@junghanacs/entwurf 0.12.10 → 0.13.1

package/scripts/lib/step-outcome.sh

@@ -0,0 +1,88 @@
+# shellcheck shell=bash
+# Release-gate STEP OUTCOME protocol — the one place PASS / SKIP / FAIL is decided.
+#
+# SOURCE-ONLY. It carries the executable bit because check-pack requires every
+# shipped `.sh` to have one (a blanket packaging invariant worth more than an
+# exception carved for this file), not because anyone should run it.
+#
+# Sourced by run.sh's `release_gate`; sourced directly by
+# scripts/check-release-gate-outcomes.ts so the truth table is exercised as a
+# unit instead of being re-derived by eye from the aggregate's output.
+#
+# WHY A PROTOCOL AND NOT A PER-BACKEND BRANCH (P1, 2026-07-31). The aggregate's
+# prose said a cut needs `LIVE=1` and `SKIP=0`, but its exit authority read only
+# the FAIL counter, so `./run.sh release-gate` returned 0 with 14 SKIPs and the
+# summary still printed "all green". Worse, a step that WAS invoked could decide
+# it lacked a prerequisite and exit 0 — Cortex with no
+# `ENTWURF_ACP_CORTEX_CONNECTION` is the measured case — and the aggregate
+# counted that as PASS. Both holes have the same shape: a skip that cannot be
+# told from an acceptance. So the fix is one outcome protocol every step speaks,
+# never a special case per backend or per prerequisite.
+#
+# EXIT CONTRACT (house style: the verdict IS the exit code)
+#   0                          PASS — the step ran and its assertions held.
+#   ENTWURF_STEP_SKIP_EXIT     SKIP — the step declined a prerequisite it does
+#                              not have. NOT acceptance, and never rounded up.
+#   anything else              FAIL — including a crash or a signal death.
+#
+# WHY 97. The repo already spends 0..4 on per-tool exit contracts that mean
+# different things per tool (meta-bridge-fresh-cut's 3 = "cut transition
+# incomplete", meta-bridge-store-doctor's 3 = "unreadable store", dev-bin's 3 =
+# "refused a foreign link"), and 126+ belongs to the shell. A LIVE smoke shells
+# out to some of those tools, so reusing a low number would let a dependency's
+# unrelated verdict surface as a false SKIP — the exact confusion this protocol
+# exists to remove. 97 is reserved for THIS cross-tool protocol and nothing else.
+ENTWURF_STEP_SKIP_EXIT=97
+
+# Classify one step's exit code. Prints PASS | SKIP | FAIL on stdout.
+entwurf_step_outcome() {
+	case "${1:-}" in
+		0) printf 'PASS\n' ;;
+		"$ENTWURF_STEP_SKIP_EXIT") printf 'SKIP\n' ;;
+		*) printf 'FAIL\n' ;;
+	esac
+}
+
+# The release authority. Returns 0 when the run may be read as a cut floor.
+#
+#   $1 MUST fail count   $2 MUST skip count   $3 cut mode (1 = a real cut)
+#
+# A FAIL always blocks. A SKIP blocks ONLY in cut mode: an ordinary diagnostic
+# `./run.sh release-gate <dir>` must stay runnable unattended and still exit 0
+# while honestly reporting what it did not call. `--cut` is the executable half
+# of "a CUT needs LIVE=1, SKIP=0" — and it needs no separate LIVE assertion,
+# because with LIVE unset every LIVE-gated step skips and the skip count blocks.
+entwurf_release_releasable() {
+	local failc="${1:-0}" skipc="${2:-0}" cut="${3:-0}"
+	[ "$failc" -gt 0 ] && return 1
+	if [ "$cut" = "1" ] && [ "$skipc" -gt 0 ]; then
+		return 1
+	fi
+	return 0
+}
+
+# The same decision as ONE greppable token, because "why is this red" must be
+# machine-readable and not parsed out of prose.
+#
+#   $1 MUST fail count   $2 MUST skip count   $3 cut mode (1 = a real cut)
+#
+# A step that RAN AND BROKE and a step that NEVER RAN are different facts, and a
+# release record that blurs them is worthless: the first is a defect to fix, the
+# second is a prerequisite to supply. So the counters are never fudged either —
+# a policy block keeps `FAIL=0 SKIP=n` and says `BLOCKED (MUST SKIP)`. Reporting
+# a synthetic `FAIL=1` for the policy block would destroy exactly the
+# distinction this protocol exists to create.
