fix(agent): retry on JSONDecodeError instead of aborting

[BrianLi009]

Problem

On a large/slow model call, the provider can return a body the client can't parse as JSON (truncated response, gateway HTML, SSE mismatch) → json.JSONDecodeError. Because json.JSONDecodeError subclasses ValueError, the retry classifier in run_agent.py treated it as a non-retryable local validation error and aborted on attempt 1 (logged as "Non-retryable client error (HTTP None)"), even though the failure is transient and usually clears on retry.

Observed in a report-generation run: a ~221K-token call returned an unparseable body and hard-failed instead of retrying.

Fix

Exclude json.JSONDecodeError from the local-validation bucket so it falls through to the normal retry/backoff path like other transient API errors. One-line change (+ comment).

is_local_validation_error = (
    isinstance(api_error, (ValueError, TypeError))
    and not isinstance(api_error, json.JSONDecodeError)
)

json is already imported. Compiles clean. Base is dev/brian/single-agent-refactor — the branch the current report pipeline runs.

🤖 Generated with Claude Code

[aman-a-shah]

Review of PR #1 — retry on JSONDecodeError

TL;DR: this PR is already fully merged into main (via PR #2, merge commit 51b4c99b, 2026-06-23). The branch tip c0a1f377 is an ancestor of origin/main, which is now 18 commits ahead. Recommend closing this PR as superseded — nothing here is lost.

The change itself (for the record)

The core fix is sound and I'd have approved it on the merits:

• Where the error arises: in the API-call except block of run_agent.py — the client fails to parse the provider's HTTP response body (truncated body, gateway HTML, SSE mismatch on large/slow calls). It is not a tool-call-arguments parse; it's a transport-level response failure, so re-sending the same request is the right remedy (no model correction needed).
• Root cause correctly identified: json.JSONDecodeError subclasses ValueError, so the is_local_validation_error bucket swallowed it and aborted on attempt 1 as a "non-retryable client error (HTTP None)".
• Bounded retry: yes — the loop is while retry_count < max_retries with max_retries = 3 and exponential backoff, so excluding JSONDecodeError from the non-retryable bucket cannot loop forever.
• Minor gap: no dedicated unit test asserting that json.JSONDecodeError is classified as retryable while plain ValueError/TypeError still abort. Moot now, but worth adding on main if this classifier gets touched again.

Note the PR's visible diff is much larger than the one-line fix because (a) the base is the stale dev/brian/single-agent-refactor branch (itself now an ancestor of main), and (b) the branch later merged origin/main into itself (cbe81506) and picked up unrelated "converge deployed state" work (httpx granular timeouts, user_notes_tool registration, web-tools/policy fixes). All of that is on main too.

CI failure diagnosis (test job, run 27239550639)

Not caused by this PR's change. The 18 failures + 1 collection error (2026-06-09) fall into three buckets, all pre-existing staleness/env issues:

1. test_run_agent.py::TestBuildApiKwargs::test_basic_kwargs — assert Timeout(connect=15.0, read=300.0, ...) == 900.0: the test wasn't updated when the branch adopted httpx.Timeout for _build_api_kwargs.
2. 17 failures in test_streaming.py / test_run_agent.py / test_context_token_tracking.py — RuntimeError: Failed to initialize OpenAI client: Missing credentials: CI env lacked dummy API keys.
3. tests/acp/test_server.py — ImportError: cannot import name 'AuthMethod' from 'acp.schema': unpinned acp dependency drifted.

Evidence that these were environmental/stale: main's own Tests run failed identically right after PR #2 merged this same branch (run 28062459400), and was then fixed on main by 1fb278ab ("Fix pre-existing CI test failures (dummy keys, acp pin, timeout assert)") and 97801e0e ("Isolate no-key tests from the CI dummy API keys"). main's Tests have been green since (latest: 2026-07-24).

Verdict & recommendation

Close as superseded — do not merge, do not rebase. Every commit on dev/brian/jsondecode-retry-fix is already in main, including the exact JSONDecodeError exclusion (now at run_agent.py:6403-6410 on main with the same explanatory comment). The failing CI here would only be fixable by merging main into the branch — which would make the PR diff empty. The base branch dev/brian/single-agent-refactor is likewise fully contained in main and can probably be deleted along with this branch after closing.

🤖 Generated with Claude Code