agentic-stack v0.19.0

v0.19.0 adds portable agentic loops and a bounded local meta-harness.

Highlights

• JSON contracts for L1 report loops and L2/L3 action loops.
• Maker → deterministic verifier → independent checker feedback lifecycle.
• Atomic resumable checkpoints, privacy-safe local events, policy gates, stagnation detection, finite budgets, and owned Git worktrees.
• loop init, validate, run, resume, status, stop, cleanup, and audit commands through both agentic-stack and ./install.sh.
• Read-only loop summaries in doctor, status, dashboard, Mission Control, and the local data-layer export.

Usage

:::bash
agentic-stack loop init /path/to/project
agentic-stack loop validate /path/to/project
agentic-stack loop run ci-sweeper "make the failing test green" /path/to/project --yes
agentic-stack loop status /path/to/project

Use agentic-stack upgrade --dry-run followed by agentic-stack upgrade --yes to add loop starters to an existing installed project without overwriting authored contracts or runtime state.

Safety boundaries

L2/L3 loops run mutating makers in owned worktrees, enforce deny-path and file count gates, require deterministic verification and independent checker approval, and stop on pause, budget, timeout, or repeated-failure decisions. Events are local-only and whitelist metadata; prompts, tasks, command arguments, stdout, stderr, and raw run IDs are not exported. The supervisor is not an operating-system sandbox; use harness-native sandboxes for stronger isolation. Schedulers should execute one bounded command and inspect its exit status before starting another.

Verification and attribution

Fresh verification evidence before tagging:

• python3 -m pytest: 166 passed.
• python3 -m harness_manager.cli loop validate --target . --json: valid ci-sweeper L2 contract.
• python3 -m harness_manager.cli loop audit --target . --json: valid audit, with zero real run evidence in the source checkout as expected.
• ./install.sh loop validate --target .: valid text dispatch.
• Fake-harness smoke: two-attempt completion, cleanup exit code 0, and the active checkout remained unchanged.
• The before/after git status --porcelain snapshots around the full suite were identical; the Claude hook suite ran against a session temp copy.

The loop design was informed by public loop-engineering discussions about bounded feedback, verifier gates, resumability, and stagnation breakers. This project implements its own Python standard-library contracts and does not copy code.