v0.9.0 — terminal UI default, security harden, native Linux install

Summary

Changes since v0.8.0 (43 commits / PRs #464–#484).

start no longer opens a browser. The default workspace is seiso tui: a Forge-shaped terminal UI with live Hugging Face Hub search, local chat, and studio command launch. The React Forge app remains available as an opt-in API (SEISO_UI=forge start or seiso forge).

This release also clears several open dependency CVEs (PyJWT instead of python-jose/ecdsa, cryptography 50), hardens tool/KB injection defenses, and fixes native Linux install/start loops.

Project direction

Seiso stays a local-first stack: chat, fine-tune, post-train, compress, and export on your hardware. Day-to-day interaction is now the terminal UI so launch does not pay for a browser tab or the Forge SPA.

Limitations — not functional yet, do not use

Treat pay/mesh/Ark as docs / scaffolding only. Do not run them for production work, real funds, or trusted multi-node jobs:

Surface	Flag	Status
Sats marketplace (seiso pay)	SEISO_ALLOW_PAY=1	Experimental. Faucet/sim only until settlement clients land.
Buzz mesh (seiso mesh)	SEISO_ALLOW_MESH=1	Experimental. Not ready for real multi-node jobs.
Ark backends	SEISO_ARK_BACKEND=bark|second|ark	Not implemented — selecting these errors until a client is bundled.

What's new

Terminal UI (default start) (#484)

• seiso tui — Forge sidebar, Dashboard, Hub, Chat, studio pages, Settings
• Hub searches Hugging Face live (catalog + GGUF repos), not only files already on disk
• /search, /download N, /open N, /unload, /run configs/…
• start / installer autostart launch the TUI (no browser)
• Optional web API: SEISO_UI=forge start

Security (#469, #470, #472)

• Replace python-jose/ecdsa with PyJWT
• Bump cryptography to 50.0.0 (CVE-2026-69247)
• Harden tool / knowledge-base injection defenses without content censorship

Native Linux install / start (#466, #467, #469)

• Harden the cross-machine native Linux install path
• Stop start from false reinstall loops
• Prefer npm for Forge UI with a safer Bun fallback

Training / review (#473, #474)

• Address 2026-08-06 full-tree review findings
• DPO cosine LR scheduler after warmup; raise code-verification test timeout floor

Housekeeping (#464, #465, #468, #471, #475, #476, #480, #482, #483)

• Dependabot: ruff 0.16.3, @xyflow/react 12.11.3, React type bumps
• Safe Python lock refresh (huggingface-hub, peft, kernels)
• Revert a broad deps bump that did not stick (#475 / #476)
• README touch-up

Full changelog: v0.8.0...v0.9.0

v0.8.0 — SEISO wordmark, −33,152 LOC, auth harden, mesh E2E

Summary

Changes since v0.7.0 (24 commits / PRs #451–#463).

Measured LOC reduction: −33,152 source lines (−18.9%) from 175,228 → 142,076.

Method: count .py / .ts / .tsx / .css / .js under seiso/, forge/, forge-ui/src/, seiso_cli/, tests/, scripts/ at tags v0.7.0 and v0.8.0 (lockfiles / dist / node_modules excluded). Cross-check: git diff --shortstat v0.7.0...v0.8.0 on those paths → 278 files changed, 4632 insertions(+), 38735 deletions(-) (net −34,103).

Driver	Δ LOC (same method, stepwise)
Adaptive-RL-Quantization extract (#459)	−34,214
Test consolidation + dead-code trim (#460–#462)	−676
Auth / mesh / macOS / wordmark / other (#451–#458, #463)	+1,738
Net v0.7.0 → v0.8.0	−33,152

PR-local simplify nets (from those PRs' own diffs): #461 ≈ −843 LOC, #462 ≈ −360 LOC.

Project direction

Seiso stays a local-first Forge + CLI stack: chat, fine-tune, post-train, compress, and export on your hardware.

This cut ships a new SEISO wordmark as the product logo across Forge, hardens local auth crypto, extends experimental mesh worker plumbing, extracts adaptive RL quantization out of tree, and continues footprint cleanup.

