Learning-driven command correction: Phase 2 & 3

Start with crates/terraphim_agent/src/learnings/hook.rs and capture.rs, then inspect the existing CLI in crates/terraphim_agent/src/main.rs and tests under crates/terraphim_agent/tests/. Review docs/src/command-rewriting-howto.md and the adapter hook references before choosing an implementation path. Done means prompt corrections are captured, compiled into a thesaurus, applied by replace, documented, and covered by the requested integration tests.

Context

Phase 1 (currently deployed) widens the OpenCode tool.execute.before hook so it pipes whitelisted package-manager commands through terraphim-agent replace, using the Terraphim Engineer role's KG. Verified mappings include:

• npm install → bun add
• npx create-react-app foo → bunx create-react-app foo
• pip install requests → uv add requests
• pip install -r requirements.txt → uv sync
• python -m pip install fastapi → uv add fastapi
• git commit -m "fix npm install bug" → git commit -m "fix bun add bug"

See docs/src/command-rewriting-howto.md for the full Phase 1 guide.

Phase 1 uses only the curated KG at ~/.config/terraphim/docs/src/kg/. Captured corrections under ~/Library/Application Support/terraphim/learnings/correction-*.md (written by terraphim-agent learn hook --learn-hook-type user-prompt-submit) are stored but not applied. This issue tracks closing that loop.

Phase 2 – Wire user-prompt-submit into agent hooks

Goal: every time the user types "use X instead of Y" / "prefer X over Y" / "use X not Y" / "switch to X" in a prompt, the existing process_user_prompt_submit parser records a CorrectionType::ToolPreference file.

Work:

1. Document which agent events map to user-prompt-submit for each adapter (Claude Code, OpenCode, Codex). OpenCode exposes something like experimental.chat.messages.transform – confirm the exact event name in @opencode-ai/plugin/dist/index.d.ts.
2. Add the hook invocation in:
   • examples/claude-code-hooks/ (shell script variant).
   • An OpenCode plugin that pipes { user_prompt } JSON into terraphim-agent learn hook --format opencode --learn-hook-type user-prompt-submit.
3. Unit tests covering the three pattern variants plus negative cases ("I prefer tea over coffee" should NOT be captured as a tool preference).
4. Docs: append a Phase 2 section to docs/src/command-rewriting-howto.md and content/docs/command-rewriting-howto.md on the terraphim.ai Zola site.

Acceptance: uttering "use uv instead of pip" in a fresh session creates one file under .terraphim/learnings/correction-*.md or the global equivalent.

Phase 3 – Compile captured corrections into a thesaurus

Goal: captured ToolPreference corrections are turned into live replace rules without manual KG authoring.

Work in crates/terraphim_agent/src/learnings/capture.rs:

/// Compile stored corrections of given types into a Thesaurus usable by `replace`.
pub fn compile_corrections_to_thesaurus(
    storage_dir: &Path,
    correction_types: &[CorrectionType],
) -> Result<Thesaurus, LearningError>

/// Export as Logseq-style KG markdown (one file per corrected term) for review / version control.
pub fn export_corrections_as_kg(
    storage_dir: &Path,
    correction_type: CorrectionType,
    output_dir: &Path,
) -> Result<usize, LearningError>

CLI subcommands in crates/terraphim_agent/src/main.rs:

terraphim-agent learn compile-thesaurus \
    --type tool-preference \
    --output ~/.config/terraphim/kg/compiled/corrections.json

terraphim-agent learn export-kg \
    --type tool-preference \
    --output ~/.config/terraphim/kg/commands/

Extend replace (or add a new --extra-thesaurus <file> flag) so hooks can merge compiled corrections over the role's KG thesaurus. Merge semantics: compiled corrections override the base KG (learnt preferences win).

Invalidation: the SQLite thesaurus cache at /tmp/terraphim_sqlite/terraphim.db (keyed thesaurus_<role>.json) must be invalidated when corrections change. Options: (a) touch the mtime of the KG directory; (b) include the corrections-dir mtime in the cache key; (c) ship a terraphim-agent cache invalidate --role <role> command.

Acceptance:

1. terraphim-agent learn capture-correction --original "pip install" --corrected "uv add" --correction-type tool-preference creates a file.
2. terraphim-agent learn compile-thesaurus --type tool-preference produces a JSON with the mapping.
3. terraphim-agent replace "pip install requests" --extra-thesaurus <compiled.json> returns uv add requests even if the base KG doesn't have that mapping.
4. Integration test runs through steps 1-3 with #[tokio::test] in crates/terraphim_agent/tests/.

Phase 3.5 – OpenCode plugin extension (optional polish)

Modify terraphim-hooks.js (and the reference script in examples/claude-code-hooks/) to:

• Detect when a new correction-*.md is written (mtime check on the learnings dir).
• Automatically run terraphim-agent learn compile-thesaurus and pass --extra-thesaurus to the next replace invocation.

This closes the feedback loop: user says "use uv instead of pip" once → captured → compiled → applied on the next bash call.

Non-goals for this issue

• Expanding the rewriteable-head whitelist beyond package managers. That is a separate conversation about risk (shell syntax is easy to corrupt).
• Multi-user/shared corrections with trust levels (the existing shared_learning module is out of scope here; see CLAUDE.md for that roadmap).

Reference files

• docs/src/command-rewriting-howto.md (this repo) – Phase 1 how-to, now on docs.terraphim.ai after next publish.
• ~/projects/personal/private_agents_settings/opencode/plugin/terraphim-hooks.js – current Phase 1 hook implementation (pushed to AlexMikhalev/private_agents_settings).
• crates/terraphim_agent/src/learnings/hook.rs – process_user_prompt_submit, LearnHookType::UserPromptSubmit.
• crates/terraphim_agent/src/learnings/capture.rs – CorrectionEvent, capture_correction.
• crates/terraphim_automata/src/builder.rs – Logseq thesaurus builder.
• crates/terraphim_automata/src/matcher.rs – Aho-Corasick replace_matches, MIN_PATTERN_LENGTH = 2.