lp: gossip

Gossip ====== Chr. Clemens Lahme 2026-07-05 Table of Contents -----------------

Purpose
System Requirements
License
Download
Configure
Build
Project File Structure
Model CSV Format
1. OpenRouter
2. llama.cpp
3. pi.dev
Glossary
Nomenclature
Base Utilities
1. datetime
2. today
3. Project Location
4. duration
5. question
OpenRouter
1. Introduction
2. Configuration
3. Workflow Overview
4. Usage Count
5. Print Answer
6. Request and Response
7. The 'ask' Command
Questions
llama.cpp
Downloading a Minimal GGUF Model
Llama Text Output To Gossip Text Format
Invoking llama.cpp Locally
1. Prerequisites
2. Usage
3. Examples
4. Behaviour
5. llama Script
Local llama.cpp Server
The llama.cpp Client
llama_call
Understanding Pi Dev
Multi Turn pi.dev Interactions
Interaction With Pi Dev
pi.dev Models Storage
pi.dev OpenRouter Models
Configured OpenRouter Models
Main Gossip Script
Print Answer
Print Question
Tags
Pi Harness Prompt
ToDo
Links

1. Purpose ---------- Gossip is a Unix command-line frontend for Large Language Models. It stores every interaction—requests as plain text, responses as structured JSON—locally in a log-based database (`gossip/db/txt` and `gossip/db/json`) for permanent, offline reference. It supports three inference backends: * OpenRouter (Cloud) The most practical default. Requires an API key and internet access. Provides instant access to a vast catalog of state-of-the-art models (Kimi 3, GLM 5.2, MiniMax 3, Nemotron 3 Ultra, etc.) with zero local compute requirements. * llama.cpp (Local) Runs models entirely offline on your hardware. Gossip can automatically fetch and build 'llama.cpp' for you (requires internet for the initial build and model download). * No GPU required: Runs efficiently on CPU. * Modest RAM is sufficient: With 64 GB RAM you can comfortably run high-capability models like Qwen 3 35B with surprising results, especially for coding. * Total privacy: Data never leaves the machine. * pi.dev (Agent / Delegation) A specialized workflow where Gossip delegates the request to a locally running `pi` agent (pi.dev). The agent drives the conversation using its own configuration (e.g. again OpenRouter). Gossip captures the resulting session, storing the agent's final answer as text alongside the other native logs. Choose OpenRouter for speed, model variety, and zero setup. Choose llama.cpp for privacy, offline work, cost control, and surprisingly strong performance on standard workstation hardware (64 GB+ RAM). Use pi.dev when you want an autonomous agent to handle multi-step tasks using your existing `pi` configuration. Or choose all three - because you can. 2. System Requirements ---------------------- Mandatory * Linux. * Ruby. * Both dash and bash shells must be available. * git. * make. * curl Included * Rlib Ruby library. * lp, a literate programming tool written in Ruby. Either OpenRouter * An OpenRouter key. * jq. * Internet access. llama.cpp * cmake. * g++. * wget in order to download a model in GGUF format. * At least one large language model downloaded in GGUF format (e.g. from Hugging Face). pi.dev 3. License ---------- Gossip is licensed under the GNU Public License (GPL) version 2, see: file://COPYING.txt 4. Download ----------- The home page of this project is at: http://techinvest.li/gossip/ You can download the source code with the following git command: git clone https://techinvest.li/git/gossip.git 5. Configure ------------ The configure checks if all system requirements are met. Run it with the following command on the command line in order to find out if anything is missing to use Gossip: ./configure Here is the script. It got already generated and checked into this project, so you can run it without getting into circular dependencies. cat > ./configure <<EOT #!/bin/sh # Do not edit this file, as it gets automatically created by lp. # configure for Gossip set -u ROOT=$(CDPATH= cd "$(dirname "$0")" && pwd) MANDATORY_MISSING=0 OPTIONAL_MISSING=0 ok() { printf ' [ ok ] %s\n' "$*"; } warn() { printf ' [ ?? ] %s\n' "$*"; } fail() { printf ' [FAIL] %s\n' "$*"; } echo echo "Gossip configure" echo "================" echo echo "Project root: $ROOT" echo # --------------------------------------------------------------------------- # Mandatory requirements # --------------------------------------------------------------------------- echo "Mandatory requirements" echo "----------------------" if [ "$(uname -s)" = "Linux" ]; then ok "Linux detected" else fail "Linux not detected (found: $(uname -s))" MANDATORY_MISSING=1 fi for shell in dash bash; do if command -v "$shell" >/dev/null 2>&1; then ok "$shell shell found: $(command -v "$shell")" else fail "$shell shell not found" MANDATORY_MISSING=1 fi done if command -v git >/dev/null 2>&1; then ok "git found: $(command -v git)" else fail "git not found" MANDATORY_MISSING=1 fi if command -v make >/dev/null 2>&1; then ok "make found: $(command -v make)" else fail "make not found" MANDATORY_MISSING=1 fi if command -v ruby >/dev/null 2>&1; then ok "Ruby found: $(ruby -v 2>&1 | sed -n '1p')" else fail "Ruby not found" MANDATORY_MISSING=1 fi if command -v curl >/dev/null 2>&1; then ok "curl found: $(command -v curl)" else fail "curl not found" MANDATORY_MISSING=1 fi # --------------------------------------------------------------------------- # Optional: OpenRouter dependencies # --------------------------------------------------------------------------- echo echo "Optional: OpenRouter dependencies" echo "---------------------------------" if command -v jq >/dev/null 2>&1; then ok "jq found: $(command -v jq)" else warn "jq not found (OpenRouter backend will be unavailable)" OPTIONAL_MISSING=1 fi if [ -f "$ROOT/etc/openrouter.rc" ]; then if LC_ALL=C grep -E -v '^[[:space:]]*(#|$)' "$ROOT/etc/openrouter.rc" \ 2>/dev/null | grep . >/dev/null 2>&1 then ok "OpenRouter key file present: etc/openrouter.rc" else warn "etc/openrouter.rc is empty or contains only comments" OPTIONAL_MISSING=1 fi else warn "etc/openrouter.rc not found (OpenRouter backend will be unavailable)" OPTIONAL_MISSING=1 fi # Internet access check using Ruby, since Ruby is mandatory. if command -v ruby >/dev/null 2>&1; then if ruby -rsocket -rtimeout -e \ 'begin; Timeout.timeout(10) { Socket.tcp("example.com", 443) { |s| } }; rescue StandardError; exit 1; end' \ >/dev/null 2>&1 then ok "Internet access (TCP connection to example.com:443)" else warn "Internet access could not be verified" OPTIONAL_MISSING=1 fi else fail "Cannot check internet because Ruby is missing" MANDATORY_MISSING=1 fi # --------------------------------------------------------------------------- # Optional: llama.cpp dependencies # --------------------------------------------------------------------------- echo echo "Optional: llama.cpp dependencies" echo "--------------------------------" if command -v cmake >/dev/null 2>&1; then ok "cmake found: $(command -v cmake)" else warn "cmake not found (llama.cpp backend will be unavailable)" OPTIONAL_MISSING=1 fi if command -v g++ >/dev/null 2>&1; then ok "g++ found: $(command -v g++)" else warn "g++ not found (llama.cpp backend will be unavailable)" OPTIONAL_MISSING=1 fi if command -v wget >/dev/null 2>&1; then ok "wget found: $(command -v wget)" else warn "wget not found (llama.cpp will not easily get a GGUF model file)" OPTIONAL_MISSING=1 fi GGUF_MODEL="${GGUF_MODEL:-}" found_model="" if [ -n "$GGUF_MODEL" ]; then if [ -f "$GGUF_MODEL" ]; then ok "GGUF model found: $GGUF_MODEL" found_model="$GGUF_MODEL" else warn "GGUF_MODEL is set but file not found: $GGUF_MODEL" OPTIONAL_MISSING=1 fi else if command -v find >/dev/null 2>&1; then for dir in "$ROOT/db/gguf" "$ROOT/opt/llama.cpp/models"; do if [ -d "$dir" ]; then candidate=$(LC_ALL=C find "$dir" -type f -name '*.gguf' 2>/dev/null \ | sed -n '1p' 2>/dev/null) if [ -n "$candidate" ]; then found_model="$candidate" break fi fi done if [ -n "$found_model" ]; then ok "GGUF model found: $found_model" else warn "No .gguf model found in db/gguf." OPTIONAL_MISSING=1 fi else warn "find not found; cannot search for GGUF models" OPTIONAL_MISSING=1 fi fi # --------------------------------------------------------------------------- # Optional: pi.dev dependencies # --------------------------------------------------------------------------- echo echo "Optional: pi.dev dependencies" echo "-----------------------------" if command -v pi >/dev/null 2>&1; then ok "pi.dev agent found: $(command -v pi)" else warn "pi command not found (pi.dev backend will be unavailable)" OPTIONAL_MISSING=1 fi # --------------------------------------------------------------------------- # Summary # --------------------------------------------------------------------------- echo if [ "$MANDATORY_MISSING" -eq 1 ]; then echo "Gossip configure: FAILED" echo "Mandatory requirements are missing. Please install them and re-run configure." exit 1 fi if [ "$OPTIONAL_MISSING" -eq 1 ]; then echo "Gossip configure: done (with optional dependency warnings)." else echo "Gossip configure: done." fi exit 0 # End of: configure EOT 6. Build -------- Gossip is distributed as a literate program: the executable scripts are not checked in directly but are generated from this document by the included 'lp' tool. To build everything, including the bin/gossip.rb command-line frontend, run: make 7. Project File Structure ------------------------- Here is part of the project directory structure: gossip +-bin | +-gossip.rb +-build +-db | +-csv | | +-model_20260726_054647.csv | | +-tags.csv | +-json | | +-response_20260726_054647.json | +-txt | +-answer_20260724_121431.txt | +-answer_20260726_054647.txt | +-question_20260726_054647.txt +-doc | +-index.txt +-lib | +-rlib.rb +-log +-opt | +-llama.cpp +-tmp db/csv contains model_YYYYMMDD_HHMMSS.csv files and db/txt contains question_YYYYMMDD_HHMMSS.txt files with question answer text content, to and from local and cloud LLMs models. The model file contains meta information about the model used. 8. Model CSV Format ------------------- The model record is a one-line CSV file stored at 'db/csv/model_<datetime>.csv'. It records which backend and model answered a given question. The '<datetime>' portion of the filename ('YYYYMMDD_HHMMSS') is the primary key that links the model record to its corresponding question, answer, and response file. There is no header line in the file - the single data line is the entire file content. Format by Backend 8.1. OpenRouter ''''''''''''''' openrouter,<model_id> Example: openrouter,deepseek/deepseek-v4-pro 8.2. llama.cpp '''''''''''''' llama.cpp,<model_basename>,version=<llama_cpp_version>,seed=<seed> Example: llama.cpp,Qwen3.6-35B-A3B-UD-Q8_K_XL.gguf,version=10075,seed=1 8.3. pi.dev ''''''''''' pi.dev,<model_id>,reasoning=<thinking_level> Example: pi.dev,inclusionai/ling-3.0-flash:free,reasoning=high 9. Glossary ----------- Large language model (llm) A neural network trained on vast amounts of text data that can understand and generate human-like language. In Gossip, LLMs are accessed either remotely via OpenRouter (cloud models like GPT, Gemini, Qwen, DeepSeek, etc.) or locally via llama.cpp (GGUF-format models such as Llama 3, Qwen 3.6, etc.). The model used for each question answer pair is recorded in 'db/csv/model_<datetime>.csv'. Inference The process of running a prompt through an LLM to generate a response. In Gossip, inference happens in two modes: a) cloud inference — an HTTPS POST to 'https://openrouter.ai/api/v1/chat/completions' with a JSON payload containing the model ID, messages, reasoning settings, temperature, and provider routing; b) local inference — a direct invocation of 'llama-cli' (from llama.cpp) with flags for single-turn, non-interactive, deterministic ('--temp 0.0 --seed 1') output. Both paths store the raw response and a cleaned-up version under 'db/txt/answer_<datetime>.txt'. Prompt The user-supplied input text sent to the LLM. In Gossip, prompts originate from three sources: a) typed on the command line ('bin/llama <model> "What is POSIX?"'), b) read from an arbitrary file ('bin/llama <model> /path/to/question.txt'), or c) piped via stdin ('echo "prompt" | bin/llama <model>'). Every prompt is immediately persisted as 'db/txt/question_<datetime>.txt' before inference begins. Request The complete HTTP payload sent to an LLM provider. In Gossip this is the JSON document POSTed to 'https://openrouter.ai/api/v1/chat/completions' (cloud) or the command-line arguments passed to 'llama-cli' (local). It contains the model identifier, the conversation messages, and inference parameters such as temperature, reasoning flags, and provider routing. Response The raw byte stream returned by the LLM provider. For OpenRouter this is the HTTP response body (JSON) including the generated message, usage statistics, and any provider-specific metadata. For llama.cpp it is the combined stdout/stderr output of the 'llama-cli' process, including the answer, thinking traces, and performance timings, thought not in JSON but normal text format. Reply Synonym for Response used when emphasising the conversational view: the assistant's turn in the dialogue. In Gossip the raw reply is first saved verbatim in a raw response file and then transformed into a converted reply - also called answer file - with cleaned text with reasoning wrapped in '<think> ... </think>' tags. JSON JavaScript Object Notation — the wire format used for all OpenRouter API requests and responses. Gossip constructs request payloads safely with 'jq' ('--slurpfile', '--argjson', '-Rs') to avoid shell-escaping issues. Responses are saved verbatim to 'db/json/response_<datetime>.json' and later pretty-printed for reading via 'bin/or_print.rb' into 'db/txt/answer_<datetime>.txt'. Pi.dev sessions are also stored as JSONL (newline-delimited JSON) under '~/.pi/agent/sessions/'. CSV Comma-Separated Values — the lightweight tabular format Gossip uses for metadata indexes. Two CSV families exist: 1. 