Limitations — not functional yet, do not use

Treat pay/mesh/Ark as docs / scaffolding only. Do not run them for production work, real funds, or trusted multi-node jobs:

Surface	Flag	Status
Sats marketplace (seiso pay)	SEISO_ALLOW_PAY=1	Experimental. Faucet/sim only until settlement clients land.
Buzz mesh (seiso mesh)	SEISO_ALLOW_MESH=1	Experimental. E2E worker path lands here; not ready for real multi-node jobs.
Ark backends	SEISO_ARK_BACKEND=bark|second|ark	Not implemented — selecting these errors until a client is bundled.

What's new

Branding (#463)

• Detailed SEISO wordmark (mascot-matched glitch style) with Local first AI subline
• Wordmark used as the logo across Forge UI (sidebar, loading, auth, HF onboarding)
• Dashboard logo sized down; subline kept readable

Auth / crypto (#452, #453)

• Plain-language local account sign-in
• Tighten Forge keygen, at-rest encryption, and NIP-49 crypto
• Simplify crypto story to one AES stack (no password path)

Mesh (experimental) (#456, #457)

• E2E worker path: import → claim → materialize → launch
• Harden Buzz→Seiso isolation residuals

Training / extract (#458, #459)

• Close train/RL surface residuals after review
• Extract adaptive RL quantization to Adaptive-RL-Quantization (largest LOC drop this release)

Housekeeping (#451, #454, #455, #460–#462)

• Forge-ui Dependabot vite/jsdom/plugin-react + digests
• macOS: sync bun.lock; install profiles work on Bash 3.2
• Consolidate grab-bag regression tests into domain suites
• Trim leftover CSS, dead helpers, stale docs, and post-extract dead code

Full changelog: v0.7.0...v0.8.0

Agent attribution

Field	Value
Buzz-agent	cursor
Nostr-npub	npub1wu0wrsnepztey2gv4s28lh2php2ndtxp2e4cw6qx2hzcw87sdtuqz8tefn
Author / Signed-off-by	Legendarylibr 117605439+Legendarylibr@users.noreply.github.com
Commit signing	Release notes updated via gh release edit; commits on the tagged range carry repo trailers

v0.7.0 — Nostr-signed mesh trust boundary, L402 pay scaffolding, training harden

Summary

Changes since v0.6.0 (26 commits / PRs #442–#447).

Project direction

Seiso stays a local-first Forge + CLI stack: chat, fine-tune, post-train, compress, and export on your hardware. The product spine is still self-hosted training and inference.

The longer arc is optional coordination and payment rails that stay opt-in and fail-closed:

1. Human UI vs agent surfaces — Forge Train stays the human path; Buzz-agent / mesh orchestration is a separate, gated surface so OS hostnames and unsigned chatter cannot become trust.
2. Nostr as the trust fabric for agents — mesh plans and Buzz-facing agent status move toward signed events only (NIP-01 / BIP-340), not ambient relay text.
3. Pay remains experimental — L402 session fund/exchange and refunds sit next to Ark scaffolding; both are not wired for real funds. Local faucet/sim only until settlement clients land.
4. Correctness over feature sprawl — this cut hardens finetune/RL, distill templates, slime/compat paths, and dependency-graph hygiene rather than expanding the default install surface.

Self-hosted Forge / local training remain the supported path. Never post pay/mesh tokens to public channels.

Limitations — not functional yet, do not use

Treat these as docs / scaffolding only. Do not run them for production work, real funds, or trusted multi-node jobs:

Surface	Flag	Status
Sats marketplace (seiso pay)	SEISO_ALLOW_PAY=1	Experimental. L402 fund/exchange + failure refunds land in this release, but live Ark / Bark–Second settlement is still not wired. Faucet (SEISO_PAY_FAUCET=1) or simulated ledger only.
Buzz mesh (seiso mesh)	SEISO_ALLOW_MESH=1	Experimental. Signed-event relay policy and agent/frontend separation land here; not ready for real multi-node jobs.
Ark backends	SEISO_ARK_BACKEND=bark|second|ark	Not implemented — selecting these errors until a client is bundled. Leave unset or use faucet/sim.

