@junghanacs/entwurf 0.14.1 → 0.14.2

package/scripts/smoke-pi-provider-state.sh

@@ -31,6 +31,14 @@ REPO_BEFORE="$(cd "$REPO_DIR" && git status --porcelain)"
 
 SB="$(mktemp -d)"
 trap 'rm -rf "$SB"' EXIT
+# This smoke drives register-pi-provider.py, which imports the sibling pi_settings_io module — and
+# CPython writes scripts/__pycache__/*.pyc for that import, INSIDE the checkout. The purity cell
+# below never caught it because those bytes are gitignored, so `git status --porcelain` stays clean
+# while the working tree is not. That difference is invisible here and load-bearing under gate
+# qualification, whose snapshot judges tree cleanliness rather than porcelain output. Write no
+# bytecode instead of cleaning it up afterwards: a smoke that leaves artifacts and then sweeps them
+# is one early `exit` away from leaving them for real.
+export PYTHONDONTWRITEBYTECODE=1
 export HOME="$SB/home"
 export XDG_DATA_HOME="$SB/xdg"
 GLOBAL="$SB/global.json"
@@ -46,8 +54,52 @@ LEGACY_CMD="$FAKE_REPO/mcp/entwurf-bridge/start.sh"
 # authoritative: drop EVERY original PATH dir that carries a real entwurf-bridge. A maintainer
 # can have both the managed dev link and the registry package shim; filtering only command -v's
 # first hit lets the second one mask the dangling-command cell.
-printf '#!/usr/bin/env bash\necho fake-bridge\n' > "$SB/bin/entwurf-bridge"
-chmod +x "$SB/bin/entwurf-bridge"
+#
+# #81: the doctor no longer stops at resolvability — it BOOTS the effective command and requires
+# the entwurf MCP tool surface back. So the healthy fake must actually speak MCP: answer the
+# `tools/list` frame (id 2) with a tools array carrying entwurf_v2. A fake that only echoed a word
+# would now (correctly) read as "resolves but does not serve MCP", which is the NEGATIVE cell
+# below — keeping the two fakes distinct is what makes this an A/B rather than a rename.
+write_mcp_fake() {   # $1 = path
+  cat > "$1" <<'FAKE'
+#!/usr/bin/env bash
+# minimal MCP stdio server: enough frames for the boot probe, nothing else.
+while IFS= read -r line; do
+  case "$line" in
+    *'"id":1'*) printf '%s\n' '{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2024-11-05","capabilities":{},"serverInfo":{"name":"fake-entwurf-bridge","version":"0"}}}' ;;
+    *'"id":2'*) printf '%s\n' '{"jsonrpc":"2.0","id":2,"result":{"tools":[{"name":"entwurf_v2"},{"name":"entwurf_self"},{"name":"entwurf_peers"},{"name":"entwurf_inbox_read"},{"name":"entwurf_register_native"},{"name":"entwurf_fresh_call"},{"name":"entwurf_resume_call"}]}}' ;;
+  esac
+done
+FAKE
+  chmod +x "$1"
+}
+# A launcher that RESOLVES and dies on exec — the shape of a relocated package-manager shim whose
+# $0-derived target is gone (observed: exit 127 with a "No such file or directory" on stderr).
+write_dead_fake() {  # $1 = path
+  cat > "$1" <<'FAKE'
+#!/usr/bin/env bash
+echo "bash: /nonexistent/../global/v11/deadbeef/node_modules/@junghanacs/entwurf/mcp/entwurf-bridge/start.sh: No such file or directory" >&2
+exit 127
+FAKE
+  chmod +x "$1"
+}
+# A launcher that boots and speaks MCP but is NOT this bridge: it serves entwurf_v2 while missing
+# entwurf_self. This is the shape of the observed #81 session (its schema had no entwurf_self) and
+# of a stale build or a foreign binary that happens to own the name — a probe keyed on one tool
+# would call it healthy, which is why identity is the exact verb SET.
+write_stale_fake() {  # $1 = path
+  cat > "$1" <<'FAKE'
+#!/usr/bin/env bash
+while IFS= read -r line; do
+  case "$line" in
+    *'"id":1'*) printf '%s\n' '{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2024-11-05","capabilities":{},"serverInfo":{"name":"stale-bridge","version":"0"}}}' ;;
+    *'"id":2'*) printf '%s\n' '{"jsonrpc":"2.0","id":2,"result":{"tools":[{"name":"entwurf_v2"},{"name":"entwurf_peers"},{"name":"entwurf_inbox_read"},{"name":"entwurf_register_native"},{"name":"entwurf_fresh_call"},{"name":"entwurf_resume_call"}]}}' ;;
+  esac
+done
+FAKE
+  chmod +x "$1"
+}
+write_mcp_fake "$SB/bin/entwurf-bridge"
 _clean_path=""
 while IFS= read -r _dir; do
   [ -n "$_dir" ] || continue
@@ -136,14 +188,91 @@ printf '{}\n' > "$PROJECT"; rm -f "$STATE"
 OUT="$(doctor)"; RC=$?
 want "doctor(user-only bare): exits 0" "[ '$RC' -eq 0 ]"
 want "doctor(user-only bare): ok — bare bin resolves" "printf '%s' \"\$OUT\" | grep -q 'bare stable bin'"
+# The green must be about BOOTING, not about resolving — anchored on the verdict word so a doctor
+# that regressed to the resolvability-only ok cannot keep this cell green.
