fix(agent-team): let gate ② track the configurable gate ① for bash / run_python_code

[D41910]

Fixes a gate ① / gate ② budget drift in the agent_team tool-result post-processor.

bash and run_python_code overflow to disk at gate ①, whose cap is
meta.max_result_chars = TOOL_EXEC_RESULT_MAX_CHARS (8K by default, and the
knob docs/tool-result-truncation-ab.md actually turns between 8000 and 50000).
The preview that comes back carries two things at its tail: the
saved read-only at <path> pointer and the verdict (exit status, the pytest
summary line).

SwarmToolResultPostProcessor gave those two tools hardcoded gate ② budgets
of 4K and 5K — below gate ①'s own preview:

_BUDGETS: ClassVar[dict[str, int]] = {
    "bash": 4_000,
    "run_python_code": 5_000,
    ...

Two consequences, both reproduced against the real code path:

1. At the default 8K cap, run_python_code already loses both. A 137,967-char
result overflows to a 7,930-char preview naming the spill file; gate ② then cuts
it to 5,116 chars and the pointer and the verdict are gone. bash survives
only by accident — _compact_bash happens to keep the tail.

2. Raising the cap buys an agent_team deployment nothing. At 50K, gate ①
returns a 49,933-char preview and gate ② hands the model 4,037 (bash) or 5,116
(run_python_code). The knob only widens the window between two caps.

### TOOL_EXEC_RESULT_MAX_CHARS=50000  tool=run_python_code
  gate(1) cap=50000  result=137967 -> preview 49933  spilled=True
    gate(2) React  (stateful_react):  49933 chars | recovery path: SURVIVES | '1 failed' verdict: kept
    gate(2) Swarm  (agent_team)   :   5116 chars | recovery path: *** LOST *** | '1 failed' verdict: *** LOST ***

Why this is an omission, not a decision

ReactToolResultPostProcessor in workflows/stateful_react_agent/_runtime.py
already carries the rule, with the reasoning written down:

if name in self._CONFIGURED_EXEC_TOOLS:
    # Gate ① is configurable. Gate ② must track it or a deployment
    # raising TOOL_EXEC_RESULT_MAX_CHARS can lose the overflow pointer.

Three further signs in this file that the two tools were expected to follow the
config, not a literal:

• the grep_search / glob_search entries right above were raised to 10K for
  exactly this reason ("Gate ① overflows these to disk at meta.max_result_chars=8K
  and puts the 'saved to , use read_file' pointer at the TAIL … 10K leaves
  the 8K body and its footer room to both survive — the value react has carried
  for this reason all along");
• test_site3_keeps_the_self_pagination_pointer is parametrised over both
  processors precisely so "a future divergence fails here instead of in a
  benchmark trajectory" — but its tool list covers only read_file,
  grep_search, glob_search;
• _compact_bash's docstring already warns that the split "silently degrades to
  'treat the whole blob as stdout'" if the producers change — this budget is the
  other half of that same contract.

So the two processors disagree on exactly the tools whose gate ① is configurable.

The fix

Wire in the same rule the React processor uses — one import, two constants, one
branch:

_CONFIGURED_EXEC_TOOLS = frozenset({"bash", "run_python_code"})
_EXEC_FOOTER_HEADROOM = 2_000
...
if name in self._CONFIGURED_EXEC_TOOLS:
    budget = get_config().tool_exec_result_max_chars + self._EXEC_FOOTER_HEADROOM
else:
    budget = self._BUDGETS.get(name, self._BUDGET_DEFAULT)

The _ELLIPSIS wording, the stderr split and the 3K stderr keep window are all
left alone.

Tests

Added to the existing tests/test_tool_result_truncation.py (no new file),
parametrised over both processors via the file's existing _processors()
helper — 4 new cases per processor:

• test_the_exec_pointer_survives_the_default_gate_1 — a gate-①-sized overflow
  must keep both the spill path and the verdict at the default cap.
• test_gate_2_follows_a_raised_exec_cap — with the cap at 50K, a 40K body must
  come through uncut, i.e. raising the knob actually widens gate ②.

Validity check — reverting only subagent_runtime.py and keeping the new
tests fails exactly 3:

FAILED test_the_exec_pointer_survives_the_default_gate_1[run_python_code-agent_team]
FAILED test_gate_2_follows_a_raised_exec_cap[bash-agent_team]
FAILED test_gate_2_follows_a_raised_exec_cap[run_python_code-agent_team]
3 failed, 38 passed

Every failure is the agent_team parameter. The stateful_react parameters pass
both before and after — that is the shape of an omission, not a redesign.

Differential fuzz, 2016 cases (12 tools × 4 body shapes × 21 sizes ×
error flag), same probe before and after via git stash:

total 2016 | unchanged 1891 | changed 125
changed by tool: {'bash': 64, 'run_python_code': 61}
VIOLATIONS: 0

• no non-exec tool changed at all (read_file / grep_search / glob_search /
  web_fetch / web_search / collect_reports / delegate_subtask / submit_report /
  unknown / empty-name: 0 changes);
• no is_error=True case changed — errors still bypass every cap;
• the React processor's output is byte-identical in all 2016 cases.

Verification

check	result
pytest tests/test_tool_result_truncation.py	41 passed (33 before)
pytest tests/ (full, 3896 collected)	3886 passed, 10 skipped, 1 failed
ruff check (CI's lint gate)	clean
pyright	0 errors, 0 warnings

The single full-suite failure is tests/test_network_tools.py::test_redirect_hops_are_validated_not_followed_blindly
— this sandbox resolves www.iana.org to 198.18.0.44, so the URL guard refuses
it. It fails identically on unmodified main (git stash → same 1 failed,
17 passed in that file), so it is not from this change.

ruff format --check reports 2 files needing reformatting, also identically on
unmodified main (this machine's ruff is newer than the repo's pin; the hunks
are at subagent_runtime.py:1357, untouched here). CI's lint gate is
ruff check, which is clean.

Not addressed here

_compact_bash computes stdout_budget = max(500, budget - len(stderr_keep) - overhead),
so a very large stderr can drive the stdout slice to the 500-char floor and elide
most of it even when the overall budget is generous. That is a separate question
from whether gate ② tracks gate ①, it behaves identically in both processors,
and changing it would alter output shape for the currently-working bash path —
so I left it alone rather than fold it in.

🤖 Prepared with Claude Code (Claude Opus 4.8, 1M context). Happy to split this
out or reword anything.

[zhanghanduo]

LGTM. Thanks for the fix.