Reduce the resident prompt to identity and policy (measured 50% vs 73% for identity-only) #1813

Overview

The production resident prompt measurably degrades task success. Delete it down to identity and policy, and measure the result.

Problem Statement

packages/agent/src/agent_guidance.rs holds roughly 2,300 tokens of prose; the assembled system prompt is 10,493 characters. On cases constructed to exercise the prompt's own guard rules, it scored 50% against 73% for a 445-character identity-only prompt.

The mechanism is dilution, not misunderstanding: under the full prompt, "List all my wine" produced node_type: "album" — a type absent from the injected context block entirely. The rule instructing verbatim id-copying was present in that prompt and did not prevent it.

The cost is structural as well as behavioral. The inference engine enforces N_CTX_MINIMUM = 16_384, justified by the tool-registered system prompt alone being roughly 6,600 tokens, and ADR-062 budgets context against total RAM. Prompt size is a hardware-requirements input.

Scope

Delete everything ADR-064 assigns to another channel:

• Argument-shape rules → already in tool schemas (agent_guidance.rs:20-28: TITLE TEMPLATE, FIELD RULES, ENUM FIELDS)
• Tool-selection guidance → tool descriptions (much of TOOL_STRATEGY_RULES)
• Rules duplicating an existing code guard → delete outright. Identical repeated tool calls, duplicate cross-turn writes, false action claims, the iteration cap, and malformed-argument runs are all already enforced in agent_loop.rs; the prose adds nothing but tokens

Retained: identity, blast-radius and deletion policy, the clarification contract, output formatting.

Explicit non-goal: do not replace the deleted block with a differently-worded resident block. The measured winner was identity-only at 445 characters. The evidence supports deletion, not substitution. Re-occupying the channel with better prose re-litigates the same failure — and rule 5 of ADR-064 says so directly. Content returns only when a specific measurement demands it, one rule at a time.

Evidence

• Full production prompt 50% vs identity-only 73%, on 10 adversarial cases, n=30 per arm
• "List all my wine" → node_type: "album", a type not in context
• Channel authority: tool schema 100%, prose 87.5%, absent 62.5%
• Adding field types to the prompt did not stop the model packing values into a free-text field — prose does not repair argument shape even when correct
• packages/agent/src/agent_guidance.rs; N_CTX_MINIMUM = 16_384 in the inference engine

Carried caveat: 2 of the 7 guard cases in that harness were badly designed — one scored 0% in every arm including the control, and one was non-monotonic. The effect direction is supported; the exact magnitude is not. Repair those cases before re-measuring.

Acceptance Criteria

• agent_guidance.rs reduced to identity and policy; deleted content is deleted, not relocated into another prose surface
• Every deleted rule is traced to the channel that now owns it, or explicitly recorded as dropped
• The two known-bad guard cases are repaired before the measurement is run
• A non-adversarial case set is added — the existing cases were built to defeat this prompt and cannot measure average-case cost
• Success rate at least matches the 73% identity-only arm; a result below the 50% production arm blocks the change
• Full agent test suite shows no regression
• Assembled prompt character count recorded before and after; N_CTX_MINIMUM re-evaluated if the reduction is large

Architecture Impact

Implements ADR-064 rule 5. May relax the minimum context window and, through ADR-062, the memory floor.

Reference harness at ~/nodespace/spike-1802-harness/ (not committed). Repair the harness before trusting any number from it — its README documents known defects, and six harness faults during the original spike each reversed a conclusion.