+entwurf_release_verdict() {
+	local failc="${1:-0}" skipc="${2:-0}" cut="${3:-0}"
+	if [ "$failc" -gt 0 ]; then
+		printf 'cut: BLOCKED (MUST FAIL)\n'
+	elif [ "$cut" = "1" ] && [ "$skipc" -gt 0 ]; then
+		printf 'cut: BLOCKED (MUST SKIP)\n'
+	elif [ "$cut" = "1" ]; then
+		printf 'cut: OK\n'
+	elif [ "$skipc" -gt 0 ]; then
+		printf 'cut: n/a (diagnostic, %s SKIP)\n' "$skipc"
+	else
+		printf 'cut: n/a (diagnostic)\n'
+	fi
+}

package/scripts/mutants/acp-augment.json

@@ -25,6 +25,112 @@
 			"timeoutSeconds": 60,
 			"signature": "[QK:AUGMENT-TRUNC-MARKER]",
 			"signatureSource": "scripts/check-acp-carrier-augment.ts"
+		},
+		{
+			"claim": "CARRIER-PROVENANCE-STATED",
+			"title": "the augment stops saying what it is — the model is back to reading a first-user prepend as its system prompt (the measured 2026-07-30 misattribution)",
+			"subject": "pi-extensions/lib/acp/augment.ts",
+			"find": [
+				"\t\"This block is prepended to the FIRST USER MESSAGE of this session. It is not your system prompt.\";"
+			],
+			"replace": ["\t\"\";"],
+			"gate": ["bash", "run.sh", "check-acp-carrier-augment"],
+			"timeoutSeconds": 60,
+			"signature": "[QK:CARRIER-PROVENANCE-STATED]",
+			"signatureSource": "scripts/check-acp-carrier-augment.ts"
+		},
+		{
+			"claim": "CARRIER-FRAME-NAMES-CLAUDE-CARRIER",
+			"title": "the claude frame stops naming its real engraving-only carrier — the model is told what its system prompt is NOT, but never what it IS",
+			"subject": "pi-extensions/lib/acp/augment.ts",
+			"find": [
+				"\t\t\t\t\t`The ${params.backend} rail does have a system-prompt carrier (\\`_meta.systemPrompt\\`), and entwurf keeps it deliberately tiny: it carries the operator engraving only — never this narrative, never AGENTS.md.`,"
+			],
+			"replace": ["\t\t\t\t\t`The ${params.backend} rail carries no system-prompt carrier at all.`,"],
+			"gate": ["bash", "run.sh", "check-acp-carrier-augment"],
+			"timeoutSeconds": 60,
+			"signature": "[QK:CARRIER-FRAME-NAMES-CLAUDE-CARRIER]",
+			"signatureSource": "scripts/check-acp-carrier-augment.ts"
+		},
+		{
+			"claim": "CARRIER-FRAME-DENIES-CORTEX-CARRIER",
+			"title": "the carrier-less rail is told it has a tiny system-prompt carrier — a flat lie, since cortex's buildSessionMeta sends no _meta at all",
+			"subject": "pi-extensions/lib/acp/augment.ts",
+			"find": [
+				"\t\t\t\t\t`The ${params.backend} rail carries no system-prompt carrier at all — entwurf sends no \\`_meta.systemPrompt\\`, so everything entwurf tells you, including any operator engraving above, arrives here as user-message text.`,"
+			],
+			"replace": [
+				"\t\t\t\t\t`The ${params.backend} rail does have a system-prompt carrier (\\`_meta.systemPrompt\\`), and entwurf keeps it deliberately tiny: it carries the operator engraving only.`,"
+			],
+			"gate": ["bash", "run.sh", "check-acp-carrier-augment"],
+			"timeoutSeconds": 60,
+			"signature": "[QK:CARRIER-FRAME-DENIES-CORTEX-CARRIER]",
+			"signatureSource": "scripts/check-acp-carrier-augment.ts"
+		},
+		{
+			"claim": "CARRIER-RAIL-DIFF-IS-SOURCE-PINNED",
+			"title": "cortex grows a session carrier while the augment still tells it there is none — the frame becomes a lie and no prose check would notice",
+			"subject": "pi-extensions/lib/acp/backend-adapter.ts",
+			"find": ["\tbuildSessionMeta() {\n\t\treturn undefined;\n\t},"],
+			"replace": ["\tbuildSessionMeta() {\n\t\treturn { systemPrompt: \"x\" };\n\t},"],
+			"gate": ["bash", "run.sh", "check-acp-carrier-augment"],
+			"timeoutSeconds": 60,
+			"signature": "[QK:CARRIER-RAIL-DIFF-IS-SOURCE-PINNED]",
+			"signatureSource": "scripts/check-acp-carrier-augment.ts"
+		},
+		{
