English | 中文

Agent Continuity Harness (ACH)

ACH keeps long-running AI work recoverable when one chat is no longer enough.

Use ACH when an agent task may drift, pause, move across windows, or need another agent to resume it without guessing from chat history. ACH starts light, then creates formal state only when continuity is actually at risk.

ACH is not a project manager and not a memory dump. It is a continuity harness: it anchors the goal, records constraints, externalizes recoverable state, and makes handoff possible.

flowchart TD
  T["Incoming task"] --> G["Guard mode"]
  G -->|"anchor goal, constraints, weak assumptions"| W["Continue normally"]
  W -->|"handoff / recovery / cross-window risk"| C["Continuity mode"]
  C --> S["Formal state root"]
  S --> A["current-goal: task axis and next step"]
  S --> B["confirmed-constraints: active constraints"]
  S --> P["pending-items: open items and impact"]
  S --> R["decisions: decisions and their basis"]
  R --> H["Resume from state, not chat memory"]

Loading

Use It When

• A task will take many turns, windows, branches, or recovery attempts.
• The goal, constraints, evidence, or decisions must survive a restart.
• Another agent must resume the work without replaying the whole conversation.
• Drift is likely and the cost of losing context is higher than the cost of keeping state.

Do not use ACH for short questions, simple edits, or tasks where the current chat is enough.

What It Does

1. Starts in guard-mode: keep the goal and constraints visible without creating heavy state.
2. Watches for continuity risk: handoff, recovery, branching, repeated failure, or cross-window work.
3. Enters continuity-mode only when needed.
4. Creates a state root that records the current route, decisions, attempts, evidence, and unresolved risks.
5. Lets a future agent resume from state instead of guessing from chat fragments.

Core Files

Every formal state root contains four recovery-core files plus a machine-readable manifest. ach init creates all of them:

File	Purpose
current-goal	Current task axis, phase, and next step
confirmed-constraints	Constraints that are confirmed and still active
pending-items	Open items, their impact, and whether they block progress
decisions	Decisions made, what they change, and their basis

Complex tasks can extend the root with optional supplemental documents — active-context, branch-attempt-ledger, artifact-provenance-index, state-relation-index — via ach add-supplemental. See the state contract.

Quick Start

Install the CLI and create state for a task:

npm install -g github:bagbag16/agent-continuity-harness
ach init my-long-task
ach status my-long-task

Or ask for ACH in conversation when continuity matters:

Use ACH for this task. Start light, but create formal state if the work needs handoff, recovery, or cross-window continuation.

Expected behavior:

• The agent keeps ordinary work lightweight.
• The agent escalates only when the task earns formal state.
• The state root becomes the source of recovery, not the raw conversation.

More: Install | Quickstart | CLI reference | FAQ

Enforcement Levels

A rule that relies on the agent remembering it is the weakest kind of rule. ACH labels every mechanism by how it is actually enforced, and is honest about which ones are still prose:

Mechanism	Level	Enforced by
State root shape, bindings, manifest integrity	gate	ach validate / preflight block handoff and resume on failure
State freshness (is reality recorded?)	derived + gate	ach reconcile computes drift from file mtimes — ground truth, not self-report; the bundled Claude Code stop gate refuses to end a session while an active task's state is stale
Artifact provenance consistency	derived	ach artifact check validates the index against active context
Mid-task checkpoint discipline	prose	SKILL.md convention — known weakness, partially compensated by the stop gate
Guard-mode → continuity-mode escalation judgment	prose	SKILL.md convention — deliberately left to judgment

Levels: gate = mechanically blocked; derived = computed from evidence the agent cannot rewrite; audit = measured after the fact; prose = relies on the agent following instructions.

Boundary

ACH manages continuity. It does not decide product strategy, replace task-specific validation, or make every task bureaucratic. If the work can finish cleanly in the current context, ACH should stay in guard mode.

ACH also does not judge whether the work is converging on its goal — that is semantic governance, not state. For autonomous loops, loop-builder designs that layer (acceptance criteria, independent supervision, stop conditions) on top of an ACH state root.

License

MIT.