gsd-pi 2.77.0-dev.58d3d4d6c → 2.77.0-dev.cfd69e714

package/packages/pi-coding-agent/dist/tests/system-prompt-skill-filter.test.d.ts

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=system-prompt-skill-filter.test.d.ts.map

package/packages/pi-coding-agent/dist/tests/system-prompt-skill-filter.test.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"system-prompt-skill-filter.test.d.ts","sourceRoot":"","sources":["../../src/tests/system-prompt-skill-filter.test.ts"],"names":[],"mappings":""}

package/packages/pi-coding-agent/dist/tests/system-prompt-skill-filter.test.js

@@ -0,0 +1,130 @@
+// @gsd/pi-coding-agent + system-prompt-skill-filter.test — coverage for the
+// optional `skillFilter` option added to buildSystemPrompt (RFC #4779). The
+// filter lets consumers narrow the <available_skills> catalog rendered into
+// the cached system prompt without touching skill loading or invocation.
+import test from "node:test";
+import assert from "node:assert/strict";
+import { buildSystemPrompt } from "../core/system-prompt.js";
+function makeSkill(name, description = `description for ${name}`) {
+    return {
+        name,
+        description,
+        filePath: `/tmp/${name}/SKILL.md`,
+        baseDir: `/tmp/${name}`,
+        source: "project",
+        disableModelInvocation: false,
+    };
+}
+function extractAvailableSkills(prompt) {
+    const start = prompt.indexOf("<available_skills>");
+    const end = prompt.indexOf("</available_skills>");
+    if (start === -1 || end === -1)
+        return "";
+    return prompt.slice(start, end + "</available_skills>".length);
+}
+// ─── Default branch (no customPrompt) ──────────────────────────────────────
+test("buildSystemPrompt: skillFilter omits filtered-out skills from <available_skills>", () => {
+    const skills = [makeSkill("alpha"), makeSkill("beta"), makeSkill("gamma")];
+    const prompt = buildSystemPrompt({
+        skills,
+        selectedTools: ["read", "Skill"],
+        skillFilter: skill => skill.name !== "beta",
+    });
+    const section = extractAvailableSkills(prompt);
+    assert.ok(section.length > 0, "catalog section should render");
+    assert.match(section, /<name>alpha<\/name>/);
+    assert.match(section, /<name>gamma<\/name>/);
+    assert.doesNotMatch(section, /<name>beta<\/name>/);
+});
+test("buildSystemPrompt: skillFilter omitted preserves pre-filter behavior (all skills render)", () => {
+    const skills = [makeSkill("alpha"), makeSkill("beta")];
+    const prompt = buildSystemPrompt({
+        skills,
+        selectedTools: ["read", "Skill"],
+    });
+    const section = extractAvailableSkills(prompt);
+    assert.match(section, /<name>alpha<\/name>/);
+    assert.match(section, /<name>beta<\/name>/);
+});
+test("buildSystemPrompt: skillFilter that rejects every skill suppresses the <available_skills> block", () => {
+    const skills = [makeSkill("alpha"), makeSkill("beta")];
+    const prompt = buildSystemPrompt({
+        skills,
+        selectedTools: ["read", "Skill"],
+        skillFilter: () => false,
+    });
+    // With zero visible skills, formatSkillsForPrompt returns an empty string,
+    // so the opening tag should not appear anywhere.
+    assert.ok(!prompt.includes("<available_skills>"));
+});
+// ─── Custom-prompt branch ──────────────────────────────────────────────────
+test("buildSystemPrompt (customPrompt): skillFilter applies to the catalog appended onto a custom prompt", () => {
+    const skills = [makeSkill("alpha"), makeSkill("beta"), makeSkill("gamma")];
+    const prompt = buildSystemPrompt({
+        customPrompt: "CUSTOM BASE",
+        skills,
+        selectedTools: ["read", "Skill"],
+        skillFilter: skill => skill.name === "alpha",
+    });
+    const section = extractAvailableSkills(prompt);
+    assert.match(section, /<name>alpha<\/name>/);
+    assert.doesNotMatch(section, /<name>beta<\/name>/);
+    assert.doesNotMatch(section, /<name>gamma<\/name>/);
+});
+// ─── Interaction with disableModelInvocation ──────────────────────────────
+test("buildSystemPrompt: skillFilter composes with disableModelInvocation (both must pass)", () => {
+    // A skill already hidden from the catalog by disableModelInvocation must
+    // remain hidden even if skillFilter would otherwise admit it. The filter
+    // narrows, it does not override the existing invisibility contract.
+    const skills = [
+        { ...makeSkill("visible"), disableModelInvocation: false },
+        { ...makeSkill("hidden"), disableModelInvocation: true },
+    ];
+    const prompt = buildSystemPrompt({
+        skills,
+        selectedTools: ["read", "Skill"],
+        skillFilter: () => true,
+    });
+    const section = extractAvailableSkills(prompt);
+    assert.match(section, /<name>visible<\/name>/);
+    assert.doesNotMatch(section, /<name>hidden<\/name>/);
+});
+// ─── Pass-through of non-filtered fields ──────────────────────────────────
+test("buildSystemPrompt: skillFilter does not affect context files or cwd rendering", () => {
+    const skills = [makeSkill("alpha")];
+    const prompt = buildSystemPrompt({
+        skills,
+        cwd: "/tmp/example",
+        contextFiles: [{ path: "CLAUDE.md", content: "project instructions" }],
+        selectedTools: ["read", "Skill"],
+        skillFilter: () => false,
+    });
+    assert.ok(prompt.includes("/tmp/example"), "cwd should still render");
+    assert.ok(prompt.includes("project instructions"), "context files should still render");
+    assert.ok(!prompt.includes("<available_skills>"), "no skill catalog when filter rejects all");
+});
+// ─── Exception safety ─────────────────────────────────────────────────────
+test("buildSystemPrompt: skillFilter that throws falls back to unfiltered list and does not propagate", (t) => {
+    // A buggy consumer predicate must not bubble out of buildSystemPrompt.
+    // If it did, _rebuildSystemPrompt could unwind mid-setTools() and leave
+    // the session with updated tools but a stale system prompt.
+    const skills = [makeSkill("alpha"), makeSkill("beta")];
+    // Suppress the console.warn the fallback emits so test output stays clean.
+    const originalWarn = console.warn;
+    const warnings = [];
+    console.warn = (...args) => { warnings.push(args.join(" ")); };
+    t.after(() => { console.warn = originalWarn; });
+    let prompt = "";
+    assert.doesNotThrow(() => {
+        prompt = buildSystemPrompt({
+            skills,
+            selectedTools: ["read", "Skill"],
+            skillFilter: () => { throw new Error("consumer bug"); },
+        });
+    });
+    const section = extractAvailableSkills(prompt);
+    assert.match(section, /<name>alpha<\/name>/, "alpha should render (fallback to unfiltered)");
+    assert.match(section, /<name>beta<\/name>/, "beta should render (fallback to unfiltered)");
+    assert.ok(warnings.some(w => w.includes("skillFilter threw") && w.includes("consumer bug")), "fallback should emit an identifying warning");
+});
+//# sourceMappingURL=system-prompt-skill-filter.test.js.map