+			"claim": "CARRIER-LEADS-ITS-OWN-BLOCK",
+			"title": "the carrier loses its leading boundary — the Claude SDK's fixed sentence swallows the operator engraving again (the measured 2026-07-31 A-join)",
+			"subject": "pi-extensions/lib/acp/engraving.ts",
+			"find": ["export const CARRIER_LEAD_SEPARATOR = \"\\n\\n\";"],
+			"replace": ["export const CARRIER_LEAD_SEPARATOR = \"\";"],
+			"gate": ["bash", "run.sh", "check-acp-carrier-augment"],
+			"timeoutSeconds": 60,
+			"signature": "[QK:CARRIER-LEADS-ITS-OWN-BLOCK]",
+			"signatureSource": "scripts/check-acp-carrier-augment.ts"
+		},
+		{
+			"claim": "CARRIER-OPT-OUT-SURVIVES-SEPARATOR",
+			"title": "the boundary is attached before the emptiness test — an emptied engraving file stops being the opt-out and becomes a whitespace carrier that still replaces the claude_code preset",
+			"subject": "pi-extensions/lib/acp/engraving.ts",
+			"find": [
+				"\tconst body = interpolate(source, params).trim();",
+				"\tif (body.length === 0) return \"\";",
+				"\treturn `${CARRIER_LEAD_SEPARATOR}${body}`;"
+			],
+			"replace": [
+				"\tconst body = `${CARRIER_LEAD_SEPARATOR}${interpolate(source, params).trim()}`;",
+				"\tif (body.length === 0) return \"\";",
+				"\treturn body;"
+			],
+			"gate": ["bash", "run.sh", "check-acp-carrier-augment"],
+			"timeoutSeconds": 60,
+			"signature": "[QK:CARRIER-OPT-OUT-SURVIVES-SEPARATOR]",
+			"signatureSource": "scripts/check-acp-carrier-augment.ts"
+		},
+		{
+			"claim": "CARRIER-STAYS-TINY",
+			"title": "the shipped carrier grows into a real prompt — size, not shape, is what routes an OAuth subscription call to metered extra usage",
+			"subject": "pi-extensions/lib/acp/prompts/engraving.md",
+			"find": ["# Engraving Here"],
+			"replace": [
+				"# Engraving Here\n\n## Operator standing orders\n\n- Address the operator by name and keep the reply language matched to the prompt.\n- Prefer surgical edits over rewrites; verify every claim you make about behavior.\n- Never bypass a commit hook, and never push without an explicit request.\n- Indentation is tabs unless the project's linter says otherwise.\n- When docs and behavior disagree, report it and fix the source of truth.\n- Long-running commands belong in tmux, not in a foreground shell.\n- Record the next concrete move before you stop working.\n\n## Project context\n\nThis checkout is the bridge itself, so treat the adapter contract in the rail doc\nas authoritative and re-read it before touching a gate. The qualification lane is\nthe oracle for every contract claim; a green assertion count proves nothing on\nits own."
+			],
+			"gate": ["bash", "run.sh", "check-acp-carrier-augment"],
+			"timeoutSeconds": 60,
+			"signature": "[QK:CARRIER-STAYS-TINY]",
+			"signatureSource": "scripts/check-acp-carrier-augment.ts"
+		},
+		{
+			"claim": "CARRIER-META-SENDS-CARRIER-VERBATIM",
+			"title": "the meta hop normalizes the carrier — the boundary is stripped back off on the wire while bridgeConfigSignature still folds the bounded value, so the A-join returns and reuse keys on a string the backend never received",
+			"subject": "pi-extensions/lib/acp/tool-surface.ts",
+			"find": ["\t\tmeta.systemPrompt = normalizedSystemPrompt;"],
+			"replace": ["\t\tmeta.systemPrompt = normalizedSystemPrompt.trim();"],
+			"gate": ["bash", "run.sh", "check-acp-carrier-augment"],
+			"timeoutSeconds": 60,
+			"signature": "[QK:CARRIER-META-SENDS-CARRIER-VERBATIM]",
+			"signatureSource": "scripts/check-acp-carrier-augment.ts"
 		}
 	]
 }

package/scripts/mutants/acp-cortex.json

@@ -0,0 +1,196 @@
+{
+	"schemaVersion": 1,
+	"lane": "acp-cortex",
+	"mutants": [
+		{
+			"claim": "CORTEX-CURATED-FOUR-ROWS",