'db/csv/model_<datetime>.csv' — one line per Q/A pair recording the backend ('gossip,openrouter' or 'gossip,llama.cpp-<version>'), the model identifier, and an optional reasoning flag; 2. 'db/csv/tags.csv' — a many-to-many tag index where each line is '<datetime>,<tag1>,<tag2>,...' enabling tag-based lookup via 'bin/tags.rb'. OpenRouter A unified API gateway that proxies requests to dozens of third-party LLM providers (Kimi, DeepSeek, Qwen, GLM, MiniMax, etc.). Gossip uses OpenRouter for all cloud inference; the API key is stored in 'etc/openrouter.rc' and requests are sent to 'https://openrouter.ai/api/v1/chat/completions'. A user account there is free and so far there have been always free models available, although they change over time, in order to start experiment. What is needed is the API key setup for Gossip to use in the etc/openrouter.rc file. llama.cpp An open source C++ inference engine for GGUF-format models that runs entirely on local hardware (CPU and or GPU). Gossip invokes its build 'llama-cli' executable for offline and private inference. The built binary lives under 'gossip/opt/llama.cpp'. pi.dev An open-source, terminal-native AI coding agent ('@earendil-works/pi-coding-agent') that runs multi-turn, tool-using sessions. Gossip can import pi.dev session logs ('~/.pi/agent/sessions/*.jsonl') to extract user/assistant exchanges and store them as first-class question/answer pairs in its own database. 10. Nomenclature ---------------- In the Gossip codebase and documentation the following terms have specific, narrower meanings than their general definitions above: question A prompt that has been persisted to 'db/txt/question_<datetime>.txt' before inference starts. Every interaction-whether typed on the command line, read from a file, piped via stdin, or imported from a pi.dev session-becomes a question file. The timestamp in the filename is the canonical primary key for the entire Q/A pair. The content of the question file can also consist of an instruction, document, or whole conversation. answer The converted reply saved to a text or markdown file answer_<datetime>.txt in location db/txt. It is the human-readable text shown to the user: the model's final answer first, followed by any reasoning/thinking content wrapped in '<think>...</think>' markers. The raw API/CLI output is not shown directly. model record A one-line CSV file 'db/csv/model_<datetime>.csv' recording the backend ('gossip,openrouter' or 'gossip,llama.cpp-<version>'), the model identifier, and routing/reasoning flags. It is the authoritative source for "which model answered this question". session (pi.dev sense) A pi.dev JSONL log file under '~/.pi/agent/sessions/'. Gossip's 'bin/pi_sessions.rb' reads these files, extracts 'message' events with 'role: user' and 'role: assistant', and converts each pair into a Gossip question/answer pair with matching timestamps. datetime The 15-character string 'YYYYMMDD_HHMMSS' produced by 'bin/datetime'. It is used as the sole identifier in all database filenames ('question_<datetime>.txt', 'model_<datetime>.csv', 'tags.csv' rows) guaranteeing lexical sort order equals chronological order. tag A free-form label stored in 'db/csv/tags.csv' as '<datetime>,<tag1>,<tag2>,...'. Tags enable fast thematic lookup via 'bin/tags.rb <tag>' without full-text search. provider (OpenRouter sense) The upstream inference provider selected via the 'provider.only' field in the OpenRouter request (e.g. 'google-ai-studio', 'deepseek', 'alibaba', 'together'). Gossip hard-codes a mapping from model ID to preferred provider in 'bin/or_ask.sh'. 11. Base Utilities ------------------ The document is written in a bottom up fashion, code that will be invoked later is mentioned first. Here is some boiler plate code that might be used throughout the project later on. 11.1. datetime '''''''''''''' The output of the datetime script will be used all over the place as an identifier and in file names to have an order of question answer pairs and other meta data. cat > ./bin/datetime <<EOT #! /bin/dash # Do not edit this file, as it gets automatically generated by lp. date '+%Y%m%d_%H%M%S' EOT 11.2. today ''''''''''' The 'today' script prints the current date in ISO YYYY-MM-DD format. It is used in other Gossip shell scripts. cat > ./bin/today <<EOT #! /bin/dash # Do not edit this file, as it gets automatically generated by lp. date '+%Y-%m-%d' EOT 11.3. Project Location '''''''''''''''''''''' We want to be able to invoke tools from gossip from any location, so inside the project it is handy to identify your own location. cat > lib/apphome.sh <<EOT if [ "$APPHOME" = "" ] then # try to locate application home directory ## resolve links - $0 may be a link to the app home PRG=$0 progname=`basename $0` while [ -h "$PRG" ] do ls=`ls -ld "$PRG"` link=`expr "$ls" : '.*-> $.*$$'` if expr "$link" : '/.*' > /dev/null then PRG="$link" else PRG="`dirname $PRG`/$link" fi done export APPHOME=`dirname "$PRG"`/.. export APPHOME=$(cd $APPHOME >/dev/null; pwd) fi EOT This is actually just a code snippet for use in other scripts. We need to copy the source into other shell scripts, otherwise we would already know the projects home location. 11.4. duration '''''''''''''' The duration script is not inlined here but copied from the Rlib project and shows in seconds or minutes how long a task has taken. 11.5. question '''''''''''''' The 'bin/question' script creates a new empty question file 'db/txt/question_<datetime>.txt' and prints its absolute path. It is used by 'bin/ask' (and can be used directly) to prepare a question file before opening an editor or writing a prompt. It can also be used to set up a Gossip question/answer database inside another project - the database files just need a 'db/txt' directory to live in. The script resolves the project root via the same 'APPHOME' logic used in 'lib/apphome.sh', so it works from any working directory. An optional 'YYYYMMDD_HHMMSS' timestamp can be passed as the first argument; otherwise the current time is used. cat > bin/question <<EOT #! /bin/dash # Do not edit this file, as it gets automatically generated by lp. if [ "$APPHOME" = "" ] then # try to locate application home directory ## resolve links - $0 may be a link to the app home PRG=$0 progname=`basename $0` while [ -h "$PRG" ] do ls=`ls -ld "$PRG"` link=`expr "$ls" : '.*-> $.*$$'` if expr "$link" : '/.*' > /dev/null then PRG="$link" else PRG="`dirname $PRG`/$link" fi done export APPHOME=`dirname "$PRG"`/.. export APPHOME=$(cd $APPHOME >/dev/null; pwd) fi DATETIME=$(date '+%Y%m%d_%H%M%S') if [ "$1" != "" ] then DATETIME="$1" fi if [ -d "${PWD}/db/txt" ] then QUESTION_DIR="${PWD}/db/txt" else QUESTION_DIR="${APPHOME}/db/txt" fi FILENAME="${QUESTION_DIR}/question_${DATETIME}.txt" touch "$FILENAME" echo "$FILENAME" # End of: question EOT Usage: # Create a new question file with current timestamp. $ ./bin/question /path/to/gossip/db/txt/question_20260705_123456.txt # Create a question file with a specific timestamp (e.g. for reproducibility). $ ./bin/question 20260705_123456 /path/to/gossip/db/txt/question_20260705_123456.txt The returned path can be passed to an editor: qf=$(./bin/question) $EDITOR $qf 12. OpenRouter -------------- 12.1. Introduction '''''''''''''''''' OpenRouter is Gossip's cloud inference backend. It provides a single, unified API endpoint ('https://openrouter.ai/api/v1/chat/completions') that routes requests to dozens of upstream model providers-Google, DeepSeek, Moonshot, NVIDIA, Z.ai, Alibaba, MiniMax, and many others. This means you get immediate access to a constantly evolving catalog of state-of-the-art models (Kimi 3, GLM 5.2, MiniMax 3, Nemotron 3 Ultra, Qwen 3.7, etc.) without managing multiple API keys, SDKs, or provider-specific quirks. 12.2. Configuration ''''''''''''''''''' You need an OpenRouter account - sign up at https://openrouter.ai/ is free and the (ever changing) free models get you quite far. What you need is an API key. After you created your account, you will get one. Create 'etc/openrouter.rc' in the Gossip project root and put a line like this: OPENROUTER_API_KEY=sk-or-v1-XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX Replace the string content with your API key. Make sure only you can read it: chmod 600 etc/openrouter.rc The file is sourced by 'or_ask.sh' and must export 'OPENROUTER_API_KEY'. Keep it out of version control (it is already in '.gitignore'). For identification of useful large language models at OpenRouter have a look at the Links section further down in this document. 12.3. Workflow Overview ''''''''''''''''''''''' Every OpenRouter interaction follows the same lifecycle: 1. Question capture - User prompt is written to 'db/txt/question_<datetime>.txt' (via 'bin/ask' or manually). 2. Payload construction - 'or_ask.sh' reads the question, builds a JSON request with 'jq' (safe against injection, quoting, and encoding issues), and writes it to a temporary file. 3. HTTP request - 'curl' POSTs the payload to OpenRouter with the API key. 4. Response logging - Raw response (headers + body) is saved to 'db/json/response_<datetime>.json'. 5. Model metadata - A one-line CSV record is written to 'db/csv/model_<datetime>.csv' recording backend, model ID, and routing. 6. Answer extraction - 'or_print.rb' parses the JSON, extracts the assistant message and optional reasoning, cleans formatting, and writes the human-readable answer to 'db/txt/answer_<datetime>.txt'. 7. Display - The answer is shown in 'less' (or printed to stdout by 'bin/ask'). All files share the same '<datetime>' prefix ('YYYYMMDD_HHMMSS'), so the question, model record, raw response, and cleaned answer are trivially correlated. 12.4. Usage Count ''''''''''''''''' First a script to see how many times the user has invoked openrouter already today. cat > ./bin/or_count.sh <<EOT #! /bin/dash # Do not edit this file, as it gets automatically generated by lp. # This script counts how many times openrouter has been invoked today. # An optional first argument selects the database directory to inspect; # it defaults to the gossip database "db". bin/or_ask.sh # passes DB_DIR here so that question/answer databases inside other # projects are counted as well (same convention as bin/question). DB_DIR=db if [ "$1" != "" ] then DB_DIR="$1" fi DATETIME=$(./bin/datetime) TODAY=$(echo $DATETIME | perl -pe 's/_.