yokai-cli 0.2.0a8

From issue tracker stories to pull requests via coding agents. A spirit that works while you sleep.

yokai

In Japanese folklore, a yokai is a spirit that operates in the background of the human world, often working at night, sometimes mischievous and sometimes helpful. This framework is the helpful kind: it watches your backlog while you sleep and leaves pull requests waiting for you in the morning.

A Python framework for spec-driven development pipelines: turn issue tracker stories into pull requests automatically, using a coding agent of your choice.

+-------------+      +----------+      +---------------+      +---------------+
|   Jira      | ---> | Router   | ---> |  Claude Code  | ---> |  Bitbucket    |
|  (story)    |      |          |      |  (agent)      |      |  (pull req)   |
+-------------+      +----------+      +---------------+      +---------------+

yokai polls your issue tracker for stories tagged with a configurable trigger label, routes each story to its target repository, runs a coding agent inside the local working tree, then commits, pushes, and opens a pull request. It posts the result back as comments on the original story so the human reviewer has full context.

Why this exists

Several commercial offerings cover the same workflow, but they all target cloud SaaS deployments (Jira Cloud, Bitbucket Cloud, GitHub). yokai started as the first open-source framework targeted at on-premise enterprise environments (Jira Data Center and Bitbucket Data Center behind firewalls and SSO, where cloud connectors do not reach), and now also ships adapters for Atlassian Cloud (Jira Cloud and Bitbucket Cloud) so the same framework can drive hybrid setups.

It is designed to be runnable from a developer laptop, with no infrastructure requirements beyond Python 3.10+, git, and the chosen coding agent CLI.

Status

Early alpha. The core orchestrator, the Jira Data Center, Bitbucket Data Center, Jira Cloud, and Bitbucket Cloud adapters, and the Claude Code adapter are working and tested. Since 0.2.0 an optional async mode is available with SQLite and Redis backends. The API is unstable and may change.

Features

• Provider-agnostic core: swap any of the issue tracker, repo hosting, coding agent, router, or storage by implementing a small interface.
• Built-in adapters for Jira Data Center, Bitbucket Data Center, Jira Cloud, Bitbucket Cloud, and Claude Code CLI.
• Two deployment modes: simple monolithic (yokai run) and scaled async (coordinator + worker(s) + result-handler) sharing a persistent queue.
• Pluggable queue backends (in-memory, SQLite, Redis) for the async mode, so a single laptop setup and a multi-host production cluster use the same code.
• Parallel processing with per-repository locking: stories on different repos run concurrently, stories on the same repo serialize.
• In-flight deduplication: a story is never picked up twice while it is being processed, even if the issue tracker label update is delayed.
• Automatic retry with exponential backoff and dead-letter queue for jobs that exceed max_attempts.
• Plugin system with lifecycle hooks: register callbacks for events like after_agent_run or on_failure without forking the framework.
• Persistent execution state via SQLite, surviving process restarts.
• Notification sinks (logger, Slack webhook, custom).
• Token redaction in all log output, including credentials embedded in Bitbucket Cloud clone URLs.
• Idempotent commands and safe failure recovery.
• PR review rework loop (async mode): when a reviewer leaves comments on a yokai-generated PR, add the ai-rework label on Jira and yokai will read the review comments, fix the code, and push to the same branch — the PR updates automatically.
• Markdown to Jira wiki markup converter: agent output (written in Markdown by Claude) is automatically converted to Jira Data Center wiki markup so comments render with proper headers, tables, bold, and code blocks.

Deployment modes

yokai can run in two modes, chosen via config:

Monolithic mode: yokai run

One process polls the tracker, runs the agent, opens PRs. This is the simplest setup and has been the default since version 0.1. It is still the recommended mode for single-developer laptop use.

Async mode: yokai coordinator + yokai worker + yokai result-handler

Three roles, three processes, a persistent queue in between. Added in 0.2.0. Use this when you want:

• Resilience: jobs survive process crashes (the queue persists).
• Scale: run multiple workers in parallel, on the same host or on different hosts.
• Separation of concerns: polling, agent execution, and PR creation can be monitored and restarted independently.

