nava-rebar 0.10.0

Event-sourced ticket system with a Jira reconciler — CLI, Python library, and MCP server.

rebar

An event-sourced, git-backed ticket store + Jira reconciler built for agent swarms — one store, exposed as a Python library, a CLI, and an MCP server.

• Three surfaces, one store — drive rebar as a CLI (rebar), a Python library (import rebar), or an MCP server (rebar-mcp).
• The tracker lives in the repo — tickets are an append-only event log on a tickets git branch; no database, no daemon, and it travels with every clone.
• Built for parallel agents — atomic claims, convergent merges, and provenance links let many agents and sessions write at once without lost work.
• Optional LLM gates — review a ticket's plan before work, its completion before close, and its code before it merges.
• Bidirectional Jira sync — a level-triggered reconciler keeps tickets and Jira in step, so teammates stay in the loop.
• Dogfooded through two independent gates — every change to rebar's own main must pass an LLM code review and CI, on Gerrit, before it lands.

Install

pipx install nava-rebar          # the `rebar` CLI (add [mcp] / [agents] for those extras)
brew install navapbc/rebar/rebar # or via Homebrew

Quickstart

Run one ticket end-to-end with the CLI or the Python library; the JSON block is the MCP server config so agents can drive the same loop over MCP. rebar --help (and rebar <command> --help) is the authoritative command reference.

# CLI: one ticket through init -> create -> ready -> claim -> close
rebar init
tid=$(rebar create task "Add a login page" | tail -1)   # capture the new ticket id
rebar ready                                              # lists it as ready to work
rebar claim "$tid" --assignee alice                     # open -> in_progress
rebar transition "$tid" in_progress closed              # in_progress -> closed

import rebar                                            # the same loop via the Python library
tid = rebar.create_ticket("task", "Add a login page")
rebar.claim(tid, assignee="alice")
rebar.transition(tid, "in_progress", "closed")

{ "mcpServers": { "rebar": { "command": "uvx", "args": ["--from", "nava-rebar[mcp]", "rebar-mcp"] } } }

That's the whole loop — init → create → ready → claim → close. The CLI and Python blocks each drive one ticket end-to-end (the same id threaded through every step, no hard-coded id); the JSON is the MCP server config — add it to your client so an agent can run the same loop via the MCP tools. State is shared through the repo so many agents (and teammates via Jira) coordinate without stepping on each other.

How it works

rebar stores tickets as an append-only event log on a dedicated tickets git orphan branch (worktree at .tickets-tracker/); ticket state is computed by replaying events, and every write auto-commits and pushes so the store is shared immediately. A level-triggered reconciler bidirectionally syncs tickets with Jira. The branch name and worktree dir are configurable (tracker.branch / tracker.dir — see Configuration). Reads stay sub-second into the thousands of tickets; for measured numbers and git-growth expectations see docs/scale-envelope.md.

Documentation lives under docs/ — start with the docs index (grouped by audience: user / operator / contributor / agent) or the day-to-day user guide.

Why rebar

If you run coding agents against a repo, you eventually want to run several at once — and the moment you do, they need a shared place to coordinate. Most trackers weren't built for that:

• They're heavy. A daemon to babysit or a local database to keep running, with dependencies thick enough that a routine upgrade can break your work tracking across machines.
• They don't travel with the code. State lives outside the repo, so a fresh clone doesn't come with its tickets.
• They fight your git history. A tracker that writes to your working branch tangles ticket churn into your source-code commits.
• They have no concurrency story. Nothing stops two agents from claiming the same work or clobbering each other's state, and concurrent edits produce merge conflicts you resolve by hand — or lose.
• They buckle at scale. Speed and usability fall off past a few hundred tickets.

rebar's answer is to make the tracker part of the repo. Tickets are an append-only event log on a dedicated tickets orphan branch (linked in through a gitignored worktree); current state is a fast, deterministic replay of that log. That single decision pays off across the board:

• Zero infrastructure, fully portable. No database, no daemon — just git and a lightweight Python install. Clone the repo and the tracker comes with it.
• No commit interference. Ticket events live on their own branch and never touch your source history. Every write auto-commits and auto-pushes, so activity is shared in real time.
• Concurrency by design. Each event gets a globally-unique filename, so parallel writes merge as a clean union, and the rare conflicting fork resolves deterministically — every clone converges with no lost data. claim is an atomic, optimistic-concurrency primitive: agents grab work without stepping on each other.
• Built to scale. The event log plus cached replay stays fast as tickets grow.

