AGPL-3.0-or-later · forever.

the journal

Notes from the source.

Long-form pedagogy on Intent as Code: why useful AI work belongs in a file, why the language locks at four verbs, and what local-first actually buys you. The archive runs from the spec opening to today.

33 live · 2 upcoming

latest · Engine · The registry re-proves everythingA workflow registry where nothing is taken on faith: every entry is content-pinned to an exact source revision and re-proven by CI (hash, certificate, advisories). The npm-of-workflows, minus the trust-me.4 min read
32 · Agents · Written by agents, reviewed by younika init teaches the language to whatever agent you already run. Then two real agents write the same workflow: one invents a schema the audit refuses, one goes green first shot. The human review is a two-line diff.5 min read31 · Language · The pipeline is a fileAn llm pipeline is a graph of model calls, tools and processes. A graph is declared, not programmed: forty lines of YAML replace the orchestration framework. Diffable, auditable, replayable.3 min read30 · Engine · The MCP server you didn't have to buildEveryone is learning how to build an MCP server. For your workflows, skip the tutorial: the nika binary already is one. Read-only tools let any agent audit a plan and learn the language from the thing that enforces it.3 min read29 · Engine · The forecast is localexplain --forecast computes duration and cost priors from your own recorded runs: stats over .nika/traces/, honest ranges at two runs, percentiles at five. Never a model call, never the network.3 min read28 · Language · The generative ai workflow, minus the mysteryStrip the buzzword and a generative ai workflow is three things: inputs you name, a model step, and effects you can point to. Declare all three in one file and the mystery evaporates. Rehearsed offline, every asset with provenance.3 min read27 · Engine · The chain of custodyA recorded run is a text file, and text files can be edited. nika trace verify recomputes the hash chain: one changed word in history breaks every line after it, and the run's printed head closes the loop.3 min read26 · Language · The agent workflow spectrumAn ai agent workflow is any pipeline where a model chooses some of the steps. Every design sits between a declared graph and a free loop. Where yours lands decides which safety you can even ask for.4 min read25 · Engine · The run that waits for youThe approval gate is a task in the file, not a Slack thread: a plan fails closed on nika:prompt, and the human's answer rides the resume.3 min read24 · Engine · The resume storykill -9 a run mid-flight, then resume it: finished work never runs twice. Durability as a file property, shown from the real journal.3 min read23 · Engine · The one-task re-runRegenerate one block without re-running the world: --task scopes a fresh run to a task and its upstream, --from re-rolls what the hashes cannot see.3 min read22 · Language · Prompts are code nowPrompt versioning without a platform: the prompt lives in the workflow file, so git diffs it, a PR reviews it, git reverts it. The engine itself names an edited prompt when two runs diverge.2 min read21 · Security · The prompt injection that goes nowherePrompt injection examples usually end with the agent taking a new action. Here the plan is authored before the model runs: the hostile note becomes data, never an action, and the boundary is checked before a token is spent.3 min read20 · Engine · The run becomes evidenceEvery journal line now carries a hash chain: verify names the first broken link, reproduce classifies every task, and the journal attests which engine wrote it. Trust, but verify. Then verify.3 min read19 · Engine · Time travel, for realEvery debugger replays the past by re-running it. Nika's past is a file, so stepping backward is free, breakpoints live in your YAML, and F5 debugs a run that already happened.2 min read18 · Engine · The editor tells the truthA green badge is a claim. We audited the editor extension against the engine the way you'd audit a client against a server, and found the badge lying. Here is the contract that keeps it honest now.3 min read17 · Security · The credentials your pipeline was breakingOpenAI and Google sign the images their APIs return. Almost every pipeline that touches those files silently converts the signature into evidence of tampering. That included ours, until this week.4 min read16 · Sovereignty · One wire, five serversThe self-hosted media world quietly standardized on OpenAI's wire shapes. That accident of history is the best sovereignty news in years, and Nika's media builtins are built on it.3 min read15 · Engine · Media are workflow citizensImages and speech now render inside workflows under the same discipline as everything else: permit-gated saves, sha256 provenance, honest warnings, real cost on the ledger, and the sovereign path first.3 min read14 · Engine · The trace you can replayEvery run leaves a flight recorder: a deterministic AI audit trail you can replay like a film, never re-execute by accident.2 min read13 · Security · The secrets lineInformation-flow, audited before it flows: how the checker proves a secret cannot leak into a prompt, a file, or a host.3 min read12 · Engine · The cost lineThe audit prices every token before it is spent: ceilings on generations, budgets on loops, and a $0.00 that means it.2 min read11 · Language · Anatomy of a verbinfer, exec, invoke, agent: what makes the four verbs genuinely different execution models, in the engine's own verdicts.2 min read10 · Sovereignty · No cloud neededOne Rust binary, your models, your files. Run LLMs locally and see what local-first actually buys you.1 min read09 · Engine · The plan you get for freeThe wiring is all you write: with: for data, after: for order. The plan falls out of the file: parallel waves, drawn before anything runs.2 min read08 · Language · Four verbs are enoughA verb is a distinct execution model, not a feature. Why the language locks at four, forever.2 min read07 · Manifesto · Intent as Code: why your AI work should be a fileChats evaporate, files compound. The case for writing AI work down as source you own: your best prompts, versioned like code.2 min read06 · Security · The blast radius is part of the filepermits: is the whole list, guardrails an agent can't talk past. Everything not on it is denied before it runs, with a named error.1 min read05 · Language · A standard library, not a plugin store28 builtins in the binary, allow-listed, nothing to install. The library grows, the language holds still.1 min read04 · Sovereignty · An open spec, a copyleft engineTwo licenses, one argument: the file must outlive every vendor, including us.2 min read03 · Origins · Starting over, on purposeWe had a working prototype. We rebuilt from scratch anyway: craft, not extraction, one gate at a time.2 min read02 · Origins · Naming the drumWhy a workflow engine carries the name of a liberation god, and a butterfly on every commit.2 min read01 · Origins · The note that started itBefore the spec, before the name, there was one evening note: the best AI work of the month had just disappeared.2 min read

this blog is a folder · content/blog on GitHub ↗ · PRs welcome · rss

03 · in the pipeline2 upcoming

Securitysoon

The advisory that names itself

When a published workflow goes bad, the registry does not unpublish history — an advisory attaches to the pinned entry in the open, and the verify protocol already told you never to install what one names.

Enginesoon

The run that never runs

nika run --dry-run shows the whole static plan and executes zero effects: waves, permits, models, the cost floor. The full rehearsal before anything is allowed to happen.

Intent as Code · 4 verbs · 28 builtins · 17 providers · written down, replayable, yours