+want "doctor(user-only bare): green says the bin BOOTS (#81), not merely resolves" "printf '%s' \"\$OUT\" | grep -q 'and it BOOTS'"
+
+# ── H-merge: the doctor normalizes what PRODUCTION normalizes, not what each FILE holds ──
+# resolveProviderConfig merges the two mcpServers maps per-name and normalizes the RESULT once. A
+# doctor that normalized each file separately would validate entries production never sees: a
+# malformed global entry the project map shadows is gone before pi ever looks, yet the per-file
+# doctor throws and calls a healthy session red. Both directions are pinned here.
+python3 -c 'import json,sys; json.dump({"entwurfProvider":{"mcpServers":{"entwurf-bridge":{"command":123}}}}, open(sys.argv[1],"w"))' "$GLOBAL"
+printf '{"entwurfProvider":{"mcpServers":{"entwurf-bridge":{"command":"entwurf-bridge"}}}}\n' > "$PROJECT"
+set +e
+OUT="$(doctor 2>&1)"; RC=$?
+set -e
+want "doctor(invalid global SHADOWED by valid project): green — production never sees the bad entry" "[ '$RC' -eq 0 ]"
+want "doctor(invalid global shadowed): EFFECTIVE scope is project" "printf '%s' \"\$OUT\" | grep -q 'EFFECTIVE (project)'"
+# An unrelated malformed server in the global map is likewise production's problem only if it
+# survives the merge — here the project map does not shadow it, so it MUST stay red.
+python3 -c 'import json,sys; json.dump({"entwurfProvider":{"mcpServers":{"entwurf-bridge":{"command":"entwurf-bridge"},"other":{"command":[]}}}}, open(sys.argv[1],"w"))' "$GLOBAL"
+printf '{}\n' > "$PROJECT"
+if doctor >/dev/null 2>&1; then die "doctor(unshadowed malformed sibling): should FAIL — the merged map is what production normalizes"; fi
+ok "doctor(unshadowed malformed sibling): FAILS — an invalid entry that SURVIVES the merge is red"
+# The effective entry itself malformed → red regardless of which scope wrote it.
+printf '{"entwurfProvider":{"mcpServers":{"entwurf-bridge":{"command":"entwurf-bridge"}}}}\n' > "$GLOBAL"
+python3 -c 'import json,sys; json.dump({"entwurfProvider":{"mcpServers":{"entwurf-bridge":{"command":123}}}}, open(sys.argv[1],"w"))' "$PROJECT"
+if doctor >/dev/null 2>&1; then die "doctor(effective malformed): should FAIL"; fi
+ok "doctor(effective malformed in project): FAILS — shadowing cannot rescue the EFFECTIVE entry"
+# restore the H-1 baseline for the cells below
+printf '{"entwurfProvider":{"mcpServers":{"entwurf-bridge":{"command":"entwurf-bridge"}}}}\n' > "$GLOBAL"
+printf '{}\n' > "$PROJECT"
+
+# ── H-boot: #81 A/B — a bare bin that RESOLVES but does not serve MCP ─────────
+# This is the whole defect in one cell: before #81 the doctor printed ok here, and a host in this
+# exact state ran Pi ACP turns with NO mcp__entwurf-bridge__* tool at all. PATH/HOME/XDG are the
+# sandbox's, so the operator's real launcher is never touched by this A/B.
+write_dead_fake "$SB/bin/entwurf-bridge"
+# `set -e` is on: a doctor that (correctly) exits non-zero here would kill the smoke before it could
+# judge the failure. Fence just the drive, never the assertions — a silently swallowed rc is how a
+# negative cell turns into decoration.
+set +e
+OUT="$(doctor 2>&1)"; RC=$?
+set -e
+want "[QK:PI-DOCTOR-BOOT-NEGATIVE] doctor(bare resolves, does not boot): FAILS instead of printing ok" "[ '$RC' -ne 0 ]"
+want "doctor(bare resolves, does not boot): names the boot failure reason" "printf '%s' \"\$OUT\" | grep -q 'does NOT serve MCP \[exited-before-tools-list\]'"
+want "doctor(bare resolves, does not boot): carries the launcher's own stderr" "printf '%s' \"\$OUT\" | grep -q 'No such file or directory'"
+# Remediation must NOT be a fix-all: entwurf may not clobber a launcher it does not own, so the
+# foreign case has to be named as the operator's repair.
+want "doctor(bare resolves, does not boot): refuses to clobber a foreign launcher" "printf '%s' \"\$OUT\" | grep -q 'will not overwrite'"
+want "doctor(bare resolves, does not boot): still distinguishes 'resolves' from 'boots'" "printf '%s' \"\$OUT\" | grep -q 'RESOLVES — probing'"
+# A bridge that BOOTS and speaks MCP but is not this one — the #81 symptom exactly.
+write_stale_fake "$SB/bin/entwurf-bridge"
+set +e
+OUT="$(doctor 2>&1)"; RC=$?
+set -e
+want "doctor(boots, wrong verb set): FAILS — booting an MCP server is not being THIS bridge" "[ '$RC' -ne 0 ]"
+want "doctor(boots, wrong verb set): names the missing verb (#81 symptom was entwurf_self)" "printf '%s' \"\$OUT\" | grep -q 'missing entwurf_self'"
+want "doctor(boots, wrong verb set): reason is the tool-set mismatch, not a boot failure" "printf '%s' \"\$OUT\" | grep -q 'tool-set-mismatch'"
 