*$//;') grep -E -i 'openrouter,' "${DB_DIR}"/csv/model_${TODAY}_??????.csv 2>/dev/null | wc -l # End of: or_count.sh EOT 12.5. Print Answer '''''''''''''''''' Following comes the Ruby script 'or_print.rb' that extracts the 'reasoning' (thinking) part and the main 'content' (answer) from the OpenRouter JSON reply, printing them clearly to standard output: cat > ./bin/or_print.rb <<EOT #! /usr/bin/env ruby # coding: utf-8 # Do not edit this file, as it gets automatically generated by lp. require "json" answer_filename = ARGV[ 0 ] if answer_filename.nil? || answer_filename.empty? puts "Usage: #{$0} <json_filename>" exit 1 end begin content = File.read( answer_filename ) json_data = JSON.parse( content ) rescue JSON::ParserError => e puts "Error parsing JSON: #{answer_filename}" exit 2 rescue Errno::ENOENT puts "File not found: #{answer_filename}" exit 3 end # ============================================================ # MODEL & COST INFORMATION # ============================================================ puts "=" * 60 puts "MODEL & USAGE" puts "=" * 60 puts model = json_data[ "model" ] provider = json_data[ "provider" ] system_fp = json_data[ "system_fingerprint" ] created = json_data[ "created" ] id = json_data[ "id" ] puts "Model: #{model}" if model puts "Provider: #{provider}" if provider puts "System Fingerprint: #{system_fp}" if system_fp puts "Request ID: #{id}" if id if created time_str = Time.at(created).strftime("%Y-%m-%d %H:%M:%S %Z") puts "Created: #{time_str}" end usage = json_data[ "usage" ] if usage puts puts "Token Usage:" puts " Prompt tokens: #{usage[ 'prompt_tokens' ]}" if usage[ 'prompt_tokens' ] puts " Completion tokens: #{usage[ 'completion_tokens' ]}" if usage[ 'completion_tokens' ] puts " Total tokens: #{usage[ 'total_tokens' ]}" if usage[ 'total_tokens' ] # Reasoning tokens (if available) if usage[ 'completion_tokens_details' ] && usage[ 'completion_tokens_details' ][ 'reasoning_tokens' ] puts " Reasoning tokens: #{usage[ 'completion_tokens_details' ][ 'reasoning_tokens' ]}" end # Cached tokens (if available) if usage[ 'prompt_tokens_details' ] && usage[ 'prompt_tokens_details' ][ 'cached_tokens' ] puts " Cached tokens: #{usage[ 'prompt_tokens_details' ][ 'cached_tokens' ]}" end puts puts "Cost:" cost = usage[ "cost" ] if cost printf " Total cost: $%.6f\n", cost else puts " Total cost: N/A" end if usage[ "cost_details" ] cd = usage[ "cost_details" ] printf " Upstream inference: $%.6f\n", cd[ "upstream_inference_cost" ] if cd[ "upstream_inference_cost" ] printf " Upstream prompt: $%.6f\n", cd[ "upstream_inference_prompt_cost" ] if cd[ "upstream_inference_prompt_cost" ] printf " Upstream completion:$%.6f\n", cd[ "upstream_inference_completions_cost" ] if cd[ "upstream_inference_completions_cost" ] end puts " BYOK: #{usage[ 'is_byok' ] ? 'yes' : 'no'}" if usage.key?( 'is_byok' ) end puts # ============================================================ # EXTRACT CHOICE # ============================================================ choices = json_data[ "choices" ] if choices.nil? || choices.empty? puts "No choices found in JSON." exit 4 end message = choices[ 0 ][ "message" ] if message.nil? puts "No message found in choices." exit 5 end # ============================================================ # REASONING (THINKING) # ============================================================ reasoning = message[ "reasoning" ] if reasoning && !reasoning.empty? puts "=" * 60 puts "THINKING / REASONING" puts "=" * 60 puts puts reasoning.gsub( /\n\n+/, "\n\n" ).strip puts end # ============================================================ # MAIN CONTENT (ANSWER) # ============================================================ content_text = message[ "content" ] if content_text && !content_text.empty? puts "=" * 60 puts "ANSWER" puts "=" * 60 puts clean_text = content_text.gsub( /\n\n+/, "\n\n" ).strip puts clean_text puts end # ============================================================ # FINISH REASON # ============================================================ finish_reason = choices[ 0 ][ "finish_reason" ] native_finish = choices[ 0 ][ "native_finish_reason" ] if finish_reason || native_finish puts "=" * 60 puts "FINISH INFO" puts "=" * 60 puts puts "Finish reason: #{finish_reason}" if finish_reason puts "Native finish reason: #{native_finish}" if native_finish puts end # If neither was present if (reasoning.nil? || reasoning.empty?) && (content_text.nil? || content_text.empty?) puts "No reasoning or content found in the message." exit 6 end EOT 12.6. Request and Response '''''''''''''''''''''''''' This script builds the JSON request, sends it to OpenRouter, receives the response, saves it to the local data store, and creates a text file out of it plus prints the location of this text file out, so the user can find and read it. cat > ./bin/or_ask.sh <<EOT #! /bin/bash # Do not edit this file, as it gets automatically generated by lp. # ============================================================================= # or_ask.sh # ============================================================================= # PURPOSE: # Safely query the OpenRouter API using JSON payloads constructed via 'jq', # with timestamped Q&A logging, and provider routing. # # ENVIRONMENT: # DATETIME : Timestamp for Q&A files. Auto-generated if unset. # DB_DIR : Database directory holding the txt/csv/json subdirectories. # Defaults to "db" (the gossip database). bin/ask sets this # to another project's db directory when the question lives # there (same convention as bin/question). # OPENROUTER_API_KEY : API key sourced from etc/openrouter.rc # # INPUT FILES: # ${DB_DIR}/txt/question_${DATETIME}.txt : Raw user prompt (must exist) # # OUTPUT FILES: # ${DB_DIR}/csv/model_${DATETIME}.csv : Model identifier & routing metadata # ${DB_DIR}/json/response_${DATETIME}.json : Raw API response (JSON + text) # ${DB_DIR}/txt/answer_${DATETIME}.txt : Converted, human readable answer # # DEPENDENCIES: # - bash (4.0+) # - jq (1.6+) # - curl # - etc/openrouter.rc (must export OPENROUTER_API_KEY) # - Helper scripts: # ./bin/datetime # ./bin/or_count.sh # ./bin/or_print.rb # ./bin/duration # # USAGE: # DATETIME=YYYYMMDD_HHMMSS ./bin/or_ask.sh <openrouter_model_id> # # ENVIRONMENT: # DATETIME : Timestamp for Q&A files. Auto-generated if unset. # OPENROUTER_API_KEY : API key sourced from etc/openrouter.rc # # INPUT FILES: # db/txt/question_${DATETIME}.txt : Raw user prompt (must exist) # # OUTPUT FILES: # db/csv/model_${DATETIME}.csv : Model identifier & routing metadata # db/json/response_${DATETIME}.json : Raw API response (JSON + text) # # WORKFLOW: # 1. Check daily usage count via or_count.sh # 2. Validate/initialize DATETIME and source API credentials # 3. Look up provider in db/csv/openrouter_models.csv (falls back to empty if not found) # 4. Read prompt from question file and safely encode it via jq # 5. Construct full JSON payload (messages, reasoning, temperature, provider) # 6. Send request to https://openrouter.ai/api/v1/chat/completions # 7. Save response & model metadata to timestamped CSV files # 8. Display answer in 'less' and print execution timing # # SAFETY & ROBUSTNESS: # - Uses 'jq -Rs' to read prompts: prevents shell injection, handles quotes, # newlines, backslashes, and Unicode safely without escaping hacks. # - 'set -e' ensures immediate exit on command failure. # - Validates question file existence before proceeding. # - Provider routing is explicit and extensible; falls back to OpenRouter defaults. # # INTEGRATION NOTES: # - Expects 'DATETIME' to be set by 'bin/set_datetime.sh' or TUI state manager # - Output format matches 'gossip' database schema (question/model/answer are # text or CSV formats) reasoning=true #reasoning=false #temperature=1.0 temperature=0.0 set -e # The database directory can be injected by the caller (bin/ask) to support # question/answer databases inside other projects (see bin/question). # It defaults to the gossip database. if [ "$DB_DIR" = "" ] then DB_DIR=db fi #mkdir -p "${DB_DIR}/txt" "${DB_DIR}/csv" "${DB_DIR}/json" printf "Today's question count: " ./bin/or_count.sh "$DB_DIR" if [ "$DATETIME" = "" ] then DATETIME=$(./bin/datetime) fi echo "DATETIME=${DATETIME}" # This sets the user key/password. source etc/openrouter.rc model="$1" echo "model=${model}" DATETIME_BEGIN=$(./bin/datetime) # Look up the provider from db/csv/openrouter_models.csv. # Format: model_id,provider # If no match is found, provider remains empty (falls back to OpenRouter defaults). # NOTE: this is gossip configuration, so it is always read from the gossip # database, independent of DB_DIR. provider="" while IFS=, read -r csv_model csv_provider; do # Skip comment/empty lines case "$csv_model" in '#'*|'') continue ;; esac if [ "$model" = "$csv_model" ]; then provider="$csv_provider" break fi done < "db/csv/openrouter_models.csv" echo "provider=${provider}" if [ "$model" = "minimax/minimax-m2.7" ] then reasoning=true fi if [ "$model" = "minimax/minimax-m3" ] then reasoning=true fi echo "reasoning=${reasoning}" QUESTION_FILENAME="${DB_DIR}/txt/question_${DATETIME}.txt" MODEL_FILENAME="${DB_DIR}/csv/model_${DATETIME}.csv" RESPONSE_FILENAME="${DB_DIR}/json/response_${DATETIME}.json" echo "RESPONSE_FILENAME=${RESPONSE_FILENAME}" ANSWER_FILENAME="${DB_DIR}/txt/answer_${DATETIME}.txt" test -f $QUESTION_FILENAME || { echo "ERROR: now file with a question found: ${QUESTION_FILENAME}" exit 2 } echo "openrouter,${model}" > $MODEL_FILENAME TMPDIR=./tmp mkdir -p $TMPDIR # Read the question text from the file. #QUESTION=$(cat "$QUESTION_FILENAME") #curl --silent https://openrouter.ai/api/v1/chat/completions \ # -H "Content-Type: application/json" \ # -H "Authorization: Bearer $OPENROUTER_API_KEY" \ # -d "{ \"model\": \"${model}\", \"messages\": [ { \"role\": \"user\", \"content\": \"${question}\" } ], \"reasoning\": { \"enabled\": ${reasoning} }, \"temperature\": ${temperature}, \"provider\": { \"only\": [ \"${provider}\" ] } }" |& tee db/txt/answer_${DATETIME}.txt # Convert bash true/false to JSON boolean for --argjson. if [ "$reasoning" = "true" ]; then reasoning_json=true else reasoning_json=false fi # Build the JSON payload with jq. JSON_FILE="$TMPDIR/openrouter_${DATETIME}.json" MESSAGE_FILE="$TMPDIR/message_${DATETIME}.json" # ------------------------------------------------------------------------- # Step 1: Build the "messages" array in a separate JSON file. # 'jq -Rs' reads the question file as a single raw string and # wraps it into a JSON array, so all special characters (quotes, # newlines, backslashes, ...) are escaped properly. The file is # written via shell redirection, so the long content never goes # on the command line. # ------------------------------------------------------------------------- jq -Rs '[{ role: "user", content: . }]' "$QUESTION_FILENAME" > "$MESSAGE_FILE" echo "MESSAGE_FILE=${MESSAGE_FILE}" # ------------------------------------------------------------------------- # Step 2: Build the full request body. # '--slurpfile messages' reads the file produced in step 1 and # binds it to '$messages' (wrapped in an outer array by slurpfile, # so $messages[0] is our messages array). The remaining parameters # are small, so passing them on the command line is safe. # ------------------------------------------------------------------------- jq -n \ --arg model "$model" \ --argjson reasoning "$reasoning_json" \ --argjson temperature "$temperature" \ --arg provider "$provider" \ --slurpfile messages "$MESSAGE_FILE" \ '{ model: $model, messages: $messages[0], reasoning: { enabled: $reasoning }, temperature: $temperature } + (if $provider != "" then { provider: { only: [$provider] } } else {} end)' \ > "$JSON_FILE" echo "JSON_FILE=${JSON_FILE}" # Send the request using the JSON file. # Both stdout and stderr are captured into the answer file. curl --silent https://openrouter.ai/api/v1/chat/completions \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $OPENROUTER_API_KEY" \ -d "@$JSON_FILE" |& tee "$RESPONSE_FILENAME" echo "Response received: the answer has been written to: ${RESPONSE_FILENAME}" #echo "openrouter,${model}" > $MODEL_FILENAME DATETIME_END=$(./bin/datetime) echo echo "DATETIME=${DATETIME}" #command="./bin/or_print.rb ${RESPONSE_FILENAME} > ${ANSWER_FILENAME}" #echo $command #echo #$command #cat db/txt/answer_${DATETIME}.txt | perl -pe 's/\\n/\n/g;' | less +G ./bin/or_print.rb $RESPONSE_FILENAME > $ANSWER_FILENAME echo less "$ANSWER_FILENAME" ./bin/duration $DATETIME_BEGIN $DATETIME_END echo "SUCCESS: $0 - $?." # End of: or_ask.sh EOT 12.7. The 'ask' Command ''''''''''''''''''''''' The 'bin/ask' script is the primary user-facing entry point for OpenRouter (cloud) inference. It handles the complete workflow: creating a timestamped question file, optionally opening your editor, sending the request, and displaying the answer. Use the EDITOR environment setting export EDITOR=nano ./bin/ask nvidia/nemotron-3-ultra-550b-a55b:free # The nano editor will open, write your question, prompt, or instruction, save # the file and exit the editor. One Liner Prompt bin/ask deepseek/deepseek-v4-pro "Explain inference in LLMs in three sentences." Here noo editor opens; the prompt is written directly to the question file. The answer is printed immediately after the request completes. Multi-Word Prompt Without Quotes bin/ask deepseek/deepseek-v4-pro What is the capital of France? All arguments after the model ID are joined with spaces to form the prompt and again the answer gets printed. And here is the script: cat > ./bin/ask <<EOT #!