package/packages/pi-coding-agent/dist/tests/system-prompt-skill-filter.test.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"system-prompt-skill-filter.test.js","sourceRoot":"","sources":["../../src/tests/system-prompt-skill-filter.test.ts"],"names":[],"mappings":"AAAA,4EAA4E;AAC5E,4EAA4E;AAC5E,4EAA4E;AAC5E,yEAAyE;AAEzE,OAAO,IAAI,MAAM,WAAW,CAAC;AAC7B,OAAO,MAAM,MAAM,oBAAoB,CAAC;AAExC,OAAO,EAAE,iBAAiB,EAAE,MAAM,0BAA0B,CAAC;AAG7D,SAAS,SAAS,CAAC,IAAY,EAAE,WAAW,GAAG,mBAAmB,IAAI,EAAE;IACvE,OAAO;QACN,IAAI;QACJ,WAAW;QACX,QAAQ,EAAE,QAAQ,IAAI,WAAW;QACjC,OAAO,EAAE,QAAQ,IAAI,EAAE;QACvB,MAAM,EAAE,SAAS;QACjB,sBAAsB,EAAE,KAAK;KAC7B,CAAC;AACH,CAAC;AAED,SAAS,sBAAsB,CAAC,MAAc;IAC7C,MAAM,KAAK,GAAG,MAAM,CAAC,OAAO,CAAC,oBAAoB,CAAC,CAAC;IACnD,MAAM,GAAG,GAAG,MAAM,CAAC,OAAO,CAAC,qBAAqB,CAAC,CAAC;IAClD,IAAI,KAAK,KAAK,CAAC,CAAC,IAAI,GAAG,KAAK,CAAC,CAAC;QAAE,OAAO,EAAE,CAAC;IAC1C,OAAO,MAAM,CAAC,KAAK,CAAC,KAAK,EAAE,GAAG,GAAG,qBAAqB,CAAC,MAAM,CAAC,CAAC;AAChE,CAAC;AAED,8EAA8E;AAE9E,IAAI,CAAC,kFAAkF,EAAE,GAAG,EAAE;IAC7F,MAAM,MAAM,GAAG,CAAC,SAAS,CAAC,OAAO,CAAC,EAAE,SAAS,CAAC,MAAM,CAAC,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC,CAAC;IAC3E,MAAM,MAAM,GAAG,iBAAiB,CAAC;QAChC,MAAM;QACN,aAAa,EAAE,CAAC,MAAM,EAAE,OAAO,CAAC;QAChC,WAAW,EAAE,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,IAAI,KAAK,MAAM;KAC3C,CAAC,CAAC;IAEH,MAAM,OAAO,GAAG,sBAAsB,CAAC,MAAM,CAAC,CAAC;IAC/C,MAAM,CAAC,EAAE,CAAC,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,+BAA+B,CAAC,CAAC;IAC/D,MAAM,CAAC,KAAK,CAAC,OAAO,EAAE,qBAAqB,CAAC,CAAC;IAC7C,MAAM,CAAC,KAAK,CAAC,OAAO,EAAE,qBAAqB,CAAC,CAAC;IAC7C,MAAM,CAAC,YAAY,CAAC,OAAO,EAAE,oBAAoB,CAAC,CAAC;AACpD,CAAC,CAAC,CAAC;AAEH,IAAI,CAAC,0FAA0F,EAAE,GAAG,EAAE;IACrG,MAAM,MAAM,GAAG,CAAC,SAAS,CAAC,OAAO,CAAC,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC;IACvD,MAAM,MAAM,GAAG,iBAAiB,CAAC;QAChC,MAAM;QACN,aAAa,EAAE,CAAC,MAAM,EAAE,OAAO,CAAC;KAChC,CAAC,CAAC;IAEH,MAAM,OAAO,GAAG,sBAAsB,CAAC,MAAM,CAAC,CAAC;IAC/C,MAAM,CAAC,KAAK,CAAC,OAAO,EAAE,qBAAqB,CAAC,CAAC;IAC7C,MAAM,CAAC,KAAK,CAAC,OAAO,EAAE,oBAAoB,CAAC,CAAC;AAC7C,CAAC,CAAC,CAAC;AAEH,IAAI,CAAC,iGAAiG,EAAE,GAAG,EAAE;IAC5G,MAAM,MAAM,GAAG,CAAC,SAAS,CAAC,OAAO,CAAC,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC;IACvD,MAAM,MAAM,GAAG,iBAAiB,CAAC;QAChC,MAAM;QACN,aAAa,EAAE,CAAC,MAAM,EAAE,OAAO,CAAC;QAChC,WAAW,EAAE,GAAG,EAAE,CAAC,KAAK;KACxB,CAAC,CAAC;IAEH,2EAA2E;IAC3E,iDAAiD;IACjD,MAAM,CAAC,EAAE,CAAC,CAAC,MAAM,CAAC,QAAQ,CAAC,oBAAoB,CAAC,CAAC,CAAC;AACnD,CAAC,CAAC,CAAC;AAEH,8EAA8E;AAE9E,IAAI,CAAC,oGAAoG,EAAE,GAAG,EAAE;IAC/G,MAAM,MAAM,GAAG,CAAC,SAAS,CAAC,OAAO,CAAC,EAAE,SAAS,CAAC,MAAM,CAAC,EAAE,SAAS,CAAC,OAAO,CAAC,CAAC,CAAC;IAC3E,MAAM,MAAM,GAAG,iBAAiB,CAAC;QAChC,YAAY,EAAE,aAAa;QAC3B,MAAM;QACN,aAAa,EAAE,CAAC,MAAM,EAAE,OAAO,CAAC;QAChC,WAAW,EAAE,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,IAAI,KAAK,OAAO;KAC5C,CAAC,CAAC;IAEH,MAAM,OAAO,GAAG,sBAAsB,CAAC,MAAM,CAAC,CAAC;IAC/C,MAAM,CAAC,KAAK,CAAC,OAAO,EAAE,qBAAqB,CAAC,CAAC;IAC7C,MAAM,CAAC,YAAY,CAAC,OAAO,EAAE,oBAAoB,CAAC,CAAC;IACnD,MAAM,CAAC,YAAY,CAAC,OAAO,EAAE,qBAAqB,CAAC,CAAC;AACrD,CAAC,CAAC,CAAC;AAEH,6EAA6E;AAE7E,IAAI,CAAC,sFAAsF,EAAE,GAAG,EAAE;IACjG,yEAAyE;IACzE,yEAAyE;IACzE,oEAAoE;IACpE,MAAM,MAAM,GAAY;QACvB,EAAE,GAAG,SAAS,CAAC,SAAS,CAAC,EAAE,sBAAsB,EAAE,KAAK,EAAE;QAC1D,EAAE,GAAG,SAAS,CAAC,QAAQ,CAAC,EAAE,sBAAsB,EAAE,IAAI,EAAE;KACxD,CAAC;IACF,MAAM,MAAM,GAAG,iBAAiB,CAAC;QAChC,MAAM;QACN,aAAa,EAAE,CAAC,MAAM,EAAE,OAAO,CAAC;QAChC,WAAW,EAAE,GAAG,EAAE,CAAC,IAAI;KACvB,CAAC,CAAC;IAEH,MAAM,OAAO,GAAG,sBAAsB,CAAC,MAAM,CAAC,CAAC;IAC/C,MAAM,CAAC,KAAK,CAAC,OAAO,EAAE,uBAAuB,CAAC,CAAC;IAC/C,MAAM,CAAC,YAAY,CAAC,OAAO,EAAE,sBAAsB,CAAC,CAAC;AACtD,CAAC,CAAC,CAAC;AAEH,6EAA6E;AAE7E,IAAI,CAAC,+EAA+E,EAAE,GAAG,EAAE;IAC1F,MAAM,MAAM,GAAG,CAAC,SAAS,CAAC,OAAO,CAAC,CAAC,CAAC;IACpC,MAAM,MAAM,GAAG,iBAAiB,CAAC;QAChC,MAAM;QACN,GAAG,EAAE,cAAc;QACnB,YAAY,EAAE,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,OAAO,EAAE,sBAAsB,EAAE,CAAC;QACtE,aAAa,EAAE,CAAC,MAAM,EAAE,OAAO,CAAC;QAChC,WAAW,EAAE,GAAG,EAAE,CAAC,KAAK;KACxB,CAAC,CAAC;IAEH,MAAM,CAAC,EAAE,CAAC,MAAM,CAAC,QAAQ,CAAC,cAAc,CAAC,EAAE,yBAAyB,CAAC,CAAC;IACtE,MAAM,CAAC,EAAE,CAAC,MAAM,CAAC,QAAQ,CAAC,sBAAsB,CAAC,EAAE,mCAAmC,CAAC,CAAC;IACxF,MAAM,CAAC,EAAE,CAAC,CAAC,MAAM,CAAC,QAAQ,CAAC,oBAAoB,CAAC,EAAE,0CAA0C,CAAC,CAAC;AAC/F,CAAC,CAAC,CAAC;AAEH,6EAA6E;AAE7E,IAAI,CAAC,iGAAiG,EAAE,CAAC,CAAC,EAAE,EAAE;IAC7G,uEAAuE;IACvE,wEAAwE;IACxE,4DAA4D;IAC5D,MAAM,MAAM,GAAG,CAAC,SAAS,CAAC,OAAO,CAAC,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC;IAEvD,2EAA2E;IAC3E,MAAM,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC;IAClC,MAAM,QAAQ,GAAa,EAAE,CAAC;IAC9B,OAAO,CAAC,IAAI,GAAG,CAAC,GAAG,IAAe,EAAE,EAAE,GAAG,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAC1E,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,GAAG,OAAO,CAAC,IAAI,GAAG,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC;IAEhD,IAAI,MAAM,GAAG,EAAE,CAAC;IAChB,MAAM,CAAC,YAAY,CAAC,GAAG,EAAE;QACxB,MAAM,GAAG,iBAAiB,CAAC;YAC1B,MAAM;YACN,aAAa,EAAE,CAAC,MAAM,EAAE,OAAO,CAAC;YAChC,WAAW,EAAE,GAAG,EAAE,GAAG,MAAM,IAAI,KAAK,CAAC,cAAc,CAAC,CAAC,CAAC,CAAC;SACvD,CAAC,CAAC;IACJ,CAAC,CAAC,CAAC;IAEH,MAAM,OAAO,GAAG,sBAAsB,CAAC,MAAM,CAAC,CAAC;IAC/C,MAAM,CAAC,KAAK,CAAC,OAAO,EAAE,qBAAqB,EAAE,8CAA8C,CAAC,CAAC;IAC7F,MAAM,CAAC,KAAK,CAAC,OAAO,EAAE,oBAAoB,EAAE,6CAA6C,CAAC,CAAC;IAC3F,MAAM,CAAC,EAAE,CACR,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,mBAAmB,CAAC,IAAI,CAAC,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC,EACjF,6CAA6C,CAC7C,CAAC;AACH,CAAC,CAAC,CAAC","sourcesContent":["// @gsd/pi-coding-agent + system-prompt-skill-filter.test — coverage for the\n// optional `skillFilter` option added to buildSystemPrompt (RFC #4779). The\n// filter lets consumers narrow the <available_skills> catalog rendered into\n// the cached system prompt without touching skill loading or invocation.\n\nimport test from \"node:test\";\nimport assert from \"node:assert/strict\";\n\nimport { buildSystemPrompt } from \"../core/system-prompt.js\";\nimport type { Skill } from \"../core/skills.js\";\n\nfunction makeSkill(name: string, description = `description for ${name}`): Skill {\n\treturn {\n\t\tname,\n\t\tdescription,\n\t\tfilePath: `/tmp/${name}/SKILL.md`,\n\t\tbaseDir: `/tmp/${name}`,\n\t\tsource: \"project\",\n\t\tdisableModelInvocation: false,\n\t};\n}\n\nfunction extractAvailableSkills(prompt: string): string {\n\tconst start = prompt.indexOf(\"<available_skills>\");\n\tconst end = prompt.indexOf(\"</available_skills>\");\n\tif (start === -1 || end === -1) return \"\";\n\treturn prompt.slice(start, end + \"</available_skills>\".length);\n}\n\n// ─── Default branch (no customPrompt) ──────────────────────────────────────\n\ntest(\"buildSystemPrompt: skillFilter omits filtered-out skills from <available_skills>\", () => {\n\tconst skills = [makeSkill(\"alpha\"), makeSkill(\"beta\"), makeSkill(\"gamma\")];\n\tconst prompt = buildSystemPrompt({\n\t\tskills,\n\t\tselectedTools: [\"read\", \"Skill\"],\n\t\tskillFilter: skill => skill.name !== \"beta\",\n\t});\n\n\tconst section = extractAvailableSkills(prompt);\n\tassert.ok(section.length > 0, \"catalog section should render\");\n\tassert.match(section, /<name>alpha<\\/name>/);\n\tassert.match(section, /<name>gamma<\\/name>/);\n\tassert.doesNotMatch(section, /<name>beta<\\/name>/);\n});\n\ntest(\"buildSystemPrompt: skillFilter omitted preserves pre-filter behavior (all skills render)\", () => {\n\tconst skills = [makeSkill(\"alpha\"), makeSkill(\"beta\")];\n\tconst prompt = buildSystemPrompt({\n\t\tskills,\n\t\tselectedTools: [\"read\", \"Skill\"],\n\t});\n\n\tconst section = extractAvailableSkills(prompt);\n\tassert.match(section, /<name>alpha<\\/name>/);\n\tassert.match(section, /<name>beta<\\/name>/);\n});\n\ntest(\"buildSystemPrompt: skillFilter that rejects every skill suppresses the <available_skills> block\", () => {\n\tconst skills = [makeSkill(\"alpha\"), makeSkill(\"beta\")];\n\tconst prompt = buildSystemPrompt({\n\t\tskills,\n\t\tselectedTools: [\"read\", \"Skill\"],\n\t\tskillFilter: () => false,\n\t});\n\n\t// With zero visible skills, formatSkillsForPrompt returns an empty string,\n\t// so the opening tag should not appear anywhere.