Iris v0.1.3

Iris Media for DSH 0.1.3

Version 0.1.3 makes media tasks safer to recover and much harder to accidentally submit twice. It is a reliability release: no new media provider is added.

Highlights

• Introduces Task/Attempt v2 across image, video, transcription, and TTS generation. One user intent owns one Task, while provider/model candidates become auditable Attempts.
• Records every Attempt before a potentially billable request. Automatic failover is allowed only after an explicit not_accepted result; timeouts, 5xx responses, disconnects, missing responses, and local persistence failures stop conservatively.
• Separates remote generation outcome from local artifact delivery. A download or filesystem failure no longer rewrites a successful generation as failed.
• Adds stable workbench states for paused observation, unknown acceptance/outcome, and unavailable artifacts, with a dedicated “Needs attention” section on desktop and mobile.
• Adds re-observe and re-deliver controls that never submit a new generation request. Informed retry requires a separate duplicate-charge confirmation, creates a linked new Task, and archives the original reminder so it is not counted twice.
• Adds acknowledge/restore controls for attention items. Acknowledgement only changes local reminder metadata; it never rewrites unknown provider facts. Legacy read-only labels now require an explicit Task v1 marker, avoiding false labels during hot-update version skew.
• Adds process-scoped state epochs/revisions and SSE backpressure coalescing so delayed snapshots and slow clients cannot roll the UI backwards.
• Adds an offline, zero-network Doctor API and CLI: dsh-iris doctor [--json], also available through npx @mokuyoaxis/dsh-iris doctor. Exit codes are 0 for healthy, 1 for warnings, and 2 for hard errors.
• Publishes architecture, security-boundary, provider-submission, task-semantics, and fault-injection documentation.

Safety notes

• Re-observe only polls an already accepted remote task.
• Re-deliver only queries and downloads an already successful result.
• Informed retry may duplicate a still-running remote job and may incur another charge. Iris requires explicit confirmation and currently exposes it only when the original text input can be reconstructed safely.
• The offline Doctor does not start DSH, contact providers, print API keys, or claim that browser/tool registration is healthy.

Compatibility

• DeepSeek Harness: >=0.1.2-rc.1 <0.1.3-0
• Node.js: >=20.10.0
• Package: @mokuyoaxis/dsh-iris

This release was developed and locally validated on Linux ARM64 under Android 16 Termux/PRoot Debian with DSH 0.1.2-rc.1. Cross-platform CI results should be confirmed from the release commit before publishing.