A lightweight PID → PL → Engineer → Auditor delivery-control framework for Claude Code: point it at a project with product intent defined, and it runs a bounded implementation-and-independent-audit loop toward a delivery candidate, while a Human retains product authority and final acceptance.
North star: PID in → autonomous Claude-native delivery → independently scrutinised product out. See docs/FORGE-NORTH-STAR.md.
Everything lives in this repository: code, doctrine, PIDs, evidence, decisions. Redis/fabric may carry live messaging between a session and its subagents during a run — it holds nothing durable and nothing this repo depends on for correctness. If it isn't in git, it didn't happen.
./ops/bootstrapInstalls nothing itself but verifies gitleaks is on PATH and wires the
committed pre-commit hook (scripts/hooks/pre-commit, activated via
core.hooksPath — not a local-only .git/hooks file, so the control is
itself part of Forge's history, not host-local convenience).
docs/
FORGE-NORTH-STAR.md core doctrine: purpose, authority model, durable truth
ARCHITECTURE.md
EVIDENCE-DOCTRINE.md when a number is actually evidence (all three roles)
OPERATING-MODEL.md
PID-TEMPLATE.md
evidence/ per-increment evidence trail
.claude/skills/forge/SKILL.md the PL-facing delivery loop
.claude/agents/
forge-engineer.md bounded implementation subagent
forge-auditor.md independent review subagent (mandatory real-browser
verification for PRODUCT_USABLE+ targets)
scripts/hooks/ committed git hooks (secret scanning)
ops/ small operator scripts, no framework
Forge does not include a local-model benchmark or delegation harness.
Claude Code's native subagent dispatch (Engineer/Auditor) already provides
an independent second set of eyes on any change; a separate delegation
layer for a cheaper or local model would duplicate that without a proven
need. See docs/FORGE-NORTH-STAR.md#anti-bloat-doctrine.