\n\tassert.ok(!prompt.includes(\"<available_skills>\"));\n});\n\n// ─── Custom-prompt branch ──────────────────────────────────────────────────\n\ntest(\"buildSystemPrompt (customPrompt): skillFilter applies to the catalog appended onto a custom prompt\", () => {\n\tconst skills = [makeSkill(\"alpha\"), makeSkill(\"beta\"), makeSkill(\"gamma\")];\n\tconst prompt = buildSystemPrompt({\n\t\tcustomPrompt: \"CUSTOM BASE\",\n\t\tskills,\n\t\tselectedTools: [\"read\", \"Skill\"],\n\t\tskillFilter: skill => skill.name === \"alpha\",\n\t});\n\n\tconst section = extractAvailableSkills(prompt);\n\tassert.match(section, /<name>alpha<\\/name>/);\n\tassert.doesNotMatch(section, /<name>beta<\\/name>/);\n\tassert.doesNotMatch(section, /<name>gamma<\\/name>/);\n});\n\n// ─── Interaction with disableModelInvocation ──────────────────────────────\n\ntest(\"buildSystemPrompt: skillFilter composes with disableModelInvocation (both must pass)\", () => {\n\t// A skill already hidden from the catalog by disableModelInvocation must\n\t// remain hidden even if skillFilter would otherwise admit it. The filter\n\t// narrows, it does not override the existing invisibility contract.\n\tconst skills: Skill[] = [\n\t\t{ ...makeSkill(\"visible\"), disableModelInvocation: false },\n\t\t{ ...makeSkill(\"hidden\"), disableModelInvocation: true },\n\t];\n\tconst prompt = buildSystemPrompt({\n\t\tskills,\n\t\tselectedTools: [\"read\", \"Skill\"],\n\t\tskillFilter: () => true,\n\t});\n\n\tconst section = extractAvailableSkills(prompt);\n\tassert.match(section, /<name>visible<\\/name>/);\n\tassert.doesNotMatch(section, /<name>hidden<\\/name>/);\n});\n\n// ─── Pass-through of non-filtered fields ──────────────────────────────────\n\ntest(\"buildSystemPrompt: skillFilter does not affect context files or cwd rendering\", () => {\n\tconst skills = [makeSkill(\"alpha\")];\n\tconst prompt = buildSystemPrompt({\n\t\tskills,\n\t\tcwd: \"/tmp/example\",\n\t\tcontextFiles: [{ path: \"CLAUDE.md\", content: \"project instructions\" }],\n\t\tselectedTools: [\"read\", \"Skill\"],\n\t\tskillFilter: () => false,\n\t});\n\n\tassert.ok(prompt.includes(\"/tmp/example\"), \"cwd should still render\");\n\tassert.ok(prompt.includes(\"project instructions\"), \"context files should still render\");\n\tassert.ok(!prompt.includes(\"<available_skills>\"), \"no skill catalog when filter rejects all\");\n});\n\n// ─── Exception safety ─────────────────────────────────────────────────────\n\ntest(\"buildSystemPrompt: skillFilter that throws falls back to unfiltered list and does not propagate\", (t) => {\n\t// A buggy consumer predicate must not bubble out of buildSystemPrompt.\n\t// If it did, _rebuildSystemPrompt could unwind mid-setTools() and leave\n\t// the session with updated tools but a stale system prompt.\n\tconst skills = [makeSkill(\"alpha\"), makeSkill(\"beta\")];\n\n\t// Suppress the console.warn the fallback emits so test output stays clean.\n\tconst originalWarn = console.warn;\n\tconst warnings: string[] = [];\n\tconsole.warn = (...args: unknown[]) => { warnings.push(args.join(\" \")); };\n\tt.after(() => { console.warn = originalWarn; });\n\n\tlet prompt = \"\";\n\tassert.doesNotThrow(() => {\n\t\tprompt = buildSystemPrompt({\n\t\t\tskills,\n\t\t\tselectedTools: [\"read\", \"Skill\"],\n\t\t\tskillFilter: () => { throw new Error(\"consumer bug\"); },\n\t\t});\n\t});\n\n\tconst section = extractAvailableSkills(prompt);\n\tassert.match(section, /<name>alpha<\\/name>/, \"alpha should render (fallback to unfiltered)\");\n\tassert.match(section, /<name>beta<\\/name>/, \"beta should render (fallback to unfiltered)\");\n\tassert.ok(\n\t\twarnings.some(w => w.includes(\"skillFilter threw\") && w.includes(\"consumer bug\")),\n\t\t\"fallback should emit an identifying warning\",\n\t);\n});\n"]}

