@sweny-ai/core 0.1.100 → 0.1.102

package/dist/claude.d.ts

@@ -44,6 +44,7 @@ export declare class ClaudeClient implements Claude {
         outputSchema?: JSONSchema;
         onProgress?: (message: string) => void;
         maxTurns?: number;
+        disallowedTools?: string[];
     }): Promise<NodeResult>;
     evaluate(opts: {
         question: string;

package/dist/claude.js

@@ -40,7 +40,7 @@ export class ClaudeClient {
         return env;
     }
     async run(opts) {
-        const { instruction, context, tools, outputSchema, onProgress, maxTurns } = opts;
+        const { instruction, context, tools, outputSchema, onProgress, maxTurns, disallowedTools } = opts;
         // Tool-call accounting (Fix #1).
         //
         // Stream-driven recording, correlated by tool_use_id:
@@ -96,6 +96,7 @@ export class ClaudeClient {
                     stderr: (data) => this.logger.debug(`[claude-code] ${data}`),
                     ...(this.model ? { model: this.model } : {}),
                     ...(Object.keys(allMcpServers).length > 0 ? { mcpServers: allMcpServers } : {}),
+                    ...(disallowedTools && disallowedTools.length > 0 ? { disallowedTools } : {}),
                 },
             });
             for await (const message of stream) {
@@ -371,6 +372,7 @@ export function buildEvaluatePrompt(question, context, choices) {
         `1. Read each choice's condition literally and match against the structured fields in the context (e.g. status, counts, enum values, boolean flags).`,
         `2. Ignore prose narrative fields ("summary", free-form rationale, conversational commentary). They are not the contract.`,
         `3. When a field's value contradicts what a prose field claims, trust the field's value.`,
+        `4. A field whose value is explicitly null means the source node DECLARED that field but did NOT emit a value. Treat null as "unknown" and do NOT match it against any specific value (do not match "is 0", "is N", "is true", "is false", or "is undefined" against a null field). Prefer a default/fallback edge when the field needed for a decision is null.`,
         `\nRespond with ONLY the choice ID, nothing else.`,
     ].join("\n");
 }