-# H-2: project STALE (user bare, project legacy) → effective=project=legacy note (green-runtime-red guard)
+write_mcp_fake "$SB/bin/entwurf-bridge"
+OUT="$(doctor)"; RC=$?
+want "doctor(after repair): the SAME unchanged doctor goes green once the launcher boots" "[ '$RC' -eq 0 ]"
+want "doctor(after repair): green names the exact verb set as the evidence" "printf '%s' \"\$OUT\" | grep -q 'exact entwurf verb set'"
+
+# The review A/B: an unowned/non-bare override shadows the healthy user bridge in production.
+# Ownership remains the operator's, but runtime truth cannot become a green note merely because
+# entwurf will not overwrite it. Keep this QK cell before the healthy legacy override: a mutant
+# that skips all non-bare runtime probes must be attributed to THIS failure, not the later positive.
+write_dead_fake "$SB/bin/dead-override"
 printf '{"entwurfProvider":{"mcpServers":{"entwurf-bridge":{"command":"entwurf-bridge"}}}}\n' > "$GLOBAL"
+python3 -c 'import json,sys; json.dump({"entwurfProvider":{"mcpServers":{"entwurf-bridge":{"command":sys.argv[2],"args":[]}}}}, open(sys.argv[1],"w"))' "$PROJECT" "$SB/bin/dead-override"
+set +e
+OUT="$(doctor 2>&1)"; RC=$?
+set -e
+want "[QK:PI-DOCTOR-PROBES-OVERRIDE] doctor(dead unowned override): FAILS runtime independently of ownership" "[ '$RC' -ne 0 ]"
+want "doctor(dead unowned override): names the exact invocation failure" "printf '%s' \"\$OUT\" | grep -q 'effective invocation does NOT serve MCP'"
+
+# H-2: project legacy override still gets an independent runtime verdict before its ownership note.
+mkdir -p "$(dirname "$LEGACY_CMD")"
+write_mcp_fake "$LEGACY_CMD"
 python3 -c 'import json,sys; json.dump({"entwurfProvider":{"mcpServers":{"entwurf-bridge":{"command":sys.argv[2]}}}}, open(sys.argv[1],"w"))' "$PROJECT" "$LEGACY_CMD"
 OUT="$(doctor)"; RC=$?
-want "doctor(project-stale): exits 0 (honest note, not a crash)" "[ '$RC' -eq 0 ]"
-want "doctor(project-stale): EFFECTIVE is project (shadows user)" "printf '%s' \"\$OUT\" | grep -q 'EFFECTIVE (project)'"
-want "doctor(project-stale): flags legacy managed path (not adopted)" "printf '%s' \"\$OUT\" | grep -q 'LEGACY managed repo path'"
+want "doctor(project-legacy): exits 0 when the exact override boots" "[ '$RC' -eq 0 ]"
+want "doctor(project-legacy): EFFECTIVE is project (shadows user)" "printf '%s' \"\$OUT\" | grep -q 'EFFECTIVE (project)'"
+want "doctor(project-legacy): runtime BOOTS before the ownership note" "printf '%s' \"\$OUT\" | grep -q 'configured override BOOTS'"
+want "doctor(project-legacy): flags legacy managed path (not adopted)" "printf '%s' \"\$OUT\" | grep -q 'LEGACY managed repo path'"
 