What's new

Mesh / Buzz agent trust (#446)

• Separate Forge frontend training from Buzz-agent training surfaces
• Harden Buzz→Seiso trust boundary; mesh relay policy becomes signed-event-only
• Sign mesh plans with Nostr NIP-01 / BIP-340; generalize signed-relay for Buzz-facing agent status
• Stop mesh announce from signing OS hostnames into Buzz; close fail-closed review gaps

Pay scaffolding (#442, #443)

• L402 session fund/exchange and job-failure refunds
• Harden pay ledger races + slime/compat path safety
• Gate quote behind SEISO_ALLOW_PAY (experimental)
• Docs: advertise L402 alongside Ark (still do-not-use)

Training / RL correctness (#444, #445)

• Finetune/RL P1s: best-metric, trust_remote_code, weight paths, HTTP prompts
• Distill eval chat-template typing; NeMo path / training gap fixes

Deps hygiene (#447)

• Ignore local uv.lock; Dependabot note so deleted-lock ghost alerts (nltk/pyasn1) cannot return
• Canonical hashed lock remains locks/python.lock (pyasn1==0.6.4; nltk not a dependency)

Full changelog: v0.6.0...v0.7.0

v0.6.0 — sats marketplace + Buzz mesh scaffolding, owner-npub auth, Lean provenance

Summary

Changes since v0.5.0 (19 commits).

Limitations — not functional yet, do not use

Treat these as docs / scaffolding only. Do not run them for production work, real funds, or trusted multi-node jobs:

Surface	Flag	Status
Sats marketplace (seiso pay)	SEISO_ALLOW_PAY=1	Scaffolding only. Live Ark pay-in / Bark–Second settlement is not wired. Faucet (SEISO_PAY_FAUCET=1) or simulated ledger only for local smoke tests.
Buzz mesh (seiso mesh)	SEISO_ALLOW_MESH=1	Experimental peer coordination scaffolding. Not ready for real multi-node jobs.
Ark backends	SEISO_ARK_BACKEND=bark|second|ark	Not implemented — selecting these errors until a client is bundled. Leave unset or use faucet/sim.

Self-hosted Forge / local training remain the supported path. Never post pay/mesh tokens to public channels.

Pay & mesh scaffolding

• Add opt-in sats marketplace modules + CLI (seiso pay) and experimental Buzz mesh coordinator + CLI (seiso mesh) (#438)
• Document Path E / marketplace / mesh with explicit not functional — do not use banners across README, docs index, CLI, and related pages (#440)

Security & auth

• Harden Forge auth around the owner npub and close SSRF/session gaps (#437)

Provenance

• Add Lean 4 formalization of Nostr dataset membership proofs + CI package-dir fix (#436)

Deps / UI

• Bump forge-ui @noble / @scure / jsdom majors with import and CI fixes (#435)

Full changelog: v0.5.0...v0.6.0

v0.5.0 — lean Forge footprint, Nostr auth/provenance, correctness hardening

Summary

Changes since v0.4.0 (96 commits).

Memory & Forge footprint

• Shrink bare Forge RSS so larger models fit: defer native runtime imports, honest Free-memory reclaim for sidecars/managed vLLM, Apple-only lean cache defaults (#430)
• Live-update Model Hub free-memory strip without changing OS headroom math (#429)
• Fix macOS unified memory reporting and startup Bun lock / doctor Hub checks (#399, #428)
• Ship Nostr relay deps with [forge] on every OS install path (#428)

Nostr auth & provenance

• Make Nostr keygen the default Forge auth path; document onboarding + default provenance (#408, #409)
• Add opt-in Nostr provenance attestation (#406)
• NIP-49 encrypted ncryptsec key backup (replaces plaintext nsec download) + hardening/tests (#414–#416)
• Harden Nostr client for NIP-01 addressable semantics (#420)
• Merkle corpus membership proofs for training provenance (#411, #412)

Security & correctness

• Path sandbox hardening (#398)
• Lock training HTTP cancel/SSE/IDOR and gpu_task restore contracts (#417)
• Multi-pass correctness audits: Compat tools honesty, pool leaks, distributed WORLD_SIZE/MASTER_*, slime, SSRF, model-pool, KB/export/job-stream (#394–#396, #418–#419, #421, #423–#426)
• Remove post-train censorship prompts (#397)
• Locked-deps CI fixes (#400)

Training & platform

• Remove dead training paths; align single vs multi-GPU configs (#422)
• Harden distributed env resolution against leftover MASTER_* and rank gates (#424)
• Unify RL-quant preset registry (#393)
• Add Buzz agent skill to orchestrate Seiso jobs (#410)

Deps / UI

• Migrate Forge UI to react-router 8.3.0 (GHSA-qwww-vcr4-c8h2) (#405)
• Bump kernels, mypy, TypeScript, testing-library (#401–#404)

Full changelog: v0.4.0...v0.5.0

SeisoLocalAI v0.4.0

SeisoLocalAI v0.4.0

Since v0.3.0 (2026-07-21). 7 PRs focused on NeMo RL integration, slime multi-GPU / GRPO correctness, training-config floors, and Forge red-team hardening.