package/dist/executor.js

@@ -166,6 +166,7 @@ export async function execute(workflow, input, options) {
                 tools: trackedTools,
                 outputSchema: node.output,
                 maxTurns: node.max_turns,
+                disallowedTools: node.disallowed_tools,
                 onProgress: (message) => {
                     safeObserve(observer, { type: "node:progress", node: currentId, message }, logger);
                 },
@@ -173,6 +174,27 @@ export async function execute(workflow, input, options) {
             // Retry only triggers on eval failure. Bail on tool/API errors.
             if (result.status !== "success")
                 break;
+            // Required-field gate: if the source node's declared output schema
+            // marks fields as `required`, those fields are the routing contract.
+            // An agent that omits a required field has violated the contract; we
+            // fail the node loudly rather than letting routing guess against a
+            // half-built data view. See the offload release-notes field bug:
+            // an optional quality_retry_count was declared in the schema but
+            // omitted by the agent, the route view silently dropped the key,
+            // and an "is 0 OR is undefined" condition ghost-matched.
+            //
+            // We only check declared `required` fields, not full JSON Schema
+            // validation. Type mismatches, formats, additionalProperties, etc.
+            // stay the workflow author's problem and don't have the same
+            // routing-correctness blast radius.
+            const missingRequired = findMissingRequiredFields(result.data, node.output);
+            if (missingRequired.length > 0) {
+                const msg = `output schema violation: required field(s) missing from emitted data: ${missingRequired.join(", ")}`;
+                result.status = "failed";
+                result.data = { ...(result.data ?? {}), error: msg };
+                logger.warn(`  ${msg}`, { node: currentId });
+                break;
+            }
             const evalResults = await evaluateAll(node.eval, result, {
                 aliases: toolAliases,
                 claude,
@@ -343,16 +365,44 @@ function buildPriorNodeContext(result) {
 function buildRouteEvalEntry(result, sourceNode) {
     const fullData = (result.data ?? {});
     const declaredProps = getDeclaredOutputProperties(sourceNode?.output);
-    const dataView = declaredProps ? pickKeys(fullData, declaredProps) : fullData;
+    // When the source node declared an `output.properties` block, the routing
+    // view is contract-shaped: every declared property is present in the
+    // view, with an explicit `null` when the agent did not emit it. The
+    // alternative (silently dropping the key) lets the LLM evaluator
+    // ghost-match conditions like "is 0", "is N", or "is undefined" against
+    // a structurally-absent field. That was the offload release-notes field
+    // bug: an optional `quality_retry_count` declared in the schema but
+    // omitted by the agent, evaluated against a condition like
+    // `quality_retry_count is 0 OR is undefined`, took the wrong branch.
+    //
+    // Returning `missing` separately lets the caller emit a loud warning so
+    // operators see the contract violation in the log stream.
+    let dataView;
+    const missing = [];
+    if (declaredProps) {
+        dataView = {};
+        for (const k of declaredProps) {
+            if (k in fullData) {
+                dataView[k] = fullData[k];
+            }
+            else {
+                dataView[k] = null;
+                missing.push(k);
+            }
+        }
+    }
+    else {
+        dataView = fullData;
+    }
     const evals = result.evals ?? [];
     if (evals.length === 0)
-        return dataView;
+        return { view: dataView, missing };
     const evalsByName = {};
     for (const e of evals) {
         evalsByName[e.name] = e;
     }
     // Spread evals first so a data-side `evals` key shadows it (back-compat).
-    return { evals: evalsByName, ...dataView };
+    return { view: { evals: evalsByName, ...dataView }, missing };
 }
 /**
  * Return the declared output property names for a node, or null when the
@@ -374,13 +424,34 @@ function getDeclaredOutputProperties(output) {
         return null;
     return new Set(keys);
 }
-function pickKeys(obj, keys) {
-    const out = {};
-    for (const k of keys) {
-        if (k in obj)
-            out[k] = obj[k];
-    }
-    return out;
+/**
+ * Return the declared `required` field names for a node's output schema,
+ * or an empty array when no required block is declared. Tolerant of
+ * malformed schemas (mirrors getDeclaredOutputProperties).
+ */
+function getDeclaredRequiredFields(output) {
+    if (!output || typeof output !== "object")
+        return [];
+    const required = output.required;
+    if (!Array.isArray(required))
+        return [];
+    return required.filter((r) => typeof r === "string");
+}
+/**
+ * Validate a node's emitted data against the declared output schema's
+ * required fields. Returns the list of required-but-missing field names.
+ *
+ * This is intentionally narrow: we do NOT do full JSON Schema validation.
+ * The routing impact comes from required fields being absent, so that's
+ * what we check. Type mismatches, format violations, additionalProperties
+ * etc. stay the workflow author's problem.
+ */
+function findMissingRequiredFields(data, output) {
+    const required = getDeclaredRequiredFields(output);
+    if (required.length === 0)
+        return [];
+    const obj = (data ?? {});
+    return required.filter((k) => !(k in obj));
 }
 function nodeSourcesToArray(ns) {
     if (!ns)
@@ -586,10 +657,28 @@ async function resolveNext(workflow, current, results, input, claude, observer,
     // always-injected `summary` prose from the reference Claude client) can
     // sway the natural-language route evaluator and override the actual
     // structured result. See `buildRouteEvalEntry` for the full rationale.
-    const context = {
-        input,
-        ...Object.fromEntries([...results.entries()].map(([k, v]) => [k, buildRouteEvalEntry(v, workflow.nodes[k])])),
-    };
+    const context = { input };
+    for (const [k, v] of results.entries()) {
+        const { view, missing } = buildRouteEvalEntry(v, workflow.nodes[k]);
+        context[k] = view;
+        if (missing.length > 0) {
+            // Loud signal for the operator. The view itself already replaces the
+            // missing keys with explicit `null`, so the LLM evaluator cannot
+            // ghost-match "is N" or "is undefined" against a structurally-absent
+            // field. The warning surfaces the contract violation in the log
+            // stream and to observer-based telemetry.
+            logger?.warn(`  route eval: source node '${k}' declared properties not in emitted data: ${missing.join(", ")}`, {
+                node: k,
+                fields: missing,
+            });
+            safeObserve(observer, {
+                type: "node:warning",
+                node: k,
+                reason: `declared output properties missing from emitted data; routing view filled them with null`,
+                fields: missing,
+            }, logger);
+        }
+    }
     const choices = conditionalEdges.map((e) => ({
         id: e.to,
         description: e.when,