package/packages/pi-coding-agent/src/core/agent-session-abort-order.test.ts

@@ -1,57 +1,133 @@
-import test, { describe } from "node:test";
+// Regression test for #4243 — abort() must be called BEFORE
+// _disconnectFromAgent() inside newSession() and switchSession() so that
+// message_end/agent_end events (and the #4216 finalization code) fire
+// before we unsubscribe from the event bus.
+//
+// Verified behaviourally: we construct a real AgentSession, wrap `abort`
+// and `_disconnectFromAgent` with call-order recording, trigger each
+// session-transition method, and assert the observed call order.
+
 import assert from "node:assert/strict";
-import { readFileSync } from "node:fs";
+import { mkdtempSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
 import { join } from "node:path";
+import { afterEach, beforeEach, describe, it } from "node:test";
+
+import { Agent } from "@gsd/pi-agent-core";
+import { AgentSession } from "./agent-session.js";
+import { AuthStorage } from "./auth-storage.js";
+import { ModelRegistry } from "./model-registry.js";
+import { DefaultResourceLoader } from "./resource-loader.js";
+import { SessionManager } from "./session-manager.js";
+import { SettingsManager } from "./settings-manager.js";
 
-const source = readFileSync(
-	join(process.cwd(), "packages/pi-coding-agent/src/core/agent-session.ts"),
-	"utf-8",
-);
+let testDir: string;
 
-describe("#4243 — abort() must be called before _disconnectFromAgent()", () => {
-	test("newSession() calls abort() before _disconnectFromAgent()", () => {
-		// Find the newSession method body where the fix was applied
-		const newSessionStart = source.indexOf("async newSession(options?:");
-		assert.ok(newSessionStart >= 0, "should find newSession method");
+async function createSession(opts: { persistSessions?: boolean } = {}): Promise<AgentSession> {
+	const agentDir = join(testDir, "agent-home");
+	const authStorage = AuthStorage.inMemory({});
+	const modelRegistry = new ModelRegistry(authStorage, join(agentDir, "models.json"));
+	const settingsManager = SettingsManager.inMemory();
+	const resourceLoader = new DefaultResourceLoader({
+		cwd: testDir,
+		agentDir,
+		settingsManager,
+		noExtensions: true,
+		noPromptTemplates: true,
+		noThemes: true,
+	});
+	await resourceLoader.reload();
 
-		// Get a window that includes the abort/disconnect section.
-		// Use 2000 chars to accommodate guard checks inserted between the two calls.
-		const window = source.slice(newSessionStart, newSessionStart + 2000);
+	// switchSession() needs a sessionFile; in-memory manager returns undefined.
+	// Use file-backed manager when the test needs to resume.
+	const sessionManager = opts.persistSessions
+		? SessionManager.create(testDir, join(testDir, "sessions"))
+		: SessionManager.inMemory(testDir);
+
+	return new AgentSession({
+		agent: new Agent(),
+		sessionManager,
+		settingsManager,
+		cwd: testDir,
+		resourceLoader,
+		modelRegistry,
+	});
+}
 