Highlights

NVIDIA NeMo RL

• New external training method method: nemo_rl (no vendoring): launches NVIDIA-NeMo/RL via uv run with recipes grpo, dpo, distillation, smoke (#387).
• CLI (seiso nemo-rl), Forge Train Studio picker, example/smoke configs, and SEISO_NEMO_RL_ROOT setup docs (#387).

Slime / GRPO / multi-GPU correctness

• DDP no_sync during grad accumulation with flush/average before mid-window optimizer steps; round-robin SGLang/vLLM generate across engine URLs (#390).
• Drop vacuous GRPO groups (<2 non-truncated rollouts); refuse biased vllm_weight_mode=full + LoRA; require vllm_base_url for cloud GPU + slime + vLLM (#390).
• Example microbatches kept as multiples of rollouts_per_prompt so GRPO groups stay intact (#388).
• HTTP rollouts honor finish_reason and server token ids; managed vLLM liveness/model probing hardened (#391).

Training-config invariants

• Validate-time product slime floors (data_gen_count, held-out auto-split), refuse empty/off data_gen, require local JSONL (or SEISO_ALLOW_TEMPLATE_SLIME), ban CI fixtures on NeMo product configs (#389).
• Held-out eval uses greedy HF generate; NeMo GRPO num_generations_per_prompt always overridden (default 4) (#389).

Forge security (red-team)

• Non-overridable security system prompt with tools; quarantine instruction-like KB chunks; Compat inference API keys are chat-only (tools need session JWT) (#392).
• Refuse remote + code-exec entirely (no ACK override); redacted chat/tool audit hashes (#392).

Dependencies & audit sweep

• pyasn1 ≥ 0.6.4 for REAL/OID DoS advisories; lock refresh (#386).
• Broader Critical/High audit fixes across compress/export/distill-RL/DPO/kernels/MoE DDP (#391).

Upgrade notes

• NeMo RL requires a recursive NeMo RL checkout and SEISO_NEMO_RL_ROOT (plus uv); not bundled in this repo.
• Product slime configs without SEISO_ALLOW_TINY_RL enforce real-data floors; template/code examples may need operator JSONL or SEISO_ALLOW_TEMPLATE_SLIME.
• Remote Forge + SEISO_ALLOW_CODE_EXEC together refuse startup; Compat /v1 tools require a session JWT (inference API keys stay chat-only).
• Multi-engine slime HTTP rollouts round-robin generate; weight sync still fans out to all engines.

Full changelog: v0.3.0...v0.4.0

SeisoLocalAI v0.3.0

SeisoLocalAI v0.3.0

Since v0.2.0 (2026-07-15). ~130 commits focused on RL/SFT correctness, real verifiable data paths, Forge safety, and kernel/chat quality.

Highlights

RL / GRPO / DPO correctness

• Outcome-first rewards: refuse format ties, field/process reward hacks, zero-spread groups, and stale vLLM/SGLang weight sync (#385).
• DeepSeek / OpenRLHF / TRL-aligned math: dual-clip, seq-mean loss, truncated/empty excluded from advantages, DPO log 2 identity and empty-completion refusal (#385).
• Code GRPO binary/dense/auto outcomes; contains-answer hardening; held-out slime eval (#367–#370).
• Real-data gates for slime / Distill-RL (Forge UI + product floors) (#371).

Safety & platform

• PENDING job cancel; fail-closed remote code-exec; DNS pin without global getaddrinfo (#374, #377, #384).
• RL-quant deploy claims require llama.cpp + external quality sidecar (#375).
• Refuse FULL/BASE export of LoRA-only checkpoints; restore inference kernels on load failure (#379, #380).
• Torch 2.13 (CVE-2025-3000); private GitHub security advisory policy (#355–#356).

Training & chat

• Unified SFT tokenization / structure-preserving prep; packing + response-mask guards (#364–#366).
• Thinking budgets / strip leaked think blocks; empty Qwen3/Ollama reply fixes (#359–#360).
• Kernel stack: fused GEMM/LoRA routing, attention/packing defaults (#358–#359).

