Describing a job in one sentence is a quick way to start it. It also makes it easy to lose track of what was decided: which file the request became, whether it was saved, whether it ran or only looked as if it had.
In v0.120.0, the command that wrote a new workflow became nika compile. It filled a built-in skeleton from explicit answers, and the README of that release was clear about the rest: "Unsupported natural language stays incomplete without a substitute workflow." Bare nika, the command with no arguments, already opened a conversation in the project, and 0.119.0 restored it when you came back.
v0.121.0, published on 25 September 2026, carries the headline "A conversation from the request to an inspectable result." Its notes describe the path in one sentence: "Open nika, describe the work, review and correct the proposed .nika workflow, save it, then choose Run explicitly."
A request that needs no model #
The session guide, as it stood at the v0.121.0 tag, has a first walkthrough that needs no API key. Every command, request and output shown below is copied from that version of the guide; they are its excerpts, not one session recorded for this post. Make a folder with a small CSV file and open the session there:
mkdir -p orders-demo/data && cd orders-demo
cat > data/orders.csv <<'EOF'
order_id,customer,status,amount
1001,Juniper Books,paid,18.00
1002,Kestrel Cafe,pending,9.50
1003,Linden Studio,paid,13.50
1004,Moss Supply,refunded,6.00
EOF
nikaType the request:
Read ./data/orders.csv, keep only the rows whose status is paid, and write them to ./out/paid.csv.Nika proposes a new workflow file and shows what it does, when it runs, what it can touch and the result of its check. This is the guide's excerpt from the plain view:
Nika proposes `compiled-workflow.nika`:
Does
1. read_source · reads a file
…
8. write_output · writes a file
Runs
when you ask (« run it ») · no schedule was asked
Can touch
external effects · none
human approval at run · none
…
check of these bytes · `compiled-workflow.nika` · clean ✔
when it runs:
· reads ./data/orders.csv
· writes ./out/paid.csv
· tools nika:assert · nika:convert · nika:jq · nika:read · nika:write
· model output estimate · $0 · no direct model task in these checked bytesNothing has run yet. /show prints the exact bytes of the proposal, and /meaning lists what Nika kept of your request, clause by clause. A sentence that is not a yes, a no or a question is read as a change, for example write the paid rows to ./out/confirmed.csv instead. The revised proposal replaces the old one and needs its own yes.
Save is one decision #
Type yes. Nika writes exactly the reviewed bytes, checks the file on disk and stops:
applied · wrote `compiled-workflow.nika`
check · `compiled-workflow.nika` · clean ✔
Saved · checked · not active · nothing has run
say « run it » to run it once (a ceiling is announced first)The third line carries the idea of this post. The 0.121.0 changelog names the rule "Save ≠ Check ≠ Run ≠ Activate". Saving writes and checks. It does not run the file, and it does not start a schedule. When a request asks for one, activating it is a separate step that proposes a change to the project file, and the changelog is blunt about what that change means: "Declared is not active". A process on the machine still has to fire it.
Saving also clears the previous run status, so a fresh save never looks as if it had already run.
Run is another #
Type run it:
running `compiled-workflow.nika` once · ceiling $0.25
…
produced · ./out/paid.csv (92 B)
read · ./data/orders.csv
cost · no model usage recordedThe run happens once, only if the saved file checks clean, under a spending ceiling announced before it starts: the amount you state, otherwise the ceiling: in the project's nika.yaml, otherwise USD 0.25. out/paid.csv holds the header and the two paid rows:
order_id,customer,status,amount
1001,Juniper Books,paid,18.00
1003,Linden Studio,paid,13.50No model was called, neither to write the workflow nor to run it. The 0.121.0 changelog describes the compiler behind this: "A compute step is stated as a typed computation, never as jq the model writes." The filter in the saved file is an ordinary jq expression, which the guide shows:
expression: '[.records[] | select(.status == "paid")]'With the same Nika version, the same request and the same input file, the guide says "this path gives the same workflow and the same output bytes". The same compiler also works without the session:
nika compile 'Read ./data/orders.csv, keep only the rows whose status is paid, and write them to ./out/paid.csv.' paid-orders.nika
nika check paid-orders.nika
nika run paid-orders.nika
nika trace verifyFour proofs, kept apart #
The notes state the design in one line: "Check, admitted access, recorded execution and the actual output remain separate proofs."
- Check describes the file: its structure and what it can touch.
- Admitted access is the route to a model that a run is allowed to use. The notes do not define the phrase further, but they say what admits a route: "the informational price catalog does not by itself authorize an OpenAI-compatible API route". A route and model that are not admitted need "a fresh, bounded Run cost review", and "unattended invocations without that review refuse before dispatch."
- Recorded execution is the run's trace, a local file of events.
/proofshows what it records and what it does not prove, andnika trace verifychecks its hash chain. - The actual output is the file you open and read.
A good result in one of them does not prove the next. A clean check does not mean the run was allowed. A verified trace does not mean the output is what you wanted.
Coming back later #
/quit closes the session. A proposal still waiting for its yes is kept. The next nika in the same folder offers /restore, which shows the original request and proposes the kept draft again, checked against the project as it is now. It calls no AI and writes nothing until you type yes. Earlier approvals, and a ceiling agreed for a saved file, do not carry over.
What this release does not do #
- The qualification is narrow, and the notes say so: it "records finite synthetic journeys on its identified macOS development candidate; it does not establish arbitrary-request correctness or Linux conversational qualification." The guide adds that the qualified installation also used a knowledge snapshot and a decision model, which public releases do not include.
- A request the deterministic compiler cannot settle needs a model that you choose. Some requests cannot be expressed at all; Nika says what stopped it and writes nothing.
- A clean check describes structure and boundary. It does not prove that the workflow does what you meant, or that a model's text is true.
- The run ceiling bounds the run's metered spend as the runtime estimates it. It is not an invoice and it does not cover authoring.
- The unknown-price review covers bounded HTTPS OpenAI-compatible text calls only. HTTP API overrides and overridden native Anthropic and Gemini gateways refuse, even in interactive runs.
- On an interactive terminal, bare
nikaopens the full-screen session, andnika --plainopens the same session as plain lines. One commit subject of 21 September mentions a--tuiflag; the notes say it "is retired", so do not use it. - Rolling back to an older binary does not migrate session history backwards. Older executables refuse the newer records.
- The guide on
mainnow describes a later release, and its output shows a few more lines than the excerpts above.
Read the v0.121.0 release, the current session guide and the qualification note for what was and was not tested.