W1: plan-tree tools — pure validated ops, eight server-side plan_* tools, tree persistence, teacher prompt fragment

- src/core/plan-tree.ts: pure tree ops (createTree/addNode/updateNode/removeNode/
  addEdge/removeEdge/setCurrent/markDone), DAG invariants (unique ids, existing
  endpoints, no self-loops/duplicates/cycles), cascade remove, version bumps on
  state changes only (true no-ops for idempotent setCurrent/markDone).
- src/agent/tools/plan.ts: plan_init (refuses re-init), plan_add_node (optional
  parents → requires edges), plan_update_node, plan_remove_node, plan_add_edge,
  plan_remove_edge, plan_set_current, plan_mark_done. Execute = getChat → op →
  saveTree → PlanToolResult with the full updated tree; PlanTreeError and missing
  chats become {ok:false, error, tree}.
- src/agent/prompts/10-plan-tree.md: tree-authoring doctrine (when: post-probe;
  grain: one reasoning step per node; discipline: set_current on start,
  mark_done on demonstrated mastery; freedom: mid-session restructure).
- tests/plan-tree.test.ts: every validation error path, version semantics,
  cascade remove, no-mutation, tool round-trips against the real chat store.
This commit is contained in:
Amos
2026-08-22 09:27:27 +03:00
parent 6d9ca54d6b
commit 1e18d804fc
4 changed files with 796 additions and 8 deletions
+226
View File
@@ -0,0 +1,226 @@
import type { PlanEdge, PlanNode, PlanTree } from "./contracts.ts";
/**
* Pure, validated plan-tree operations (vision §2 "the formal plan tree").
*
* Every function takes a tree and returns a NEW tree — never mutates its
* input. All validation throws PlanTreeError with a message written for the
* agent (it flows straight back as a tool-result error), stating exactly what
* is wrong and which id/edge is at fault.
*
* Invariants enforced everywhere:
* - Node ids are unique and non-empty; every edge endpoint references an
* existing node.
* - The structure is a DAG: no self-loops, no duplicate edges, no cycles.
* - `version` increments on every successful STATE-CHANGING op. True no-ops
* (setCurrent to the already-current node, markDone of an already-done
* node) return the input tree unchanged — version drives UI re-animation,
* so a non-change must not tick it. createTree always starts at version 1.
*/
/** Error thrown by every op on invalid input. Message is agent-readable. */
export class PlanTreeError extends Error {
constructor(message: string) {
super(message);
this.name = "PlanTreeError";
}
}
/** An empty, never-stored tree used as the seed before plan_init runs. */
export function emptyTree(): PlanTree {
return { nodes: [], edges: [], currentNodeId: null, doneNodeIds: [], version: 0 };
}
// ─── validation helpers ──────────────────────────────────────────────────────
function assertValidNode(node: PlanNode): void {
if (!node || typeof node.id !== "string" || node.id.trim() === "") {
throw new PlanTreeError("Every node needs a non-empty `id` (a stable slug like 'chain-rule-intuition').");
}
if (typeof node.title !== "string" || node.title.trim() === "") {
throw new PlanTreeError(`Node "${node.id}" needs a non-empty \`title\`.`);
}
}
function nodeMap(nodes: PlanNode[]): Map<string, PlanNode> {
const map = new Map<string, PlanNode>();
for (const node of nodes) {
assertValidNode(node);
if (map.has(node.id)) {
throw new PlanTreeError(`Duplicate node id "${node.id}" — node ids must be unique.`);
}
map.set(node.id, node);
}
return map;
}
function requireNode(map: Map<string, PlanNode>, id: string, what: string): void {
if (!map.has(id)) {
throw new PlanTreeError(`${what} references unknown node id "${id}" — it is not in the tree.`);
}
}
function edgeKey(from: string, to: string): string {
return `${from}\u0000${to}`;
}
/**
* Would adding edge from→to create a cycle? Walk forward from `to` along
* existing edges; if we can reach `from`, the new edge closes a loop.
*/
function createsCycle(edges: PlanEdge[], from: string, to: string): boolean {
const adjacency = new Map<string, string[]>();
for (const e of edges) {
const list = adjacency.get(e.from);
if (list) list.push(e.to);
else adjacency.set(e.from, [e.to]);
}
const stack = [to];
const seen = new Set<string>();
while (stack.length > 0) {
const current = stack.pop()!;
if (current === from) return true;
if (seen.has(current)) continue;
seen.add(current);
for (const next of adjacency.get(current) ?? []) stack.push(next);
}
return false;
}
/** Full edge validation against existing nodes/edges: endpoints exist, no self-loop, no duplicate, acyclic. */
function assertValidEdge(from: string, to: string, nodes: PlanNode[], existingEdges: PlanEdge[]): void {
const map = nodeMap(nodes); // also re-validates nodes on createTree
requireNode(map, from, "Edge `from`");
requireNode(map, to, "Edge `to`");
if (from === to) {
throw new PlanTreeError(`Edge "${from}" → "${to}" is a self-loop — a node can't require itself.`);
}
if (existingEdges.some((e) => e.from === from && e.to === to)) {
throw new PlanTreeError(`Edge "${from}" → "${to}" already exists — duplicate edges are not allowed.`);
}
if (createsCycle(existingEdges, from, to)) {
throw new PlanTreeError(
`Edge "${from}" → "${to}" would create a cycle: "${to}" already leads (directly or transitively) back to "${from}". ` +
`"from" must be understood before "to" — the tree is a DAG.`,
);
}
}
// ─── ops ─────────────────────────────────────────────────────────────────────
/** Build a fresh tree (version 1) from node and edge lists, validating all invariants. */
export function createTree(nodes: PlanNode[], edges: PlanEdge[]): PlanTree {
const map = nodeMap(nodes);
for (const e of edges) {
requireNode(map, e.from, "Edge `from`");
requireNode(map, e.to, "Edge `to`");
}
const accepted: PlanEdge[] = [];
for (const e of edges) {
// Each edge is checked against the ones already accepted, so any
// violation names the culprit edge.