/bin/bash # Do not edit this file, as it gets automatically generated by lp. set -e SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)" PROJECT_DIR="$(dirname "$SCRIPT_DIR")" # Work from the project directory so relative paths in helper scripts work. #cd "$PROJECT_DIR" print_usage() { echo "Usage: $0 <openrouter_model_id> [prompt text...]" echo "" echo "Interactively asks an OpenRouter model a question." echo "" echo "If no prompt text is given, opens the editor to write one." echo "If prompt text is given, uses it directly." echo "If a prompt file is given, reads it as the prompt." echo "If a question file (question_<datetime>.txt) is given, reuses it." echo "" echo "Like bin/question this script supports question/answer databases" echo "inside other projects: if the current directory contains a db/txt" echo "directory, that project's database is used. If a reused question" echo "file lives in another project's db/txt, the model, response and" echo "answer files are written to the analog locations next to it." } if [ "$1" = "--help" ] || [ "$1" = "-h" ]; then print_usage exit 0 fi if [ $# -lt 1 ]; then print_usage exit 1 fi model="$1" shift REUSE_QUESTION=0 QUESTION_FILE="" PROMPT_FILE="" USE_FILE_PROMPT=0 # Database selection, same convention as bin/question: a db/txt directory # in the current working directory makes the current project the database # location; otherwise the gossip database is used. if [ -d "${PWD}/db/txt" ]; then DB_DIR="$(readlink -f "${PWD}/db")" else DB_DIR="$(readlink -f "${PROJECT_DIR}/db")" fi # An existing file argument is resolved before changing to the project # directory, so relative file names work from any location. if [ $# -eq 1 ] && [ -f "$1" ]; then CANON_FILE="$(readlink -f "$1")" base="${CANON_FILE##*/}" # Is this a question_<datetime>.txt file inside a db/txt directory? # It may belong to the gossip project or to any other project. if [[ "$base" =~ ^question_[0-9]{8}_[0-9]{6}\.txt$ ]] \ && [ "$(basename "$(dirname "$CANON_FILE")")" = "txt" ] \ && [ "$(basename "$(dirname "$(dirname "$CANON_FILE")")")" = "db" ] then REUSE_QUESTION=1 DATETIME="${base#question_}" DATETIME="${DATETIME%.txt}" QUESTION_FILE="$CANON_FILE" # The question file's location decides where the model, response and # answer files are written to (the analog locations next to it). DB_DIR="$(dirname "$(dirname "$CANON_FILE")")" echo "Reusing existing question file: $QUESTION_FILE" else # Existing file, but not a question file -> use it as prompt input. USE_FILE_PROMPT=1 PROMPT_FILE="$CANON_FILE" fi fi # Work from the project directory so relative paths in helper scripts work. cd "$PROJECT_DIR" # Generate datetime and save the question content unless we reuse an # existing question file. if [ "$REUSE_QUESTION" -ne 1 ] then DATETIME=$(./bin/datetime) QUESTION_FILE="${DB_DIR}/txt/question_${DATETIME}.txt" if [ "$USE_FILE_PROMPT" -eq 1 ] then cp "$PROMPT_FILE" "$QUESTION_FILE" echo "Saved question to: $QUESTION_FILE (from file $PROMPT_FILE)" elif [ $# -gt 0 ] then # Prompt from command line arguments. printf '%s\n' "$*" > "$QUESTION_FILE" echo "Saved question to: $QUESTION_FILE" else # Create an empty question file and open the editor. This mirrors # bin/question, but into the database selected above. touch "$QUESTION_FILE" echo "Created question file: $QUESTION_FILE" ${EDITOR:-vi} "$QUESTION_FILE" fi fi # Query OpenRouter. DB_DIR tells or_ask.sh where to write # the model, response and answer files. DATETIME="$DATETIME" DB_DIR="$DB_DIR" ./bin/or_ask.sh "$model" # Show model record echo echo "Model record:" cat "${DB_DIR}/csv/model_${DATETIME}.csv" echo # Show answer echo "Answer:" echo cat "${DB_DIR}/txt/answer_${DATETIME}.txt" echo "${DB_DIR}/txt/answer_${DATETIME}.txt" # End of: ask EOT Usage: # Interactive: opens editor for the prompt bin/ask qwen/qwen3.7-plus # One-liner prompt on the command line bin/ask qwen/qwen3.7-plus "What is POSIX?" # With a specific model bin/ask deepseek/deepseek-v3.2 "Explain inference in LLMs." # With existing questions file - overwrites existing answer. bin/ask deepseek/deepseek-v3.2 ./db/txt/questions_20260727_132051.txt The script: 1. Generates a timestamp and creates a question file ('db/txt/question_<datetime>.txt') - either empty (opened in '$EDITOR') or pre-filled with the command-line prompt. 2. Calls 'or_ask.sh' with the model and the same timestamp so all files stay in sync. 3. Prints the model record from 'db/csv/model_<datetime>.csv'. 4. Prints the answer from 'db/txt/answer_<datetime>.txt'. 13. Questions ------------- The bin/questions.rb script prints out the number, short date, short time, and question one liner in chronological order in a table format cat > bin/questions.rb <<EOT #! /usr/bin/env ruby # Do not edit this file, as it gets automatically generated by lp. $: << File.dirname( __FILE__ ) + '/../lib' require 'rlib' require 'stack' require 'term' term = Term.new columns = term.cols # ------------------------------------------------------------------------- # Help / usage # ------------------------------------------------------------------------- def print_usage puts <<~USAGE Usage: #{$0} [--help|-h] Lists all recorded questions in chronological order with their timestamps and the model that answered them. Database location (DB_DIR) -------------------------- The script uses the same database discovery convention as bin/ask and bin/question: * If the current working directory contains a db/txt directory, that project's database is used. This allows any project that uses Gossip as its LLM frontend to have its own isolated question/answer history. * Otherwise, the Gossip project's own database (gossip/db) is used. This means you can run 'bin/questions.rb' from inside any project that has a db/txt subdirectory (created by bin/question or bin/ask) and see only that project's Q&A history. Run it from the Gossip root (or any directory without db/txt) to see the global Gossip history. Output format ------------- Each line shows: <no>. <YY-MM-DD HH:MM> <model> <first line of question> USAGE exit 0 end if ARGV.include?('--help') || ARGV.include?('-h') print_usage end # ------------------------------------------------------------------------- # Determine the database directory (DB_DIR) using the same convention as # bin/ask: if the current working directory contains a db/txt directory, # use that project's database; otherwise use the gossip project's database. # ------------------------------------------------------------------------- if File.directory?(File.join(Dir.pwd, 'db', 'txt')) db_dir = File.realpath(File.join(Dir.pwd, 'db')) else project_dir = File.realpath(File.join(File.dirname(__FILE__), '..')) db_dir = File.join(project_dir, 'db') end question_dir = File.join(db_dir, 'txt') model_dir = File.join(db_dir, 'csv') txt_files = Dir.glob(File.join(question_dir, 'question_*.txt')) txt_files.sort! total = txt_files.length width = total.to_s.length # First pass: collect model names to determine the maximum width. model_names = [] txt_files.each do |txt_filename| basename = File.basename(txt_filename, ".txt") timestamp_str = basename.sub("question_", "") model_name = "" model_csv = File.join(model_dir, "model_#{timestamp_str}.csv") if File.exist?(model_csv) line = Rlib.readfile(model_csv).strip parts = line.split(',') backend = parts[0] model_id = parts[1] || "" if backend == "openrouter" || backend == "pi.dev" # Remove the lab name before the first '/'. model_id = model_id.sub(/^[^\/]+\//, '') # Remove optional ':free' suffix at the end. model_id = model_id.sub(/:free\z/, '') model_name = model_id elsif backend == "llama.cpp" model_name = model_id else model_name = model_id end end # Beautify some model name(s). model_name.sub!( /-550b-a55b$/, '' ) model_name.sub!( /Q8_0$/, 'Q8' ) model_names << model_name end # Determine padding width from the longest model name (minimum 4 for "model"). max_model_len = model_names.map { |n| n.length }.max max_model_len = [max_model_len, 4].max # Second pass: print the table with the model column. txt_files.each_with_index do |txt_filename, index| # Extract timestamp from filename: question_YYYYMMDD_HHMMSS.txt basename = File.basename(txt_filename, ".txt") timestamp_str = basename.sub("question_", "") # Parse YYYYMMDD_HHMMSS if timestamp_str =~ /^(\d{4})(\d{2})(\d{2})_(\d{2})(\d{2})(\d{2})$/ year_short = $1[2,2] # YY month = $2 day = $3 hour = $4 minute = $5 datetime_short = "#{year_short}-#{month}-#{day} #{hour}:#{minute}" else datetime_short = "??-??-?? ??:??" end number = sprintf( "%#{width}d", index + 1 ) # Pad the model name so all rows align. model_name = model_names[index] model_padded = sprintf( "%-#{max_model_len}s", model_name ) question = "#{number}. #{datetime_short} #{model_padded} " content = Rlib.readfile( txt_filename ) push content split lines = pop question << lines.join( " " ) if question.length > columns question = question[ 0...columns ] end term.puts question end exit 0 # End of: questions.rb EOT 14. llama.cpp ------------- In addition to using OpenRouter remotely to access an LLM, we are also able to use llama.cpp locally. To build and install 'llama.cpp' locally (as your regular user, not root) so multiple versions can coexist over time, Gossip provides a script that clones the upstream repository, checks out a pinned release tag, verifies the commit hash, and builds a static 'llama-cli' binary under 'gossip/opt/llama.cpp-<version>/'. A symlink 'gossip/opt/llama.cpp' always points to the active version. cat > bin/make_llama.cpp.sh <<EOT #! /bin/dash # Do not edit this file, it gets automatically generated. set -e if [ "$(basename "$PWD")" != "gossip" ]; then echo "ERROR: we are not inside the project's 'gossip' directory." exit 2 fi test -d opt || mkdir opt test -d build || mkdir build cd build if [ -d llama.cpp ] then cd llama.cpp git checkout master git fetch origin master git merge origin/master else git clone --single-branch --branch master https://github.com/ggerganov/llama.cpp.git cd llama.cpp fi # Older versions are kept here for historical tracability. #VERSION=5097 #COMMIT=fe5b78c89670b2f37ecb216306bed3e677b49d9f #SHA256=6d19a23ea19980cb0ca73e6faaef274ba12423bf7ab83b4665369379914e3dfb VERSION=10075 COMMIT=76f46ad29d61fd8c1401e8221842934bf62a6064 SHA256=dc20e99158cd2c5ba8bc09476ea5378cf448656fbf593f41b2ea3628678f4bf8 TAG=b$VERSION git checkout $TAG # Check the correct commit is checked out. git log | head -n 1 | grep $COMMIT || { echo "ERROR: wrong commit."; exit 1; } # Check the hash of this commit (sha1 is not enough). git cat-file commit $COMMIT | sha256sum | grep $SHA256 \ || { echo "ERROR: from git commit content!"; exit 1; } # And build in parallel. export MAKEFLAGS=-j$(nproc) umask 0022 test -d build && rm -rf build echo cmake cmake --fresh -B build -DBUILD_SHARED_LIBS=OFF -DLLAMA_CURL=OFF echo build cmake --build build --config Release -j $(nproc) #sudo chown -R root:root build #sudo mv -iv build /opt/llama.cpp-$VERSION mv -iv build ../../opt/llama.cpp-$VERSION cd ../../opt test -e llama.cpp && rm llama.cpp ln -s llama.cpp-$VERSION llama.cpp # This is software that needs regular updates, so we leave the git repository in # place for later reuse. # All installed files are in its own directory under gossip/opt. echo "SUCCESS: $0 - $?." EOT 15. Downloading a Minimal GGUF Model ------------------------------------ To test llama.cpp locally without downloading massive files, you can use a minimal model. The Qwen 3.5 0.8B model is only about 0.8 GB and is perfect for testing purposes and as an example of how to use llama.cpp locally. You can download it directly from Hugging Face using `wget`: wget -O db/gguf/Qwen3.5-0.8B-Q8_0.gguf \ https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF/resolve/main/Qwen3.5-0.8B-Q8_0.gguf For further reference see the Links section further down in how to identify interesting open-weight models at Hugging Face. 16. Llama Text Output To Gossip Text Format ------------------------------------------- The default llama-cli output has some escape codes and boilerplate text which should be removed. We also prefer the full answer first and then followed in a enclosed thinking section with reasong steps. This output should be converted to something like this: ''' [The answer...] <the original think tag> [The thinking...] </the original think tag> ''' Here is the script: cat > bin/convert_llama_output.rb <<EOT #! /usr/bin/env ruby # Convert llama.cpp output to Gossip format # Usage: ruby convert_llama_output.rb <input_file> [output_file] require 'json' def convert_llama_output(input_path, output_path = nil) content = File.read(input_path) # Extract thinking section (between [Start thinking] and [End thinking]) thinking_match = content.match(/^\[Start thinking\](.*?)^\[End thinking\]/m) thinking = thinking_match ? thinking_match[1].strip : "" # Clean up thinking: remove trailing checkmark/emoji and whitespace thinking = thinking.sub(/[✅\u2705\u2714\u2713]+$/, '').strip # Extract answer - it's after [End thinking] and before the performance stats line # The answer starts after [End thinking] and ends before "[ Prompt:" or "Exiting..." after_thinking = content.split(/^\[End thinking\]/m).last || "" # Remove the performance line and "Exiting..." answer = after_thinking .sub(/^\[ Prompt:.*?\]\s*/m, '') .sub(/^Exiting\.\.\.\s*$/m, '') .strip # Build output in Gossip format output = "#{answer}\n\n<think>\n#{thinking}\n</think>\n" if output_path File.write(output_path, output) puts "Written to #{output_path}" else puts output end end if __FILE__ == $0 if ARGV.empty? || ARGV.include?('-h') || ARGV.include?('--help') puts "Usage: #{$0} <input_file> [output_file]" puts "Converts llama.cpp CLI output to Gossip format with <think>/</think> markers." exit 0 end input_file = ARGV[0] output_file = ARGV[1] unless File.exist?(input_file) warn "ERROR: Input file not found: #{input_file}" exit 1 end convert_llama_output(input_file, output_file) end EOT 17. Invoking llama.cpp Locally ------------------------------ 'bin/llama' is a thin wrapper around the local 'llama-cli' binary. It records the prompt and answer in 'gossip/db/txt', runs the model in a single, non-interactive turn, and reformats the raw output (e.g. from Qwen 3.6) into cleaned up text. 17.1. Prerequisites ''''''''''''''''''' * 'llama.cpp' has been built by 'bin/make_llama.cpp.sh'. This leaves a working tree at 'gossip/opt/llama.cpp' and the 'llama-cli' binary at 'gossip/opt/llama.cpp/bin/llama-cli'. * A model file in GGUF format is available. 17.2. Usage ''''''''''' /path/to/gossip/bin/llama <model_file> [prompt text ...] /path/to/gossip/bin/llama <model_file> <prompt_file> /path/to/gossip/bin/llama <model_file> db/txt/question_YYYYMMDD_HHMMSS.txt echo "prompt" | /path/to/gossip/bin/llama <model_file> The script locates the Gossip project directory from its own path, so it may be called from any working directory. 17.3. Examples '''''''''''''' Ask a new question: /path/to/gossip/bin/llama \ /path/to/gossip/db/gguf/llama-3.2-3b-instruct-q4_k_m.gguf \ "What is POSIX?" Re-ask a previously recorded question with a different model: /path/to/gossip/bin/llama \ /path/to/gossip/db/gguf/Qwen3.6-35B-A3B-UD-Q8_K_XL.gguf \ /path/to/gossip/db/txt/question_20260705_123456.txt Use any file for the prompt input: /path/to/gossip/bin/llama \ /path/to/gossip/db/gguf/Qwen3.6-35B-A3B-UD-Q8_K_XL.gguf \ /my/notes/big_question.txt A copy of the input text will then be created in the archive location, e.g. as: /path/to/gossip/db/txt/question_20260705_123456.txt 17.4. Behaviour ''''''''''''''' * Model path. The first argument is the model file. If it cannot be found relative to the current directory, the script also tries to resolve it relative to the project directory. * Prompt source. The prompt can be supplied in three ways: * As trailing arguments after the model path. They are joined with a single space. * As a single existing file name. The file is passed to 'llama-cli' with '-f'. This is useful for long prompts. * Via standard input if no prompt arguments are given. * Reusing a recorded question. If the single file argument is a canonical 'db/txt/question_YYYYMMDD_HHMMSS.txt' file inside the project, the script reuses that question file and writes the answer to the matching 'db/txt/answer_YYYYMMDD_HHMMSS.txt'. This lets a previously stored question be re-asked to a different model without creating a new question file. * Storing results. For a new prompt, a timestamp is generated with 'bin/datetime'. The prompt is saved as 'db/txt/question_<datetime>.txt' and the raw model output as 'db/txt/answer_<datetime>.txt'. The information about the model as well as the llama-cli version that has been used is also conserved in the db/csv/model_<datetime>.csv file. * Reformatting: the raw answer is converted to the Gossip format (answer first, then any reasoning inside the original think tags) and the '.txt' file is replaced by the converted text. * Non-interactive invocation. 'llama-cli' is invoked with '--single-turn', '--simple-io', '--no-display-prompt', '--no-escape', '--log-disable', and '--temp 0.0', so it returns a single response and exits. 17.5. llama Script '''''''''''''''''' And here is the script: cat > bin/llama <<EOT #! /bin/bash # Do not edit this file, it gets automatically generated. set -euo pipefail # Determine the project directory (parent of the directory containing this script) SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)" PROJECT_DIR="$(dirname "$SCRIPT_DIR")" #echo "SCRIPT_DIR=${SCRIPT_DIR}" #echo "PROJECT_DIR=${PROJECT_DIR}" # Use DB_DIR from caller if provided (e.g. from bin/ask), # otherwise default to the project's own database. if [ -z "${DB_DIR:-}" ]; then DB_DIR="$PROJECT_DIR/db" fi mkdir -p "$DB_DIR/csv" mkdir -p "$DB_DIR/txt" # Path to llama-cli binary LLAMA_CLI="$PROJECT_DIR/opt/llama.cpp/bin/llama-cli" if [ ! -x "$LLAMA_CLI" ]; then echo "ERROR: llama-cli not found or not executable at $LLAMA_CLI" >&2 echo "Run $PROJECT_DIR/bin/make_llama.cpp.sh to build it first." >&2 exit 1 fi # Model handling: first parameter is the model file, remaining parameters form the prompt if [ $# -lt 1 ]; then echo "ERROR: no model file provided as first argument." >&2 echo "Usage: $0 <model_file> [prompt text... | prompt_file]" >&2 exit 1 fi MODEL="$1" shift # If MODEL is not found relative to CWD, try relative to PROJECT_DIR if [ ! -f "$MODEL" ] && [ -f "$PROJECT_DIR/$MODEL" ]; then MODEL="$PROJECT_DIR/$MODEL" fi USE_FILE_PROMPT=0 REUSE_QUESTION=0 PROMPT_FILE="" PROMPT="" DATETIME="" QUESTION_FILE="" RESPONSE_FILE="" ANSWER_FILE="" # Canonical path to the question directory (respects DB_DIR). QUESTION_DIR="$(readlink -f "$DB_DIR/txt")" # Check if exactly one argument remains and it is an existing file if [ $# -eq 1 ] && [ -f "$1" ]; then CANON_FILE="$(readlink -f "$1")" # Is this file already one of our question_<datetime>.txt files? if [[ "$CANON_FILE" == "$QUESTION_DIR"/question_*.txt ]]; then base="${CANON_FILE##*/}" if [[ "$base" =~ ^question_[0-9]{8}_[0-9]{6}\.txt$ ]]; then REUSE_QUESTION=1 DATETIME="${base#question_}" DATETIME="${DATETIME%.txt}" QUESTION_FILE="$CANON_FILE" RESPONSE_FILE="$DB_DIR/txt/response_${DATETIME}.txt" ANSWER_FILE="$DB_DIR/txt/answer_${DATETIME}.txt" PROMPT_FILE="$CANON_FILE" USE_FILE_PROMPT=1 echo "Reusing existing question file: $QUESTION_FILE" fi fi # Existing file, but not a known question file -> use it as prompt input if [ "$REUSE_QUESTION" -ne 1 ]; then USE_FILE_PROMPT=1 PROMPT_FILE="$1" fi elif [ $# -gt 0 ]; then # All remaining parameters concatenated with a space creates the total prompt PROMPT="$*" else # If no additional parameters, read prompt from stdin PROMPT=$(cat) fi # If we are not reusing an existing question file, generate a new timestamp and save the prompt if [ "$REUSE_QUESTION" -ne 1 ]; then if [ -z "$DATETIME" ]; then DATETIME=$("$PROJECT_DIR/bin/datetime") fi if [ -z "$DATETIME" ]; then echo "ERROR: failed to generate datetime" >&2 exit 1 fi QUESTION_FILE="$DB_DIR/txt/question_${DATETIME}.txt" RESPONSE_FILE="$DB_DIR/txt/response_${DATETIME}.txt" ANSWER_FILE="$DB_DIR/txt/answer_${DATETIME}.txt" if [ "$USE_FILE_PROMPT" -eq 1 ]; then cat "$PROMPT_FILE" > "$QUESTION_FILE" echo "Saved question to: $QUESTION_FILE (from file $PROMPT_FILE)" else if [ -z "$PROMPT" ]; then echo "ERROR: no prompt provided" >&2 exit 2 fi printf '%s\n' "$PROMPT" > "$QUESTION_FILE" echo "Saved question to: $QUESTION_FILE" fi fi # Verify the model file exists if [ ! -f "$MODEL" ]; then echo "ERROR: model not found at $MODEL" >&2 exit 2 fi # Determine llama.cpp version from the opt/llama.cpp symlink target. # The target is expected to look like llama.cpp-<VERSION>, e.g. llama.cpp-10075. LLAMA_CPP_VERSION="unknown" if [ -L "$PROJECT_DIR/opt/llama.cpp" ] then LLAMA_CPP_TARGET=$(readlink "$PROJECT_DIR/opt/llama.cpp" || true) LLAMA_CPP_BASENAME=$(basename "$LLAMA_CPP_TARGET") LLAMA_CPP_VERSION=${LLAMA_CPP_BASENAME#llama.cpp-} if [ "$LLAMA_CPP_VERSION" = "$LLAMA_CPP_BASENAME" ] then LLAMA_CPP_VERSION="unknown" fi fi MODEL_BASENAME=$(basename "$MODEL") # Invoke llama-cli and save output to answer file # using options to surpress an interactive loop. echo "Invoking llama-cli..." if [ "$USE_FILE_PROMPT" -eq 1 ]; then "$LLAMA_CLI" \ -m "$MODEL" \ -f "$PROMPT_FILE" \ --temp 0.0 \ --no-display-prompt \ --log-disable \ --simple-io \ --single-turn \ --no-escape \ --seed 1 \ 2>&1 | tee "$RESPONSE_FILE" else "$LLAMA_CLI" \ -m "$MODEL" \ -p "$PROMPT" \ --temp 0.0 \ --no-display-prompt \ --log-disable \ --simple-io \ --single-turn \ --no-escape \ --seed 1 \ 2>&1 | tee "$RESPONSE_FILE" fi echo "Saved answer to: $ANSWER_FILE" # Write the model metadata CSV for this question/answer pair. CSV_FILE="$DB_DIR/csv/model_${DATETIME}.csv" printf 'llama.cpp,%s,version=%s,seed=1\n' "$MODEL_BASENAME" "$LLAMA_CPP_VERSION" > "$CSV_FILE" echo "Saved model info to: $CSV_FILE" # Optionally convert the output to Gossip format #if command -v ruby >/dev/null 2>&1 && [ -f "$PROJECT_DIR/bin/convert_llama_output.rb" ]; then #CONVERTED_FILE="${ANSWER_FILE}.tmp" ruby "$PROJECT_DIR/bin/convert_llama_output.rb" "$RESPONSE_FILE" "$ANSWER_FILE" #mv "$CONVERTED_FILE" "$ANSWER_FILE" # The llama-cli text output and stripped markdown files are too close to be # worth being kept both, so we remove the initial response file. rm "$RESPONSE_FILE" echo "Converted output to gossip format: ${ANSWER_FILE}" #fi EOT 18. Local llama.cpp Server -------------------------- The bin/llama script invokes llama.cpp in batch mode. That means, the large language model has to be loaded into memory for each request. Standard is to access the large language model via an inference server. cat > ./bin/start_server.sh <<EOT #! /bin/dash # Do not edit this file, as it gets automatically generated by lp. # Used e.g. for the Qwen 3.6 35B and 27B models. # 2026-07-24 set -e echo "START: $0 - $$." LAYERS=99 echo "HOST=$(hostname)" echo "pwd=${PWD}" export MODELFILENAME="$1" echo "MODEL=${MODELFILENAME}" BASEMODEL=${MODELFILENAME##*/} echo "${BASEMODEL}" > db/csv/llama-server_model.csv THREADS=$(nproc) echo "THREADS=${THREADS}" LOGFILENAME=log/gossip.log echo "logfile=${LOGFILENAME}" server="opt/llama.cpp/bin/llama-server" echo \ $server -v --temp 0.0 -s 1 -t $THREADS -ngl $LAYERS -cram 0 -m $MODELFILENAME $server -v --temp 0.0 -s 1 -t $THREADS -ngl $LAYERS -cram 0 -m $MODELFILENAME > $LOGFILENAME 2>&1 # End of: start_server.sh EOT This script must be started from the Gossip project directory. 