post-edit hook silently does nothing on Claude Code: tool_input.file_path is not parsed #110

The PostToolUse hook parses the wrong payload shape for Claude Code, so both features it drives have been silently inert: agent-edit provenance records nothing, and the post-edit blast-radius warning never fires.

Evidence

Same file, same hook, two payload shapes:

$ printf '{"session_id":"s1","file_path":"<abs>/cmd/doctor.go"}' | codemap hook post-edit
   Imports 1 hub(s): internal/buildinfo/version.go
  recorded: 1

$ printf '{"session_id":"s1","tool_name":"Edit","tool_input":{"file_path":"<abs>/cmd/doctor.go"}}' | codemap hook post-edit
  recorded: 0

The second is what Claude Code actually sends for Edit/Write. It produces no output and no record.

Corroborating: across a four-day session in this repo where I made a dozen-plus Edit/Write tool calls and shipped four merged PRs, .codemap/agent_edits.jsonl was never created. Every UserPromptSubmit hook that session printed no agent edits recorded this session — accurately reporting its own failure.

Root cause

parseHookFilePathsErr (cmd/hooks.go:1798) handles exactly two structured shapes:

if filePath, ok := data["file_path"].(string); ok { ... }        // top-level
if toolInput, ok := data["tool_input"].(map[string]any); ok {
    if command, ok := toolInput["command"].(string); ok { ... }  // Codex apply_patch
}

tool_input.file_path matches neither. The regex fallback that would have caught it only runs when json.Unmarshal fails — and a well-formed payload parses fine, so it never executes. Valid JSON in the unhandled shape is the one case that silently returns nothing.

Impact

• Edit provenance (fix: make codemap's model of agent sessions truthful #84) — agent_edits.jsonl is never written on Claude Code. showWorkingSetSummary and showSessionProgress therefore always fall back to disk-churn counts, which is the exact conflation that feature was built to remove.
• Post-edit blast radius — the "you just edited a file that N others import" warning never fires. This is the behavior CLAUDE.md instructs agents to rely on before editing.

Fix

Accept tool_input.file_path (and, for symmetry, tool_input.notebook_path) alongside the existing shapes. Worth also handling tool_response.filePath, which some tools echo back.

Regression tests should cover each payload shape end-to-end rather than only the parser, since the parser was arguably "working" — it just never saw the shape it needed.

Cross-reference

Same class as #109: a feature reporting success (or in this case, honestly reporting emptiness) while doing nothing. Both are cases where the check and the reality diverged silently.