On top of that foundation, rebar adds what parallel agent work actually needs:

• Bidirectional Jira sync — agents work in rebar, teammates work in Jira, and a level-triggered reconciler keeps the two in step. To run it automatically in CI, see docs/jira-sync-setup.md (the GitHub Actions reconcile-bridge + heartbeat-canary setup).
• Conflict-aware scheduling — tickets record their file impact, so next-batch hands parallel agents work that won't collide on the same files.
• Scratch space — an invisible per-ticket channel for subagents to pass notes to one another.
• Structural quality gates — clarity, acceptance-criteria, dispatch-readiness, and repo-wide health checks keep work dispatch-ready.
• LLM review gates (optional) — review an agent's plan before work starts, its completion before the ticket closes, and its code before it merges. Plan-review and code-review share one four-pass kernel — a finder cites evidence, a separate verifier tests each claim with atomic yes/no questions, and a deterministic policy (never the model) decides what blocks — so a review coaches with grounded, cited findings rather than a black-box score. A passing plan or completion review leaves an HMAC-signed attestation: a machine-checkable signal of rigorous agentic development, not vibe-coding.
• Provenance links — discovered_from ties emergent work back to the ticket that surfaced it.
• One store, three interfaces — drive it from the CLI, a Python library, or the MCP server.

Requirements

System prerequisites:

• Python ≥ 3.11
• git — required (the store is a git orphan branch + worktree). The engine is pure in-process Python; bash and jq are not required at runtime.
• No external lock binary is required. Write serialization uses a two-window lock built entirely from the Python standard library — a fcntl.flock(LOCK_EX) advisory lock plus an atomic mkdir lock (src/rebar/_store/lock.py) — so there is no dependency on util-linux's flock binary (or any other external tool). The mkdir window keeps mutual exclusion holding even where fcntl.flock is unreliable (e.g. some network filesystems).
• acli (Atlassian CLI) — only for live Jira reconciliation.

Python dependencies. A base install (pip install nava-rebar) — the rebar CLI, the import rebar library, and the lean workflow engine — pulls only three runtime dependencies: pyyaml (the workflow DSL loader), jsonschema (the schema-registry + workflow input/output-contract validator), and referencing (the JSON Schema $ref resolver jsonschema builds on); the engine core and reconciler are otherwise stdlib-only. Everything else is an optional extra, lazy-imported so the base stays light (CI enforces that):

• Optional runtime capabilities — install what you serve:
  • [mcp] — the rebar-mcp server (mcp>=1.9).
  • [agents] — the LLM agent-operations framework + agentic workflow steps (rebar review, the code_review workflow): the provider-agnostic pydantic-ai runtime (pydantic-ai-slim[anthropic]) plus json-repair.
• Development & authoring extras — not needed to run or serve rebar:
  • [eval] — prompt evaluation (rebar prompt eval) with Inspect AI; an authoring/CI capability.
  • [tracing] — an OpenTelemetry OTLP trace sink (write-only; never read back into a rebar decision), for diagnostics.
  • [dev] — the test/lint/type tooling (pytest, ruff, mypy, hatchling). pip install -e '.[dev]' also self-references [agents] so the validation tests run rather than skip, and is required to run the full test suite (the interface-parity tests import the MCP server, so they error — not skip — without mcp).
  • Node/npm — needed only for the workflow visual editor's front-end: rebuilding its vendored bundle (src/rebar/llm/workflow/editor_assets/, the bpmn-js editor) and running the faithful editor E2E tier (tests/e2e/, which drives the real bpmn-io libraries). Both are developer-only — the built bundle is committed/shipped and the E2E tier self-skips when Node is absent — so neither the base install nor the default test suite needs Node. See docs/workflow-editor.md.

