Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@src/openhuman/providers/compatible.rs`:
- Around line 1000-1008: The helper parse_native_response currently returns a
successful ProviderChatResponse with text None when api_response.choices is
empty; change parse_native_response to return a Result<ProviderChatResponse,
Error> (or existing error type) and propagate an Err when
choices.into_iter().next() is None so empty choices are treated as an error;
update callers (e.g., chat() and the other occurrence around
parse_native_response usage at the second site) to handle the Result (propagate
or map the error) to keep the contract consistent and avoid silently accepting
invalid payloads.
- Around line 333-343: OpenHumanUsage currently defines non-optional u64 token
fields (input_tokens/output_tokens/total_tokens/cached_input_tokens) with
#[serde(default)] which deserializes missing fields to 0 and causes
oh_usage.map(|u| u.input_tokens) to return Some(0) and block fallback; change
those fields to Option<u64> (e.g., input_tokens: Option<u64>) and update
extract_usage to use oh_usage.and_then(|u| u.input_tokens) / and_then for each
token so missing keys yield None and allow the fallback chain to proceed; also
add a debug/trace log in extract_usage noting which branch provided the usage
(OpenHuman vs standard source) when making the branch decision so flow changes
are observable.

---

Nitpick comments:
In `@src/openhuman/agent/harness/session/transcript.rs`:
- Around line 207-213: The current escape_content and unescape_content functions
perform a simple replace between MSG_CLOSE and MSG_CLOSE_ESCAPED but are not
fully reversible if the original content already contains the literal escape
sequence (<!--\/MSG-->), which will be transformed on round-trip; add a concise
code comment above the functions escape_content and unescape_content documenting
this edge case (mention the exact sequences MSG_CLOSE and MSG_CLOSE_ESCAPED and
that an input containing MSG_CLOSE_ESCAPED will be corrupted on unescape), and
optionally note that a more robust fix would be to use a non-ambiguous encoding
or a two-pass escaping scheme if full reversibility is later required.

In `@src/openhuman/agent/harness/subagent_runner.rs`:
- Around line 432-444: TranscriptMeta is being populated with zeroed usage and
cost fields; update the pipeline to thread actual usage from run_inner_loop into
the transcript: modify run_inner_loop (and its callers) to return or expose the
ChatResponse.usage (or an aggregated Usage struct) instead of discarding it, sum
input/output/token usage and charged amount across sub-agent provider responses,
and then set TranscriptMeta.input_tokens, output_tokens, cached_input_tokens (if
applicable), and charged_amount_usd from that aggregated Usage when constructing
TranscriptMeta in subagent_runner.rs so transcripts reflect real usage for
debugging and accounting.

In `@src/openhuman/providers/compatible.rs`:
- Around line 1066-1099: In extract_usage, add a structured debug trace that
logs which source provided token counts (openhuman vs standard usage) and
whether billing metadata was present: inspect oh, std_usage, oh_usage, and
oh_billing and emit a development-level logger call (with a stable prefix) that
records booleans like from_openhuman, from_standard, and has_billing and the
numeric token values (input_tokens, output_tokens, cached_input_tokens) but no
sensitive payloads; ensure the log is placed after computing the token variables
and before returning the ProviderUsageInfo so the instrumentation reflects the
branch decision used to populate ProviderUsageInfo.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@src/openhuman/agent/harness/session/transcript.rs`:
- Around line 207-223: The current escape_content/unescape_content pair is not
bijective because a pre-existing literal MSG_CLOSE_ESCAPED becomes MSG_CLOSE
during unescape; fix by making the scheme reversible: in escape_content
(function escape_content) first escape backslashes (e.g., replace "\" with an
escaped marker) and then escape MSG_CLOSE (MSG_CLOSE -> MSG_CLOSE_ESCAPED), and
in unescape_content perform the exact reverse order (first unescape
MSG_CLOSE_ESCAPED -> MSG_CLOSE, then unescape the backslash marker -> "\"), or
alternatively switch to a length-delimited encoding (prefix content with its
byte length) inside escape_content and decode accordingly in unescape_content;
update references to MSG_CLOSE and MSG_CLOSE_ESCAPED and ensure both functions
are exact inverses so every input round-trips exactly.
- Around line 166-175: resolve_new_transcript_path currently races by returning
a suggested filename (max+1) without reserving it; change the allocation to
atomically reserve a new transcript file by attempting to create the file with
O_EXCL style semantics in a retry loop: in resolve_new_transcript_path (or a new
helper e.g., reserve_new_transcript_path) compute sanitized name and index, then
try to open/create the candidate path using create_new
(OpenOptions::new().write(true).create_new(true)) or equivalent, and if
create_new fails because the file exists, increment index and retry until
success (with a reasonable cap/backoff); return the successfully created PathBuf
so the later fs::write in subagent_runner (calls at lines referenced) can safely
write without clobbering another concurrent writer. Ensure errors propagate via
Result and clean up partially created files on failure if needed.
- Around line 268-320: The parser currently treats malformed message tags as
end-of-input by using `break`, causing truncated files to load silently; in
`parse_messages()` replace those `break` branches (specifically the `else {
break; }` after failing to find `MSG_OPEN_SUFFIX`/role, and the final
double-`find` that falls back to `break` when `MSG_CLOSE` isn't found) with hard
errors (e.g., `anyhow::bail!` or `Err(anyhow!(...))`) that include context (the
current `search_from`/content_start index, the expected token like
`MSG_OPEN_PREFIX`/`MSG_OPEN_SUFFIX`/`MSG_CLOSE`, and a short snippet of the
remaining `raw`), so callers get an explicit parse error instead of a partial
`Ok(messages)` return; keep the successful message pushes using
`unescape_content` and `ChatMessage` unchanged.