Save and resume

Reference note: The release-pinned local workflow guide retains older optional-extra installation instructions and describes forms intake as future work. Those statements conflict with the 1.3.0 package metadata and released GUI. Use Get started and Dependencies and reference for 1.3.0 setup. Treat that older guide as an input-limit/design reference.

Before you begin

Use the selected environment from Get started. For a disposable example, prepare the trainer fixture before following the numbered actions. Set TASK to its actual canonical saved-task directory. In a POSIX shell, for the example path below, use TASK='/absolute/path/to/training-form/saved-task'; replace the parent path with the unused absolute directory you chose for the fixture. For real work, use your existing task's canonical directory instead.

What survives a new form session?

Prerequisite: Harness 1.3.0 with a registered editable draft missing its goal and success criteria. TASK = its canonical directory.

1. Click Continue form.
2. In “What should this work accomplish?”, enter “Review quartz retention documentation”.
3. Click Save answers. Leave the success question empty.
4. Stop the listener with Ctrl-C. Run: python -m attune_harness.gui --task "$TASK" --edit --launch-json
5. Open the new private launch_url. Click Continue form.

Expected: “Saved answers” retains the goal above the empty success question.

Limit: Reopening a form continues intake. Eligible execution resumes through the CLI.

Use the same task directory in both sessions, and a fresh private launcher link after restarting. Opening a new form or restarting expires the earlier form checkpoint. Saved answers remain in the task record. Unsaved text is not part of this result. This process ends with the draft still draft; it grants no execution authority. The later tutorial sections teach explicit saved memory, project scope, revision and current execution evidence. See Save memories and task intent for that separate CLI store.

Prepare a disposable form example

Prepare a separate disposable fixture for each process. The following Python uses the released work contract to create an intake draft with no goal or success criteria. Run it with the active Harness 1.3.0 interpreter. Replace training-form with an unused absolute directory. It makes no model calls.

from pathlib import Path
import json
from attune_harness.work_contract import SIGNALS, create_work

root = Path('/absolute/path/to/training-form')
root.mkdir(parents=True, exist_ok=False)
project = root / 'project'
project.mkdir()
(project / 'source.py').write_text('def value():\n    return 1\n')
(project / 'plan.md').write_text('Keep saved tasks linked to their current decision.\n')
config = root / 'participants.json'
config.write_text(json.dumps({'schema_version': 1, 'participants': {
    'local': {'adapter': 'deterministic', 'tools': [], 'max_turns': 1,
              'max_tool_calls': 0}}}))
budget = {'max_operations': 20, 'max_attempts': 1, 'max_output_bytes': 10000}
create_work(project, config, directory=root / 'saved-task',
    intent={'goal': None, 'context': ['Training example'],
            'scope': ['source.py'], 'constraints': ['No implementation or model calls'],
            'acceptance': [], 'questions': []},
    signals={**dict.fromkeys(SIGNALS, False), 'existing_artifact': None},
    assignments=[{'role': 'planner', 'participant': 'local',
                  'output_contract': 'Bounded plan with observable checks',
                  'budgets': budget}],
    inputs=['source.py'], artifact='plan.md', budget=budget)

Launch with --task /absolute/path/to/training-form/saved-task --edit --launch-json. This sample uses the Python API only to prepare training data. It does not imply a released browser “New task” control.

Verification scope

All four processes ran against installed Harness 1.3.0 on macOS in isolated synthetic tasks. The two browser sequences used fresh headless Chromium contexts. Navigation reached accepted; continuity retained its saved goal after a listener restart and remained draft. Planning returned created, draft, then accepted; it changed no sample source and created no build run. Retrieval returned two matching Markdown sources, and verification passed the single supported link claim with semantic checks absent.

These observations establish those local sample outcomes. They do not qualify Windows, real participant quality, paid-provider behavior or all possible projects. No released code changed for this training update. The broader workspace navigation, specification editor and browser execution/resume designs are outside the 1.3.0 forms boundary.

The released CLI guide, local workflow guide and saved memory guide retain the implementation baseline independently of current development docs.

Continue execution deliberately

Saving form answers records intake. Accepting intent records a scope decision. Neither action starts a build. Harness 1.3.0 browser forms cannot dispatch builds or resume execution.

1. In the same selected environment, inspect the saved task: attune-harness status "$TASK" --format markdown.
2. Read its state, evidence and next action. Resolve stale scope or uncertain effects before continuing; do not blindly repeat a submission.
3. If the task is eligible and you have the required current authority, continue through the CLI's documented resume route. Read attune-harness resume --help and inspect the participant registry first. External/native execution requires its applicable grants and can incur provider costs.

Expected result: Inspection explains the saved state; it does not execute work. An authorized resume may dispatch participants or effects. Use a new form listener to reopen intake instead of invoking execution resume for that purpose.

Explicit CLI saved memory is a separate store from form answers. See the released saved-memory guide.