See Install and Tests.

Install

rebar ships from one Python package — PyPI distribution nava-rebar (the import package and commands stay rebar / rebar-mcp). Pick the channel that fits. (System prerequisites in all cases: git and python3 (≥ 3.11); write serialization uses a built-in fcntl.flock + mkdir lock with no external binary; acli only for live Jira reconciliation.)

Homebrew (CLI)

brew install navapbc/rebar/rebar
# or: brew tap navapbc/rebar && brew install rebar

Installs the rebar CLI (and the rebar library inside the formula's venv). For the MCP server via Homebrew users, install the [mcp] extra with pipx/uvx below.

PyPI — pipx / pip

Runtime (prod) — install what you'll run:

pipx install nava-rebar              # isolated CLI on PATH: rebar (+ lean workflow engine)
pip  install nava-rebar              # library: import rebar  (runtime deps: pyyaml, jsonschema)
pip  install 'nava-rebar[mcp]'       # + MCP server: rebar-mcp
pip  install 'nava-rebar[agents]'    # + LLM agent ops + agentic workflow steps (rebar.llm)
pip  install 'nava-rebar[eval]'      # + prompt evaluation: `rebar prompt eval` (Inspect AI)
pip  install 'nava-rebar[tracing]'   # + OTLP trace sink (write-only)
pip  install 'nava-rebar[agents,eval,tracing]'   # the union, if you want it all

The base install runs scripted workflows (rebar workflow new/validate/show/run) with no extra; agentic workflow steps and rebar review need [agents]. Authoring a workflow visually — rebar workflow edit <file>, a local bpmn-js editor that round-trips the diagram back to the IR — also needs no extra and no Node/npm: the editor front-end ships pre-built in the wheel and is served locally (no CDN). For what the engine is for — when to author a workflow vs a bespoke op, the YAML DSL, the three-pass review pattern, and the prompt-library + eval seam — see docs/workflow-engine.md; for visual editing specifically see docs/workflow-editor.md.

The [agents] extra adds the optional LLM agent-operations framework (rebar.llm) — tool-using agents that review tickets/code and emit structured findings, over library / CLI (rebar review) / MCP. It is multi-provider (Claude and ChatGPT out of the box, plus Gemini and OpenAI-compatible local servers like LMStudio/Ollama via REBAR_LLM_MODEL/REBAR_LLM_MODEL_PROVIDER/ REBAR_LLM_BASE_URL) and is never required by core rebar — none of the LLM stack is installed or imported unless you opt into this extra (CI enforces it); see docs/llm-framework.md.

MCP server — from the MCP Registry

Listed in the MCP Registry as io.github.navapbc/rebar. Registry-aware MCP clients can add it by that name; or register it directly in your client config (zero pre-install via uvx):

{
  "mcpServers": {
    "rebar": {
      "command": "uvx",
      "args": ["--from", "nava-rebar[mcp]", "rebar-mcp"],
      "env": { "REBAR_ROOT": "/path/to/your/repo" }
    }
  }
}

(Already pip/pipx-installed nava-rebar[mcp]? Use "command": "rebar-mcp" instead.) Server flags: REBAR_MCP_READONLY=1 exposes only read tools; reconcile is dry-run unless REBAR_MCP_ALLOW_JIRA_SYNC=1. Both flags accept any case-insensitive truthy value — 1, true, or yes (surrounding whitespace tolerated); anything else (incl. unset) is off.

