Skip to content

pg_hardstorage_llm_chat.md

Latest commit

 

History

History
76 lines (61 loc) · 5.09 KB

pg_hardstorage_llm_chat.md

File metadata and controls

76 lines (61 loc) · 5.09 KB
Raw
title pg hardstorage llm chat
description CLI reference for the pg hardstorage llm chat command.
tags
cli
reference

pg_hardstorage llm chat

Interactive chat session — REPL with /show-* transparency commands

Synopsis

Drop into an interactive chat session. Each line of input is sent to the configured LLM as a user turn; tool calls fire automatically; the assistant's response streams back to the terminal.

Slash commands (typed at the prompt): /help show this list /exit, /quit leave the session /clear drop the conversation history (keep system prompt) /show-context running session metadata (skill, provider, tokens) /show-tools tool surface available to the active skill /show-skill active skill name + version + source path /show-budget tokens used + remaining

The default skill is 'ask' (general-purpose Q&A). Switch with --skill (e.g. --skill incident for incident-response mode).

pg_hardstorage llm chat [flags]

Options

      --audit-repo llm export-session <id> --repo <url>   capture every llm.* event into the named repo's audit chain (export later via llm export-session <id> --repo <url>)
  -h, --help                                              help for chat
      --history-key-file string                           path to a 32-byte hex file overriding the local-KEK-derived history key (use when no kek.bin exists)
      --mode string                                       execution mode: read-only (default; execute_command refuses every invocation) | advise+execute (gated invocation per the safety stack) (default "read-only")
      --no-history                                        don't write an encrypted transcript to <state>/llm/conversations/. Default: record when a local KEK is available
      --principal string                                  operator principal for per-principal history isolation (default: $USER, or 'anonymous')

Options inherited from parent commands

      --airgapped airgapped: strict                                                       refuse outbound endpoints (LLM providers, sinks, OTLP collectors) outside loopback / RFC1918 / explicit airgap.allowlist. Also enabled by PG_HARDSTORAGE_AIRGAPPED=1 or airgapped: strict in the config file.
  -c, --config string                                                                     path to config file (default: XDG/FHS lookup)
      --cpu-profile go tool pprof <path>                                                  write a pprof CPU profile to this path for the duration of the command (go tool pprof <path> to analyse). Off when empty.
      --endpoint string                                                                   provider endpoint override
      --mem-profile string                                                                write a pprof heap profile to this path at command exit. Off when empty.
      --model string                                                                      model id override
      --no-color                                                                          disable ANSI color in text output
      --on-error-llm                                                                      on a structured-error failure, drop into the matching LLM helper skill (auto_on_error trigger). Also enabled by PG_HARDSTORAGE_ON_ERROR_LLM=1.
      --otel-endpoint string                                                              OpenTelemetry OTLP/HTTP endpoint (e.g. http://otel-collector:4318); empty disables tracing
      --otel-stdout                                                                       also export OpenTelemetry traces to stderr (useful for dev)
  -o, --output string                                                                     output format: text|json|ndjson|yaml|template|csv|markdown|html|tap|junit|pdf (default: text on TTY, json off-TTY)
      --profile-port go tool pprof http://127.0.0.1:6060/debug/pprof/profile?seconds=30   if non-zero, expose net/http/pprof on 127.0.0.1:<port> for live profiling of long-running commands (e.g. go tool pprof http://127.0.0.1:6060/debug/pprof/profile?seconds=30). Off when zero.
      --provider string                                                                   LLM provider override (default: env / config / 'openai' if key in scope, else 'mock')
  -q, --quiet                                                                             suppress non-essential output
      --skill string                                                                      skill to use (ask, explain, restore, incident) (default "ask")
      --template string                                                                   Go text/template applied when --output template (or implied if --template is set without --output)

SEE ALSO

  • pg_hardstorage llm - LLM helper — grounded assistant with skill files, mandatory preview, full audit