19. The llama.cpp Client ------------------------ We assume the llama.cpp server has been started as it waiting at http://localhost:8080/ for requests. The script to send requests and receive responses is what bin/ask_curl.rb is about. cat > ./bin/ask_curl.rb <<EOT #!/usr/bin/env ruby # Do not edit this file, as it gets automatically generated by lp. # This script encapsulates the JSON payload creation, 'curl' invocation, # response parsing, and answer file writing. $: << File.dirname(__FILE__) + '/../lib' require 'rlib' require 'json' # Usage: ask_curl.rb <prompt_file> <seed> <temp> <repeat_penalty> <output_file> prompt_file = ARGV[0] seed = ARGV[1] temp = ARGV[2] repeat_penalty = ARGV[3] output_file = ARGV[4] json_request_file = prompt_file.sub( /\/txt\//, '/json/' ).sub( /\/question_/, '/request_' ).sub( /\.txt$/, '.json' ) json_response_file = json_request_file.sub( /\/request_/, '/response_' ) # Read the prompt (question + formatting) prompt = Rlib.readfile_assert(prompt_file).chomp # Prepare JSON payload data_json = { "prompt" => prompt, "seed" => seed.to_i, "temperature" => temp.to_f, "repeat_penalty" => repeat_penalty.to_f }.to_json Rlib.writefile_assert(json_request_file, data_json) # Execute curl command = "curl --silent --request POST --url 'http://127.0.0.1:8080/completion' --header 'Content-Type: application/json' -d @#{json_request_file}" puts command json_output = Rlib.output_assert(command) #json_filename = prompt_file.gsub( /csv/, "json" ) #json_filename.sub!( /question_/, 'reply_' ) Rlib.writefile_assert( json_response_file, json_output ) # Parse and display response h = JSON.parse(json_output) puts "=====" h.keys.each do |key| next if key == "content" puts puts "key=#{key}" puts h[key] puts "===" end puts puts "=====" answer = h["content"] puts answer puts "=====" # Save the answer Rlib.writefile_assert(output_file, answer) puts "File written: #{output_file}" puts "SUCCESS: #{__FILE__} - 0." # End of: ask_curl.rb EOT 20. llama_call -------------- bin/llama_call submits prompts to a running local llama.cpp server (started separately via 'bin/start_server.sh') and archives the resulting question/answer pair in the Gossip database, mirroring the interface and storage conventions of 'bin/llama'. The model basename is read from 'db/csv/llama-server_model.csv', which 'bin/start_server.sh' writes when the server is started. Prompts are accepted the same way as 'bin/llama': trailing arguments joined by a space, a single file argument, a canonical 'question_<datetime>.txt' file for reuse, piped stdin, or an interactive editor when no arguments are given. cat > ./bin/llama_call <<EOT #! /bin/bash # Do not edit this file, as it gets automatically generated. set -euo pipefail # Determine the project directory (parent of the directory containing this script) SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)" PROJECT_DIR="$(dirname "$SCRIPT_DIR")" print_usage() { echo "Usage: $0 [prompt text...]" echo " $0 <prompt_file>" echo " $0 db/txt/question_YYYYMMDD_HHMMSS.txt" echo " echo \"prompt\" | $0" echo " $0 # Opens interactive editor" echo "" echo "Submits a prompt to a running local llama.cpp server and archives the" echo "question/answer pair in the Gossip database." echo "" echo "Modes:" echo " - Arguments: joined with spaces as the prompt" echo " - File arg: prompt read from file" echo " - Question file: reuses existing Q/A pair (matches timestamp)" echo " - No args: opens \$EDITOR interactively" echo " - Piped stdin: reads prompt from pipe" echo "" echo "The model is determined from db/csv/llama-server_model.csv, which is" echo "written by bin/start_server.sh when the server is started." echo "" echo "Like bin/ask this script supports question/answer databases inside" echo "other projects: if the current directory contains a db/txt directory," echo "that project's database is used. If a reused question file lives in" echo "another project's db/txt, the answer and response files are written" echo "to the analog locations next to it." } # Safely check for help flag (avoids unbound variable with set -u) if [ "${1:-}" = "--help" ] || [ "${1:-}" = "-h" ]; then print_usage exit 0 fi # Use DB_DIR from caller if provided (e.g. from bin/ask), # otherwise default to the project's own database. if [ -z "${DB_DIR:-}" ]; then DB_DIR="$PROJECT_DIR/db" fi mkdir -p "$DB_DIR/csv" mkdir -p "$DB_DIR/txt" mkdir -p "$DB_DIR/json" # Verify that the llama-server is reachable. #if ! curl --silent --fail http://127.0.0.1:8080/health > /dev/null 2>&1; then # if ! curl --silent --fail http://127.0.0.1:8080/ > /dev/null 2>&1; then # echo "ERROR: llama-server does not appear to be running at http://127.0.0.1:8080/" >&2 # echo "Start it first with: $PROJECT_DIR/bin/start_server.sh <model.gguf>" >&2 # exit 1 # fi #fi # Server model is always specified by the Gossip configuration. SERVER_MODEL_FILE="${PROJECT_DIR}/db/csv/llama-server_model.csv" MODEL_BASENAME="" if [ -f "$SERVER_MODEL_FILE" ]; then MODEL_BASENAME="$(tr -d '\n' < "$SERVER_MODEL_FILE")" echo "Using model from $SERVER_MODEL_FILE: $MODEL_BASENAME" else echo "ERROR: no model basename provided and $SERVER_MODEL_FILE not found." >&2 exit 1 fi # Initialize flags USE_FILE_PROMPT=0 REUSE_QUESTION=0 PROMPT_FILE="" PROMPT="" DATETIME="" QUESTION_FILE="" ANSWER_FILE="" JSON_FILE="" # Canonical path to the question directory (respects DB_DIR). QUESTION_DIR="$(readlink -f "$DB_DIR/txt")" # 1. Check if a single argument is a known question file if [ $# -eq 1 ] && [ -f "$1" ]; then CANON_FILE="$(readlink -f "$1")" if [[ "$CANON_FILE" == "$QUESTION_DIR"/question_*.txt ]]; then base="${CANON_FILE##*/}" if [[ "$base" =~ ^question_[0-9]{8}_[0-9]{6}\.txt$ ]]; then REUSE_QUESTION=1 DATETIME="${base#question_}" DATETIME="${DATETIME%.txt}" QUESTION_FILE="$CANON_FILE" ANSWER_FILE="$DB_DIR/txt/answer_${DATETIME}.txt" JSON_FILE="$DB_DIR/json/response_${DATETIME}.json" PROMPT_FILE="$CANON_FILE" USE_FILE_PROMPT=1 echo "Reusing existing question file: $QUESTION_FILE" fi fi fi # 2. Handle piped/redirected stdin (non-terminal) if [ "$REUSE_QUESTION" -ne 1 ] && [ ! -t 0 ]; then PROMPT=$(cat) if [ -z "$PROMPT" ]; then echo "ERROR: empty prompt from stdin" >&2 exit 2 fi # Falls through to new prompt handling below # 3. Handle no arguments -> interactive editor elif [ "$REUSE_QUESTION" -ne 1 ] && [ $# -eq 0 ]; then DATETIME=$("$PROJECT_DIR/bin/datetime") QUESTION_FILE="$DB_DIR/txt/question_${DATETIME}.txt" ANSWER_FILE="$DB_DIR/txt/answer_${DATETIME}.txt" JSON_FILE="$DB_DIR/json/response_${DATETIME}.json" touch "$QUESTION_FILE" echo "Created question file: $QUESTION_FILE" ${EDITOR:-vi} "$QUESTION_FILE" USE_FILE_PROMPT=1 PROMPT_FILE="$QUESTION_FILE" REUSE_QUESTION=1 # 4. Handle regular file argument (not a question file) elif [ "$REUSE_QUESTION" -ne 1 ] && [ $# -eq 1 ] && [ -f "$1" ]; then USE_FILE_PROMPT=1 PROMPT_FILE="$1" # 5. Handle command-line arguments elif [ "$REUSE_QUESTION" -ne 1 ] && [ $# -gt 0 ]; then PROMPT="$*" fi # Generate new prompt/question if not reusing one if [ "$REUSE_QUESTION" -ne 1 ]; then if [ -z "$DATETIME" ]; then DATETIME=$("$PROJECT_DIR/bin/datetime") fi if [ -z "$DATETIME" ]; then echo "ERROR: failed to generate datetime" >&2 exit 1 fi QUESTION_FILE="$DB_DIR/txt/question_${DATETIME}.txt" ANSWER_FILE="$DB_DIR/txt/answer_${DATETIME}.txt" JSON_FILE="$DB_DIR/json/response_${DATETIME}.json" if [ "$USE_FILE_PROMPT" -eq 1 ]; then cat "$PROMPT_FILE" > "$QUESTION_FILE" echo "Saved question to: $QUESTION_FILE (from file $PROMPT_FILE)" else if [ -z "$PROMPT" ]; then echo "ERROR: no prompt provided" >&2 exit 2 fi printf '%s\n' "$PROMPT" > "$QUESTION_FILE" echo "Saved question to: $QUESTION_FILE" fi fi # Determine llama.cpp version from the opt/llama.cpp symlink target. LLAMA_CPP_VERSION="unknown" if [ -L "$PROJECT_DIR/opt/llama.cpp" ]; then LLAMA_CPP_TARGET=$(readlink "$PROJECT_DIR/opt/llama.cpp" || true) LLAMA_CPP_BASENAME=$(basename "$LLAMA_CPP_TARGET") LLAMA_CPP_VERSION=${LLAMA_CPP_BASENAME#llama.cpp-} if [ "$LLAMA_CPP_VERSION" = "$LLAMA_CPP_BASENAME" ]; then LLAMA_CPP_VERSION="unknown" fi fi # Invoke ask_curl.rb to talk to the server. echo "Invoking llama-server via ask_curl.rb..." ruby "$PROJECT_DIR/bin/ask_curl.rb" \ "$QUESTION_FILE" \ 1 \ 0.0 \ 1.0 \ "$ANSWER_FILE" # Write the model metadata CSV for this question/answer pair. CSV_FILE="$DB_DIR/csv/model_${DATETIME}.csv" printf 'llama.cpp,%s,version=%s,seed=1\n' "$MODEL_BASENAME" "$LLAMA_CPP_VERSION" > "$CSV_FILE" echo "Saved model info to: $CSV_FILE" echo "Saved answer to: $ANSWER_FILE" echo "Saved response JSON to: $JSON_FILE" echo "SUCCESS: $0 - $?." # End of: llama_call EOT 21. Understanding Pi Dev ------------------------ Pi takes the directory the the user is when he or she invokes the pi command to be the project directory. As the user activity get logged per session, different sessions for different projects accumulate under ~/.pi/agent/sessions. In that directory the initial absolute user location is encrypted by replacing in the absolute path '/' with '-' and adding '-' to the front. The pi.dev application directory is, in my case, located at: cd .local/share/pi-node/node-*/lib/node_modules/@earendil-works/pi-coding-agent 22. Multi Turn pi.dev Interactions ---------------------------------- Example session lp investigation 2026-07-20: projects (locations in ~/.pi/agent/sessions/) sessions session / jsonl events event type session thinking_level_change model_change message role user assistant text thinking toolCall toolResult messages user assistant toolcall toolcall toolresult toolresult assistant toolcall toolcall toolcall toolcall toolresult toolresult toolresult toolresult assistant toolcall toolcall toolcall toolresult toolresult toolresult assistant toolcall toolcall toolresult toolresult assistant answer toolcall toolresult toolcall toolresult assistant answer toolcall toolcall toolresult toolresult assistant answer 23. Interaction With Pi Dev --------------------------- Let's read for what has happened in the pi.dev environment so far. The object hierarchy structure looks like: session (directory) jsonl (file) event (hash) session (attributes: timestamp, dir) message The data structure is jsonl files, one for each session. In each file there are messages, each message has a type. cat > bin/pi_sessions.rb <<EOT #! /usr/bin/env ruby # Do not edit this file, as it gets automatically generated by lp. $: << File.dirname( __FILE__ ) + '/../lib' require 'json' require 'rlib' require 'stack' require 'term' def read_jsonl( jsonl_filename ) ret_val = [] File.foreach( jsonl_filename ) do |line| line = line.strip next if line.empty? # skip blank lines begin record = JSON.parse(line) ret_val << record rescue JSON::ParserError => e warn "Bad line: #{e.message}" exit 2 end end return ret_val end require 'time' def get_datetime( ts_ms ) utc_time = Time.at(ts_ms / 1000r).utc ret_val = utc_time.iso8601(3) ret_val.gsub!( /[\-:]/, '' ) ret_val.sub!( /T/, '_' ) ret_val.sub!( /\..*$/, '' ) return ret_val end dir = "#{ENV[ 'HOME' ]}/.pi/agent/sessions/" if !File.exist?