Private-repo fetch credentials (code-reading gates)

The LLM code-reading gates (review_plan, verify_completion, review_ticket, review_code, scan_spec) default to attested mode: they git fetch the verified ref from origin and read an immutable snapshot at the pinned SHA — never the server's mutable checkout. So a server pointed (REBAR_ROOT) at a private repository needs read credentials to fetch: a git credential helper, a deploy key, or a token in the server's clone. With no credentials, attested mode fails closed with a descriptive, actionable error (it never hangs on a prompt — GIT_TERMINAL_PROMPT=0); source=local (read the in-place checkout, never signed) is the back-out that needs no fetch. Full semantics, the HMAC trust model, and the snapshot env knobs (REBAR_GATE_TMPDIR, the disk-space watermark, the EFS/NFS flock caveat) are in docs/repo-snapshot-gates.md.

From source

git clone https://github.com/navapbc/rebar && cd rebar
pip install .              # library + CLI (runtime deps: pyyaml, jsonschema)
pip install '.[mcp]'      # + MCP server (FastMCP)
# Developing rebar itself — the full dev environment (test/lint/type tooling +
# the agents stack so the LLM validation tests RUN, not skip):
pip install -e '.[dev]'

Contributing changes? GitHub is a read-only mirror — main only advances via Gerrit's two-vote gate (LLM-Review + Verified/CI). New contributors: start with the friendly walkthrough docs/your-first-change.md; the full reference is CONTRIBUTING.md (clone from Gerrit, push to refs/for/main, then land it with a plain Gerrit Submit once both votes pass — main is Rebase-If-Necessary, so Gerrit rebases onto the tip and submits server-side).

Packaging note — why rebar installs unpacked to disk. The library, CLI, MCP server, and the whole read/write core run in-process in Python. The one component that runs as a subprocess is the Jira reconciler, which ships under src/rebar/_engine/ as package data (python -m rebar_reconciler, plus the jira-capability-probe.py script and the alias wordlist): it is launched and read from the filesystem as real on-disk files, so the package must be installed unpacked to a real directory and zipimport / zip-safe bundles (zipapp, shiv, PEX, Lambda zips) are unsupported. Every standard install satisfies this: pip/pipx wheels (hatchling builds unpacked), editable installs, and Homebrew all land real files. engine_dir() asserts the engine dir is present on disk at the first reconciler call and fails loudly otherwise.

Advanced (optional) — gate commits with self-hosted code review. Not needed for standard rebar use. If you want every commit to main automatically LLM-reviewed before it can land, you can self-host Gerrit + the rebar review-bot on AWS (the bot imports the same rebar.llm review kernel the MCP server exposes) and demote GitHub to a read-only mirror that only advances via Gerrit after the LLM-Review vote passes. See docs/gerrit-aws-setup.md for the server setup. (This repo runs exactly that setup — see the contributor note above and CONTRIBUTING.md.)

CLI

The complete, always-current command reference — every subcommand with its usage — is docs/cli-reference.md, generated from the CLI's own help data (so it can never drift from the code). The essentials to get moving: rebar init → rebar create <type> "<title>" → rebar ready → rebar claim <id> --assignee <you> → rebar transition <id> <current> <target>.

Run rebar help (or rebar --help / -h) for the subcommand overview, and rebar <subcommand> --help (or rebar help <subcommand>) for a specific subcommand's usage — --help prints usage and never executes the command. Help is only recognized as the first argument after the subcommand, so a --help/-h/help that appears inside a free-text parameter (title, comment body, search query, …) is treated as literal text, not a help request.

Repo root is resolved from REBAR_ROOT, falling back to the git toplevel of the working directory.

Structured output. Every data-returning command emits machine-readable JSON via the canonical --output json flag (short -o json; --output llm gives a token-minified shape for show/list/ready). Each distinct JSON shape is documented by a JSON Schema and validated across the CLI, library, and MCP in CI. See docs/output-schemas.md for the per-command contract and the schema source-of-truth.

