convez.ai
MCP Server
kagura-ai

kagura-engineer

kagura-ai/kagura-engineer

Ver no GitHub
BomPluginScore · de 0 a 100

Autonomous coding harness over Claude Code + Kagura Memory. Ships a doctor (dependency-chain diagnostics) and setup (one-command environment bootstrap) CLI; task pipeline in progress.

Instalação

$npx convez-ai kagura-ai/kagura-engineer

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
1
Acompanhando
0
Saúde
87%
Último commit
há 3 meses
Versão
v0.7.0
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

kagura-engineer

Part of the Kagura Memory Cloud offering. Licensed under Apache-2.0 — © 2026 Kagura AI.

An autonomous coding harness over Claude Code and Kagura Memory Cloud.

The long-term goal is a memory-backed actor that executes real, resumable coding tasks (see Roadmap). Shipping today: doctor and setup stand up the environment, and run / review / goal drive GitHub issues to PRs through a memory-grounded loop. It's an early 0.x harness, not a finished actor. Memory Cloud is the recommended backbone (free to start), with an offline SQLite fallback for the basic loop.


Install

Requires Python ≥ 3.11.

As a tool (recommended)

Published on PyPI. uv will also fetch a suitable Python for you.

# uv (also bootstraps Python 3.11 if needed)
uv tool install kagura-engineer

# or pipx
pipx install kagura-engineer

# or plain pip
pip install kagura-engineer

The review command shells out to the separate kagura-code-reviewer console script. Pull it in alongside the harness with the review extra:

uv tool install "kagura-engineer[review]"

(kagura-engineer setup can also bootstrap it later; without it, review degrades to a clean FAIL gate.)

To install straight from the repository instead — e.g. an unreleased commit:

uv tool install git+ssh://git@github.com/kagura-ai/kagura-engineer.git

Pin a version with a tag: pip install kagura-engineer==0.1.0.

For development (from a checkout)

pip install -e ".[dev]"     # editable install + pytest

Either way, this exposes the kagura-engineer CLI (entry point kagura_engineer.cli:app).

As a Claude Code plugin

This repo also ships a thin skill-plugin wrapper (.claude-plugin/ + skills/) so the harness is installable and discoverable from inside Claude Code. The skills (kagura-engineer:doctor, :setup, :run, :review, :goal) are thin — they shell out to the CLI installed above and surface its output; no harness logic is duplicated. Install the CLI first, then add the plugin from this repo as a marketplace source.

It is referenced by the umbrella kagura-plugins marketplace (rule: reference, don't vendor — the plugin lives here and kagura-plugins/marketplace.json points at it), where kagura-engineer takes its place as a Tier-2 Harness: multi-phase, stateful/resumable, repo-mutating, PR-creating, and HITL-gated.


Configuration

Every command reads a repo.yaml (override with --config / -c):

profile: coding                                   # required
memory_cloud_url: https://memory.kagura-ai.com    # required for cloud backend
workspace_id: ws_xxxxxxxx                          # required for cloud backend — Memory Cloud scope
context_id: 00000000-0000-0000-0000-000000000000  # required for cloud backend — context within the workspace
ollama_url: http://localhost:11434                 # optional (default shown)
memory_backend: cloud                              # optional: cloud | local (default: cloud)
local_memory_path: .kagura/memory.db               # optional (used only when backend=local)
memory_mcp_config: .mcp.json                       # optional: override the auto-discovered <repo>/.mcp.json
review:
  models: [qwen2.5-coder:7b, haiku]               # optional (default: [])
  max_loops: 3                                      # optional (default: 3)
  code_review: auto                                 # optional: auto | always | never (default: auto)
  effort: medium                                    # optional: low | medium | high (default: medium)

workspace_id → context_id → memory is the Memory Cloud filter hierarchy. A missing required field, unparseable YAML, or an unreadable file fails with a clean error and exit code 2. With memory_backend: local the three Cloud-only fields may be omitted — an offline repo.yaml is just profile + memory_backend: local.

Memory backend. Memory Cloud is the recommended default and is free to start. Authenticate with either of two equivalent credentials — run honours both, env-first:

  • export KAGURA_API_KEY=... — a workspace API key. Explicit and CI-friendly.
  • kagura auth login — installs the kagura CLI and writes an OAuth profile to ~/.kagura/credentials.json. Used automatically when KAGURA_API_KEY is unset.