+			"title": "dropping a curated row silently shrinks the GLG-decided 4-row cortex surface",
+			"subject": "pi-extensions/lib/acp/models.ts",
+			"find": [
+				"export const SUPPORTED_CORTEX_MODEL_IDS = [",
+				"\t\"cortex-auto\",",
+				"\t\"cortex-claude-opus-5\",",
+				"\t\"cortex-claude-sonnet-5\",",
+				"\t\"cortex-openai-gpt-5.4\",",
+				"] as const;"
+			],
+			"replace": [
+				"export const SUPPORTED_CORTEX_MODEL_IDS = [",
+				"\t\"cortex-auto\",",
+				"\t\"cortex-claude-sonnet-5\",",
+				"\t\"cortex-openai-gpt-5.4\",",
+				"] as const;"
+			],
+			"gate": ["bash", "run.sh", "check-acp-cortex"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:CORTEX-CURATED-FOUR-ROWS]",
+			"signatureSource": "scripts/check-acp-cortex.ts"
+		},
+		{
+			"claim": "CORTEX-AUTH-NARROW-CREDENTIAL-CACHE",
+			"title": "widening the auth passthrough to the whole cortex/cache leaks operator tool_outputs/tip_history into the child",
+			"subject": "pi-extensions/lib/acp/overlay.ts",
+			"find": [
+				"\tcortexLinkIfExists(",
+				"\t\tjoin(realSnowflake, \"cortex\", \"cache\", \"credential_cache\"),",
+				"\t\tjoin(cortexDir, \"cache\", \"credential_cache\"),",
+				"\t);"
+			],
+			"replace": [
+				"\trmSync(join(cortexDir, \"cache\"), { recursive: true, force: true });",
+				"\tcortexLinkIfExists(join(realSnowflake, \"cortex\", \"cache\"), join(cortexDir, \"cache\"));"
+			],
+			"gate": ["bash", "run.sh", "check-acp-cortex"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:CORTEX-AUTH-NARROW-CREDENTIAL-CACHE]",
+			"signatureSource": "scripts/check-acp-cortex.ts"
+		},
+		{
+			"claim": "CORTEX-AUTOUPDATE-OFF",
+			"title": "flipping autoUpdate back on lets the cortex runtime replace itself on launch mid-rail (D4)",
+			"subject": "pi-extensions/lib/acp/overlay.ts",
+			"find": ["\treturn `${JSON.stringify({ autoUpdate: false }, null, \"\\t\")}\\n`;"],
+			"replace": ["\treturn `${JSON.stringify({ autoUpdate: true }, null, \"\\t\")}\\n`;"],
+			"gate": ["bash", "run.sh", "check-acp-cortex"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:CORTEX-AUTOUPDATE-OFF]",
+			"signatureSource": "scripts/check-acp-cortex.ts"
+		},
+		{
+			"claim": "CORTEX-BRIDGE-DUAL-HOME",
+			"title": "dropping the bridge-only real-HOME restore leaves the garden store cut off — tools reach the model but see an empty garden (D10)",
+			"subject": "pi-extensions/lib/acp/overlay.ts",
+			"find": ["\t\tif (stdio.name === CORTEX_DUAL_HOME_BRIDGE_SERVER) env.HOME = realHome;"],
+			"replace": ["\t\tvoid realHome;"],
+			"gate": ["bash", "run.sh", "check-acp-cortex"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:CORTEX-BRIDGE-DUAL-HOME]",
+			"signatureSource": "scripts/check-acp-cortex.ts"
+		},
+		{
+			"claim": "CORTEX-MCP-NON-STDIO-FAILLOUD",
+			"title": "silently dropping a non-stdio server advertises a config the session does not have (D9 fail-loud)",
+			"subject": "pi-extensions/lib/acp/overlay.ts",
+			"find": [
+				"\t\tif (\"type\" in server && (server.type === \"http\" || server.type === \"sse\")) {",
+				"\t\t\tthrow new Error(",
+				"\t\t\t\t`entwurf: cortex mcp.json projection cannot represent ${server.type} server ` +",
+				"\t\t\t\t\t`${JSON.stringify(server.name)} — only stdio entries are measured through ` +",
+				"\t\t\t\t\t`$SNOWFLAKE_HOME/cortex/mcp.json (CP0 D9). Remove it from entwurfProvider.mcpServers ` +",
+				"\t\t\t\t\t`for cortex models or front it with a stdio bridge.`,",
+				"\t\t\t);",
+				"\t\t}"
+			],
+			"replace": [
+				"\t\tif (\"type\" in server && (server.type === \"http\" || server.type === \"sse\")) {",
+				"\t\t\tcontinue;",
+				"\t\t}"
+			],