( dir ) puts "Directory not found: #{dir}" exit 1 end script_dir = File.dirname( File.expand_path( __FILE__ ) ) project_dir = File.dirname( script_dir ) command = "find #{dir} -type f -name '*.jsonl' | sort -n" output = Rlib.output( command ) jsonl_files = output.split( /\n/ ) event_types = {} thinking_level = nil router = nil model = nil new_question_answer = false answer = "" think = "" jsonl_files.each do |jsonl_filename| puts "jsonl_filename=#{jsonl_filename}" events = read_jsonl( jsonl_filename ) puts "events=#{events.length}" datetime = "" events.each do |event| puts "event_type=#{event[ "type" ]}" # Event fields. #puts event.inspect #exit 3 event.keys.each do |key| printf "%s, ", key end puts if event[ "type" ] == "message" message = event[ "message" ] timestamp = message[ "timestamp" ] #temp_datetime( timestamp ) puts "message=#{message.inspect}" content = message[ "content" ] if message[ "role" ] == "user" Rlib.assert( content.length == 1, \ "ERROR: content assumption is wrong: " \ + "#{content.length}" ) content = content[ 0 ] #puts content.inspect timestamp = message[ "timestamp" ] #puts timestamp datetime = get_datetime( timestamp ) #puts datetime txt_filename = "#{project_dir}/db/txt/question_#{datetime}.txt" #exit 3 puts "\nQuestion #{datetime}:\n\n" question = content[ content[ "type" ] ] puts question puts Rlib.assert( router != nil ) Rlib.assert( model != nil ) model_content = "pi.dev,#{model}" \ + ",reasoning=#{thinking_level}\n" model_filename = txt_filename.sub( /question/, "model" ).sub( /\/txt\//, '/csv/' ).sub( /\.txt$/, '.csv' ) if File.exist?( txt_filename ) == false Rlib.writefile( txt_filename, question ) puts "Wrote: #{txt_filename}" new_question_answer = true else new_question_answer = false end if File.exist?( model_filename ) == false Rlib.writefile( model_filename, model_content ) puts "Wrote: #{model_filename}" end answer = "" think = "" elsif message[ "role" ] == "assistant" puts "assistant.content=#{content.inspect}" content.each do |next_content| if next_content[ "type" ] == "text" local_answer = next_content[ "text" ] # If not likely will just invoke another tool and there is nothing to say. if answer != " " #puts "content=" #puts next_content.inspect puts "\nAnswer #{datetime}:\n\n" if answer > "" answer << "\n\n" end answer << local_answer puts answer puts txt_filename = "#{project_dir}/db/txt/answer_#{datetime}.txt" if new_question_answer answer_think = String.new( answer ) if think > "" answer_think << "\n\n<think>\n#{think}\n</think>\n" end Rlib.writefile( txt_filename, answer_think ) puts "Wrote: #{txt_filename}" end end elsif next_content[ "type" ] == "toolCall" #puts next_content.inspect puts "\nToolcall:\n\n" local_tool = next_content[ "name" ] if local_tool == "read" print "less '" print next_content[ "arguments" ][ "path" ] puts "'" think << "\n\ncat #{next_content[ 'arguments' ][ 'path' ]}" else print local_tool print " " puts next_content[ "arguments" ][ "command" ] think << "\n\n#{local_tool} #{next_content[ 'arguments' ][ 'command' ]}" end puts elsif next_content[ "type" ] == "thinking" puts "assistant.thinking" puts next_content.inspect local_think = next_content[ "thinking" ] if think > "" think << "\n\n" end think << local_think puts "<think>\n#{think}\n</think>\n" else # Everything else is a surprise. Rlib.assert( false ) end end elsif message[ "role" ] == "toolResult" # They can be quite long and we skip them for now. puts "Skipping tool result." else #puts content.inspect puts message[ "role" ].inspect puts "what?" exit 3 puts content[ 0 ][ "type" ].inspect puts content[ content[ "type" ] ] end elsif event[ "type" ] == "session" dir = event[ "cwd" ] elsif event[ "type" ] == "thinking_level_change" thinking_level = event[ "thinkingLevel" ] elsif event[ "type" ] == "model_change" router = event[ "provider" ] model = event[ "modelId" ] else puts "ERROR: we got something else than a session or a thinking level" \ " change or message: #{event.inspect}" exit 2 end event_types[ event[ "type" ] ] = \ event_types[ event[ "type" ] ].to_i + 1 end puts "#------------------------------" end puts event_types.inspect exit 0 EOT 24. pi.dev Models Storage ------------------------- Location: pi stores all available models in two JSON files at ~/.pi/agent/ (or at $PI_CODING_AGENT_DIR/agent/ if that environment variable is set). Files: 1. models.json - Missing by default - the path is: ~/.pi/agent/models.json - It contains user configured provider settings and model definitions/overrides. 2. models-store.json - Cached overlay of dynamically refreshed model catalogs fetched from pi.dev. - Format: JSON, keyed by provider ID, each entry containing a "models" array of model objects with id, name, provider, cost, contextWindow, etc. - Default path: ~/.pi/agent/models-store.json - This file is populated by polling pi.dev at https://pi.dev/api/models/providers/{providerId} every 4 hours (REMOTE_CATALOG_REFRESH_INTERVAL_MS). 3. Built-in model source (shipped with the @earendil-works/pi-ai npm package): - Per-provider JSON files in the providers/data/ directory (e.g., amazon-bedrock.json, anthropic.json, openai.json, etc.). - A .manifest.json file listing all provider JSON files with content hashes and a generation timestamp (schemaVersion: 3). - These are bundled into models.generated.js at build time. Format: All model data is stored as JSON. 25. pi.dev OpenRouter Models ---------------------------- What follows is the Ruby script 'bin/pi_models_openrouter.rb' prints out all available OpenRouter LLM models. 1. Reads '~/.pi/agent/models-store.json' 2. Extracts the "openrouter" key which contains an array of models 3. Filters models where the "input" array contains "text" 4. Sorts the remaining models alphabetically by "id" 5. Prints the numbered models with all info: id, input, cost (input, output, cacheRead, cacheWrite), contextWindow, maxTokens Here is an example of The JSON structure for models-store.json: ''' { "openrouter": [ { "id": "model-id", "input": ["text", "image"], "cost": { "input": 0.01, "output": 0.02, "cacheRead": 0.001, "cacheWrite": 0.002 }, "contextWindow": 128000, "maxTokens": 4096 } ] } ''' cat > bin/pi_models_openrouter.rb <<EOT #! /usr/bin/env ruby # Do not edit this file, as it gets automatically generated by lp. feature_free_only = false if [ "free", "--free" ].include?( ARGV[ 0 ] ) feature_free_only = true end $: << File.dirname( __FILE__ ) + '/../lib' require 'json' require 'rlib' require 'term' models_store_filename = "#{ENV[ 'HOME' ]}/.pi/agent/models-store.json" if !File.exist?( models_store_filename ) puts "File not found: #{models_store_filename}" exit 1 end content = Rlib.readfile( models_store_filename ) models_store = JSON.parse( content ) openrouter = models_store[ "openrouter" ] if openrouter.nil? puts "No 'openrouter' key found in #{models_store_filename}" exit 2 end models = openrouter[ "models" ] #puts "models=#{models.length}" # Filter models that support text input text_models = models.each do |model| model[ "input" ].include?( "text" ) end #puts "text_models=#{text_models.length}" # Sort alphabetically by id text_models.sort_by! { |model| model[ "id" ] } count = 0 text_models.each_with_index do |model, index| cost = model[ 'cost' ] cost_average = (cost[ 'input' ].to_f + cost[ 'output' ].to_f) / 2.0 if (cost_average != 0.0) && feature_free_only next end count += 1 number = count puts "#{number}. #{model[ 'id' ]}" #puts " input: #{model[ 'input' ].join( ', ' )}" #puts " cost:" #puts " input: #{cost[ 'input' ]}" #puts " output: #{cost[ 'output' ]}" cost_average = ((cost[ 'input' ].to_f + cost[ 'output' ].to_f) / 2.0).to_s.sub( /\.0$/, '' ) puts " cost =#{cost[ 'input' ]}/#{cost[ 'output' ]}/#{cost_average}" #puts " cacheRead: #{cost[ 'cacheRead' ]}" #puts " cacheWrite: #{cost[ 'cacheWrite' ]}" puts " context=#{model[ 'contextWindow' ]}" puts " output =#{model[ 'maxTokens' ]}" puts end exit 0 # End of: pi_models_openrouter.rb EOT 26. Configured OpenRouter Models -------------------------------- Here is a Ruby script that prints beautified all in db/csv/openrouter_models.csv configured models: cat > bin/openrouter_models.rb <<EOT #!/usr/bin/env ruby # coding: utf-8 # Do not edit this file, as it gets automatically generated by lp. $: << File.dirname( __FILE__ ) + '/../lib' require 'rlib' require 'term' # Resolve the project root directory from this script's location, # so the script works regardless of the current working directory. script_dir = File.dirname( File.expand_path( __FILE__ ) ) project_dir = File.dirname( script_dir ) term = Term.new columns = term.cols csv_filename = File.join( project_dir, "db/csv/openrouter_models.csv" ) unless File.exist?( csv_filename ) puts "File not found: #{csv_filename}" exit 1 end models = [] File.foreach( csv_filename ) do |line| line = line.strip next if line.empty? next if line.start_with?( '#' ) parts = line.split( ',' ) next if parts.length < 2 model_id = parts[ 0 ].strip provider = parts[ 1 ].strip next if model_id.empty? # Split model_id into lab and model name at the first '/'. if model_id.include?( '/' ) lab, model_name = model_id.split( '/', 2 ) else lab = "" model_name = model_id end models << [ lab, model_name, provider ] end if models.empty? puts "No models configured in #{csv_filename}" exit 0 end # Column widths max_lab_len = [ models.map { |m| m[ 0 ].length }.max, 3 ].max max_model_len = [ models.map { |m| m[ 1 ].length }.max, 4 ].max max_provider_len = [ models.map { |m| m[ 2 ].length }.max, 8 ].max # Print header puts " Lab".ljust( max_lab_len ) + " " + " Model".ljust( max_model_len ) + " " + " Provider" puts " " + "-" * max_lab_len + " " + "-" * max_model_len + " " + "-" * max_provider_len # Print rows models.each_with_index do |( lab, model_name, provider ), index| number = sprintf( "%#{models.length.to_s.length}d", index + 1 ) line = "#{number}. #{lab.ljust( max_lab_len )} #{model_name.ljust( max_model_len )} #{provider}" line = line[ 0...columns ] if line.length > columns term.puts line end #puts #puts "#{models.length} model(s) configured." exit 0 # End of: openrouter_models.rb EOT How it works: 1. Reads 'db/csv/openrouter_models.csv' line by line. 2. Skips empty lines and comment lines (starting with '#'). 3. Splits each line on the first comma to extract 'model_id' and 'provider'. 4. Splits 'model_id' at the first '/' into 'lab' (the upstream provider, e.g. 'deepseek') and 'model_name' (the specific model, e.g. 'deepseek-v4-pro'). 5. Calculates column widths based on the longest entries for aligned output. 6. Prints a numbered, aligned table with a header and separator line showing Lab, Model, and Provider columns. 7. Truncates lines that exceed the terminal width (using 'term.cols'). 27. Main Gossip Script ---------------------- The gossip.rb script is the main entrance for the user to access different functionality from the command line. cat > bin/gossip.rb <<EOT #! /usr/bin/env ruby # Do not edit this file, as it gets automically generated by lp. $: << File.dirname( __FILE__ ) + '/../lib' require 'json' require 'rlib' require 'stack' require 'term' if ARGV.include?( "help" ) || ARGV.include?( "--help" ) puts "Usage: gossip [COMMAND]" puts puts "Gossip is an LLM frontend for the command line." puts puts "Commands:" puts " ask <modelid> [prompt...] Ask an OpenRouter model a question." puts " questions List local questions stored under db/txt." print " question [datetime] Create a new empty question file under db/txt" puts " and print its name." puts " models [free|configured] List OpenRouter models." puts " pisessions Read and process pi.dev sessions from: " + "~/.pi/agent/sessions/" puts " help, --help Display this help message and exit." puts " version, --version Print out the application name and version." exit 0 elsif ARGV.include?( "version" ) || ARGV.include?( "--version" ) command = "git log | grep -E '^commit ' | wc -l" output = Rlib.output( command ) if output puts "gossip #{output}" else puts "ERROR: git is not available." exit 1 end exit 0 end #feature = "pisessions" feature = "questions" if ARGV[ 0 ] == "pisessions" feature = "pisessions" elsif ARGV[ 0 ] == "question" feature = "question" elsif ARGV[ 0 ] == "ask" feature = "ask" elsif ARGV[ 0 ] == "models" feature = "models" end BIN_DIR = File.dirname( File.expand_path( __FILE__ ) ) if feature == "questions" system "#{BIN_DIR}/questions.rb" elsif feature == "question" system "#{BIN_DIR}/question #{ARGV[ 1..