Skip to content

feat(agent-interface): bind exact candidate experiments #59

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
drewstone merged 1 commit into from
Jul 16, 2026
Merged

feat(agent-interface): bind exact candidate experiments #59

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 5 additions & 0 deletions .changeset/add-candidate-benchmark-task.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,5 @@
---
"@tangle-network/agent-interface": minor
---

Add one immutable candidate experiment shared by evaluation and execution, with exact baseline and candidate bundles, benchmark tasks, repetitions, seeds, and Runtime receipts.
2 changes: 1 addition & 1 deletion packages/agent-interface/package.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,7 +46,7 @@
],
"scripts": {
"build": "tsc -p tsconfig.json",
"check-types": "tsc --noEmit",
"check-types": "tsc --noEmit && tsc --noEmit -p tsconfig.test.json",
"clean": "rm -rf dist tsconfig.tsbuildinfo",
"test": "vitest run",
"test:watch": "vitest",
Expand Down
4 changes: 1 addition & 3 deletions packages/agent-interface/src/agent-candidate-code-schema.test.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,9 +21,7 @@ const instructionDelivery = { kind: "argv-append" as const };

describe("candidate code and execution schemas", () => {
it("distinguishes disabled, proposer no-op, and real changes", () => {
expect(() =>
agentCandidateCodeSchema.parse({ kind: "disabled", reason: "control" }),
).not.toThrow();
expect(() => agentCandidateCodeSchema.parse({ kind: "disabled" })).not.toThrow();
expect(() =>
agentCandidateCodeSchema.parse({
kind: "no-op",
Expand Down
1 change: 0 additions & 1 deletion packages/agent-interface/src/agent-candidate-code-schema.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,7 +62,6 @@ const executableSchema = z
export const agentCandidateCodeDisabledSchema = z
.object({
kind: z.literal("disabled"),
reason: z.enum(["control", "not-applicable"]),
})
.strict() satisfies z.ZodType<AgentCandidateCodeDisabled>;

Expand Down
242 changes: 58 additions & 184 deletions packages/agent-interface/src/agent-candidate-execution-plan-schema.test.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,37 +47,20 @@ function planFixture() {
};
return {
kind: "agent-candidate-execution-plan-material" as const,
bundleDigest: candidateSha("1"),
executionId: "execution-1",
attempt: {
number: 1,
maxAttempts: 1,
retryPolicy: "pre-model-infrastructure-only" as const,
},
task: {
benchmark: "pier",
benchmarkVersion: "0.3",
taskId: "task-1",
splitDigest: candidateSha("2"),
instruction: {
encoding: "utf8" as const,
sha256: candidateSha("3"),
byteLength: 42,
delivery: {
kind: "utf8-file" as const,
env: "TANGLE_CANDIDATE_TASK_PATH" as const,
path: "/tangle/input/task.txt" as const,
},
},
repository: {
identity: "tangle-network/agent-runtime",
rootIdentity: "tangle-network/agent-runtime",
baseCommit: "1".repeat(40),
baseTree: "2".repeat(40),
},
outcome: { kind: "workspace" as const },
workspace: workspace("src/task.ts", "4"),
runCell: {
kind: "agent-candidate-run-cell" as const,
experimentDigest: candidateSha("0"),
arm: "candidate" as const,
bundleDigest: candidateSha("1"),
suiteDigest: candidateSha("9"),
taskDigest: candidateSha("2"),
taskIndex: 0,
repetition: 0,
seed: 42,
attempt: 1,
digest: candidateSha("3"),
},
executionId: "execution-1",
workspaces: {
taskRoot: "/work/task",
candidateRoot: "/work/candidate",
Expand All @@ -91,6 +74,19 @@ function planFixture() {
},
harness: "codex" as const,
harnessVersion: "0.1.0",
instructionDelivery: {
kind: "utf8-file" as const,
env: "TANGLE_CANDIDATE_TASK_PATH" as const,
path: "/tangle/input/task.txt" as const,
},
limits: {
timeoutMs: 120_000,
maxSteps: 50,
maxModelCalls: 12,
maxInputTokens: 100_000,
maxOutputTokens: 20_000,
maxCostUsd: 5,
},
container: {
source: "evaluator-task-container" as const,
image: "pier-task:0.3",
Expand Down Expand Up @@ -118,15 +114,6 @@ function planFixture() {
},
],
},
grader: {
name: "fixture-grader",
version: "1.0.0",
artifact: {
locator: { kind: "s3" as const, bucket: "test-artifacts", key: "grader" },
sha256: candidateSha("a"),
byteLength: 1,
},
},
launch: {
executable: "node",
args: [{ kind: "public" as const, value: "dist/agent.js" }],
Expand All @@ -139,14 +126,6 @@ function planFixture() {
cwd: { workspace: "task" as const, path: "." },
},
memory: { mode: "disabled" as const },
limits: {
timeoutMs: 60_000,
maxSteps: 200,
maxModelCalls: 20,
maxInputTokens: 200_000,
maxOutputTokens: 20_000,
maxCostUsd: 20,
},
network: { mode: "disabled" as const },
};
}
Expand All @@ -157,6 +136,7 @@ describe("agentCandidateExecutionPlanMaterialSchema", () => {
kind: "agent-profile-workspace-plan" as const,
digest: candidateSha("1"),
material: {
sourceProfileDigest: candidateSha("0"),
harness: "codex" as const,
files: [
{
Expand Down Expand Up @@ -223,36 +203,8 @@ describe("agentCandidateExecutionPlanMaterialSchema", () => {
).toThrow(/Unrecognized key/);
});