assertValidEdge(e.from, e.to, nodes, accepted);
accepted.push({ from: e.from, to: e.to });
}
return { nodes: nodes.map((n) => ({ ...n })), edges: accepted, currentNodeId: null, doneNodeIds: [], version: 1 };
}
/** Add a node, optionally creating edges parent → new node for each id in `parents`. */
export function addNode(tree: PlanTree, node: PlanNode, parents?: string[]): PlanTree {
assertValidNode(node);
if (tree.nodes.some((n) => n.id === node.id)) {
throw new PlanTreeError(`Duplicate node id "${node.id}" — a node with that id already exists in the tree.`);
}
const parentList = parents ?? [];
if (parentList.includes(node.id)) {
throw new PlanTreeError(`Node "${node.id}" can't list itself as a parent.`);
}
const newEdges: PlanEdge[] = [];
for (const parent of parentList) {
// Validates existence (including the new node itself as `to`),
// self-loop, duplicate (against tree edges plus the new edges being
// added), and acyclicity.
assertValidEdge(parent, node.id, [...tree.nodes, node], [...tree.edges, ...newEdges]);
newEdges.push({ from: parent, to: node.id });
}
return {
...tree,
nodes: [...tree.nodes, { ...node }],
edges: [...tree.edges, ...newEdges],
version: tree.version + 1,
};
}
/** Update a node's title and/or summary. The id must exist; ids are immutable. */
export function updateNode(tree: PlanTree, id: string, patch: { title?: string; summary?: string }): PlanTree {
const index = tree.nodes.findIndex((n) => n.id === id);
if (index === -1) {
throw new PlanTreeError(`Cannot update node "${id}" — it is not in the tree.`);
}
const current = tree.nodes[index]!;
const next: PlanNode = { ...current };
if (patch.title !== undefined) next.title = patch.title;
if (patch.summary !== undefined) next.summary = patch.summary;
assertValidNode(next);
const nodes = tree.nodes.slice();
nodes[index] = next;
return { ...tree, nodes, version: tree.version + 1 };
}
/**
* Remove a node and cascade: every edge touching it (both directions),
* currentNodeId (→ null if it pointed here), and its doneNodeIds entry.
*/
export function removeNode(tree: PlanTree, id: string): PlanTree {
if (!tree.nodes.some((n) => n.id === id)) {
throw new PlanTreeError(`Cannot remove node "${id}" — it is not in the tree.`);
}
return {
...tree,
nodes: tree.nodes.filter((n) => n.id !== id),
edges: tree.edges.filter((e) => e.from !== id && e.to !== id),
currentNodeId: tree.currentNodeId === id ? null : tree.currentNodeId,
doneNodeIds: tree.doneNodeIds.filter((d) => d !== id),
version: tree.version + 1,
};
}
/** Add a `requires` edge from → to. Rejects self-loops, duplicates, and cycles. */
export function addEdge(tree: PlanTree, from: string, to: string): PlanTree {
assertValidEdge(from, to, tree.nodes, tree.edges);
return { ...tree, edges: [...tree.edges, { from, to }], version: tree.version + 1 };
}
/** Remove an existing edge. Removing a nonexistent edge is an error. */
export function removeEdge(tree: PlanTree, from: string, to: string): PlanTree {
const index = tree.edges.findIndex((e) => e.from === from && e.to === to);
if (index === -1) {
throw new PlanTreeError(`Cannot remove edge "${from}" → "${to}" — no such edge exists.`);
}
return { ...tree, edges: tree.edges.filter((e) => e.from !== from || e.to !== to), version: tree.version + 1 };
}
/**
* Set currentNodeId ("where we are"). Requires the node exists. Does NOT
* touch doneNodeIds. Setting the already-current node is a true no-op (no
* version bump).
*/
export function setCurrent(tree: PlanTree, id: string): PlanTree {
if (!tree.nodes.some((n) => n.id === id)) {
throw new PlanTreeError(`Cannot set current node "${id}" — it is not in the tree.`);
}
if (tree.currentNodeId === id) return tree;
return { ...tree, currentNodeId: id, version: tree.version + 1 };
}
/**
* Mark a node done. Requires the node exists; idempotent — marking an
* already-done node is a true no-op (no version bump).
*/
export function markDone(tree: PlanTree, id: string): PlanTree {
if (!tree.nodes.some((n) => n.id === id)) {
throw new PlanTreeError(`Cannot mark node "${id}" done — it is not in the tree.`);
}
if (tree.doneNodeIds.includes(id)) return tree;
return { ...tree, doneNodeIds: [...tree.doneNodeIds, id], version: tree.version + 1 };
}