-1 ].join( ' ' )}" elsif feature == "ask" #command = "#{BIN_DIR}/ask #{ARGV[ 1..-1 ].join( ' ' )}" #puts command.inspect #exit 3 system "#{BIN_DIR}/ask #{ARGV[ 1..-1 ].join( ' ' )}" elsif feature == "pisessions" dir = "#{ENV[ 'HOME' ]}/.pi/agent/sessions/" if File.exist?( dir ) command = "ls -1 #{dir}" output = Rlib.output( command ) puts output end system "#{BIN_DIR}/pi_sessions.rb" elsif feature == "models" if ARGV[ 1 ] == "configured" system "#{BIN_DIR}/openrouter_models.rb" else system "#{BIN_DIR}/pi_models_openrouter.rb #{ARGV[ 1..-1 ].join( ' ' )}" end end exit 0 # End of: gossip.rb EOT 28. Print Answer ---------------- In the following script date and time is used for input as in the list of questions for easy copying and pasting: Alternatively the date time timestamp to select the answer file and to print it out. cat > ./bin/print_answer.sh <<EOT #! /bin/dash # Do not edit this file, as it gets automatically generated by lp. DATE=$1 TIME=$2 # This script prints out the text anser according to the date and time given as # separate parameters. # Alternative usages: # ./bin/print_answer.sh 23-10-25 14:30 # ./bin/print_answer.sh 20231025_143022 if [ $# -eq 1 ]; then DATETIME="${1}" elif [ $# -eq 2 ]; then DATETIME="20$(echo "${DATE}_${TIME}" | perl -pe 's/[-:]//g;')??" else echo "Usage: $0 <YYYYMMDD_HHMMSS> | <YY-MM-DD> <HH:MM>" exit 2 fi TXT_FILE=db/txt/answer_${DATETIME}.txt #MD_FILE=db/txt/answer_${DATETIME}.txt if [ -f $TXT_FILE ]; then cat $TXT_FILE echo echo $TXT_FILE #elif [ -f $MD_FILE ]; then # cat $MD_FILE # echo # echo $MD_FILE else echo "ERROR: no answer file found for ${DATETIME}" >&2 exit 1 fi # End of: print_answer.sh EOT cat > bin/print_answer.rb <<EOT #! /usr/bin/env ruby # Do not edit this file, as it gets automatically generated by lp. $: << File.dirname( __FILE__ ) + '/../lib' require 'rlib' if ARGV.length != 1 puts "Usage: #{$0} <number>" puts " Positive n: n-th oldest answer (1 = first)." puts " Negative n: n-th newest answer (-1 = last)." exit 1 end index = ARGV[0].to_i dir = "db/txt" txt_files = Dir.glob( "#{dir}/answer_*.txt" ) #md_files = Dir.glob( "#{dir}/answer_*.md" ) #all_files = (txt_files + md_files).sort all_files = txt_files all_files.sort! n = all_files.length if n == 0 puts "Error: no answer files found in #{dir}/" exit 1 end resolved_index = if index > 0 index - 1 elsif index < 0 n + index # Ruby-style negative indexing else -1 # index == 0 is invalid; will be caught below end if resolved_index < 0 || resolved_index >= n puts "Error: only #{n} answer(s) available, index #{index} is out of range" exit 1 end filename = all_files[resolved_index] content = Rlib.readfile( filename ) if content.nil? puts "Error: cannot read #{filename}" exit 1 end print content puts puts filename exit 0 # End of: print_answer.rb EOT 29. Print Question ------------------ The questions.rb file lists all available questions in chronological order. The script bin/print_question.rb a) Takes a number as an argument (1-based index), b) prints the corresponding question from the database, while the c) oldest question is number 1 and the newest is number n. d) At the end the question file name is printed so the user can quickly copy and paste the name for further processing. cat > bin/print_question.rb <<EOT #! /usr/bin/env ruby # Do not edit this file, as it gets automatically generated by lp. $: << File.dirname( __FILE__ ) + '/../lib' require 'rlib' if ARGV.length != 1 puts "Usage: #{$0} <number>" puts " Oldest question is 1, newest is n." exit 1 end index = ARGV[0].to_i if index < 1 puts "Error: index must be >= 1" exit 1 end dir = "db/txt" txt_files = Dir.glob( "#{dir}/question_*.txt" ) txt_files.sort! if index > txt_files.length puts "Error: only #{txt_files.length} question(s) available" exit 1 end txt_filename = txt_files[ index - 1 ] content = Rlib.readfile( txt_filename ) print content puts txt_filename exit 0 # End of: print_question.rb EOT 30. Tags -------- The Tag database file is 'db/csv/tags.csv' and here is an example content to demonstrate its standard CSV file format. The number of tags for a single question model answer triplet is open ended, just add more tags to the same line each separated by a comma: ''' 20250512_174554,gossip 20250614_223442,c 20260709_073007,emacs,git ''' We need a tool to get an overview of all tags and to drill down a little bit to be able to find a certain question for a given tag. Here's a Ruby script to read the tags CSV file, which builds an internal hash of all tags and prints out some overview with most common tags on top. If an optional tag name has been provided, the list of all questions tagged accordingly is printed out in reverse chronological order. The script: 1. Reads 'db/csv/tags.csv' line by line 2. Splits each line by comma - first field is the datetime, rest are tags 3. Builds a hash where each tag maps to an array of datetimes 4. Sorts each datetime array chronologically 5. Outputs the result (you can also 'require' this and use 'read_tags_csv' directly in other scripts) 6. Added argument handling at the beginning of the main execution block 7. When a tag argument is provided ('ARGV[0]'): - Looks up the tag in the tags hash - If found, prints each question in a single line (newest first) - If not found, shows an error message and exits with status 1 Usage Examples: # List all tags with counts (original behavior) $ bin/tags.rb 1. gossip 42 2. emacs 35 3. git 28 The program: cat > bin/tags.rb <<EOT #! /usr/bin/env ruby # Do not edit this file, as it gets automatically generated by lp. $: << File.dirname( __FILE__ ) + '/../lib' require 'rlib' require 'term' def read_tags_csv( csv_filename ) tags_hash = {} return tags_hash unless File.exist?( csv_filename ) File.foreach( csv_filename ) do |line| line = line.strip next if line.empty? fields = line.split( ',' ) next if fields.length < 2 datetime = fields[ 0 ] tags = fields[ 1..-1 ] tags.each do |tag| tag = tag.strip next if tag.empty? tags_hash[ tag ] ||= [] tags_hash[ tag ] << datetime end end # Sort the datetime arrays for each tag (newest first) tags_hash.each_value do |datetimes| datetimes.sort! datetimes.reverse! end return tags_hash end # Main execution csv_file = "db/csv/tags.csv" tags = read_tags_csv( csv_file ) # Check if a tag argument was provided if ARGV[0] tag = ARGV[0] if tags.key?(tag) datetimes = tags[tag] # Calculate width for right-aligned numbering num_width = datetimes.length.to_s.length # Get terminal width for text truncation term = Term.new columns = term.cols datetimes.each_with_index do |dt, index| num = (index + 1).to_s.rjust(num_width) # Convert YYYYMMDD_HHMMSS to YY-MM-DD HH:MM year_short = dt[2,2] # YY from YYYY month = dt[4,2] # MM day = dt[6,2] # DD hour = dt[9,2] # HH minute = dt[11,2] # MM short_dt = "#{year_short}-#{month}-#{day} #{hour}:#{minute}" # Read and process question content txt_filename = "db/txt/question_#{dt}.txt" if File.exist?(txt_filename) content = Rlib.readfile(txt_filename) # Split content into lines and join with spaces (mimicking questions.rb behavior) lines = content.split(/\n/) question_content = lines.join(" ") else question_content = "[Question file not found]" end # Build full line and truncate to terminal width full_line = "#{num}. #{short_dt} #{question_content}" if full_line.length > columns full_line = full_line[0...columns] end puts full_line end else warn "ERROR: Tag '#{tag}' not found in #{csv_file}" exit 1 end else # Original behavior: list all tags with counts # Build array of [tag, count] and sort by count desc, then tag asc tag_counts = tags.map { |tag, datetimes| [tag, datetimes.length] } tag_counts.sort_by! { |tag, count| [-count, tag] } # Determine column widths for alignment max_num_width = tag_counts.length.to_s.length max_tag_width = tag_counts.map { |tag, _| tag.length }.max || 0 max_count_width = tag_counts.map { |_, count| count.to_s.length }.max || 0 # Print numbered, aligned output tag_counts.each_with_index do |(tag, count), index| num = (index + 1).to_s.rjust(max_num_width) tag_col = tag.ljust(max_tag_width) count_col = count.to_s.rjust(max_count_width) puts "#{num}. #{tag_col} #{count_col}" end end exit 0 EOT Key enhancements when a tag is specified: 1. For each timestamp: - Reads the corresponding question file ('db/txt/question_<datetime>.txt') - Processes content by splitting lines and joining with spaces (matching 'questions.rb' behavior) - Handles missing question files gracefully 2. Formats output as: '[right-aligned-number]. [YY-MM-DD HH:MM] [question-content]' 3. Truncates long lines to fit terminal width (using 'term.cols') 4. Maintains right-aligned numbering and newest-first ordering Example output: $ bin/tags.rb emacs 1. 26-07-09 07:30 What is the best way to configure Emacs for Ruby development? 2. 26-06-15 14:22 How do I enable line numbers in Emacs? 3. 26-05-20 09:15 Is there a way to sync Emacs settings across machines? 31. Pi Harness Prompt --------------------- As of 2026-07-31: ''' You are an expert coding assistant operating inside pi, a coding agent harness. You help users by reading files, executing commands, editing code, and writing new files. Available tools: read bash edit write In addition to the tools above, you may have access to other custom tools depending on the project. Guidelines: Use read to examine files instead of cat or sed. Use bash for file operations like ls, rg, find. Use edit for precise and single changes. Use multiple edits for multiple changes in one file. Use write only for new files or complete rewrites. Pi documentation (read only when the user asks about pi itself, its SDK, extensions, themes, skills, or TUI): - Main documentation: ... - Additional docs: ... - Examples: ... (extensions, custom tools, SDK) - When reading pi docs or examples, resolve docs/... under Additional docs and examples/... under Examples, not the current working directory - When asked about: extensions (docs/extensions.md, examples/extensions/), themes (docs/themes.md), skills (docs/skills.md), prompt templates (docs/prompt-templates.md), TUI components (docs/tui.md), keybindings (docs/keybindings.md), SDK integrations (docs/sdk.md), custom providers (docs/custom-provider.md), adding models (docs/models.md), pi packages (docs/packages.md), environment variables (docs/environment-variables.md) - When working on pi topics, read the docs and examples, and follow .md cross-references before implementing - Always read pi .md files completely and follow links to related docs (e.g., tui.md for TUI API details) ''' 32. ToDo -------- . Print answer does not find pi text answers, as it is looking for a markdown file. . Edit in pi sessions should keep their changes in the answer file. . How do pi edits work? . The today script has been taken out of another script and now is nowhere used. We wait, till it will be used again, otherwise it needs to be removed. . Make llama_call invoke servers other than locally and on port 8080. . Invoke llama and llama_call from gossip or ask? 33. Links --------- Two upstream catalogs are worth keeping open in a browser tab whenever you extend Gossip with a new model or sanity-check a price tag: https://openrouter.ai/models OpenRouter's full, live model index. Lists every model the gateway can route to — commercial, closed-weight, and open-weight alike — together with each model's release date (typically one to two days later than the upstream provider's own announcement) and the current per-million-token price for input and output. Use this page to discover a new model, confirm its exact OpenRouter model id (the value you pass to `bin/ask <modelid>`), and copy a provider hint into `db/csv/openrouter_models.csv`. https://huggingface.co/models?sort=trending&search=gguf Hugging Face's trending GGUF models, ordered by popularity. This is the place to look when you want a fresh open-weight model to run locally via the llama.cpp backend (`bin/llama <model.gguf>` or `bin/start_server.sh <model.gguf>`). The trending view surfaces the community's current favourites — the candidates most likely to give surprisingly strong results on local hardware.