Show HN: TERMy – A fast terminal assistant that does not use LLMs

TEXT START: I love research and development, you may have heard of me because of PJON (Padded Jittering Operative Network).

The Dissection

This is a valid small-systems engineering report disguised as a manifesto. TERMy is a hand-authored command grammar with synonym tables, regex extraction, matching logic, templates, and shell adapters. Its real achievement is narrow but useful: it removes predictable chores from an LLM loop, runs locally in milliseconds, and makes behavior inspectable. The project is a phrasebook attached to a shell, not a mind.

The text then upgrades implementation details into inflated claims. Adding JSON entries is called learning, but it is manual programming without generalization. Static response and thinking strings are presented as intelligence, but they are templates. A permission prompt is treated as inherent safety, although it is only an authorization gate. The final leap—from an efficient terminal utility to democratic, liberated conversational agents—is product rhetoric layered over a credible niche tool.

The Core Fallacy

The central error is confusing bounded automation with an escape from AI-driven automation. TERMy works because someone has already enumerated the intent space. When novelty, ambiguity, context, exception handling, or multi-step planning appears, the finite grammar breaks down. The architecture itself concedes this by routing unresolved cases to an LLM.

This does not contradict P1. It implements P1 more efficiently: use cheap deterministic automation for easy cognitive tasks and reserve expensive models for the harder residue. It attacks token cost and latency, not ownership of productive assets or the mass employment-wage-consumption circuit.

The safety claim is also structurally false. Asking a human to approve a command does not prove that the command is correct, narrow, understandable, or harmless. It merely transfers the final liability to the user. Dataset errors, parser bugs, shell differences, stale assumptions, unsafe interpolation, and chained commands remain live failure modes.

Hidden Assumptions

• Natural-language variation can be captured by finite vocabularies and matching heuristics.
• User intent is explicit rather than contextual, ambiguous, or compositional.
• Variables such as filenames can be safely extracted with simple regular expressions.
• Users can meaningfully inspect and evaluate generated shell commands.
• Community-authored datasets will remain consistent, current, secure, and auditable.
• Local execution and low cost automatically produce democratization.
• Replacing LLM calls for trivial tasks materially changes who owns or controls automation.
• The operating system, shell, permissions, tools, and environment behave as the dataset expects.

Social Function

Classification: partial truth, transition management, and techno-libertarian ideological anesthetic.

The partial truth is solid: using a large model to activate a virtual environment or list files is wasteful. A deterministic front-end can reduce cost, latency, and model improvisation. The transition-management function is stronger: TERMy is naturally a preprocessing layer that future copilots can absorb before escalating difficult requests to larger models.

The anesthetic appears in the claims about freedom and democracy. Cheap, local software can broaden access, but it does not redistribute AI capital, restore bargaining power, or preserve human necessity. No-LLM branding is anti-hype engineering, not an escape from displacement. It makes automation cheaper and more ubiquitous.

The Verdict

TERMy is a useful deterministic terminal interface and a plausible component of future agent stacks. It is not a general NLU system, a self-learning agent, an inherently safe executor, or a structural alternative to AI dominance.

Under the Discontinuity Thesis, it is not a lifeboat. It is a sharper tool for removing another layer of human-needed work. Its best-case future is niche transition infrastructure: a fast local router that handles known commands and forwards the rest to sovereign-controlled models. Technically valid. Economically non-subversive. Rhetorically overextended.