it("requires the split, fixed disjoint roots, and evaluator model access", () => {
it("requires fixed disjoint roots and evaluator model access", () => {
const plan = planFixture();
const { splitDigest: _splitDigest, ...taskWithoutSplit } = plan.task;
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
task: taskWithoutSplit,
}),
).toThrow();
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
task: {
...plan.task,
repository: {
...plan.task.repository,
baseTree: "2".repeat(64),
},
},
}),
).toThrow(/one object format/);
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
task: {
...plan.task,
instruction: { ...plan.task.instruction, byteLength: 0 },
},
}),
).toThrow();
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
Expand All @@ -262,71 +214,45 @@ describe("agentCandidateExecutionPlanMaterialSchema", () => {
},
}),
).toThrow(/disjoint/);
const { access: _access, ...modelWithoutAccess } = plan.model;
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
model: modelWithoutAccess,
workspaces: { ...plan.workspaces, taskRoot: "/tangle/input" },
}),
).toThrow();
});

it("accepts a bounded exact output instead of a Git result", () => {
const plan = planFixture();
const outputPlan = {
...plan,
task: {
...plan.task,
workspace: {
...plan.task.workspace,
material: { ...plan.task.workspace.material, files: [] },
},
outcome: {
kind: "output" as const,
mediaType: "application/json",
maxBytes: 1_048_576,
},
},
};
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse(outputPlan),
).not.toThrow();
const { repository: _repository, ...taskWithoutRepository } = outputPlan.task;
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...outputPlan,
task: taskWithoutRepository,
}),
).not.toThrow();
).toThrow(/task file delivery must be outside/);
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...outputPlan,
task: {
...outputPlan.task,
outcome: { ...outputPlan.task.outcome, mediaType: "Application/JSON" },
...plan,
launch: {
...plan.launch,
env: {
TANGLE_CANDIDATE_TASK_PATH: {
kind: "public",
value: "/tmp/task.txt",
},
},
},
}),
).toThrow();
).toThrow(/fixed evaluator-owned path/);
const { access: _access, ...modelWithoutAccess } = plan.model;
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...outputPlan,
task: {
...outputPlan.task,
outcome: { ...outputPlan.task.outcome, maxBytes: 64 * 1024 * 1024 + 1 },
},
...plan,
model: modelWithoutAccess,
}),
).toThrow();
});

it("requires repository provenance only for workspace outcomes", () => {
it("rejects task-owned fields so they cannot drift from the signed suite", () => {
const plan = planFixture();
const { repository: _repository, ...taskWithoutRepository } = plan.task;
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
task: taskWithoutRepository,
}),
).toThrow(/require task repository provenance/);
for (const field of ["task", "grader", "limits"] as const) {
expect(
agentCandidateExecutionPlanMaterialSchema.safeParse({
...plan,
[field]: {},
}).success,
).toBe(false);
}
});

it("rejects ambiguous roots and silently omitted profile behavior", () => {
Expand All @@ -341,6 +267,7 @@ describe("agentCandidateExecutionPlanMaterialSchema", () => {
}

const profilePlan = {
sourceProfileDigest: candidateSha("0"),
harness: "codex",
files: [],
env: {},
Expand Down Expand Up @@ -398,7 +325,7 @@ describe("agentCandidateExecutionPlanMaterialSchema", () => {
expect(() => agentCandidateResolvedModelSchema.parse(withoutEffort)).toThrow();
});

it("freezes only exact model-gateway domains and disables them for zero-call plans", () => {
it("freezes only exact model-gateway domains", () => {
const plan = planFixture();
for (const domain of [
"*.tangle.tools",
Expand Down Expand Up @@ -436,75 +363,22 @@ describe("agentCandidateExecutionPlanMaterialSchema", () => {
},
}),
).toThrow(/lexicographically sorted/);
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
limits: { ...plan.limits, maxModelCalls: 0 },
model: {
...plan.model,
access: { ...plan.model.access, network: { mode: "disabled" } },
},
}),
).not.toThrow();
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
model: {
...plan.model,
access: { ...plan.model.access, network: { mode: "disabled" } },
},
}),
).toThrow(/require one frozen gateway allowlist/);
});

it("freezes counted attempts, retry policy, and tool-loop steps", () => {
it("keeps only the exact attempt number in the run cell", () => {
const plan = planFixture();
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
attempt: { ...plan.attempt, number: 2 },
}),
).toThrow(/cannot exceed/);
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
attempt: { number: 1, maxAttempts: 2, retryPolicy: "none" },
}),
).toThrow(/exactly one attempt/);
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
limits: { ...plan.limits, maxSteps: 0 },
runCell: { ...plan.runCell, attempt: 0 },
}),
).toThrow();
});

it("binds exact task-instruction delivery outside both workspaces", () => {
const plan = planFixture();
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
launch: { ...plan.launch, env: {} },
}),
).toThrow(/signed fixed task path/);
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
workspaces: { ...plan.workspaces, taskRoot: "/tangle" },
}),
).toThrow(/outside both workspaces/);
expect(() =>
agentCandidateExecutionPlanMaterialSchema.parse({
...plan,
task: {
...plan.task,
instruction: {
...plan.task.instruction,
delivery: { kind: "argv-append" },
},
},
runCell: { ...plan.runCell, maxAttempts: 2 },
}),
).toThrow(/cannot expose an instruction path/);
).toThrow(/Unrecognized key/);
});

it("requires active workspace bytes and fresh reset evidence", () => {
Expand Down
Loading
Loading