+			"gate": ["bash", "run.sh", "check-acp-cortex"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:CORTEX-MCP-NON-STDIO-FAILLOUD]",
+			"signatureSource": "scripts/check-acp-cortex.ts"
+		},
+		{
+			"claim": "CORTEX-OVERLAY-SESSION-SCOPED",
+			"title": "a static shared overlay dir lets two residents overwrite one mcp.json/envelope",
+			"subject": "pi-extensions/lib/acp/overlay.ts",
+			"find": ["\tconst scopeDir = join(overlaysRoot, cortexOverlayScopeId(params.scopeKey));"],
+			"replace": ["\tconst scopeDir = join(overlaysRoot, \"shared\");"],
+			"gate": ["bash", "run.sh", "check-acp-cortex"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:CORTEX-OVERLAY-SESSION-SCOPED]",
+			"signatureSource": "scripts/check-acp-cortex.ts"
+		},
+		{
+			"claim": "CORTEX-HOME-PRESENCE-REFUSED",
+			"title": "a truthy check instead of presence lets an empty-string CORTEX_HOME slip past the refusal (D3 E2b nuance)",
+			"subject": "pi-extensions/lib/acp/backend-adapter.ts",
+			"find": ["\t\tif (\"CORTEX_HOME\" in process.env) {"],
+			"replace": ["\t\tif (process.env.CORTEX_HOME) {"],
+			"gate": ["bash", "run.sh", "check-acp-cortex"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:CORTEX-HOME-PRESENCE-REFUSED]",
+			"signatureSource": "scripts/check-acp-cortex.ts"
+		},
+		{
+			"claim": "CORTEX-LAUNCH-NO-MODEL-PIN",
+			"title": "re-introducing the launch -m pin creates a second model authority that drifts from per-turn set-model",
+			"subject": "pi-extensions/lib/acp/backend-adapter.ts",
+			"find": [
+				"\tresolveLaunch({ config }) {",
+				"\t\tconst settings = config.adapterSettings as CortexAdapterSettings | undefined;",
+				"\t\tconst connection = settings?.cortexConnection?.trim() || undefined;",
+				"\t\tconst selectionArgs: string[] = [];"
+			],
+			"replace": [
+				"\tresolveLaunch({ config, nativeModelId }) {",
+				"\t\tconst settings = config.adapterSettings as CortexAdapterSettings | undefined;",
+				"\t\tconst connection = settings?.cortexConnection?.trim() || undefined;",
+				"\t\tconst selectionArgs: string[] = [\"-m\", nativeModelId];"
+			],
+			"gate": ["bash", "run.sh", "check-acp-cortex"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:CORTEX-LAUNCH-NO-MODEL-PIN]",
+			"signatureSource": "scripts/check-acp-cortex.ts"
+		},
+		{
+			"claim": "CORTEX-ENFORCE-SET-MODEL",
+			"title": "enforcing a constant instead of the native id silently answers with the wrong model (CP0-M)",
+			"subject": "pi-extensions/lib/acp/backend-adapter.ts",
+			"find": [
+				"\t\tawait setConfig.call(connection, { sessionId: acpSessionId, configId: \"model\", value: nativeModelId });",
+				"\t},",
+				"",
+				"\t// A connection change must invalidate a reused session (rail: Adapter contract). Flat,"
+			],
+			"replace": [
+				"\t\tawait setConfig.call(connection, { sessionId: acpSessionId, configId: \"model\", value: \"auto\" });",
+				"\t},",
+				"",
+				"\t// A connection change must invalidate a reused session (rail: Adapter contract). Flat,"
+			],
+			"gate": ["bash", "run.sh", "check-acp-cortex"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:CORTEX-ENFORCE-SET-MODEL]",
+			"signatureSource": "scripts/check-acp-cortex.ts"
+		},
+		{
+			"claim": "CORTEX-OVERLAY-KEY-IS-SESSION-KEY",
+			"title": "re-deriving the overlay scope from ambient env drops opts.sessionId — two same-process/cwd sessions alias one overlay/mcp.json",
+			"subject": "pi-extensions/lib/acp/backend-adapter.ts",
+			"find": ["\t\tconst overlay = ensureCortexDualHomeOverlay({", "\t\t\tscopeKey: sessionKey,"],
+			"replace": ["\t\tconst overlay = ensureCortexDualHomeOverlay({", "\t\t\tscopeKey: piSessionId ?? \"static\","],
