interlinked-setup

interlinked-setup — install, operate & troubleshoot the harness

Hook project discovery walks up from the tool call's cwd to a configured ancestor: .interlinked/config.json, guard-rules.json, harness.pid, or a socket marks a project. A descendant directory containing only hook receipts or prediction state does not shadow that ancestor. An explicitly marked nested project retains its own identity. If no configured ancestor exists within the bounded walk, the nearest .interlinked/ remains the fallback; directory presence alone does not prove a live daemon.

The local guard-prediction oracle defaults to shadow, independently of the optional Supermodel graph protocol. Shared .interlinked/config.json accepts harness.guard_prediction.mode: "enforced" | "shadow" | "off". Use interlinked-harness for exact prediction/reconciliation receipts and limitations. Receipts are required only in explicit enforced mode. Edit-time function token, cyclomatic and cognitive cap crossings are advisory by default; shared harness.edit_metrics.mode: "enforced" restores refusal, while "advisory" states the default explicitly. Cap values and protected baselines do not change. Native cold fallback can run this AST oracle; standalone generated hooks need the daemon and report NOT CHECKED without it. Physical file size is advisory in both warm and cold paths; rebuilding must not revive the old line-count block.

Claude compiler batching activates only after a session sends native PostToolBatch. After daemon restart, edits use the ordinary compiler path until that boundary is observed again; previously queued work remains durable and is checked at Stop/commit. Keep the installed PostToolBatch hook current. It delivers repair context rather than cancelling the agent loop for ordinary compiler findings. See interlinked-verify.

Skill refresh preserves modified installed copies. If they still describe sub-cap complexity slew blocks, compare them with canonical skills/*/SKILL.md, retain local customizations/backups, and reconcile deliberately. Restarting a daemon cannot erase guidance already loaded into a model's conversation.

Non-gating native hook commands use a compact missing-runtime diagnostic and exit 1; gating hooks retain the self-contained fail-closed fallback and exact repair allowlist. After rebuilding, use interlinked install-hooks --refresh --preserve-mode to refresh installed command text, then verify interlinked harness status --json shows the serving build. Keep PostToolUse scoped to mutating tools; do not add PostToolUseFailure duplicates.

Provider qualification must exercise native edit routes, advisory file-size delivery, and completion after ordinary Stop advisories. Claude and Codex keep those advisories on stderr; explicit repair blocks still request continuation. Codex rewrites include permissionDecision: allow. Retain registrations and native transcripts: an encoded response alone does not prove native enforcement.

Before behavioral qualification, use interlinked tests readiness <language> --json, then run a real project suite. Python readiness probes the selected environment for pytest, pytest-cov and coverage; it does not install or infer passing tests. The hook installer also ignores the named generated payload-key, enforcement-status and coverage-runtime-estimate outputs. These ignore entries never exempt policy files or make runtime-input caching sound.

For an externally frozen policy, keep the policy files on read-only supervisor mounts and pin their parent directories as mount points; a read-only file under a replaceable parent is insufficient. Keep daemon sockets, atomic receipt writes, proposals and intended ratchet state writable. Qualify CLI mutations, direct replacement, local overrides and daemon restart under the actual non-root agent identity before claiming immutable effective policy. A same-UID daemon and --by labels do not provide an authorization boundary.

Interlinked is a local, offline-first guard layer for AI coding agents. Repository config, logs, and daemon state live under <repo>/.interlinked/; installation also writes selected providers' settings, managed bridge files, and installed skill directories at their provider-specific project or user locations. A local daemon ("the harness", a Unix-socket server) evaluates each event the installed runner exposes to its hook surface. This skill covers turning it on, keeping the daemon healthy, configuring it, and turning it off. The remote server is optional — hooks, guard, and activity capture work with zero network.