Explore a language of intent to replace IOA TOML specs

[nerdsane]

What this is

A design exploration, not an implementation. This PR carries a living brainstorm document — docs/plans/2026-06-10-intent-language-exploration.md — examining what should succeed the .ioa.toml spec format as the shared medium of intent between humans and agents.

Framing

The current format is a serialization of a good semantic model (I/O automata), but the surface is a phrasebook, not a grammar: a closed inventory of string-embedded utterances with no productive syntax. The exploration works through three lenses:

• Linguistic — guards as felicity conditions (Austin), actions as frames (Fillmore), modality as grammar (deontic/alethic/evidential), pidgin-to-creole as the trajectory of human-agent formats
• PL-theoretic — P, Quint, Statecharts, the embedded-DSL lesson from LLM-fluent host languages, deep vs. surface structure
• Artistic — score vs. performance languages, Sol LeWitt's instruction latitude, binding vs. interpretive registers in musical notation

Decisions so far (from live brainstorm with Rita)

1. The human side of the channel is oral — humans speak, never write; the machine owns the written canonical form (amanuensis model)
2. Approval is risk-tiered — continuous conversational grounding for routine changes, explicit readback ritual for consequential ones
3. Core + projections — machine-owned formal core, speakable English projection as the confirmation/safety surface, conversation as evidential provenance
4. Weight as grammar — irreversible/wide-blast-radius clauses require syntactically heavier forms (iconicity); the approval tier derives from the clause itself

A first three-register sketch (canonical core / speakable projection / readback moment) is in §6 of the document.

Status

Ideation in progress. Converges into an ADR if/when a direction is settled. No code changes.

🤖 Generated with Claude Code

[rita-aga]

Superseded by #339 — same branch and commit, recreated so the PR is authored by rita-aga.