+---------------+    enqueue    +-----------------+
| Coordinator   | ------------> |  Job Queue      |
| (polls Jira)  |               |  (SQLite/Redis) |
+---------------+               |                 |
                                |                 |
+---------------+    dequeue    |                 |
| Worker(s)     | <------------ |                 |
| (run agent)   |               |                 |
+---------------+    write      |                 |
       |          ------------> |  Result Store   |
       v                        |                 |
+---------------+    read       |                 |
| ResultHandler | <------------ |                 |
| (commit + PR) |               +-----------------+
+---------------+

Backends for the queue:

• SQLite (default): single file, no external services, good for single-host deployments.
• Redis: multi-host, production-grade. Install with the [redis] extra.
• In-memory: tests and experiments only.

See docs/async_mode.md for the full operational guide.

Quickstart

1. Install

pip install yokai-cli
# or with Redis support:
pip install yokai-cli[redis]

You also need:

• Python 3.10 or later
• git
• The CLI of your chosen coding agent (e.g. Claude Code: npm install -g @anthropic-ai/claude-code)

2. Generate a starter config

yokai init --output config.yaml

Edit config.yaml and fill in your Jira and Bitbucket details. Tokens should be passed via environment variables and referenced as ${VAR_NAME} in the file.

3. Set credentials

export JIRA_USERNAME=your.username
export JIRA_TOKEN=your-jira-personal-access-token
export BITBUCKET_USERNAME=your.username
export BITBUCKET_TOKEN=your-bitbucket-http-access-token

The Bitbucket token must have repository write permission. Read-only tokens will fail at the push step.

4. Tag a story and run

In Jira, add the label ai-pipeline to a story in the Backlog status. Make sure the story has a component that matches one of the entries in your routing.components map, or add a label like repo:my-repo.

Then start yokai. Pick one mode:

Monolithic (simplest):

yokai run --config config.yaml

Async on one host (more resilient):

# in three separate terminals
yokai coordinator    --config config.yaml
yokai worker         --config config.yaml
yokai result-handler --config config.yaml

Either way, it polls Jira, clones the target repo, runs Claude Code, opens a pull request, and posts two comments back on the Jira story.

5. Inspect history

yokai status --config config.yaml        # legacy SQLite execution store
yokai queue-status --config config.yaml  # async queue state (jobs, workers, dead-letters)

PR review rework loop

When a reviewer leaves comments on a yokai-generated pull request, yokai can automatically fix the code based on the review feedback.

How it works

1. yokai processes a story and opens a PR (the normal flow).
2. A reviewer comments on the PR in Bitbucket.
3. Someone adds the label ai-rework to the Jira story.
4. The coordinator picks it up, finds the open PR on Bitbucket by matching the story key in the branch name, and reads all review comments.
5. The worker checks out the existing branch (no new branch), runs Claude Code with a rework-specific prompt that includes every review comment with file path and line number, then commits and pushes to the same branch.
6. The PR updates automatically with the new commit.
7. yokai removes ai-rework and restores ai-done on Jira.

The rework loop can be repeated: if the reviewer leaves more comments, add ai-rework again.

Label flow

ai-pipeline → ai-processing → ai-done
                                  ↓  (reviewer adds ai-rework)
                            ai-rework → ai-processing → ai-done
                                                            ↓  (repeat if needed)
                                                      ai-rework → ...

Configuration

The rework label is configurable (default ai-rework):

issue_tracker:
  rework_label: ai-rework

Architecture

The core of the framework is a small set of abstract interfaces:

Interface	Responsibility	Built-in implementation
IssueTracker	search, comment, label stories	JiraDataCenterTracker, JiraCloudTracker
RepoHosting	clone, branch, commit, push, open PR	BitbucketDataCenterHosting, BitbucketCloudHosting
CodingAgent	run an AI agent in a working tree	ClaudeCodeAgent
StoryRouter	resolve a story to a repository	ComponentMapRouter, LabelPrefixRouter, ChainRouter
NotificationSink	post events to humans	LoggerNotificationSink, SlackWebhookSink
ExecutionStore	persist story execution state (legacy mode)	InMemoryExecutionStore, SqliteExecutionStore

The monolithic Pipeline depends only on these interfaces. Concrete adapters are constructed by factory.build_pipeline(config) from a FrameworkConfig loaded from YAML.

The async mode adds four more interfaces in yokai.queue:

Interface	Responsibility	Built-in implementation
JobQueue	enqueue, dequeue (with lease), update status	InMemoryBackend, SqliteBackend, RedisBackend
ResultStore	store and retrieve agent results	same as above
WorkerRegistry	track live workers via heartbeats	same as above
CoordinatorLock	leader-election lock for coordinator HA	same as above

These are wrapped around the existing adapters by the yokai.queue_adapters bridge layer, so async mode automatically supports every combination (Jira DC/Cloud x Bitbucket DC/Cloud) that legacy mode supports.

