[drknowhow]

Why

48% of c3_shell commands were cat, head, tail, sed, grep, rg, find and ls over project files (7,278 events measured 2026-09-04) — reads and searches the code-intelligence tools answer within budget, indexed, and with the index's exclusions. Cod's ruling: advisory only; refusing on size pushes agents to the native shell and loses every control. S4, the last phase of the shell remediation (campaign c3-shell-remediation).

What

• cli/tools/shell_nudge.py: bypass_hint(cmd, project_path) — at most ONE [c3_shell:hint] line naming the equivalent call: c3_read(file_path, lines=[a,b]) for cat/head/sed -n 'a,bp', c3_compress(mode='map') then c3_read for tail, c3_search(action='code'|'exact', query, path, ignore_case) for grep/rg (identifier → code, otherwise exact), c3_search(action='files', query) for find -name / recursive ls. Never for paths outside the project, ~/$VAR paths, node_modules/dist/build/.git/.c3; a plain ls is not a search. The command is never rewritten or refused.
• render_shell_response appends the line after the hint section; telemetry detail.hint records the kind so the follow rate and the false-positive rate can be measured from .c3/tool_telemetry.jsonl before anything is promoted.

Verification

15 tests (reads, sed ranges, grep identifier vs regex vs -e, pipes, outside-project and variable paths, finds, silence for everything else, wiring + telemetry kind); shell suites 82 green; harness 26/26 unchanged; ruff clean; full suite in CI.

https://claude.ai/code/session_018VQLfSMFw9Lfd9mGntt19H