fix(agents): the waiting brief matches the harness (v0.448.2)

[vikasprogrammer]

Follow-up to #834 (which made Monitor governed — this brief now recommends it).

Two false premises, both measured

• The brief prescribed for i in $(seq 1 12); do … sleep 5; done. claude-code now REFUSES a foreground sleep: Blocked: sleep N followed by …. To wait for a condition, use Monitor with an until-loop… Do not chain shorter sleeps to work around this block. Fleet transcripts show agents hitting that block repeatedly.
• Its central argument was a ~5-minute prompt-cache TTL. Cache writes on these runs are essentially all 1-hour: instapods 8,793 ephemeral_1h vs 0 ephemeral_5m (~8.8k responses), expresstech ~20k vs 12, instawp all 1h. The "one 10-minute wait costs more than twenty polls" claim doesn't hold.
• Also: long-running-thing & → run_in_background: true ("No & needed", survives across turns, notifies on exit); the 2-minute limit is now a kill or an auto-background depending on build.

Kept: the ~2-minute foreground cut-off, "make each wait earn its turn", "do not read this as batch less", the SINGLE-call scoping, both-lane wiring.

The retired TTL argument (with its original measurement) is preserved in the source comment so it isn't reinstated later. waiting-brief-test.cjs updated to pin the NEW claims and to assert the 5-minute argument does not come back; context-injection-test.cjs learns that timeout_ms is a tool field, not a tool.

🤖 Generated with Claude Code