 # H-3: state-owned DRIFT → FAIL (state says managed but effective is foreign)
 printf '{"entwurfProvider":{"mcpServers":{"entwurf-bridge":{"command":"/foreign/thing"}}}}\n' > "$GLOBAL"

package/scripts/smoke-resident-garden-guard.sh

@@ -70,7 +70,7 @@ fi
 SESSIONS_BASE="$RGG_AGENT_DIR/sessions"
 cleanup() { rm -rf "$RGG_AGENT_DIR"; }
 trap cleanup EXIT
-# Default target stays pi-native (openai-codex/gpt-5.4), but honor the SAME target
+# Default target stays pi-native (openai-codex/gpt-5.6-luna), but honor the SAME target
 # knob as the other live smokes: ENTWURF_LIVE_TARGET="<provider>/<model>". On
 # acp-on-v2 the ACP `entwurf` provider is back, and `./run.sh smoke-acp-rgg-live`
 # drives THIS runner against it (deterministic half) — the citizen logic (record +
@@ -83,7 +83,7 @@ if [ -n "${ENTWURF_LIVE_TARGET:-}" ]; then
 	esac
 else
 	_rgg_provider="${ENTWURF_LIVE_PROVIDER:-openai-codex}"
-	_rgg_model="${ENTWURF_LIVE_MODEL:-gpt-5.4}"
+	_rgg_model="${ENTWURF_LIVE_MODEL:-gpt-5.6-luna}"
 fi
 PROVIDER="${SMOKE_RGG_PROVIDER:-$_rgg_provider}"
 MODEL="${SMOKE_RGG_MODEL:-$_rgg_model}"