When KAGURA_API_KEY is set it takes precedence; otherwise the kagura auth login profile is used. doctor and setup both check that one of these resolves and guide you if neither does — a reachable host with no credential is flagged, not silently passed. With a credential in place run/review are grounded immediately. Memory Cloud is the primary store and the moat. The local SQLite backend is the offline/dev tier: it implements the same MemoryClient interface with approximations of the rich features (tag-overlap explore for graph, importance-bump feedback, decay for Sleep-adjacent maintenance), while Memory Cloud provides the full Hebbian reinforcement, neural-graph, and server-side Sleep consolidation.

When the cloud backend is active, critical writes (savepoint remember and set_state) that fail during a Cloud outage are buffered to a local write-ahead log and replayed to Cloud on the next run — so Cloud stays the source of truth without losing a run's progress to a transient outage.

The one genuinely planned Plan 5+ item is memory auto-store / failure-mode learning (see docs/superpowers/plans/2026-06-08-memory-auto-store.md).

For offline or CI use, memory_backend: local switches the run/review grounding to an offline SQLite store (local_memory_path, stdlib sqlite3 — no API key, no network). It implements the same client Protocol; offline recall is a keyword-overlap match (no embeddings).

In-task memory MCP. Beyond string-injecting recalled memory into each headless claude -p prompt, the run/fix phases also attach the kagura-memory recall/remember tools (--mcp-config, additive) so the model can recall during the task. kagura-engineer setup generates <repo>/.mcp.json for you (no hand-authoring) via the kagura-memory SDK: with a kagura auth login profile it writes the refresh-aware stdio form (kagura-mcp --profile <p>, no secret baked in); with only KAGURA_API_KEY it writes the static-token url form. run/review auto-discover that <repo>/.mcp.json; set memory_mcp_config only to point somewhere else. By default setup writes the .mcp.json only — pass --full to additionally install the SDK's memory hooks + skills (interactive Claude Code wiring; off by default so autonomous runs are not flooded with per-edit remember calls). The server's tools must be permitted in your Claude settings; recalled content is treated as untrusted reference.


Commands

kagura-engineer doctor

Checks the dependency chain and prints a status table (or --json). Each check is isolated — one failing check never aborts the rest of the run.

CheckVerifies
gitgit on PATH, inside a work tree
claude-codeclaude on PATH + version, auth source (API key / subscription)
ghgh on PATH and authenticated
ollamadaemon reachable at ollama_url, review.models present (tag-aware match)
haikuan Anthropic auth source resolves (env key or subscription cache)
memorybackend-aware: memory-cloud reachable, or (when memory_backend: local) memory-local SQLite writable — host/credentials never echoed
gh-issue-driventhe gh-issue-driven plugin is installed (the workflow run/goal drive)

Statuses: OK / WARN / FAIL.

kagura-engineer doctor
kagura-engineer doctor --json
kagura-engineer doctor -c path/to/repo.yaml

Exit codes: 0 all OK/WARN · 1 any FAIL. A missing/invalid repo.yaml no longer refuses with exit 2: doctor degrades to a synthetic config FAIL row plus the config-free checks, and exits 1.

kagura-engineer setup

Resolves the same dependencies end-to-end: installs what's missing (via the platform package manager) and bootstraps auth. Idempotent and re-runnable.

Steps run in canonical order:

git → claude-code → gh → ollama → ollama-models → memory-cloud

Statuses: OK / SKIPPED / NEEDS_USER / FAIL. Interactive actions (a login that can't be automated) surface as NEEDS_USER.

kagura-engineer setup                  # full run
kagura-engineer setup --dry-run        # preview only; no side effects
kagura-engineer setup --fix gh         # run a single step
kagura-engineer setup --no-input       # never prompt; fail loudly on user-action steps
kagura-engineer setup --json

Exit codes (Plan 2 design doc §1.6): 0 all OK/SKIPPED · 1 any FAIL (wins over 2) · 2 any NEEDS_USER, or a config / unknown --fix error.