-		// Find the abort and _disconnectFromAgent calls
-		const abortIdx = window.indexOf("await this.abort();");
-		const disconnectIdx = window.indexOf("this._disconnectFromAgent();");
+/**
+ * Wrap two methods on the same object so their call order is recorded.
+ * Returns the recording array — assertions use index lookups.
+ */
+function recordCallOrder<O extends object>(
+	target: O,
+	methods: Array<keyof O>,
+): string[] {
+	const order: string[] = [];
+	for (const method of methods) {
+		const name = String(method);
+		const original = (target as any)[name] as (...args: unknown[]) => unknown;
+		if (typeof original !== "function") {
+			throw new Error(`recordCallOrder: ${name} is not a function on target`);
+		}
+		(target as any)[name] = function (this: O, ...args: unknown[]) {
+			order.push(name);
+			return original.apply(this, args);
+		};
+	}
+	return order;
+}
+
+describe("#4243 — abort() must run before _disconnectFromAgent()", () => {
+	beforeEach(() => {
+		testDir = mkdtempSync(join(tmpdir(), "agent-session-abort-"));
+	});
+
+	afterEach(() => {
+		rmSync(testDir, { recursive: true, force: true });
+	});
 
-		assert.ok(abortIdx >= 0, "newSession should call await this.abort()");
-		assert.ok(disconnectIdx >= 0, "newSession should call this._disconnectFromAgent()");
+	it("newSession() invokes abort() before _disconnectFromAgent()", async () => {
+		const session = await createSession();
+		const order = recordCallOrder(session as any, ["abort", "_disconnectFromAgent"]);
+
+		const ok = await session.newSession();
+		assert.equal(ok, true);
+
+		const abortIdx = order.indexOf("abort");
+		const disconnectIdx = order.indexOf("_disconnectFromAgent");
+		assert.ok(abortIdx >= 0, `newSession should call abort(); order=${order.join(",")}`);
+		assert.ok(
+			disconnectIdx >= 0,
+			`newSession should call _disconnectFromAgent(); order=${order.join(",")}`,
+		);
 		assert.ok(
 			abortIdx < disconnectIdx,
-			"abort() must be called BEFORE _disconnectFromAgent() so that message_end/agent_end events fire before unsubscribing from the event bus",
+			`abort() must run before _disconnectFromAgent(); order=${order.join(",")}`,
 		);
 	});
 
-	test("newSession() references #4243 in the abort/disconnect comment", () => {
-		const idx = source.indexOf("#4243");
-		assert.ok(idx >= 0, "source should reference issue #4243 for the abort-order fix");
-	});
-
-	test("switchSession() calls abort() before _disconnectFromAgent()", () => {
-		// Find the switchSession method body
-		const switchStart = source.indexOf("async switchSession(sessionPath:");
-		assert.ok(switchStart >= 0, "should find switchSession method");
+	it("switchSession() invokes abort() before _disconnectFromAgent()", async () => {
+		const session = await createSession({ persistSessions: true });
+		// Seed a session file to switch to (switchSession reads from the session manager).
+		await session.newSession();
+		const sessionFile = session.sessionFile;
+		assert.ok(typeof sessionFile === "string" && sessionFile.length > 0, "need a session file to switch to");
 
-		// Get a window that includes the abort/disconnect section
-		const window = source.slice(switchStart, switchStart + 800);
+		const order = recordCallOrder(session as any, ["abort", "_disconnectFromAgent"]);
 
-		// Find the abort and _disconnectFromAgent calls
-		const abortIdx = window.indexOf("await this.abort();");
-		const disconnectIdx = window.indexOf("this._disconnectFromAgent();");
+		const ok = await session.switchSession(sessionFile);
+		assert.equal(ok, true);
 
-		assert.ok(abortIdx >= 0, "switchSession should call await this.abort()");
-		assert.ok(disconnectIdx >= 0, "switchSession should call this._disconnectFromAgent()");
+		const abortIdx = order.indexOf("abort");
+		const disconnectIdx = order.indexOf("_disconnectFromAgent");
+		assert.ok(abortIdx >= 0, `switchSession should call abort(); order=${order.join(",")}`);
+		assert.ok(
+			disconnectIdx >= 0,
+			`switchSession should call _disconnectFromAgent(); order=${order.join(",")}`,
+		);
 		assert.ok(
 			abortIdx < disconnectIdx,
-			"abort() must be called BEFORE _disconnectFromAgent() in switchSession so that events fire before unsubscribing",
+			`abort() must run before _disconnectFromAgent() in switchSession; order=${order.join(",")}`,
 		);
 	});
 });

