Parlay

Parlay is a small, persistent terminal chat for local AI models, written in Perl and built for Unix-like systems.

Machine-readable local discovery is available as XML:

perl bin/parlay capabilities

The command reports version, archive schema, host, path, runner, model, synthesizer, and download-consent capabilities without invoking runners, installing software, or downloading models.

For a compact human-readable version check:

parlay --version

Purpose

Parlay exists so an operator can keep a long-running local conversation with one local model and use that conversation to improve the working context over time.

The unchanged conversation history is the source of truth. A gist is only a derived, versioned artifact.

Parlay is an expert discussion space, not a programming worker. Participants may use code snippets as precise technical examples, in the same spirit as formulas or logical notation. Those snippets must be valid for the named language and must not be treated as verified merely because they look plausible. Execution, test running, and task orchestration remain Hydra responsibilities.

The intended long-term shape is a structured conversation between generalist perspectives and real specialists. Generalists can explore concepts and propose translations between domains; specialists can say what is possible, impossible, underspecified, or surprisingly feasible inside their own domain. The synthesis should make cross-domain possibilities visible that a single participant might not have found alone. That is Parlay's practical meaning of artificial creativity: disciplined, auditable emergence from bounded perspectives.

MVP 0.1

Parlay 0.1 is intentionally small:

• terminal chat with exactly one configured local agent
• append-only conversation event storage
• resumable conversations
• SQLite as the local canonical store
• FTS5 for fast full-text lookup
• versioned gists derived from the conversation archive
• operator-visible correction loop for bad gists
• offline operation on a local machine

Parlay 0.2: Multi-perspective responses

The first 0.2 increment can ask multiple explicit local agents for independent perspectives, archive every response, and use a local agent to synthesize one operator-facing answer:

printf '%s\n' 'What should I consider?' | perl bin/parlay respond-multi \
  --db var/test/parlay.sqlite --conversation work \
  --agent ollama:qwen3:8b --agent apfel \
  --synthesizer ollama:qwen3:8b

perl bin/parlay chat --db var/test/parlay.sqlite --conversation work \
  --agent ollama:qwen3:8b --agent apfel \
  --synthesizer ollama:qwen3:8b --rounds 2

respond-multi also accepts --prompt or --prompt-file. All required local binaries are checked before a turn starts. A failed perspective or synthesis rolls back the complete turn rather than leaving a partial archive.

Role profiles can be configured in PARLAY_CONFIG by mapping a supported role name to a local agent, then listing that role as an agent:

generalist=apfel
expert(perl)=ollama:qwen3:8b
skeptic=apfel
synthesizer=ollama:qwen3:8b
agent=generalist
agent=expert(perl)
agent=skeptic

Supported role names are generalist, expert(topic), skeptic, renderer, and moderator; expert topics use lowercase letters, digits, and hyphens, starting with a letter. The role shapes the participant prompt and archive actor name, while the concrete runner and model identity remain stored separately. A role profile does not grant tool use, shell access, or execution authority.

The default is the original one-round response. Add --rounds 2 for a bounded expert room: each agent first answers independently, then reads only the other agents' compact gists and contributes one short follow-up. The REPL supports /rounds 1 and /rounds 2; PARLAY_ROUNDS=2 selects the initial mode. Only the final synthesis is printed prominently, while show and search expose the complete append-only background timeline.

Moderation is enabled explicitly with --moderator deterministic. The configured --rounds and --max-full-answer-grants are treated as hard limits and compact synthesis instructions are recorded. An optional local policy helper can instead be selected with --moderator apfel or --moderator ollama:<small-model> (or PARLAY_MODERATOR in the default REPL). Its plain-text decision is clamped to those same limits, archived as a parlay-moderator system event, and never interpreted as commands or tool use.

Initialization

Initialize a SQLite database from the schema:

perl bin/parlay init --db var/test/parlay.sqlite

Quick Install

On a machine that can reach the Forgejo repository:

git clone https://git.rpdo.net/rpdo-net/parlay.git
cd parlay
sh install.sh

The installer is intentionally small. It checks for git, perl, and the required SQLite Perl modules, clones or fast-forwards Parlay into ~/.local/share/parlay, and writes a parlay wrapper to ~/.local/bin.

To update an installed Parlay later:

parlay --update

The updater refuses dirty local changes in the installed checkout and only fast-forwards the configured ref.

On first use, Parlay never downloads an Ollama model silently. If the configured model is missing in an interactive REPL, Parlay asks before running ollama pull; choose explain this to me for the short reason. In non-interactive commands, a missing model fails with a reason and no download is started.

New installs also write a small user config if none exists. Apple Silicon prefers apfel; Linux with a detected modern GPU and Ollama prefers ollama:qwen3:8b; CPU-oriented hosts with llama-cli and a readable GGUF model prefer llama.cpp:/absolute/model.gguf. This selection only configures Parlay; it does not install binaries or download models.

Override paths or the source repository with environment variables:

PARLAY_INSTALL_DIR="$HOME/Projects/parlay" \
PARLAY_BIN_DIR="$HOME/.local/bin" \
PARLAY_REPO_URL="https://git.rpdo.net/rpdo-net/parlay.git" \
sh install.sh

Initial Hydra Registration

Local Hydra project id: parlay

Initial local path:

/Users/ralf/Projects/hydra/workshop/parlay

Initial lifecycle state: incubating

Publication policy: local-only

Privacy: private

Read-only Hydra Status Context

Parlay can import an explicitly supplied Hydra status JSON file as an append-only system event. The importer never discovers, invokes, or writes to Hydra. See Read-only Hydra Status Source.

Design Rule

Keep it simple. Parlay must become useful before it becomes clever.

Project Documents

• MVP 0.1
• Storage
• Roadmap
• Strategic Papers