DeepSeek-Reasonix

You will like this if your model of choice is DeepSeek and you start tasks and walk away. The daily trait is the checkpoint: every turn is saved outside git, so a run that went wrong at hour three rewinds to hour two without touching your commit history. Plan mode holds writes until you approve, and the same engine shows up as a TUI, a desktop app, a local web UI and a VS Code panel.

Where it hurts is confidence: no benchmark, and a community team rather than a vendor. Pick it for long unattended DeepSeek runs. Pick Aider if you want numbers behind the claims.

The risk is coupling. The README's first feature is cache-aware context maintenance, and the two-model mode keeps executor and planner in separate sessions so each stays cache-stable. That is engineering for one provider's billing behaviour. If DeepSeek changes how the prefix cache is priced or invalidated, the tool's reason to exist changes with it, and every other provider gets an agent tuned for someone else.

The consequence: keep a second agent installed and do not build habits around the cache discount. What it does right: task contracts and a pause policy are written down, with a tool contract published for regression review, more than most harnesses manage.

The design is stated in terms of the model's cache. 1. Startup injects a small, stable environment summary so the prefix does not vary between turns. 2. Stale tool output is snipped before any summary compaction, so pruning precedes summarisation. 3. Executor and planner can run as two models in separate sessions, each with its own stable prefix. 4. Session memory retrieval is documented as a Context Engine v2.

Verification is delegated to the user through per-call permissions; the agent does not check its own work. No benchmark is published. The observation: this is the only tool on the board whose architecture document reads like a cost model.

There is no cap table to analyse, which is its own answer. The project is a community repository under esengine, the support section lists PayPal and WeChat Pay, and the README says donations buy no feature priority. The name borrows a model vendor's brand, and the landing page tells you to bring your own DeepSeek key; so its distribution is DeepSeek's, which it does not control.

Moat: none; the cache trick is reproducible by anyone. Pricing power: zero by design. Likely outcome: absorbed into a vendor's official CLI, or maintained as long as its top contributors stay interested. Position: pass as an investment, fine as a dependency you can fork.

The demo is a six-hour run that can be rewound. For sixty seats the software is free and the meter is DeepSeek tokens; the landing page's own example shows an eighteen-minute session at four cents, which is the number finance will remember and the one engineering will not reproduce. There is no SSO, SCIM or audit log. CI fit is good: one static binary and a run subcommand.

The editor extension is published under an individual's marketplace identity rather than an organisation, support is a Discord with a help channel in two languages, and there is no vendor to sign anything. Not yet.

MIT and Go, which is not my first language but is one I can read in an evening. Providers, agent, tools and plugins are all declared in reasonix.toml, and any OpenAI-compatible endpoint is a config entry rather than a code change. reasonix mcp add takes stdio, SSE and streamable HTTP servers into one registry, skills are markdown under ~/.reasonix/skills, and Extension Protocol v1 sidecars can intercept runtime events and add providers.

make cross gives me six static binaries with CGO off, so it runs on a box with nothing else installed. The README never says the word Ollama, which I noticed. Otherwise this is mine to bend.