Summary
PraisonAI gateway bots expose only five built-in slash commands (/help, /status, /new, /stop, /whoami). Users running a long-lived bot cannot switch the model mid-conversation, see their token/cost usage, compress an over-long conversation to stay within the context window, or queue follow-up messages — all common, high-value controls for a production chat bot. The underlying capabilities largely exist in core (per-call model override, context optimisation/compaction, usage tracking via hooks); they are simply not wired to an in-chat control surface.
Current behaviour
# src/praisonai/praisonai/bots/_commands.py — _initialize_builtin_commands()
self.register("help", {"description": "Show help message", "builtin": True})
self.register("status", {"description": "Show bot status", "builtin": True})
self.register("new", {"description": "Reset conversation session", "builtin": True})
self.register("stop", {"description": "Cancel current agent task", "builtin": True})
self.register("whoami", {"description": "Show your user info and permissions", "builtin": True})Custom commands can be registered programmatically via CommandRegistry.register(...), and access is already gated by CommandAccessPolicy — but there are no built-ins for model switching, usage/cost, context compression or message queuing. The enabling primitives already exist in core:
- context compaction/optimisation:
praisonaiagents/context/optimizer.py (SummarizeOptimizer, LLMSummarizeOptimizer), context/manager.py
- token/cost tracking:
BEFORE_LLM/AFTER_LLM hooks in praisonaiagents/hooks/types.py
- per-call model override on
Agent.chat
Desired behaviour
A standard, opt-in set of built-in gateway commands, each gated by the existing CommandAccessPolicy:
/model <name> — switch the LLM for this session (session-scoped override)/usage — show tokens used and estimated cost for the session/compress — summarise earlier turns to free context window/queue <text> — enqueue a follow-up to run after the current turn- (and
/retry, /undo where transcript state allows)
Layer placement
- Primary layer: wrapper — these are gateway/bot control-surface commands wired into the
CommandRegistry.
- Why not core: the underlying operations (compaction, usage, model override) belong in core and already exist; the chat command surface is a wrapper concern.
- Why not tools: these are operator controls over the session, not agent-callable task integrations.
- Why not plugins: they are first-class, always-available bot UX, not optional cross-cutting policy.
- Secondary touch (optional): core — expose small
Session.usage_snapshot() and Session.compact() conveniences so the wrapper handlers stay thin.
- 3-way surface (CLI + YAML + Python): yes — the commands live in chat; YAML
CommandAccessPolicy gates them; Python registers/overrides handlers via CommandRegistry.
Proposed approach
- Extension point: add built-in handlers in
bots/_commands.py that call existing core APIs; keep them opt-out via config.
- Minimal API sketch:
@bot.on_command("model")
async def _model(ctx, name):
ctx.session.set_model(name) # session-scoped override
return f"Model switched to {name} for this conversation."
@bot.on_command("usage")
async def _usage(ctx):
u = ctx.session.usage_snapshot() # tokens + cost from hooks
return f"Tokens: {u.total} · Est. cost: ${u.cost:.4f}"Resolution sketch
# Before (today)
/model gpt-4o -> "Unknown command"
/usage -> "Unknown command"
/compress -> "Unknown command"
# After (proposed)
/model gpt-4o -> "Model switched to gpt-4o for this conversation."
/usage -> "Tokens: 18,204 · Est. cost: $0.0431"
/compress -> "Summarised 42 earlier messages; ~12k tokens freed."
Severity
High — for a long-running, multi-user gateway these controls are table stakes; their absence forces users to start a new session (/new) and lose context simply to change model or recover from context-window overflow.
Validation
Read bots/_commands.py (_initialize_builtin_commands registers exactly five commands; CommandAccessPolicy already present). Confirmed via search that no /model, /usage, /cost, /compress or /queue handler exists anywhere under bots/. Verified the enabling primitives already exist in core (context/optimizer.py, context/manager.py, hooks/types.py LLM token hooks, per-call model override on Agent.chat).
Summary
PraisonAI gateway bots expose only five built-in slash commands (
/help,/status,/new,/stop,/whoami). Users running a long-lived bot cannot switch the model mid-conversation, see their token/cost usage, compress an over-long conversation to stay within the context window, or queue follow-up messages — all common, high-value controls for a production chat bot. The underlying capabilities largely exist in core (per-call model override, context optimisation/compaction, usage tracking via hooks); they are simply not wired to an in-chat control surface.Current behaviour
Custom commands can be registered programmatically via
CommandRegistry.register(...), and access is already gated byCommandAccessPolicy— but there are no built-ins for model switching, usage/cost, context compression or message queuing. The enabling primitives already exist in core:praisonaiagents/context/optimizer.py(SummarizeOptimizer,LLMSummarizeOptimizer),context/manager.pyBEFORE_LLM/AFTER_LLMhooks inpraisonaiagents/hooks/types.pyAgent.chatDesired behaviour
A standard, opt-in set of built-in gateway commands, each gated by the existing
CommandAccessPolicy:/model <name>— switch the LLM for this session (session-scoped override)/usage— show tokens used and estimated cost for the session/compress— summarise earlier turns to free context window/queue <text>— enqueue a follow-up to run after the current turn/retry,/undowhere transcript state allows)Layer placement
CommandRegistry.Session.usage_snapshot()andSession.compact()conveniences so the wrapper handlers stay thin.CommandAccessPolicygates them; Python registers/overrides handlers viaCommandRegistry.Proposed approach
bots/_commands.pythat call existing core APIs; keep them opt-out via config.Resolution sketch
Severity
High — for a long-running, multi-user gateway these controls are table stakes; their absence forces users to start a new session (
/new) and lose context simply to change model or recover from context-window overflow.Validation
Read
bots/_commands.py(_initialize_builtin_commandsregisters exactly five commands;CommandAccessPolicyalready present). Confirmed via search that no/model,/usage,/cost,/compressor/queuehandler exists anywhere underbots/. Verified the enabling primitives already exist in core (context/optimizer.py,context/manager.py,hooks/types.pyLLM token hooks, per-call model override onAgent.chat).