fix(sight): don't auto-overwrite invalid JSON configs

[zhangtaibo]

What & Why

Fixes #1502.

ensure_default_agents_config (added in 98158d2 to fix #1427) silently overwrote invalid JSON config files with the embedded default, suppressing the documented WARN Failed to load config from {path:?}: {e}, using embedded defaults log line.

Root cause: extract_schema_version internally does serde_json::from_str::<VersionProbe>(json).ok() — for invalid JSON it returns None. Then None < Some(CURRENT_SCHEMA_VERSION) is true, so the function backed up the file to .bak.<ts> and overwrote it with the default. By the time load_from_file ran, the file was valid JSON → parse_json_rules succeeded → the config_load_err warn branch in AgentSight::new (src/unified.rs:186) never fired.

Reproduction (on 0.8.0 RPM, before this fix)

echo 'NOT_VALID_JSON_AT_ALL{{{' > /tmp/agentsight_invalid_cfg_test.json
timeout 8 agentsight trace --config /tmp/agentsight_invalid_cfg_test.json > /tmp/invalid_config_test.log 2>&1
cat /tmp/invalid_config_test.log
# empty — no "JSON parse error" warn
ls /tmp/agentsight_invalid_cfg_test.json*
# original silently backed up to .bak.<ts>, file rewritten with default schema_version=2

This breaks agentic-os-tests/agentsight-tests/tests/test_config_validation.py::TestInvalidConfigHandling::test_invalid_json_config_returns_error and TestConfigFileParsingErrors::test_config_truncated_json_error (both failing on the 2026-07-15 regression run).

Fix

In ensure_default_agents_config (src/agentsight/src/config.rs), probe JSON validity with serde_json::from_str::<serde_json::Value> before calling extract_schema_version. If the parse fails, return Ok(()) without touching the file — this lets the caller's load_from_file → parse_json_rules surface JSON parse error: <serde error> and the warn-fallback path fire. Auto-upgrade is only meaningful for valid JSON whose schema_version is outdated.

let content = std::fs::read_to_string(path)
    .with_context(|| format!("Failed to read existing config at {path:?}"))?;

// If the file is not valid JSON at all, leave it untouched so the caller's
// load_from_file → parse_json_rules surfaces "JSON parse error: ..." and
// AgentSight::new emits its `Failed to load config from {path:?}: {e},
// using embedded defaults` warn log. Auto-upgrade is only meaningful for
// *valid* JSON whose schema_version is outdated; silently overwriting an
// invalid file masks the parse error and breaks the documented fallback
// contract (see issue #1502).
if serde_json::from_str::<serde_json::Value>(&content).is_err() {
    return Ok(());
}

let on_disk_version = extract_schema_version(&content);
// ... existing auto-upgrade logic unchanged ...

This preserves the #1427 auto-upgrade behavior for valid-but-stale configs (still backed up + overwritten when schema_version < CURRENT_SCHEMA_VERSION) while restoring the parse-error warn log for invalid JSON.

Tests

Added 2 regression tests in config::tests:

• ensure_default_agents_config_leaves_invalid_json_untouched — writes "NOT_VALID_JSON_AT_ALL{{{", calls ensure_default_agents_config, asserts file content unchanged and no .bak file created.
• ensure_default_agents_config_leaves_truncated_json_untouched — same with a truncated-JSON variant.

All existing ensure_default_agents_config_* tests still pass:

running 6 tests
test config::tests::ensure_default_agents_config_leaves_truncated_json_untouched ... ok
test config::tests::ensure_default_agents_config_leaves_invalid_json_untouched ... ok
test config::tests::ensure_default_agents_config_preserves_current_config ... ok
test config::tests::ensure_default_agents_config_creates_file_when_missing ... ok
test config::tests::ensure_default_agents_config_upgrades_old_schema_version ... ok
test config::tests::ensure_default_agents_config_upgrades_stale_config ... ok

test result: ok. 6 passed; 0 failed; 0 ignored; 0 measured; 1083 filtered out

End-to-end verification on test env (47.239.61.17)

Built agentsight release binary with this fix, scp'd to test env, ran manual repro:

$ echo 'NOT_VALID_JSON_AT_ALL{{{' > /tmp/agentsight_invalid_cfg_test.json
$ timeout 8 /tmp/agentsight-fixed trace --config /tmp/agentsight_invalid_cfg_test.json > /tmp/invalid_config_test.log 2>&1
$ cat /tmp/invalid_config_test.log
[2026-07-15T13:55:17.835Z WARN  agentsight::unified:186] Failed to load config from "/tmp/agentsight_invalid_cfg_test.json": Failed to parse config from "/tmp/agentsight_invalid_cfg_test.json": JSON parse error: expected value at line 1 column 1, using embedded defaults
$ cat /tmp/agentsight_invalid_cfg_test.json
NOT_VALID_JSON_AT_ALL{{{
$ ls /tmp/agentsight_invalid_cfg_test.json*
/tmp/agentsight_invalid_cfg_test.json

The expected WARN ... JSON parse error: ... using embedded defaults log line now fires, and the invalid file is left untouched (no .bak backup). Same behavior verified for truncated JSON ({"cmdline": {"allow": [{"rule": ["*cosh*").

Related

• Fixes [sight] agentsight 0.8.0: ensure_default_agents_config silently overwrites invalid JSON configs — "JSON parse error" warn log never fires #1502
• [sight] bug: sight rpm升级最新版本后，由于config.json格式不兼容，导致无法采集token数据 #1427 (closed) — original auto-upgrade feature request that introduced the regression
• [sight] agentsight 0.8.0: serve schema migration overwrites user cmdline/https/codex_offsets with built-in defaults — user customizations lost on upgrade #1496 (open) — separate but related: agentsight serve schema migration overwrites user customizations for valid v1 configs (different code path, same theme of schema-migration logic silently clobbering user state)

[jfeng18]

LGTM. Fix is minimal and correct: invalid JSON → early return without touching the file, letting the downstream load_from_file warn-fallback fire as documented. Three discriminating tests cover invalid/truncated/binary cases. Consistent with our root-cause analysis of #1502.