Landscape-sweep worker: pass an intent on every ctx_execute; keep ctx_search at its default limit (containment M3/M5)

[seathatflowsinourveins]

What

A follow-up to #398, on output containment for the landscape sweep's Claude workers.

The token lane's per-agent count on smoke 3 (wf_c7bc3ee8-c16) found two things:

• Fetch routing is fixed: WebFetch 0, and 3 of 3 fetches went through ctx_fetch_and_index.
• Containment is not:
  • 45 of 134 tool results were over 5 KB, carrying 79.6% of result bytes;
  • ctx_execute had 27 of 52 over 5 KB, and ctx_search had 11 of 20.

The #381 targets are:

• M3: at most 1 non-exception result over 5 KB per child, and at most 20% of bytes;
• M5: at most 10% of ctx results over 5 KB.

Fix, using context-mode's own contract. The worker body now has a Context rule:

• pass an intent on every ctx_execute and ctx_batch_execute call;
• read with ctx_search at its default limit of 3, with specific queries;
• print derived answers only.

Per its schema, the intent parameter makes output over about 5 KB get "indexed into knowledge base and returns section titles + previews — not full content".

The change is in adoption/agents/claude/landscape-sweep-worker.md and its byte-identical .claude/agents/ copy.

The token lane owns the host token-lanes block, which carries the same rule, and the enforcement fallback: a PreToolUse refusal of ctx_execute without intent. That ships only if the next one-layer smoke shows the prose rule alone does not bring M3/M5 to target.

Verification

• python3 -m unittest tests.test_install_claude_profile tests.test_landscape_sweep_harness: 167 run, OK. This includes the project-scope byte-identity test.
• python3 scripts/validate.py: passed.
• Integration: the next one-layer smoke, measured with the token lane's counter. That smoke runs under the user's Gate A, which holds the full sweep anyway.

SOTA sources

• context-mode 1.0.169, the ctx_execute and ctx_batch_execute tool schemas: the intent parameter indexes output over ~5 KB and returns previews. The ctx_search default limit is 3.
• The PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381 E2E preregistration's M3 and M5 containment targets.

🤖 Generated with Claude Code

https://claude.ai/code/session_012qwXNtFyUhY5x2dtrm7tG5