soseek 0.2.0 — compact evidence and saved-source windows

soseek 0.2.0 makes everyday CLI results smaller and easier for agents to consume while keeping the full saved evidence available.

• Compact results are now the default. Use --full for full source text, successful attempts, routing, raw payloads, and original citation offsets. Explicit --fields projects the full stored result; title/url/published_at aliases are supported.
• doctor now prints only access checks; --verbose includes capability schemas. providers gives a short operation catalog; operations PROVIDER retains schemas.
• JSON and non-TTY use suppress routine progress. --progress enables it; errors, partial warnings, usage uncertainty and recovery IDs remain visible.
• Standard Exa search requests query-relevant highlights in the same search call, without an automatic second read. Explicit contents and native-call input remain under caller control.
• runs source RUN_ID SOURCE_ID --offset 0 --max-chars 4000 --json reads existing evidence in Unicode windows without a paid provider call. CLI omission and incomplete stored content are distinguished.

Migration: automation that consumes attempts/routing/raw/citation spans must add --full. HTTP result and archive schemas remain full. --max-chars is a displayed text budget, not a hard JSON/token or upstream spending limit.

Offline replay of the same five saved samples estimated token reductions of 99.41% for doctor, 45.37% for search, 76.24% for research, 40.49% for a short page read, and 97.45% for native Grok search. Estimates use tiktoken o200k_base, not provider billing. Research savings include shortened page excerpts, not only redundant metadata. See docs/OUTPUT-EFFICIENCY.md for methodology and limits.

Validation: Go race tests and vet, frontend build and safety tests, evaluation runner tests, config checks, and CI Docker build passed. Production health reports 0.2.0. A real Exa query returned three snippets in one upstream search (reported USD 0.007); default JSON stderr was empty. Four saved-source windows reconstructed the first snippet exactly with unchanged usage. The public admin JS/CSS match the build; CLI and Skill were installed/synced through jstarctl.

Archives include clients for macOS/Linux/Windows, arm64 and amd64. Linux archives also contain soseekd with the management panel embedded. macOS arm64 and Linux amd64 were exercised; other targets were cross-compiled. SHA256SUMS-0.2.0 covers the six archives.

soseek 0.1.0

soseek 0.1.0 provides a self-hostable search and research service with a thin terminal client. The cloud server owns provider credentials, execution, budgets and evidence; clients use individual access tokens.

Included:

• Go soseekd server and soseek CLI, authenticated discovery, OpenAPI, durable runs, SSE, cancellation and idempotent recovery.
• Exa, Tavily, Firecrawl, Serper and Grok adapters; explicit provider/channel choice, multi-provider fusion and page reading.
• A bounded Grok research planner, local inspection of saved long documents, cited answers, and an explicit fixed workflow.
• Single-instance SQLite persistence, private user data, online backups, Docker/systemd examples and an Agent Skill.

Assets contain macOS/Linux/Windows CLI builds for ARM64 and AMD64. Linux archives also contain the server. The separate Linux AMD64 Docker archive can be loaded with gzip -dc soseek_0.1.0_linux_amd64_image.tar.gz | docker load. Check files against SHA256SUMS-0.1.0; BUILD-INFO-0.1.0.json identifies the source and server/image digests.

From a source checkout, sh scripts/install.sh --version 0.1.0 installs the CLI. Configure your own server and individual token as described in the usage guide and self-hosting guide.

Validation includes Go race tests and vet, six client targets, a 24-question live Exa/Tavily/merged baseline (72/72 tasks succeeded), real page reading, a completed Grok research sample, multi-provider research with explicit evidence gaps, public identity isolation/SSE/cancellation, and independent backup restore. See the acceptance record and baseline.

This is a single-instance release, without multi-node scheduling or high availability. Native Grok search on the tested third-party endpoint sometimes exceeded its requested internal tool limit; such runs retain evidence and actual usage, return partial, and stop later calls. Model citations still require semantic review. Credits and unknown monetary charges remain distinct from known USD costs. These observations do not establish universal search quality or upstream reliability.