Forge — replicable AI-SDLC agent operating system

A vendored, installable copy of the MockServer AI-SDLC system (DVRR: Decompose · Verify · Review · Reintegrate) — per-session git-worktree isolation, a fail-closed commit/merge gate chain, adversarial review with a binding constitution, risk-scaled autonomy, operator kill-switch, and a golden-task eval harness for the AI components themselves.

Operating manual: OPERATING-MANUAL.md — install flow, daily workflows with diagrams, review system, safety rails, troubleshooting.

Research and decision record: docs/ (start with docs/README.md; the adoption decision and installer design are in docs/06-discussion-outcome.md). Upstream provenance and deliberate deviations: system/UPSTREAM (Apache-2.0, mock-server/mockserver-monorepo).

Install into a repo

From the target repository root:

/forge-init          # in Claude Code — full install + per-project initialization

or mechanically only (leaves gates fail-closed until init):

<forge-checkout>/bin/forge-install.sh \
    [--project-name NAME] [--branch BRANCH] [--target DIR]

(<forge-checkout> = wherever this repo is cloned; the installer resolves its own payload location.)

/forge-init = forge-install.sh (copy + token/branch substitution + .gitignore + manifest) plus the judgment layer: stack detection → validation tables, Gate-1 test command + blast-radius suite, project review triggers, AGENTS.md project sections, agent/skill project context, eval fixtures + baselines — finishing with the system's own gate chain on the install diff (control-class change: review-final + evals + explicit human approval; never auto-committed).

Personal-engine mode

When the DVRR engine is already installed in the user's personal opencode config (~/.config/opencode/agents/DVRR.md plus ~/.config/opencode/dvrr/rules/), /forge-init defaults to personal-engine mode. This keeps the engine in user scope and drafts per-repo context in .dvrr/context.pending/. Each file requires explicit user approval before moving to .dvrr/context/ and completing the local .dvrr/forge-manifest. It skips forge-install.sh, repo payloads and eval baselines; /.dvrr/ is added to .gitignore if needed. Init must run from a normal agent, not DVRR. A repo without a commit skill/process keeps DVRR commits disabled.

Pass --repo-install to force the legacy full repo install, or --personal-engine to request the user-scope engine layout; the installed region oracle must still validate. Without an installed engine or a flag, the default is repo-install. Both flags together require choosing one before any write.

Layout

Path	What
system/engine/	Tier-1 payload — byte-close to upstream for diffable syncs
system/template/	Tier-2 payload — upstream files with {{FORGE_*}} tokens and <!-- FORGE:REGION ... --> blocks that /forge-init fills (defaults fail closed)
system/seeds/	Tier-3 source material — eval-task templates, per-stack validation snippets, brownfield exploration protocol
system/UPSTREAM	provenance manifest: upstream commit, deviations, sync procedure
system/LICENSE-upstream	Apache-2.0 license of the upstream repo
bin/forge-install.sh	mechanical installer (idempotent)
docs/	research bundle + decision records

The canonical /forge-init command is versioned at commands/forge-init.md; install it to user scope on each machine so it works in any repo. The files reference the forge checkout via a {{FORGE_ROOT}} token — substitute your checkout path at copy time (run from this repo's root):

• Claude Code: sed "s|{{FORGE_ROOT}}|$PWD|g" commands/forge-init.md > ~/.claude/commands/forge-init.md
• Codex and Kimi Code: cp -R skills/forge-init ~/.agents/skills/ && sed -i '' "s|{{FORGE_ROOT}}|$PWD|g" ~/.agents/skills/forge-init/SKILL.md — both read the cross-tool ~/.agents/skills/ directory; invoke as $forge-init in Codex or /skill:forge-init (shorthand /forge-init) in Kimi (list with /skills).

(All load at session startup — restart the session after copying. The skill body is generated from commands/forge-init.md; keep the two in sync.)

Model routing (decision, 2026-07-09)

Harness	Strong (implementer, review-final, security-auditor, debugger)	Weak (everything else)
opencode	zai/glm-5.2	minimax/MiniMax-M3
Claude Code	fable	opus
Codex	gpt-5	gpt-4o / gpt-4o-mini (upstream kept)
Kimi Code	k3	k3 (single model family — no per-role routing; effort via global [thinking])

Per-agent temperatures (opencode) and reasoning efforts (Claude) are kept verbatim from upstream. Codex is a first-class third harness (forge-original, no upstream counterpart): per-agent model/effort in .codex/agents/*.toml, kill-switch deny-list as execpolicy rules in .codex/rules/forge.rules, Stop-hook telemetry in .codex/hooks.json, and /commit + /worktree-merge as skills in .agents/skills/ (invoke as $commit / $worktree-merge, or Codex picks them implicitly from their descriptions). Codex loads a repo's .codex/ layer only after the operator trusts the repo (fail-closed until then); skills in .agents/skills/ are outside that trust gate.

Kimi Code CLI is the fourth harness (forge-original, 2026-07-18): it has no custom sub-agent registry, so the 11 DVRR roles ship as project skills in .kimi-code/skills/ (bodies verbatim from .claude/agents/); reviewer roles run on the built-in read-only explore sub-agent. Kimi has no project-level config, so the kill-switch deny rules + Stop-hook telemetry ship as .kimi-code/config-snippet.toml, merged once into ~/.kimi-code/config.toml (fail-open until merged — the snippet's deny-listed commands are universally destructive, so the global merge is safe beyond forge repos). ZCode (Z.ai's desktop ADE) is a companion harness, advisory tier: it reads AGENTS.md natively but has no project-level agents/commands/config and no enforced read-only sub-agents — import .agents/skills/ via its Settings → Skills UI; run gate-chain commits from a CLI harness. Changing routing after install is a control-class change inside the installed system (evals + review-final + human approval).

Syncing with upstream

system/engine/ is kept byte-identical to upstream wherever possible; deliberate edits carry a forge: modified from upstream marker. To sync: clone upstream, diff -r against engine/ (path-mapped), review hunks, pull what applies, bump the hash in system/UPSTREAM.