Repo-wide health with validate. rebar validate takes no ticket id — it scans the whole store and prints an overall tracker-health score (1-5, exit 0-4) bucketed into critical / major / minor / warning findings (--output json, --terse, --verbose, --fix). Passing it a ticket id errors. (rebar also has per-ticket structural gates that each take an <id> and verify a ticket is shaped like dispatchable work — every type needs an ## Acceptance Criteria checklist. See the ticket template and gate reference in docs/plan-review-criteria-guide.md.)

Links. rebar link <id1> <id2> <relation> requires a relation; the six relations are blocks, depends_on, relates_to, duplicates, supersedes, discovered_from. rebar unlink <source> <target> takes no relation argument — it is pair-scoped and removes the most-recently-created link between that ordered pair, one per call, so to remove multiple links between the same pair you call unlink repeatedly. Note that blocking links (blocks/depends_on) may be promoted up the parent hierarchy when created (see below), so unlink must target the promoted (ancestor) endpoint to remove such a link.

Ticket work also leaves an HMAC-signed attestation — a machine-checkable proof that a gate ran and that its verified steps are unaltered. For most projects this is produced automatically by the code-review, plan-review, and completion-verifier gates, so you never sign by hand. To sign manifests yourself or customize the process, see docs/manifest-signing.md.

Hierarchy promotion of blocking links

For blocking dependencies only (blocks, depends_on), rebar promotes the link endpoints up the parent hierarchy so the dependency sits between tickets at a comparable level (epic↔epic, story↔story, task/bug↔task/bug). When it does so it emits a REDIRECT: A→B promoted to … note. Non-blocking relations (relates_to, duplicates, supersedes, discovered_from) are linked exactly as given, with no promotion.

The store auto-commits and auto-pushes every write

Every rebar write (create, edit, transition, claim, link, …) auto-commits its event to the tickets branch and auto-pushes that branch to origin/tickets whenever an origin remote exists. Local ticket activity is therefore shared with the remote immediately — including test/scratch tickets, so be deliberate when working against a repo with a shared tickets remote. The push is best-effort: with no origin remote nothing is pushed, and a push failure (e.g. non-fast-forward it cannot auto-merge, or no network) never fails the write — it leaves the local commit intact and the branch diverged. rebar fsck reports PUSH_PENDING when the local tickets branch is ahead of origin/tickets, so unpushed activity is observable. See docs/concurrency.md for the push/merge-retry algorithm.

Running locally, offline, or read-only. This auto-sync is configurable when you don't want the store talking to a remote (the full key set and env names are in docs/config.md):

• sync.push (env REBAR_SYNC_PUSH) — always (default) pushes each write synchronously; async pushes in the background so per-write network latency doesn't serialize a batch; off keeps commits local and never pushes (fsck still surfaces PUSH_PENDING).
• sync.pull (env REBAR_SYNC_PULL) — on (default) lets reads fetch from the remote (the freshness policy below); off gives a pure-local replay (offline work, tight loops, or right after a write that already synced). Pass --no-pull to a single read subcommand for the same effect (e.g. rebar list --no-pull).
• mcp.readonly (env REBAR_MCP_READONLY=1) — serves only read tools over MCP, so no writes — and therefore no commits or pushes — happen at all.

How big can it get? Reads stay sub-second into the thousands of tickets; writes are bounded by the per-event git commit (~25–30/s). See docs/scale-envelope.md for representative measured numbers, git-growth expectations, and the compaction/maintenance commands, and docs/import-export.md for bulk NDJSON export/import.

Reads share one freshness policy across CLI, library, and MCP

Every read — show, list, ready, search, deps — first runs a throttled (≤1/min), best-effort git fetch + reconverge of the local tickets branch with origin/tickets, so a read reflects collaborators' pushes within at most a minute. This is one contract shared by all three interfaces: CLI, library (rebar.list_tickets(), …), and the MCP read tools all resolve through a single read implementation. (Previously only CLI reads synced, leaving MCP — the primary agent surface — with the stalest reads; that divergence is gone.) To skip this fetch for a pure-local replay, set sync.pull=off or pass --no-pull — see Running locally, offline, or read-only above. Only the network fetch/merge is affected; the local reduce/cache path is unchanged. See docs/concurrency.md.

