Repository navigation
Open
Description
Tracking issue for the next phase of Stele. Each item is an issue; the PR that implements it says Closes #N.
Before starting #5 (done)
- Final review of PR Close remaining containment gaps, fix review findings, add macOS/Windows support #4 at its current head, which includes the review fixes and macOS/Windows support.
- Merge PR Harden Stele parser containment baseline #3 (containment baseline).
- Point PR Close remaining containment gaps, fix review findings, add macOS/Windows support #4 at
main, check the combined result, and merge it.
Sandbox model
Bubblewrap is Linux-only, so each workload runs on the backend that fits it:
| Workload | Linux | macOS / Windows |
|---|---|---|
| Lightweight, deterministic extractors | Wasm/WASI (#7) | Wasm/WASI (#7), the same way |
| Heavy ML parsers (MinerU, Marker, Docling) | bubblewrap + seccomp (#6), or OCI with gVisor (#8) | OCI container through the Docker Desktop / Podman VM (#8) |
- Parsers never get network access. Model weights are fetched in a separate trusted step, and parsers run offline (Sandbox backend interface and capability matching #5).
- No unsandboxed fallback. If no backend can satisfy a parser's requirements, Stele refuses to run it (Sandbox backend interface and capability matching #5).
Dependency tree
#5 Backend interface
├─ #6 Linux: bubblewrap + seccomp + Landlock
├─ #7 Wasm/WASI
├─ #8 OCI (runc / gVisor / GPU)
│ ├─ #9 MinerU
│ └─ #10 Marker + Docling
└─ #16 Snapshot archive + content-addressed store
└─ #12 Ledger redesign
└─ #13 Durable dispatch + invalidation
└─ #14 Replay (also needs #7)
#11 Telemetry + diagnostics ← depends on #6, #7, #8
Layers (phase 1: complete)
| Layer | Issues |
|---|---|
| Execution abstraction | [x] #5 |
| Isolation engines | [x] #6 · [x] #7 · [x] #8 |
| Heavy parser integration | [x] #9 · [x] #10 |
| Runtime evidence | [x] #11 |
| Immutable evidence foundation | [x] #16 |
| Provenance and state model | [x] #12 |
| Delivery and invalidation evidence | [x] #13 |
Replay (REPRODUCED / EQUIVALENT / DIVERGED / UNREPLAYABLE) |
[x] #14 · [x] #29 per-field tolerances · [x] #30 recorded run conditions |
Shipped in PR #27 without a separate issue:
- Append-only, hash-chained event log, with ledger verification and external anchors
- Canonical extraction contract (
stele.extraction/v1) and trusted resolver - ChatGPT export adapter: streamed, keeping every conversation branch, run through the Wasm splitter
Next
| Item | Issue |
|---|---|
| Identity contract: typed, self-verifying references (Source, Snapshot, Observation, Record, Anchor, Event) | [x] #42 (PR #45) |
| Full-corpus production verification and sign-off. The baseline is today's bubblewrap-only setup. | [ ] #43 |
Follow-ups
| Item | Issue |
|---|---|
| Parser images: GPU validation, Marker OCR, arm64 builds, hash-pinned torch | [ ] #37 |
OCI: honour writable_scratch=False; tell a parser's own exit 125–127 apart from an engine failure |
[ ] #38 |
| Ledger migration: re-verify records instead of leaving them invalidated for missing evidence | [ ] #39 |
| Extraction: page and bbox normalizers from the parsers' layout JSON | [ ] #40 |
Deferred
- Firecracker / Kata Containers, only if Stele becomes multi-tenant
- A native Windows sandbox (AppContainer)
- Networked parser execution: only ever as an operator override with its own evidence event, never as a parser manifest option
Metadata
Metadata
Assignees
Labels
No labels
Activity
Add the ChatGPT export adapter: every branch, one conversation at a time
Add the hash-chained event log and ledger verification