Skip to content

RPG OS v0.9.2

Choose a tag to compare

View all tags
@github-actions github-actions released this 08 Sep 08:39
v0.9.2
a77cd1a

RPG OS v0.9.2 — optional announced autosave

This experimental release adds an opt-in, turn-driven autosave policy while preserving the existing campaign formats, state ownership, evidence boundaries and protected-write procedure. The policy invokes the ordinary complete-present CHECKPOINT; it does not add a background service, a second save format or an invisible private memory store.

Announce, checkpoint, verify

With the policy explicitly accepted, a meaningful unsaved change or substantial scene boundary can schedule a checkpoint. The standard backstops are 15 completed PLAY replies since verified persistence and a supported context reading of at least 65%. Context readings must come from actual model-visible host telemetry or an explicit operator report, never an estimate or an assumed ability to read the UI.

The GM announces the checkpoint one response before automatic execution. Several triggers coalesce into one pending notice. After the next completed play response, it enters ADMIN, preserves the complete present through the existing recovery-protected CHECKPOINT, verifies the writes, and reports the save identity and evidence limitation briefly. Nothing runs while the chat is idle. The pending PC choice and fictional time remain unchanged.

The player can request immediate CHECKPOINT or full SAVE, delay/resume autosave, or disable it. A manual full save is never downgraded. Recovery, active scene handover and an in-progress write prevent concurrent autosave. Failure preserves the recovery marker and does not claim success or reset the pending operation.

Save before switching

The outgoing context must save and verify before a planned model, reasoning-setting, provider, chat or GM change. The warning does not encourage switching to a stronger model before saving. This is a conservative continuity rule, not a promise about any host's internal context-transfer behavior.

A checkpoint protects current state but does not archive new dialogue. Use a full CLOSE or the existing scene-handover procedure before discarding source access; disclose the evidence gap if the operator elects checkpoint-only transfer. The incoming GM reloads authoritative saved records rather than reconstructing missing hidden facts or rolls.

Integration and verification scope

  • ADMIN/AUTOSAVE.md defines selection, triggers, notice, execution, overrides, failures and transitions.
  • LAW, BOOTSTRAP, NEW_GAME, RECALIBRATE and CLOSE_CONTRACT route the policy through the existing permission and persistence boundaries.
  • TOOLS/autosave.py is an optional deterministic, read-only scheduling helper. It evaluates supplied observations; it does not inspect the model, read campaign state, write files or execute saves.
  • TOOLS/test_autosave.py covers scheduler decisions, notice ordering, overrides, source-labelled telemetry, blocked operations, route integration and synthetic checkpoint/evidence metadata.
  • ADMIN/TEST_AUTOSAVE.md contains ten separately unrun live behavioral cases. Automated tests do not establish model compliance, complete hidden-state retention or cross-host reliability.
  • Both repository workflows run the new suite alongside the existing structural, handover, packaging, exact-source, search, evidence and agent-instruction suites. Release packaging still checks the exact committed tree and SHA-256.

See the release's GitHub Actions runs for the actual tested commit, platform skips and publication result. Do not interpret this release note as a claim that a future run has passed.

Compatibility

Autosave is off unless selected for that campaign. Installing v0.9.2 alone does not enable it or edit a separately stored campaign. Existing v0.7.3, v0.8.0, v0.8.1, v8.1.1, v0.9.0 and v0.9.1 record formats remain compatible through the protected upgrade procedure. Accepted agreements, private facts, individual/controller state, PC control, real-randomness requirements and the v0.9.1 prospective-deepening rule remain intact.

There is no new mandatory agreement clause, metadata field, commit kind, personality vector, autosave file or startup test suite. Ordinary PLAY stays read-only; accepted autosave enters ADMIN for the actual writes. Repeated checkpoints do not replace full CLOSE or independent backups.

To select the standard policy in a campaign, say: “Enable the standard announced autosave policy in ADMIN/AUTOSAVE.md for this campaign.”

Quick start · Autosave procedure · Behavioral cases · Installation and upgrades · Verification

Assets 4
Loading