hearting

Hearting (formerly agent_setting) is a local-first workflow layer for coding agents. Rather than configuring a single runtime, it defines shared contracts once and projects them onto the native skill, agent, hook, mode, and command surfaces that Claude Code, Codex, and OpenCode each discover. The stated goal is to carry research, planning, implementation, and verification through one consistent cycle across all three. Key mechanisms described in the README: - Routed capabilities: 12 entry routers over 26 capabilities. Before substantial work, the agent proposes a five-field route card (task, reason, route, scope, completion) for approval. - Intensity ladder: direct, quick, standard, strong, thorough, and adversarial levels select the stage graph and dispatch depth; verification rigor is derived from intensity. - Sealed cross-harness dispatch: at standard and above, each stage runs as a separately launched session with a sealed role, model profile, and disjoint write scope. Parallel groups of 2-4 legs start in one transaction, and legs spread across harness families by default. Dispatch depth 3 is forbidden. - Per-node model tiers: deep, balanced-deep, light, and mini are sealed per node at compile time; adapters map them to concrete models without naming vendors in shared contracts. - Fleet: a live dashboard over the attempt registry showing interactive sessions and dispatched workers from all three runtimes in one tree, with state, harness, sealed profile, context gauge, and token accounting. `fleet --once` and `fleet --json` work without curses. - Guards: 39 hooks, 5 of them hard blocks, denying write scope, spec read, artifact root, git state, and memory path violations before the tool call. - Artifact system: research/analyze-project to spec to plans for code, and research to draft to refine for documents, under a project-wide .agent_reports/ root. - Memory: SQLite with FTS5 across sessions, projects, and runtimes; changed decisions are superseded rather than deleted. Installation is via a one-line release script that installs an integrity-checked archive and registers a daily user-level update check where supported. The installer does not touch runtime credentials, sessions, logs, or databases. When Codex is present, it also installs a reversible protected ingress so interactive codex commands enter the harness-managed App Server, while non-interactive commands pass through to the real CLI. `hearting update` stages and verifies a new release before switching the active pointer and rolls back on failure. A `compute-hosts` operator launcher reads a user-owned YAML config seeded once as a commented template. Runtime support covers Claude Code, Codex, and OpenCode in both linked (live-debug) and packaged (immutable bundle) projection modes. Unsupported surfaces are marked SKIP with a reason. Development workflows include a single generate/check entrypoint, generated-projection drift tests, and repository Git hooks that run the same checks as CI. Documentation covers usage, installation layout, release policy, capabilities, roles, routing, operations, memory, and hooks. Licensed under MIT.