+			"gate": ["bash", "run.sh", "check-acp-cortex"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:CORTEX-OVERLAY-KEY-IS-SESSION-KEY]",
+			"signatureSource": "scripts/check-acp-cortex.ts"
+		},
+		{
+			"claim": "ACP-OVERLAY-SESSIONKEY-WIRED",
+			"title": "passing a cwd-derived key instead of resolveSessionKey's value silently merges distinct sessions' overlay scopes at the production call site",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": [
+				"\t\t\tconst overlay = adapter.ensureOverlay({ cwd, modelId: model.id, nativeModelId, config, sessionKey });"
+			],
+			"replace": [
+				"\t\t\tconst overlay = adapter.ensureOverlay({ cwd, modelId: model.id, nativeModelId, config, sessionKey: `cwd:${cwd}` });"
+			],
+			"gate": ["bash", "run.sh", "check-acp-session-reuse"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:ACP-OVERLAY-SESSIONKEY-WIRED]",
+			"signatureSource": "scripts/check-acp-session-reuse.ts"
+		},
+		{
+			"claim": "CORTEX-PROVIDER-SIX-ROW-SURFACE",
+			"title": "filtering the cortex rows out of the REAL provider entry drops the whole backend from the operator's model list while models.ts and check-acp-cortex stay green",
+			"subject": "pi-extensions/acp-provider.ts",
+			"find": ["\t\tmodels: allCuratedModels(),"],
+			"replace": ["\t\tmodels: allCuratedModels().filter((m) => !m.id.startsWith(\"cortex-\")),"],
+			"gate": ["bash", "run.sh", "check-acp-provider-surface"],
+			"timeoutSeconds": 300,
+			"signature": "[QK:CORTEX-PROVIDER-SIX-ROW-SURFACE]",
+			"signatureSource": "scripts/check-acp-provider-surface.ts"
+		}
+	]
+}

package/scripts/mutants/acp-prompt-lifecycle.json

@@ -0,0 +1,100 @@
+{
+	"schemaVersion": 1,
+	"lane": "acp-prompt-lifecycle",
+	"mutants": [
+		{
+			"claim": "PROMPT-STALL-NOT-KILLED",
+			"title": "an absolute wall clock re-planted on the in-flight prompt — a turn that is still working is cut for being long",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": ["\t\treturn await Promise.race([session.connection.prompt(promptArgs), lifecycle]);"],
+			"replace": [
+				"\t\treturn await Promise.race([\n\t\t\tsession.connection.prompt(promptArgs),\n\t\t\tlifecycle,\n\t\t\tnew Promise<never>((_, reject) => setTimeout(() => reject(new Error(\"prompt timed out after 100ms\")), 100)),\n\t\t]);"
+			],
+			"gate": ["bash", "run.sh", "check-acp-prompt-lifecycle"],
+			"timeoutSeconds": 180,
+			"signature": "[QK:PROMPT-STALL-NOT-KILLED]",
+			"signatureSource": "scripts/check-acp-prompt-lifecycle.ts"
+		},
+		{
+			"claim": "ABORT-SENDS-PROTOCOL-CANCEL",
+			"title": "abort stops asking the agent to cancel its own turn — back to killing a child mid-tool instead of the protocol ending",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": ["\t\t\tsession.connection.cancel?.({ sessionId: promptArgs.sessionId });"],
+			"replace": ["\t\t\tvoid promptArgs;"],
+			"gate": ["bash", "run.sh", "check-acp-prompt-lifecycle"],
+			"timeoutSeconds": 180,
+			"signature": "[QK:ABORT-SENDS-PROTOCOL-CANCEL]",
+			"signatureSource": "scripts/check-acp-prompt-lifecycle.ts"
+		},
+		{
+			"claim": "ABORT-ESCALATION-BOUNDED",
+			"title": "the post-abort grace ignores the bound it was given — an abort against a wedged agent waits on a clock the caller does not control",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": ["\t\tgraceTimer = setTimeout(escalateAbort, opts.graceMs);"],
+			"replace": ["\t\tgraceTimer = setTimeout(escalateAbort, 5_000);"],
+			"gate": ["bash", "run.sh", "check-acp-prompt-lifecycle"],
+			"timeoutSeconds": 180,
+			"signature": "[QK:ABORT-ESCALATION-BOUNDED]",
+			"signatureSource": "scripts/check-acp-prompt-lifecycle.ts"
+		},
+		{
