete stage (npm)
bin/resume.sh                        # …or from a checkout
bin/resume.sh REQ-<id> --go          # Kiro continues from exactly where Claude stopped

Handoff granularity is per pipeline stage (last completed artifact/commit) — not a live mid-sentence continuation.

Tool mapping (Claude Code → Kiro)

Read/Grep/Glob → read/grep/glob · Write/Edit → write · Bash → shell · Agent (subagent) → delegate · Task* → todo · mcp__x__y → @x/y

Release-decision bench

bench/run-matrix.sh                                       # dry-run: plan + cost estimate
bench/run-matrix.sh --tools claude,kiro --versions 0.19.1,0.20.0 --n 3 --go
node bench/compare.mjs                                    # matrix + release recommendation

Runs the same corpus through every (tool × version) cell, scores identically, and recommends the best tool+version. Spends real tokens/credits/Claude-usage — only runs with --go.

Regenerate on eng-org updates

node bin/generate.mjs                # re-vendors the newest in-use eng-org core
bin/doctor.sh                        # confirms Claude/Kiro parity

Stops at production-ready — merge and deploy stay human-controlled (Constitution §H.47).