forked from bchanot/claude
Bastien Chanot
6bfc0543e5
After a plan/reflection is elaborated and before it executes, three fresh blind sub-agents (correctness / robustness / simplicity) attack it on the big model; the main loop RE-THINKS every aspect a BLOCKER lands (a named plan change, or [deferred]) and re-challenges once if the plan materially changed. Advisory into each skill's existing human gate — the human stays the decider. - lib/challenge-plan.md — reusable phase: fail-safe (never fail open), severity-driven (any single-lens BLOCKER = must-address), RE-THINK loop - agents/plan-challenger.md — challenger role (read-only, big-model per BDR-066) - lib/tests/plan-challenger.test.sh — 41-assertion structure lock - wired into 11 orchestrators: ship-feature/init-project/feat/bugfix (build-plan), onboard/audit-delta/code-clean (proposals), seo/geo/harden/web-validate (fix-bundle) Hardened by dogfooding: 3 blind challengers reviewed this feature's own v1 plan and caught 4 BLOCKERs (fail-open, consensus-buries-lone-finding, wrong model tier vs BDR-066, false on-disk-plan premise) — all fixed here.
48 lines
2.2 KiB
Bash
48 lines
2.2 KiB
Bash
#!/usr/bin/env bash
|
|
# lib/tests/plan-challenger.test.sh — structure lock: the plan-challenger agent,
|
|
# the reusable lib/challenge-plan.md phase, and every reflection orchestrator that
|
|
# wires it (3-way adversarial plan challenge, BDR-066). STATIC only — the agent's
|
|
# adversarial behavior needs a live model (manual smoke), not a CI gate.
|
|
set -u
|
|
R="$(cd "$(dirname "$0")/../.." && pwd)"
|
|
pass=0; fail=0
|
|
ok() { pass=$((pass+1)); }
|
|
ko() { fail=$((fail+1)); printf 'FAIL %s\n' "$1"; }
|
|
has() { if grep -qF "$2" "$R/$1"; then ok; else ko "$1 missing: $2"; fi; }
|
|
fm_lacks() { if awk 'NR<=10' "$R/$1" | grep -qF "$2"; then ko "$1 frontmatter must NOT contain: $2"; else ok; fi; }
|
|
|
|
A="agents/plan-challenger.md"
|
|
L="lib/challenge-plan.md"
|
|
|
|
# 1) agent shape
|
|
has "$A" "name: plan-challenger"
|
|
has "$A" "tools: Read, Grep, Glob, Bash"
|
|
fm_lacks "$A" "model: sonnet" # BDR-066: audit judgment → NOT sonnet-pinned
|
|
has "$A" "CHALLENGE — LENS:" # load-bearing verdict grammar
|
|
has "$A" "VERDICT: SOLID | CONCERNS(n) | FATAL(n)"
|
|
has "$A" "correctness"
|
|
has "$A" "robustness"
|
|
has "$A" "simplicity"
|
|
has "$A" "Report-only"
|
|
|
|
# 2) reusable phase — the mechanism lives here (one canonical include)
|
|
has "$L" 'subagent_type="plan-challenger"'
|
|
has "$L" "BDR-066" # challengers on the big model
|
|
has "$L" "a mute verifier is NEVER a PASS" # fail-safe (never fail open)
|
|
has "$L" "Severity-driven" # any single-lens BLOCKER = must-address
|
|
has "$L" "RE-THINK" # findings re-plan the aspect, not just noted
|
|
has "$L" "correctness | robustness | simplicity"
|
|
has "$L" "CHALLENGE SUMMARY"
|
|
has "$L" "build-plan" # the three KINDs
|
|
has "$L" "proposals"
|
|
has "$L" "fix-bundle"
|
|
|
|
# 3) every reflection orchestrator wires the phase + carries a challenge summary
|
|
for s in ship-feature init-project feat bugfix onboard audit-delta code-clean seo geo harden web-validate; do
|
|
has "skills/$s/SKILL.md" "lib/challenge-plan.md"
|
|
has "skills/$s/SKILL.md" "CHALLENGE SUMMARY"
|
|
done
|
|
|
|
printf 'plan-challenge structure lock: %d pass, %d fail\n' "$pass" "$fail"
|
|
[ "$fail" -eq 0 ]
|