package/packages/pi-coding-agent/src/core/agent-session-model-switch.test.ts

@@ -1,21 +1,93 @@
-import test from "node:test";
+// Regression test: explicit model switches must cancel any in-flight retry
+// BEFORE applying the new model. Otherwise stale provider backoff errors
+// from the previous model can continue to land after the switch.
+//
+// Verified behaviourally: construct a real AgentSession, wrap
+// `_retryHandler.abortRetry` and `agent.setModel` to record call order,
+// invoke setModel(), and assert the observed order.
+
 import assert from "node:assert/strict";
-import { readFileSync } from "node:fs";
+import { mkdtempSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
 import { join } from "node:path";
+import { afterEach, beforeEach, describe, it } from "node:test";
+
+import { Agent } from "@gsd/pi-agent-core";
+import { getModel } from "@gsd/pi-ai";
+import { AgentSession } from "./agent-session.js";
+import { AuthStorage } from "./auth-storage.js";
+import { ModelRegistry } from "./model-registry.js";
+import { DefaultResourceLoader } from "./resource-loader.js";
+import { SessionManager } from "./session-manager.js";
+import { SettingsManager } from "./settings-manager.js";
+
+let testDir: string;
+
+async function createSession(): Promise<AgentSession> {
+	const agentDir = join(testDir, "agent-home");
+	const authStorage = AuthStorage.inMemory({});
+	// Seed a runtime anthropic API key so modelRegistry.isProviderRequestReady()
+	// returns true and setModel() doesn't throw on missing credentials.
+	authStorage.setRuntimeApiKey("anthropic", "sk-test-not-used");
+	const modelRegistry = new ModelRegistry(authStorage, join(agentDir, "models.json"));
+	const settingsManager = SettingsManager.inMemory();
+	const resourceLoader = new DefaultResourceLoader({
+		cwd: testDir,
+		agentDir,
+		settingsManager,
+		noExtensions: true,
+		noPromptTemplates: true,
+		noThemes: true,
+	});
+	await resourceLoader.reload();
+
+	return new AgentSession({
+		agent: new Agent(),
+		sessionManager: SessionManager.inMemory(testDir),
+		settingsManager,
+		cwd: testDir,
+		resourceLoader,
+		modelRegistry,
+	});
+}
+
+describe("AgentSession — explicit model switch cancels retry before applying new model", () => {
+	beforeEach(() => {
+		testDir = mkdtempSync(join(tmpdir(), "agent-session-model-switch-"));
+	});
+
+	afterEach(() => {
+		rmSync(testDir, { recursive: true, force: true });
+	});
+
+	it("setModel() calls _retryHandler.abortRetry() before agent.setModel()", async () => {
+		const session = await createSession();
+
+		const order: string[] = [];
+		const retryHandler = (session as any)._retryHandler;
+		const originalAbortRetry = retryHandler.abortRetry.bind(retryHandler);
+		retryHandler.abortRetry = () => {
+			order.push("abortRetry");
+			return originalAbortRetry();
+		};
+
+		const agent = (session as any).agent;
+		const originalSetModel = agent.setModel.bind(agent);
+		agent.setModel = (model: unknown) => {
+			order.push("setModel");
+			return originalSetModel(model);
+		};
+
+		const newModel = getModel("anthropic", "claude-3-5-sonnet-20241022");
+		await session.setModel(newModel, { persist: false });
 
-const source = readFileSync(join(process.cwd(), "packages/pi-coding-agent/src/core/agent-session.ts"), "utf-8");
-
-test("agent-session: explicit model switches cancel retry before applying new model", () => {
-	const start = source.indexOf("private async _applyModelChange(");
-	assert.ok(start >= 0, "missing _applyModelChange");
-	const window = source.slice(start, start + 900);
-	const abortIdx = window.indexOf("this._retryHandler.abortRetry();");
-	const setModelIdx = window.indexOf("this.agent.setModel(model);");
-
-	assert.ok(abortIdx >= 0, "_applyModelChange should cancel any in-flight retry");
-	assert.ok(setModelIdx >= 0, "_applyModelChange should set the new model");
-	assert.ok(
-		abortIdx < setModelIdx,
-		"retry cancellation must happen before applying the new model to prevent stale provider retries",
-	);
+		const abortIdx = order.indexOf("abortRetry");
+		const setIdx = order.indexOf("setModel");
+		assert.ok(abortIdx >= 0, `setModel should cancel in-flight retry; order=${order.join(",")}`);
+		assert.ok(setIdx >= 0, `setModel should call agent.setModel; order=${order.join(",")}`);
+		assert.ok(
+			abortIdx < setIdx,
+			`retry cancellation must happen before applying the new model; order=${order.join(",")}`,
+		);
+	});
 });

package/packages/pi-coding-agent/src/core/agent-session-tool-refresh.test.ts

@@ -1,64 +1,133 @@
-// GSD-2 — Regression tests for #3616: tool list persistence across newSession() calls
-// Copyright (c) 2026 Jeremy McSpadden <jeremy@fluxlabs.net>
+// Regression test for #3616: newSession() must restore the full tool set
+// when cwd is unchanged.
+//
+// The bug: extensions may narrow the active tool list via setActiveTools()
+// during a session. Without a refresh in the else branch of newSession(),
+// the narrowed set persists into the next session — breaking auto-mode
+// subagent sessions that expect a full tool palette.
+//
+// Verified behaviourally: construct an AgentSession, wrap _refreshToolRegistry
+// to record its args, call newSession() with cwd unchanged, and assert that
+// a refresh was requested with includeAllExtensionTools: true.
 
-import test, { describe } from "node:test";
 import assert from "node:assert/strict";
-import { readFileSync } from "node:fs";
+import { mkdtempSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
 import { join } from "node:path";
+import { afterEach, beforeEach, describe, it } from "node:test";
 
-const source = readFileSync(
-	join(process.cwd(), "packages/pi-coding-agent/src/core/agent-session.ts"),
-	"utf-8",
-);
+import { Agent } from "@gsd/pi-agent-core";
+import { AgentSession } from "./agent-session.js";
+import { AuthStorage } from "./auth-storage.js";
+import { ModelRegistry } from "./model-registry.js";
+import { DefaultResourceLoader } from "./resource-loader.js";
+import { SessionManager } from "./session-manager.js";
+import { SettingsManager } from "./settings-manager.js";
 
-describe("#3616 — newSession() must restore full tool set", () => {
-	test("newSession() calls _refreshToolRegistry with includeAllExtensionTools when cwd is unchanged", () => {
-		// Find the newSession method
-		const newSessionStart = source.indexOf("async newSession(options?:");
-		assert.ok(newSessionStart >= 0, "should find newSession method");
+let testDir: string;
 
