Foundry

Foundry is an event-driven workflow engine for engineering automation. It runs quality gates, focused coding tasks, durable multi-cycle campaigns, maintenance, dependency remediation, releases, and scheduled operational workflows across a registry of Git repositories.

Foundry has two processes:

• foundryd is the long-running daemon that owns workflow state, routes events, runs task blocks, and records traces.
• foundry is the command-line client used by people, scripts, and agents.

The two communicate over gRPC. Every workflow is observable as an event chain, and mutating work can be simulated with dry-run throttle.

Why Foundry

Traditional automation expresses a whole workflow as one script. Foundry expresses reusable steps as task blocks connected by typed events. This gives the system a few useful properties:

• the same workflow can be triggered manually, on a schedule, or by another workflow;
• independent projects and event branches can run concurrently;
• every block execution and emitted event is retained in a trace;
• observer blocks still run in dry-run mode while mutators simulate success;
• coding work is checked against repository intent and mechanical gates before it can land;
• broader missions can continue across multiple evidence-reviewed tasks without relying on a stale, pre-cut backlog.

Install

Homebrew

brew tap svetzal/tap
brew install foundry

From source

Building from source requires Rust 1.85 or newer and protoc.

git clone https://github.com/svetzal/foundry.git
cd foundry
./install.sh

The installer builds both binaries and installs them to ~/.cargo/bin. On macOS it also applies stable ad-hoc signing identifiers so privacy grants survive future upgrades.

Quick Start

Start the daemon:

foundryd

Register a Git repository in another terminal:

foundry registry add \
  --name my-project \
  --path /absolute/path/to/my-project \
  --stack rust \
  --agent codex \
  --repo owner/my-project \
  --branch main \
  --iterate \
  --maintain \
  --push

Derive and save quality gates:

foundry gates --init my-project
foundry validate my-project

Run one concrete coding task:

foundry task my-project \
  "Add a --quiet flag and prove it suppresses progress output"

Inspect the result:

foundry history --project my-project
foundry trace <event-id> --verbose

Choose the Right Workflow

Need	Command
Execute one concrete, immediately actionable objective	foundry task <project> "<objective>"
Complete a broader mission whose next step depends on current evidence	foundry campaign …
Improve a project against its charter	foundry iterate <project>
Detect intent drift without changing code	foundry scout <project>
Check quality gates without changing code	foundry validate <project>
Run registered maintenance work	foundry run [--project <name>]
Diagnose and remediate a failing GitHub Actions pipeline	foundry pipeline <project>
Run an agent-driven release	foundry release <project>

Prefer these convenience commands to raw foundry emit; they set up event streaming, progress display, and trace rendering for the workflow.

Tasks

A task is one concrete coding objective. Foundry:

1. creates an isolated Git worktree;
2. checks repository intent and resolves project gates;
3. runs the coding agent;
4. executes required gates in the worktree;
5. asks a separate reviewer for a typed verdict; and
6. lands or preserves the result.

The verdict is one of complete, remainder, defect, blocked_on_decision, or runner_error. Complete work lands when its required gates pass. Converging remainder work may also land when at least one required gate ran and all required gates passed; its finite gaps become the next campaign objective. Defects, decisions, and runner failures never land.

Any work that does not land is preserved on a named remote branch or, when a push is unavailable, in a Git bundle under ~/.foundry/preserved/.

Campaigns

A campaign is a durable mission, not a queue of pre-written tasks. On every cycle, Foundry inspects the delivered trunk, any accumulated preserved work, the last typed task result, owner decisions, recent objective history, and the campaign's done evidence. It then chooses exactly one outcome:

• done — evidence proves the mission complete;
• advance — dispatch one next objective; or
• escalate — stop for budget, policy, or human judgement.