Valid --fix targets: git, claude-code, gh, ollama, ollama-models, memory-cloud.

kagura-engineer run

The memory-grounded agent loop. run <issue#> verifies the environment, recalls relevant memory, isolates a worktree, drives gh-issue-driven start→ship via headless claude -p (HITL gate on red/unknown verdicts), and opens a PR — persisting a savepoint to Memory Cloud.

kagura-engineer run 42                 # drive issue #42 to a PR
kagura-engineer run 42 --no-remember   # recall but don't persist
kagura-engineer run 42 --unattended    # don't pause on green/yellow (red still halts)
kagura-engineer run 42 --json

Exit codes: 0 PR reached · 1 hard fail · 2 blocked (guard or gate halt — resumable by re-running).

The report (final summary line and the --json review object) records which code-review provider/model the run actually used — the brain's in-phase /code-review, so provider is the resolved brain backend at the brain endpoint. A run that halted before the implement phase reviewed nothing, shown as review: none ran / "review": null so it stays distinguishable.

Whether that in-phase /code-review runs at all is framed by review.code_review in repo.yaml: auto (default) lets the brain decide — the implement prompt documents the criteria (run on large diffs, risk-bearing layers like auth/config-parsing/subprocess/persistence, or behaviour changes without tests; skip small mechanical/docs-only diffs) — while always/never force it on/off for repos that want CI-like reproducibility. review.effort (low/medium/high) is the effort hint passed to /code-review. Under never the report's review record stays null — the policy guaranteed no review ran.

kagura-engineer review

Launches the separate kagura-code-reviewer on a PR or branch, reads its machine-readable JSON envelope (never scrapes Markdown), and gates on the verdict. Recalled memory is passed to the reviewer as an untrusted, reference-only --context-file; the raw report is written to .kagura/review.json so you can read the full findings. This is v1 (review + gate) — the auto-review/fix loop is a later plan.

kagura-engineer review                 # review HEAD against main
kagura-engineer review feat/x          # review a branch
kagura-engineer review 42              # review PR #42 (resolved to its branch)
kagura-engineer review --base develop  # diff against a different base
kagura-engineer review --json          # machine-readable report
kagura-engineer review --fix           # auto-fix loop (Plan 4b)

Exit codes: 0 green/yellow (or nothing to review) · 1 could not review (reviewer infra error) · 2 red (blocking findings — resumable).

With --fix, a red verdict triggers the auto-fix loop: claude -p reads the persisted findings, fixes the blocking ones and commits, then re-reviews — repeating up to review.max_loops times. The reviewer stays bounded (it only emits findings); the actor does the edits. A review that couldn't run (infra error) never triggers a fix.

--fix commits to (and re-reviews) the currently checked-out branch, so check out the branch you want fixed before running it — for review <PR#> --fix that means the PR's head branch.

kagura-engineer goal

Drive a whole milestone to PRs: enumerate its open issues (via gh) and run the run loop over each, in order. It auto-continues while issues ship and halts at the first blocked/failed issue (resumable by re-running — already-shipped issues resume cleanly).

kagura-engineer goal v0.3              # drive milestone "v0.3" to PRs
kagura-engineer goal v0.3 --unattended # don't pause on green/yellow (red still halts)
kagura-engineer goal v0.3 --json

Exit codes: 0 all issues shipped · 1 hard fail · 2 blocked (an issue's gate halted — resolve it, then re-run).

Headless auth (run / review --fix / goal)

These commands spawn headless claude -p subprocesses, which need a valid Anthropic credential in the environment. Two options, in recommended order:

  1. A Claude Pro/Max subscription (recommended) — run claude once to claude login. run/goal fan out many claude -p phases per issue, so a flat-rate subscription is dramatically cheaper than metered API billing for an autonomous loop. Caveat: heavy runs can hit subscription rate limits — if you drive a whole milestone unattended (CI/cron), use an API key instead.
  2. ANTHROPIC_API_KEY — a metered API key. Best for unattended CI where no interactive claude login seat exists. Must be a real value (an empty string is treated as unset).

doctor's claude-code check only verifies the binary launches, not that auth works, so a bad credential surfaces as a phase that can't produce a verdict.

