feat(hook): auto-capture code areas the agent edits

[pszymkowiak]

Summary

Closes #196.

Adds a hook-driven, no-MCP equivalent of Context-Engine's
record_code_area: every time Claude Code / Codex / Gemini / Copilot
calls Edit / Write / MultiEdit / NotebookEdit, the existing
PostToolUse hook (icm hook post) extracts tool_input.file_path and
upserts a row in a new code_areas table. Same (project, file_path)
increments touch_count instead of duplicating, so the table grows in
file count, not edit count.

What's new

• Schema: new code_areas(id, project, file_path, description, session_id, tool_name, touch_count, first_touched_at, last_touched_at) table with UNIQUE(project, file_path) constraint
  driving the upsert.
• Store API:
  • SqliteStore::upsert_code_area(project, file_path, description, session_id, tool_name) — ON CONFLICT bumps touch_count,
    refreshes last_touched_at / session_id / tool_name, and only
    overwrites description when the caller passes Some (preserves
    the most recent meaningful hint via COALESCE).
  • SqliteStore::list_code_areas(project, in_file, since, limit) —
    filter by exact or suffix path; ordered by last_touched_at DESC.
  • SqliteStore::code_area_count().
• Hook integration: extract_tool_input_file_path() covers three
  shapes seen across Claude Code 1.x / 2.x, Codex, and Gemini
  (tool_input.file_path, top-level file_path,
  tool_input.arguments.file_path). cmd_hook_post calls
  upsert_code_area before the extract counter — independent of
  the throttle, errors swallowed so telemetry never blocks the hook.
• CLI: new icm code-areas command with --in-file, --project,
  --since (ISO-8601), --limit, --format {table,json}.

Why no MCP

Per the issue thread: tying this to the PostToolUse hook gives 100%
coverage across every AI tool already icm init'd. No new MCP tool
to expose, no need for the agent to comply via system prompt. The
existing hook chain already runs after every tool call.

Notes

• description is None in the MVP. Once feat: LLM-summarized wake-up briefing (auto-detect invoker, cheap defaults, configurable) #165 (LLM-summarized
  briefing) ships, the same provider infrastructure can feed a diff
  summary into description opt-in.
• No new dependencies. Hook + transcript + sqlite plumbing was already
  in place — the patch reuses all of it.
• Source MCP tool (Context-Engine-AI/Context-Engine) is
  source-available proprietary; ICM ships an Apache-2.0 equivalent
  built from scratch.

Tests

• 7 new unit tests in icm-store (upsert idempotency, touch_count
  increment, description COALESCE behaviour, project + path-suffix
  filters, since filter, last_touched_at DESC ordering, count).
• 513 tests pass across the workspace, cargo clippy --workspace --all-targets -- -D warnings clean.
• Manual smoke against the release binary:
  • 3 hook payloads (Edit / Write / MultiEdit) → 2 unique paths
    captured.
  • auth.rs re-touched → touch_count = 2.
  • Bash payload → correctly ignored (no code-area row).
  • --in-file filter and --format json both verified.

Sample output

$ icm code-areas
Project              Hits   Last touched         Path
--------------------------------------------------------------------------------
icm                  2      2026-05-31 09:26:47  src/auth.rs
icm                  1      2026-05-31 09:26:47  src/login.rs

$ icm code-areas --in-file src/auth.rs --format json
[
  {
    "id": 1,
    "project": "icm",
    "file_path": "src/auth.rs",
    "description": null,
    "session_id": null,
    "tool_name": "Write",
    "touch_count": 2,
    "first_touched_at": "2026-05-31T09:26:47.203993069+00:00",
    "last_touched_at": "2026-05-31T09:26:47.211044445+00:00"
  }
]

Test plan

• cargo build --workspace
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace (513 passed)
• Manual hook smoke (3 tool variants + 1 non-edit)
• icm code-areas end-to-end with --in-file, --format json

Closes #196.