Create a definition:

{
  "name": "retrieval-identity-v1",
  "project": "my-project",
  "mission": "Prove both retrieval entrypoints preserve raw response identity.",
  "context_paths": ["docs/retrieval-contract.md"],
  "done_evidence": [
    {
      "kind": "gate",
      "command": "cargo test retrieval_identity",
      "required": true,
      "artifacts": ["tests/retrieval_identity.rs"]
    },
    {
      "kind": "review",
      "statement": "Both real entrypoints are covered without masking response IDs."
    }
  ],
  "budget": { "max_cycles": 8 },
  "escalation": ["A public compatibility decision requires an owner."],
  "authorized_by": "Owner Name",
  "agent_provider": "codex"
}

Then add and advance it:

foundry campaign add ./retrieval-identity-v1.json
foundry campaign show retrieval-identity-v1
foundry campaign advance retrieval-identity-v1

Task results request the next campaign advance automatically. Use the control commands when intervention is required:

foundry campaign pause retrieval-identity-v1
foundry campaign resume retrieval-identity-v1
foundry campaign decide retrieval-identity-v1 \
  --decision "Preserve the existing public response type."
foundry campaign resume retrieval-identity-v1 --add-cycles 2
foundry campaign complete retrieval-identity-v1 \
  --reason "Owner-reviewed production evidence confirms the mission shipped."
foundry campaign cancel retrieval-identity-v1 \
  --reason "Superseded by the streaming rewrite." --now

Campaign done-evidence gates run during formation against delivered trunk. They are distinct from the project's task gates, which run inside each isolated task worktree. Required campaign gates must therefore be runnable from the repository checkout. Put deployment or remote-host assertions in an owner-reviewed review statement, or make the remote probe optional.

See Tasks and Campaigns for the lifecycle, evidence model, landing rules, provider pauses, budget extensions, recovery, and cycle-level observability.

Scheduled Work

foundryd includes declarative cron-like sentinels. The canonical schedules run nightly maintenance, a daily commit digest, a periodic operations digest, and a nightly supply-chain scan.

foundry sentinel list
foundry sentinel show nightly-maintenance
foundry sentinel disable nightly-maintenance
foundry sentinel enable nightly-maintenance

Sentinel definitions live in ~/.foundry/sentinels.json. Enable and disable commands notify the running daemon immediately.

Observability

Every event carries trace, span, causation, and scatter/gather context. Campaign task-side events also carry the campaign name and cycle number, so a cycle can be reconstructed from the event stream without inferring boundaries from timestamps.

foundry status
foundry watch --project my-project
foundry history --project my-project
foundry trace <event-id> --verbose

Persistent data is stored below ~/.foundry/ by default:

Path	Contents
registry.json	Registered projects and enabled actions
campaigns.json	Durable campaign definitions and state
events/YYYY-MM.jsonl	Append-only workflow events
traces/YYYY-MM-DD/	Renderable workflow traces
worktrees/	Disposable task worktrees
preserved/	Git bundle fallbacks for work that did not land
sentinels.json	Scheduled workflow triggers

Online registry and campaign commands are daemon-authoritative. Use --offline only for deliberate direct-file recovery while foundryd is stopped.

Dry Runs

Throttle controls mutating task blocks without removing the rest of the workflow:

foundry run --project my-project --throttle dry_run

Observers execute normally. Mutators produce their simulated success events without changing repositories or external systems.

Documentation

The complete documentation is an mdBook under book/:

• Getting Started
• Project Registry
• Tasks and Campaigns
• Workflow Formations
• CLI Reference
• Architecture

Build it locally with:

mdbook build book

Development

Foundry is a Rust workspace containing:

• foundry-sdk — event, payload, gate, campaign, and task-block contracts;
• foundry-engine — event routing, propagation, retries, and gathering;
• foundry-blocks — production workflow blocks;
• foundryd — the daemon and gRPC service; and
• foundry-cli — the command-line client and renderers.

Run the required quality gates:

cargo fmt --all -- --check
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace
RUSTDOCFLAGS="-D warnings" cargo doc --workspace --no-deps
cargo deny check
mdbook build book

See AGENTS.md for repository conventions and release procedure.

Licence

Foundry is released under the MIT Licence.