Build

Sample mode

Run Konkyo's real agent logic from a fresh clone, with no keys and no services.

git clone <the repository> konkyo && cd konkyo
corepack enable && pnpm install
pnpm konkyo ask --sample        # the first recorded question
pnpm konkyo samples             # all of them
pnpm konkyo ask --sample 2      # by number, name, or the exact question
SAMPLE_MODE=true pnpm konkyo    # the same as --sample, with no .env at all

No .env, keys, database or Docker are needed: the hackathon's validation flow (rm .env, then SAMPLE_MODE=true) runs exactly this.

What it does

A recorded run kept, for each outside boundary, what came back: OpenAI's and Nemotron's HTTP replies, the rows of each Zetaris statement, the library search, and the H-MEM recall. Sample mode hands those back in place of the services, and every step of Konkyo's own logic runs again on them: planning, citation parsing, the code check of each quote, the Verifier's verdicts and rewrites, the matrix, your World Model (written to a scratch folder; a space's isn't replayed), and a new trace in logs/.

Replies are matched by what was asked (the request body, the SQL), never by order, so concurrent agents replay correctly, and never by account, so a run recorded on one account replays anywhere. A question that wasn't recorded is refused rather than invented. When a stalled call was asked twice (a hedged request), the replay takes the copy that answered; a request that got no reply is recorded as such and fails the same way on replay; a model the recorded run never asked is skipped, since which models were busy depended on that run's timing.

The examples

examples/ holds five recorded runs, each with its question, the answer as Markdown, the trace, and the replies:

  1. A cited answer across two papers.
  2. The agree/disagree matrix, with two Verifier rewrites (step by step).
  3. A false premise, corrected from the source.
  4. A textbook and a lecture transcript together, with a Verifier rewrite.
  5. Nothing in the library covers it, so Konkyo says so.

The sources are openly licensed (CC BY papers; OpenStax and MIT OpenCourseWare under CC BY-NC-SA).

On this page