Cleanup

• Remove RL-quant markdown research-report stage; unused Forge DB tables; dead helpers (#372–#373, #382).
• July 2026 codebase review report (#376).

Upgrade notes

• Slime HTTP rollouts require *_sync_weights: true (debug-only env escape).
• Default auto_stop_metric is outcome_reward_mean; loss_aggregation defaults to seq_mean; clip_ratio_c defaults to 3.0.
• Distill-RL requires rollout_temperature > 0 and grpo_group_size >= 2.
• Product slime / Distill-RL paths expect real verifiable data (CI uses SEISO_ALLOW_TINY_RL=1).

Full changelog: v0.2.0...v0.3.0

SeisoLocalAI v0.2.0

SeisoLocalAI v0.2.0

Second release since v0.1.0 (2026-06-19). ~700 commits and 90+ merged PRs focused on making native Linux NVIDIA chat stable, long replies finish reliably, training/SLIME work, and the product story clearer.

Highlights

Long-form chat (multi-pass, OOM-safe)

• Dynamic multi-pass auto-continue so long answers (songs, papers, essays) finish without one huge completion that can OOM (#340 and earlier long-reply work: #323–#326).
• Per-pass completion stays OOM-clamped (often ~512–768 tokens on native Linux NVIDIA); total reply budget scales with request, long-form intent, and free VRAM.
• Linear-decay context packing for continue passes: free headroom grows as the draft grows; recent assistant tail is preferred over a full-context pack.
• Continue on mid-sentence EOS (not only finish_reason=length) — fixes live unfinished cuts like after “again” / “longer”.
• Empty continue pass: one strong-cue retry.
• Smoother Forge UI streaming for multi-pass: adaptive reveal, continue-gap cue, live caret (#340).

Native Linux NVIDIA inference

• Extensive OOM and hang guards for llama/Ollama sidecar paths (context clamp, batch caps, VRAM budgeting, fail-closed GGUF isolation).
• Ollama-first / sidecar routing for GGUF chat; improved VRAM planning and tok/s without relaxing safety guards.
• Native Linux KV pipeline and TTFT improvements; pin chat context after preload so multi-turn reuses KV size.
• Packing and load hot-path performance for large GGUF / train startup.

System monitoring (native Linux)

• Correct CPU brand from /proc/cpuinfo (no more arch-only x86_64 label).
• AMD k10temp / broader thermal sensors for live CPU temperature (#339).

Training / SLIME

• Distributed SLIME training wired through Forge UI (#335).
• SLIME sampling aligned with upstream design; optional per-token loss (#336, #337).
• Train tab no longer runs dataset prep on open (#327).
• Train/inference hot-path optimizations.

Security & reliability

• RAG trust-boundary and audit correlation hardening (#332).
• Inference generation ownership, preload locks, job runtime / kernel boundaries.
• Local sign-in reset for forgotten passwords (#291).
• CI: setuptools ≥83 for pip-audit, assorted lint cleanups.

Docs & product framing

• Clearer project description: local-first platform for chat, fine-tune, quantize, compress, and export (README + package metadata, #341 / #334).
• Install locks, platform docs, and citations updates.

Upgrade notes

• Prefer a fresh git pull (or reinstall via start) on native Linux + NVIDIA.
• Long replies use multi-pass continues; you may see “Reply incomplete — continuing (N/M)…” during generation — that is expected, not a failure.
• Optional env (defaults are usually fine):
  • SEISO_CHAT_AUTO_CONTINUE_MAX — max extra passes (-1 = auto)
  • SEISO_CHAT_AUTO_CONTINUE_TOTAL_TOKENS — cumulative output ceiling (default 32768)

Full change range

• Previous: v.0.1.0
• This release: v0.2.0 @ main
• Compare: v.0.1.0...v0.2.0

## SeisoLocalAI v0.1.0

SeisoLocalAI v0.1.0

Initial public release of SeisoLocalAI, a local AI forge for installing, launching, and managing local model workflows.

Highlights

• One-command Linux/macOS installer.
• Forge web UI with local onboarding.
• Local model chat and model hub workflows.
• Training, compression, export, and quantization tools.
• Installer doctor for guided troubleshooting.
• Stable install completion screen with the Forge URL.

Notes

This is the first versioned release. Expect rapid iteration as the installer, Forge UI, and local model workflows continue to improve.