+			"claim": "CHILD-EXIT-DIAGNOSED",
+			"title": "a mid-prompt child death stops reporting HOW it ended — the operator is back to an undiagnosable connection error",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": [
+				"\tconst parts = [\n\t\texit.code !== null ? `exit code ${exit.code}` : undefined,\n\t\texit.signal ? `signal ${exit.signal}` : undefined,\n\t].filter(Boolean);\n\treturn parts.length > 0 ? parts.join(\", \") : \"no exit status\";"
+			],
+			"replace": ["\treturn \"no exit status\";"],
+			"gate": ["bash", "run.sh", "check-acp-prompt-lifecycle"],
+			"timeoutSeconds": 180,
+			"signature": "[QK:CHILD-EXIT-DIAGNOSED]",
+			"signatureSource": "scripts/check-acp-prompt-lifecycle.ts"
+		},
+		{
+			"claim": "REUSE-CARRIES-CHILD-DIAGNOSTICS",
+			"title": "the reuse path drops the child's stderr tail again — exactly the live sonnet failure that showed only \"ACP connection closed\"",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": ["\t\t\tfinishError(err, aborted, session.stderrTail);"],
+			"replace": ["\t\t\tfinishError(err, aborted);"],
+			"gate": ["bash", "run.sh", "check-acp-prompt-lifecycle"],
+			"timeoutSeconds": 180,
+			"signature": "[QK:REUSE-CARRIES-CHILD-DIAGNOSTICS]",
+			"signatureSource": "scripts/check-acp-prompt-lifecycle.ts"
+		},
+		{
+			"claim": "PROMPT-ERROR-NOT-TRANSIENT",
+			"title": "a prompt-phase lifecycle failure worded as a timeout — pi reclassifies it as transient and cold-replays the whole prompt",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": ["\t\t\t\"this turn has no answer\","],
+			"replace": ["\t\t\t\"the prompt timed out\","],
+			"gate": ["bash", "run.sh", "check-acp-prompt-lifecycle"],
+			"timeoutSeconds": 180,
+			"signature": "[QK:PROMPT-ERROR-NOT-TRANSIENT]",
+			"signatureSource": "scripts/check-acp-prompt-lifecycle.ts"
+		},
+		{
+			"claim": "IDLE-DEATH-ANNOUNCED",
+			"title": "a backend session that died between turns is cleared but never announced — the next turn respawns silently and the operator never learns the session is gone",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": [
+				"\t\t\tconst priorEnd = takeUnreportedChildEnd(sessionKey);\n\t\t\tif (priorEnd) {\n\t\t\t\tpushAcpLifecycleNotice(\n\t\t\t\t\tstate,\n\t\t\t\t\t`previous ${adapter.backend} session ended between turns (${describeChildEnd(priorEnd)}) — opening a new one`,\n\t\t\t\t);\n\t\t\t}"
+			],
+			"replace": ["\t\t\ttakeUnreportedChildEnd(sessionKey);"],
+			"gate": ["bash", "run.sh", "check-acp-prompt-lifecycle"],
+			"timeoutSeconds": 180,
+			"signature": "[QK:IDLE-DEATH-ANNOUNCED]",
+			"signatureSource": "scripts/check-acp-prompt-lifecycle.ts"
+		},
+		{
+			"claim": "RETIRED-TEARDOWN-NOT-ANNOUNCED",
+			"title": "a teardown we performed ourselves is reported as a death — every ordinary turn-scoped turn cries wolf about its own cleanup",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": ["\t} else if (!session.retiring) {"],
+			"replace": ["\t} else {"],
+			"gate": ["bash", "run.sh", "check-acp-prompt-lifecycle"],
+			"timeoutSeconds": 180,
+			"signature": "[QK:RETIRED-TEARDOWN-NOT-ANNOUNCED]",
+			"signatureSource": "scripts/check-acp-prompt-lifecycle.ts"
+		}
+	]
+}

package/scripts/mutants/acp-stop-reason.json

@@ -0,0 +1,80 @@
+{
+	"schemaVersion": 1,
+	"lane": "acp-stop-reason",
+	"mutants": [
+		{
+			"claim": "STOP-REFUSAL-NOT-SUCCESS",
+			"title": "a model refusal reported as a finished answer — the turn has no content, yet pi seals it done/stop",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": [
+				"\t\tcase \"refusal\":\n\t\t\treturn {\n\t\t\t\tstopReason: \"error\",\n\t\t\t\trawStopReason: stopReason,\n\t\t\t\terrorMessage: \"ACP backend stopped with: refusal (the model declined to answer; the turn is incomplete)\",\n\t\t\t};"
