Repository navigation
openraid v1.1.3
Use model capacity and swarm parallelism fully, with durable completion votes and simpler file creation.
Improvements
- Model-sized context and output budgets. Omitted limits adopt the selected
model's catalog capacity across providers, including reasoning output. Explicit
overrides and saved-session intent are preserved; model changes recalculate
implicit defaults. #20
#21 - Earlier, richer compaction. Compaction starts around 75% of available input
capacity. Continuation summaries scale with context instead of stopping at
2,048 tokens, while respecting supported output limits.
#23 - Agent-scaled concurrency. Default provider admission is
max(32, agents);
command and PTY capacity ismax(4, agents). Explicit caps stay in effect.
#22
#24 - Exclusive native file creation.
write_filetakespathandcontent,
creates missing parents, and atomically refuses existing targets. Use
apply_patchfor subsequent edits.
#25 - Persistent completion evidence. Done workers park without provider calls
or peer-history compaction. Peer chatter preserves votes and consensus grace;
new owner steering and roster controls invalidate prior completion evidence.
#26
Using the updated defaults
Unknown model limits fall back to 32,000 context and 16,384 output tokens. Output
that would consume the whole context is reduced to one quarter of context to keep
usable input headroom. Saved implicit budgets and concurrency follow the selected
model and launch size; explicit overrides remain tracked separately.
Downloads
| Platform | Archive |
|---|---|
| Linux x86-64 | openraid-v1.1.3-linux-x86_64.tar.gz |
| Windows x86-64 | openraid-v1.1.3-windows-x86_64.zip |
| macOS Intel | openraid-v1.1.3-darwin-x86_64.tar.gz |
| macOS Apple Silicon | openraid-v1.1.3-darwin-arm64.tar.gz |
Extract the matching archive and verify it against SHA256SUMS below.
Verification
Regression coverage exercises configured and bundled model capacities, explicit
overrides, remembered launches, shared thinking headroom, long-context summaries,
concurrency scaling, competing file creators, and voted-worker behavior under
continuous peer traffic and new owner instructions. Four-platform release CI
gates publication on Rust and Node tests, formatting, warnings-denied Clippy,
optimized builds, checksums, and a 500-worker offline smoke.
Local acceptance passed 354 Rust tests and 25 Node tests, with no failures
or ignored/skipped tests. Formatting, warnings-denied Clippy, and whitespace
checks passed.