-		// Get the method body (up to the next top-level method)
-		const methodBody = source.slice(newSessionStart, newSessionStart + 3000);
+async function createSession(): Promise<AgentSession> {
+	const agentDir = join(testDir, "agent-home");
+	const authStorage = AuthStorage.inMemory({});
+	const modelRegistry = new ModelRegistry(authStorage, join(agentDir, "models.json"));
+	const settingsManager = SettingsManager.inMemory();
+	const resourceLoader = new DefaultResourceLoader({
+		cwd: testDir,
+		agentDir,
+		settingsManager,
+		noExtensions: true,
+		noPromptTemplates: true,
+		noThemes: true,
+	});
+	await resourceLoader.reload();
 
-		// Verify the cwd-changed branch rebuilds tools
-		assert.ok(
-			methodBody.includes("if (this._cwd !== previousCwd)"),
-			"should have cwd-change guard",
-		);
+	return new AgentSession({
+		agent: new Agent(),
+		sessionManager: SessionManager.inMemory(testDir),
+		settingsManager,
+		cwd: testDir,
+		resourceLoader,
+		modelRegistry,
+	});
+}
+
+describe("#3616 — newSession() restores narrowed tool set when cwd unchanged", () => {
+	beforeEach(() => {
+		testDir = mkdtempSync(join(tmpdir(), "agent-session-tool-refresh-"));
+	});
+
+	afterEach(() => {
+		rmSync(testDir, { recursive: true, force: true });
+	});
 
-		// Verify the else branch exists and refreshes tools with includeAllExtensionTools
-		const elseIdx = methodBody.indexOf("} else {");
-		assert.ok(elseIdx >= 0, "should have else branch for cwd-unchanged case");
+	it("calls _refreshToolRegistry with includeAllExtensionTools: true when cwd unchanged", async () => {
+		const session = await createSession();
+		// Pin _cwd so newSession()'s `process.cwd()` branch takes the
+		// cwd-unchanged path. The production code compares `this._cwd !==
+		// previousCwd`; we force equality by setting _cwd to current cwd.
+		(session as any)._cwd = process.cwd();
+
+		const refreshCalls: Array<{ includeAllExtensionTools?: boolean }> = [];
+		const originalRefresh = (session as any)._refreshToolRegistry.bind(session);
+		(session as any)._refreshToolRegistry = (options?: { includeAllExtensionTools?: boolean }) => {
+			refreshCalls.push(options ?? {});
+			return originalRefresh(options);
+		};
+
+		const ok = await session.newSession();
+		assert.equal(ok, true);
 
-		const elseBranch = methodBody.slice(elseIdx, elseIdx + 800);
 		assert.ok(
-			elseBranch.includes("_refreshToolRegistry"),
-			"else branch should call _refreshToolRegistry",
+			refreshCalls.length > 0,
+			"newSession() should invoke _refreshToolRegistry in the cwd-unchanged branch",
 		);
 		assert.ok(
-			elseBranch.includes("includeAllExtensionTools: true"),
-			"else branch should pass includeAllExtensionTools: true to restore narrowed tools",
+			refreshCalls.some((o) => o.includeAllExtensionTools === true),
+			`at least one _refreshToolRegistry call must pass includeAllExtensionTools: true; observed=${JSON.stringify(refreshCalls)}`,
 		);
 	});
 
-	test("newSession() references #3616 in the else-branch comment", () => {
-		const idx = source.indexOf("#3616");
-		assert.ok(idx >= 0, "source should reference issue #3616 for the tool restore fix");
+	it("agent.reset() does not clear _state.tools (tools persist across reset)", () => {
+		// Structural invariant protecting #3616: if reset() starts clearing
+		// tools, newSession()'s refresh becomes the only defense against loss.
+		// Assertion is behavioural — seed tools, call reset(), observe survival.
+		const agent = new Agent();
+		const tool = {
+			name: "test_tool",
+			description: "x",
+			schema: { type: "object", properties: {}, additionalProperties: false } as any,
+			execute: async () => ({ content: [] }),
+		};
+		(agent as any)._state.tools = [tool];
+		agent.reset();
+		assert.deepEqual(
+			(agent as any)._state.tools,
+			[tool],
+			"Agent.reset() must preserve _state.tools",
+		);
 	});
 
-	test("agent.reset() does not clear _state.tools (tools persist across reset)", () => {
-		// This is a structural invariant — if reset() starts clearing tools,
-		// the newSession() refresh becomes the only defense against tool loss.
-		const agentSource = readFileSync(
-			join(process.cwd(), "packages/pi-agent-core/src/agent.ts"),
-			"utf-8",
-		);
-		const resetStart = agentSource.indexOf("reset()");
-		assert.ok(resetStart >= 0, "should find reset() method");
-		const resetBody = agentSource.slice(resetStart, resetStart + 400);
+	it("takes the cwd-changed branch (rebuilds runtime) when cwd differs", async () => {
+		const session = await createSession();
+		// Force the cwd-changed branch: set _cwd to something that won't equal process.cwd().
+		(session as any)._cwd = join(testDir, "some", "other", "cwd");
+
+		let buildRuntimeCalled = false;
+		let buildRuntimeIncludedAll = false;
+		const originalBuild = (session as any)._buildRuntime.bind(session);
+		(session as any)._buildRuntime = (options?: { includeAllExtensionTools?: boolean }) => {
+			buildRuntimeCalled = true;
+			if (options?.includeAllExtensionTools === true) buildRuntimeIncludedAll = true;
+			return originalBuild(options);
+		};
+
+		const ok = await session.newSession();
+		assert.equal(ok, true);
+
+		assert.ok(buildRuntimeCalled, "cwd-changed branch must rebuild the tool runtime");
 		assert.ok(
-			!resetBody.includes("tools"),
-			"reset() should NOT touch _state.tools — tools are managed by agent-session",
+			buildRuntimeIncludedAll,
+			"cwd-changed branch must rebuild with includeAllExtensionTools: true",
 		);
 	});
 });