The on-disk store is not human-readable — read it with rebar

The tickets branch is rebar's internal storage format, not a document for people to read. Each ticket is a directory of append-only JSON event files (${hlc}-${uuid}-${TYPE}.json); the current state of a ticket is what you get by replaying those events through the reducer. Two consequences follow:

• It isn't laid out in order. Event files are named by a Hybrid Logical Clock
  • UUID and merge across clones as a union, so the files for one ticket are not a top-to-bottom narrative — they are an unordered set that only becomes meaningful after the reducer sorts and folds them. A single EDIT/STATUS/TAG_DELTA file in isolation tells you a delta, not the ticket.
• The current state is computed, never stored. Nothing on the branch holds the compiled "current" ticket except a local, rebuildable .cache.json (gitignored). Reading the raw files by hand will mislead you — a later event may supersede an earlier one, a SNAPSHOT may fold many away, and concurrent forks resolve by a deterministic rule you'd have to apply yourself.

So don't cat the .tickets-tracker/ worktree to find out where a ticket stands — use the read commands, which run the reducer for you: rebar show <id>, rebar list, rebar deps <id>, rebar search <query> (CLI), the matching library calls (rebar.show_ticket(...)), or the MCP read tools.

For reference, docs/sample-ticket-log.jsonl is a small synthetic event log (one event per line) showing what the underlying data actually looks like — a two-agent epic + child tickets exercising create/claim/comment/link/tag/file-impact/sign/transition. Note that its lines are deliberately not in timestamp order: that is the point. The event body schema is documented in docs/event-schema.md.

Python library

import rebar

rebar.init_repo(repo_root="/path/to/repo")
tid = rebar.create_ticket("story", "Add login page", priority=2)
ticket = rebar.show_ticket(tid)                 # TicketState
tickets = rebar.list_tickets(status="open")     # list[TicketState]
try:
    rebar.transition(tid, "open", "in_progress")
except rebar.ConcurrencyError:
    ...                                          # ticket changed since last read

result = rebar.reconcile("dry-run")              # Jira sync (non-mutating)

# Cryptographic attestation (environment-bound HMAC):
rebar.sign_manifest(tid, ["unit tests: PASS", "security review: clean"])
verdict = rebar.verify_signature(tid)            # {"verified": True, "verdict": "certified", ...}

# Native, in-process reads (no subprocess):
from rebar import reduce_all_tickets, reduce_ticket

Typed return contract. The schema-backed rebar.* functions are annotated with TypedDicts in rebar.types (e.g. TicketState, TransitionResult, ClaimResult), so a type checker knows which keys a return value carries. These are derived from the canonical JSON Schemas and describe the guaranteed keys — returns stay plain dicts and the runtime shape is open (extra keys may appear), so this is a floor, not a closed universe. Import them for annotations/TypedDict access:

from rebar.types import TicketState, TransitionResult

t: TransitionResult = rebar.transition(tid, "open", "in_progress")

Stable exception surface. rebar.RebarError (base) and its subclass rebar.ConcurrencyError are the public exceptions. RebarError carries .returncode (the underlying engine exit code) and .stderr; ConcurrencyError (exit 10) means a status-dependent op (transition/claim/reopen) lost an optimistic-concurrency race — re-read and retry, don't force. Catch RebarError to handle any rebar failure uniformly.

What's stable to depend on. rebar is versioned 0.x; see docs/api-stability.md for the per-surface stability matrix (CLI, --output json schemas, the rebar.* facade, MCP tools, the event wire format, and config keys) and what "may change before 1.0" means for each.

