fix: reject whitespace-only task_summary in orchestrator context capture (#459)

[Deez-Automations]

Fixes #459.

The bug

OrchestratorAgent._capture_agent_context decides whether to store an upstream agent's task_summary for propagation to downstream agents (via _enrich_with_prior_context, which injects every stored (agent_label, summary) pair verbatim into the next agent's task description under a "Prior workflow context" header). The guard was if summary: — a whitespace-only string (" ") is truthy in Python, so it passed the guard and got stored and propagated as pure noise.

Per the issue: this wastes tokens, and the guard being satisfiable by a non-empty-but-content-free string is a soft spot a prompt-injection payload could exploit to slip something past it undetected.

Scope check

Verified against current source before writing anything: the issue's line numbers (421-422) match exactly, and this is the sole call site of this truthiness pattern in the file — the two other task_summary reads (an event-emission call, already [:200]-truncated, and a logging call) don't gate storage, so neither needed the fix. All 6 delegation methods (onboarding, invoice, fraud, payments, system_maintenance, communication agents) funnel through this one method, so the single fix covers all 6.

Fix

# before
if summary:

# after
if isinstance(summary, str) and summary.strip():

Went one step past the issue's own suggested summary and summary.strip(). Security review caught that .strip() alone would raise AttributeError on a truthy non-string task_summary — today that's unreachable in practice (schema-typed as "string" under a strict LLM tool-call contract in base.py's complete_task schema), verified independently against that schema before applying, but relying on an external API contract alone for local crash-safety doesn't match this repo's own boundary-validation convention. Added the isinstance check so a violation of that invariant fails safe (summary just isn't stored) instead of turning into an AttributeError that burns one of the agent's limited delegation retries.

A real summary with incidental leading/trailing whitespace is still stored verbatim — only the truthiness check changed, the stored value itself is untouched.

Tests

9 new tests in tests/unit/agents/test_orchestrator_whitespace_context.py (no prior test file existed for OrchestratorAgent at all): whitespace-only, empty-string, missing-key, tab/newline-only, non-string, a real summary, a real summary with incidental whitespace kept verbatim, plus two integration tests through _enrich_with_prior_context confirming the polluted context never reaches a downstream task description. Confirmed RED before the fix (4/8 failed, exactly the whitespace-related cases), GREEN after.

Test plan

• 9 new unit tests pass
• Broader regression (tests/unit/agents/, tests/unit/ctf/) run — same 3 pre-existing failures in test_specialized_agents.py (unrelated agent classes), already confirmed unrelated in a sibling PR's testing
• Local merge-tested solo against fresh origin/main — clean
• Local merge-tested pairwise against Fix: orchestrator confirms payment to vendor even when it fails #574, Fix: DNS-based SSRF bypass in guardrail webhook URL validation #575, Fix: guardrail payload corruption before signing + missing after_tool for complete_task #576, Fix: Dark Lab tool override silently drops parameters/inputSchema #577, Fix: profile share-card cache key omits fields that affect the render #578, fix: guard vendor_id/invoice_id=None in chat assistant tool callables (#418) #579 — all clean. Fix: orchestrator confirms payment to vendor even when it fails #574 also touches orchestrator.py; verified the auto-merge is genuinely non-overlapping (different, unrelated methods), not just conflict-free

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.