+			],
+			"replace": ["\t\tcase \"refusal\":\n\t\t\treturn { stopReason: \"stop\", rawStopReason: stopReason };"],
+			"gate": ["bash", "run.sh", "check-acp-stop-reason"],
+			"timeoutSeconds": 120,
+			"signature": "[QK:STOP-REFUSAL-NOT-SUCCESS]",
+			"signatureSource": "scripts/check-acp-stop-reason.ts"
+		},
+		{
+			"claim": "STOP-TURNBUDGET-NOT-SUCCESS",
+			"title": "an exhausted per-turn request budget reported as success — the answer is cut short but pi seals it done",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": [
+				"\t\tcase \"max_turn_requests\":\n\t\t\treturn {\n\t\t\t\tstopReason: \"error\",\n\t\t\t\trawStopReason: stopReason,\n\t\t\t\terrorMessage:\n\t\t\t\t\t\"ACP backend stopped with: max_turn_requests (the backend's per-turn request budget was exhausted; the turn is incomplete)\",\n\t\t\t};"
+			],
+			"replace": ["\t\tcase \"max_turn_requests\":\n\t\t\treturn { stopReason: \"stop\", rawStopReason: stopReason };"],
+			"gate": ["bash", "run.sh", "check-acp-stop-reason"],
+			"timeoutSeconds": 120,
+			"signature": "[QK:STOP-TURNBUDGET-NOT-SUCCESS]",
+			"signatureSource": "scripts/check-acp-stop-reason.ts"
+		},
+		{
+			"claim": "STOP-UNKNOWN-NOT-SUCCESS",
+			"title": "the collapse-to-stop default returns: any reason this build has never seen becomes a clean success",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": [
+				"\t\tdefault:\n\t\t\treturn {\n\t\t\t\tstopReason: \"error\",\n\t\t\t\trawStopReason: stopReason,\n\t\t\t\terrorMessage: `ACP backend stopped with an unrecognized reason: ${stopReason}`,\n\t\t\t};"
+			],
+			"replace": ["\t\tdefault:\n\t\t\treturn { stopReason: \"stop\", rawStopReason: stopReason };"],
+			"gate": ["bash", "run.sh", "check-acp-stop-reason"],
+			"timeoutSeconds": 120,
+			"signature": "[QK:STOP-UNKNOWN-NOT-SUCCESS]",
+			"signatureSource": "scripts/check-acp-stop-reason.ts"
+		},
+		{
+			"claim": "STOP-ABSENT-NOT-SUCCESS",
+			"title": "a prompt result carrying NO stop reason treated as a finished turn instead of an error",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": [
+				"\t\tcase undefined:\n\t\t\treturn {\n\t\t\t\tstopReason: \"error\",\n\t\t\t\terrorMessage: \"ACP backend ended the turn without a stop reason\",\n\t\t\t};"
+			],
+			"replace": ["\t\tcase undefined:\n\t\t\treturn { stopReason: \"stop\" };"],
+			"gate": ["bash", "run.sh", "check-acp-stop-reason"],
+			"timeoutSeconds": 120,
+			"signature": "[QK:STOP-ABSENT-NOT-SUCCESS]",
+			"signatureSource": "scripts/check-acp-stop-reason.ts"
+		},
+		{
+			"claim": "STOP-RAW-PRESERVED",
+			"title": "the raw ACP reason is dropped on the way out, so no consumer can see what the backend actually said",
+			"subject": "pi-extensions/lib/acp/backend.ts",
+			"find": ["\t\tif (verdict.rawStopReason !== undefined) state.output.rawStopReason = verdict.rawStopReason;\n"],
+			"replace": [""],
+			"gate": ["bash", "run.sh", "check-acp-stop-reason"],
+			"timeoutSeconds": 120,
+			"signature": "[QK:STOP-RAW-PRESERVED]",
+			"signatureSource": "scripts/check-acp-stop-reason.ts"
+		},
+		{
+			"claim": "STOP-PENDING-SEED",
+			"title": "the stream state seeds success: an in-flight turn already reads as a finished stop before any terminal arrives",
+			"subject": "pi-extensions/lib/acp/event-mapper.ts",
+			"find": ["\t\tstopReason: \"pending\","],
+			"replace": ["\t\tstopReason: \"stop\","],
+			"gate": ["bash", "run.sh", "check-acp-stop-reason"],
+			"timeoutSeconds": 120,
+			"signature": "[QK:STOP-PENDING-SEED]",
+			"signatureSource": "scripts/check-acp-stop-reason.ts"
+		}
+	]
+}