Nested-in-Claude-Code gotcha: if you run kagura-engineer from inside a Claude Code session, the inherited ANTHROPIC_API_KEY is that session's internal token and is invalid for a standalone claude -p. Drop it so the child falls back to your claude login:

env -u ANTHROPIC_API_KEY kagura-engineer run 42

Project layout

kagura-engineer/
├── README.md                  # this file
├── pyproject.toml
├── docs/plan/                 # design docs (plan-2-setup.md, …)
├── src/kagura_engineer/
│   ├── cli.py                 # typer app: doctor / setup / run / review / goal / eval
│   ├── config.py              # repo.yaml loader + Config (pydantic)
│   ├── proc.py                # shared subprocess helper
│   ├── doctor/                # Plan 1 — checks, registry, result, render
│   │   ├── checks.py
│   │   ├── registry.py        # run_all(cfg) → [CheckResult], per-check isolation
│   │   ├── result.py          # Status (OK/WARN/FAIL), CheckResult
│   │   └── render.py          # table + json
│   ├── setup/                 # Plan 2 — step orchestrator
│   │   ├── __init__.py        # STEP_NAMES, build_plan, run_plan → SetupReport
│   │   ├── auth.py            # resolve_anthropic_auth (shared with doctor)
│   │   ├── install.py         # run_install helper + stderr_tail
│   │   ├── platform.py        # OS / package-manager / WSL detection
│   │   ├── result.py          # StepStatus, StepResult, SetupReport
│   │   ├── git.py · claude.py · gh.py · ollama.py · memory_cloud.py
│   │   └── render.py
│   ├── run/                   # Plan 3 — memory-grounded agent loop
│   │   ├── memory.py          # MemoryClient Protocol + KaguraCloudClient
│   │   ├── local_memory.py    # Plan 5 — offline SQLite backend
│   │   └── gate.py · workflow.py · worktree.py · result.py · render.py
│   ├── review/                # Plan 4 — reviewer launch + verdict gate
│   │   └── reviewer.py · envelope.py · loop.py · fixer.py · context.py · …
│   ├── goal/                  # milestone driver over run
│   │   └── issues.py · render.py · result.py
│   └── eval/                  # #57 — A/B: memory-grounded uplift (run loop ×2)
│       └── metrics.py · result.py · render.py
└── tests/                     # pytest (pythonpath = src)

Development

pip install -e ".[dev]"
pytest                         # full suite

pyproject.toml sets pythonpath = ["src"], so import kagura_engineer resolves under pytest without an editable install.

Releasing

The version lives in src/kagura_engineer/__init__.py — the canonical source, read by hatch for the wheel. The Claude Code plugin manifests .claude-plugin/plugin.json and .claude-plugin/marketplace.json mirror it. /gh-issue-driven:tag bumps the two manifests; bump __init__.py (and the literal in tests/test_version.py) to match in the same release. tests/test_plugin.py asserts __init__.py, plugin.json, and marketplace.json are all equal, so a skew fails CI.


Roadmap

The bootstrap CLI exists to stand up the environment for the actual product: a memory + actor harness. The defining capability is the combination, not the parts — a stateless agent or a memory-less actor doesn't get there.

  • Cost-aware planning — recall past similar tasks' real cost/failure modes and budget the plan accordingly.
  • Long-running task resume — checkpoint task state to Memory Cloud; resume cleanly in a fresh context after the window dies.
  • Failure-mode learning — every failure becomes a memory with a prevents edge to its fix, surfaced preemptively next time. Recurring-failure cost → 0.
  • Sub-agent dispatch with memory handoff — children receive context as memory IDs, not prompt text, keeping the parent context small.

Claude is driven via the Claude Code CLI; a Pro/Max subscription is inherited for self-hosted use, with ANTHROPIC_API_KEY (BYOK) as the multi-tenant fallback. Memory Cloud is the persistent backbone, consumed as the primary MCP server (recall / remember / create_edge / explore / …).

Explicit non-goals: not a chat interface for Memory Cloud, not a fine-tuned or domain model, not a chat-ingestion source, not a memory analyzer. The job is autonomous task execution with persistent memory — nothing more.


Related repositories

RepoRoleRelationship
memory-cloudPersistence + MCP serverThe backbone. Primary MCP.
kagura-memory-python-sdkPrimitive SDKUsed by the memory MCP client wrapper.
kagura-memory-ai-workerChat ingestionProduces memories the harness later reads.
kagura-memory-dataset-workerExport + fine-tuneIndependent; may export harness-produced memories.
kagura-embeddings-workerSovereign embeddingsIndirect — recall quality depends on the workspace's embeddings lane.

License

Apache License 2.0 — © 2026 Kagura AI. See LICENSE for the full text and NOTICE for attribution.

Ver o original no GitHub

Bom para

Relacionados