To add support for a different system (GitHub Issues, GitLab, Linear, Aider, OpenCode, etc.), implement the relevant interface and register the new builder. See docs/writing_an_adapter.md.

Concurrency

Monolithic mode uses a ThreadPoolExecutor to process multiple stories in parallel up to max_parallel_stories. To prevent two stories from trampling each other's working tree on the same repo, each repository has its own lock. Two stories on different repositories run truly in parallel; two stories on the same repo serialize through the lock.

A separate in-flight registry tracks stories that have been submitted to the pool but have not yet had their tracker label updated, so the polling loop never submits the same story twice.

Async mode achieves parallelism by running multiple worker processes. The queue backend handles the mutual exclusion atomically: a job is dequeued exactly once, and the coordinator re-queues it only if the worker's lease expires. Dedup of in-flight stories is done at the queue level via per-story keys.

Hooks

The monolithic pipeline emits 9 lifecycle events. Plugins register callbacks for the events they care about. A failing callback never breaks the pipeline, only logs the exception.

Event	When it fires	Payload keys
before_process	Story acquired by worker	story
after_resolve_repo	Repository resolved	story, repo_slug
after_clone	Working tree ready	story, repo_path
before_agent_run	About to invoke agent	story, repo_path, prompt
after_agent_run	Agent finished	story, agent_result
after_commit	Local commit created	story, commit
after_push	Branch pushed	story, branch_name
after_pull_request	Pull request opened	story, pull_request
on_success	Full flow succeeded	story, pull_request
on_failure	Any error in the flow	story, error

See examples/example_plugin.py for a working plugin. Since 0.2.0, hooks are emitted in both monolithic and async modes. Plugins written against the legacy Pipeline (using pipeline._hooks.register(...)) work unchanged in async mode thanks to a compatibility shim in async_factory - they receive a small object with ._hooks just like a real Pipeline.

Configuration reference

The full configuration is a single YAML file. See examples/enterprise_data_center.yaml for an annotated example of the legacy monolithic mode.

Sections:

• issue_tracker - connection and filtering for the issue source
• repo_hosting - connection and branch policy for the repo host
• agent - coding agent command and timeouts
• routing - how to resolve stories to repositories
• orchestrator - polling and parallelism settings (monolithic mode)
• storage - execution state persistence for monolithic mode (memory or sqlite)
• queue - optional. Enables async mode. Fields: backend (sqlite/memory/redis), db_path, redis_url, and sub-sections for coordinator, worker, result_handler. Omit this section to keep only the monolithic yokai run mode.
• plugins - list of dotted import paths to plugin install functions

Environment variable references like ${VAR_NAME} are expanded at load time. Missing variables raise a clear configuration error.

CLI reference

Command	Mode	What it does
yokai init	-	Write a starter YAML to stdout or a file
yokai run	monolithic	Run the single-process polling orchestrator
yokai status	monolithic	List recent executions from the SQLite store
yokai coordinator	async	Poll the tracker and enqueue jobs
yokai worker	async	Dequeue and run the coding agent
yokai result-handler	async	Commit, push, open PR, comment
yokai queue-status	async	Show queue counts, live workers, dead-letters
yokai queue-retry <job-id>	async	Re-enqueue a dead-lettered or failed job

Development

Clone the repo and install in editable mode with dev extras:

git clone https://github.com/inkman97/yokai
cd yokai
pip install -e ".[dev,redis]"

Run the test suite:

pytest

The test suite (~680 tests) has unit tests with HTTP mocking for the Jira and Bitbucket adapters, parallelism tests using fake in-memory adapters, an integration test that exercises real git operations against a local bare repository (no network needed), a full contract test suite for the three queue backends (in-memory, SQLite, Redis via fakeredis), and end-to-end tests for the PR review rework loop.

Contributing

This project is maintained as a side effort. Contributions are welcome, especially:

• Additional issue tracker adapters (Linear, GitHub Issues)
• Additional repo hosting adapters (GitHub, GitLab)
• Additional coding agent adapters (Aider, OpenCode, Cursor CLI)
• Additional queue backends (RabbitMQ, PostgreSQL)
• Bug reports from real on-premise enterprise deployments
• Improvements to documentation

Please open an issue first if you plan a substantial change.

License

MIT