MCP server

rebar-mcp          # stdio transport

Exposes ticket operations as MCP tools. The complete tool reference, grouped by gate tier (read-only / LLM-gated / write-gated), is docs/mcp-reference.md (generated from the server's own registrars). reconcile defaults to dry-run (live requires REBAR_MCP_ALLOW_JIRA_SYNC=1). Set REBAR_MCP_READONLY=1 to expose only the read tools (no write/mutation tools). To register it in an MCP client (registry name io.github.navapbc/rebar, or a direct uvx config), see Install → MCP server above.

Maintainers: the registry manifest lives in server.json; publish/update it with the mcp-publisher CLI (see docs/releasing.md). The registry verifies PyPI-package ownership via this annotation (kept in this README, which is the PyPI long description):

mcp-name: io.github.navapbc/rebar

License

Apache-2.0 — see LICENSE.

Configuration

rebar reads TOML config from [tool.rebar] in pyproject.toml or a standalone rebar.toml (nearest up-tree, stopping at .git), falling back to a user config at ~/.config/rebar/config.toml (honoring $XDG_CONFIG_HOME). Precedence, highest first: rebar -c SECTION.KEY=VALUE / CLI flag > REBAR_<SECTION>_<KEY> env > project config > user config > built-in default. rebar config prints the resolved values and which layer each came from.

[tool.rebar]
verify.require_completion_verification_for_close = true  # gate work-ticket close on a PASS
                                           # completion verdict (signed onto the ticket);
                                           # fail-closed. Default false.
ticket.display_mode = "auto"               # auto | canonical | alias | short
compact.threshold   = 10
sync.push = "always"                       # always | async | off
sync.pull = "on"                           # on | off
mcp.readonly = false
scratch.base_dir = ""                      # default <repo>/.rebar/scratch
tracker.dir    = ".tickets-tracker"        # store worktree/symlink dir (env REBAR_TRACKER_DIR)
tracker.branch = "tickets"                 # orphan branch the event log lives on (env REBAR_TRACKER_BRANCH)

The full key set, the REBAR_<KEY> env names, and deprecation aliases are in docs/config.md.

When the close gate is enabled, closing a story/epic requires a certified signature made at the current HEAD — sign a manifest of verified steps (rebar sign <id> '[...]') then rebar transition <id> closed; re-sign if HEAD moved, or bypass with --force-close=<reason>.

rebar keeps its writable state under .rebar/ at the repo root. The scratch store defaults to <repo>/.rebar/scratch/ (override with scratch.base_dir / REBAR_SCRATCH_BASE_DIR), and one-shot migration stamps are written under .rebar/ as well.

Tests

Run the suite from an environment with the [dev] extra installed (a venv is recommended); the interface-parity tests import the MCP server, so a bare interpreter without the mcp extra will error rather than skip.

python -m venv .venv && source .venv/bin/activate
pip install -e '.[dev]'                       # editable + pytest, mcp, ruff, mypy
pytest -m "not integration"                   # the single entry point (CI runs this)
pytest tests/interfaces                       # interface-parity tier only
pytest tests/scripts                          # engine/reconciler tier only

pytest is the single entry point. The engine is pure in-process Python (the bash engine and its .sh suites were removed in the bash→Python migration — see docs/bash-migration.md). CI (.github/workflows/test.yml) runs pytest -m "not integration" on Ubuntu and macOS for every push and PR. The integration tier (live Jira / network) is excluded from that default run; run it explicitly with credentials via pytest -m integration.

The Python suite is sub-divided by concern:

• tests/scripts, tests/unit — the in-process engine (reducer, graph, reconciler).
• tests/interfaces — proves the library, CLI, and MCP interfaces behave identically over one git-backed store:
  • test_parity.py runs each operation through all three interfaces (and a cross-interface coherence check: write via one, read via the others);
  • test_surface.py pins the per-interface capability surface (e.g. MCP has no init; there is no classify);
  • test_library.py / test_cli.py / test_mcp.py cover per-interface specifics (typed exceptions, exit-code passthrough, read-only/live gates).