package/dist/schema.d.ts

@@ -799,6 +799,7 @@ export declare const nodeZ: z.ZodObject<{
     skills: z.ZodDefault<z.ZodArray<z.ZodString, "many">>;
     output: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
     max_turns: z.ZodOptional<z.ZodNumber>;
+    disallowed_tools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
     rules: z.ZodOptional<z.ZodUnion<[z.ZodArray<z.ZodUnion<[z.ZodString, z.ZodObject<{
         inline: z.ZodString;
     }, "strict", z.ZodTypeAny, {
@@ -1197,6 +1198,7 @@ export declare const nodeZ: z.ZodObject<{
     skills: string[];
     output?: Record<string, unknown> | undefined;
     max_turns?: number | undefined;
+    disallowed_tools?: string[] | undefined;
     rules?: (string | {
         inline: string;
     } | {
@@ -1285,6 +1287,7 @@ export declare const nodeZ: z.ZodObject<{
     skills?: string[] | undefined;
     output?: Record<string, unknown> | undefined;
     max_turns?: number | undefined;
+    disallowed_tools?: string[] | undefined;
     rules?: (string | {
         inline: string;
     } | {
@@ -1423,6 +1426,7 @@ export declare const workflowZ: z.ZodObject<{
         skills: z.ZodDefault<z.ZodArray<z.ZodString, "many">>;
         output: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
         max_turns: z.ZodOptional<z.ZodNumber>;
+        disallowed_tools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
         rules: z.ZodOptional<z.ZodUnion<[z.ZodArray<z.ZodUnion<[z.ZodString, z.ZodObject<{
             inline: z.ZodString;
         }, "strict", z.ZodTypeAny, {
@@ -1821,6 +1825,7 @@ export declare const workflowZ: z.ZodObject<{
         skills: string[];
         output?: Record<string, unknown> | undefined;
         max_turns?: number | undefined;
+        disallowed_tools?: string[] | undefined;
         rules?: (string | {
             inline: string;
         } | {
@@ -1909,6 +1914,7 @@ export declare const workflowZ: z.ZodObject<{
         skills?: string[] | undefined;
         output?: Record<string, unknown> | undefined;
         max_turns?: number | undefined;
+        disallowed_tools?: string[] | undefined;
         rules?: (string | {
             inline: string;
         } | {
@@ -2198,6 +2204,7 @@ export declare const workflowZ: z.ZodObject<{
         skills: string[];
         output?: Record<string, unknown> | undefined;
         max_turns?: number | undefined;
+        disallowed_tools?: string[] | undefined;
         rules?: (string | {
             inline: string;
         } | {
@@ -2323,6 +2330,7 @@ export declare const workflowZ: z.ZodObject<{
         skills?: string[] | undefined;
         output?: Record<string, unknown> | undefined;
         max_turns?: number | undefined;
+        disallowed_tools?: string[] | undefined;
         rules?: (string | {
             inline: string;
         } | {
@@ -2478,6 +2486,7 @@ export declare function parseWorkflow(raw: unknown): {
         skills: string[];
         output?: Record<string, unknown> | undefined;
         max_turns?: number | undefined;
+        disallowed_tools?: string[] | undefined;
         rules?: (string | {
             inline: string;
         } | {
@@ -2920,6 +2929,14 @@ export declare const workflowJsonSchema: {
                         readonly minimum: 1;
                         readonly description: "Max AI model turns for this node. When absent, the executor's default applies.";
                     };
+                    readonly disallowed_tools: {
+                        readonly type: "array";
+                        readonly items: {
+                            readonly type: "string";
+                            readonly minLength: 1;
+                        };
+                        readonly description: "Built-in tool names the agent cannot use at this node (e.g. ['Bash']). Forwarded to the Claude Agent SDK's disallowedTools, which removes the tools from the model context entirely.";
+                    };
                     readonly rules: {
                         readonly $ref: "#/$defs/NodeSources";
                         readonly description: "Per-node rules. Additive by default; set { only: true } to block cascade.";

package/dist/schema.js

@@ -194,6 +194,7 @@ export const nodeZ = z
     skills: z.array(z.string()).default([]),
     output: jsonSchemaZ.optional(),
     max_turns: z.number().int().min(1).optional(),
+    disallowed_tools: z.array(z.string().min(1)).optional(),
     rules: nodeSourcesZ.optional(),
     context: nodeSourcesZ.optional(),
     eval: z.array(evaluatorZ).min(1).optional(),
@@ -631,6 +632,11 @@ export const workflowJsonSchema = {
                         minimum: 1,
                         description: "Max AI model turns for this node. When absent, the executor's default applies.",
                     },
+                    disallowed_tools: {
+                        type: "array",
+                        items: { type: "string", minLength: 1 },
+                        description: "Built-in tool names the agent cannot use at this node (e.g. ['Bash']). Forwarded to the Claude Agent SDK's disallowedTools, which removes the tools from the model context entirely.",
+                    },
                     rules: {
                         $ref: "#/$defs/NodeSources",
                         description: "Per-node rules. Additive by default; set { only: true } to block cascade.",

package/dist/types.d.ts

@@ -279,6 +279,15 @@ export interface Node {
     requires?: NodeRequires;
     /** Node-local retry on eval failure (with optional autonomous reflection). */
     retry?: NodeRetry;
+    /**
+     * Built-in tool names the agent must NOT have access to during this node.
+     * Forwarded to the Claude Agent SDK's `disallowedTools` option, which
+     * removes the named tools from the model's context entirely (not just
+     * blocked-via-permission). Use this to scope a node — e.g. an implement
+     * node that should commit but must not shell out `gh pr create` or
+     * `git push`. Names follow the SDK's tool naming (e.g. `Bash`, `Write`).
+     */
+    disallowed_tools?: string[];
 }
 /** An edge connecting two nodes */
 export interface Edge {
@@ -400,6 +409,23 @@ export type ExecutionEvent = {
     attempt: number;
     reason: string;
     preamble: string;
+}
+/**
+ * Non-fatal contract violation surfaced during node execution. Currently
+ * emitted when a source node declares output properties that the agent
+ * did not emit. The route eval view fills the missing keys with `null`
+ * so the LLM evaluator can't ghost-match conditions like "is undefined"
+ * or "is 0", but the operator still wants a loud signal in the log
+ * stream.
+ *
+ * `fields` lists the declared property names that were missing from the
+ * emitted data. `reason` is a short human-readable summary.
+ */
+ | {
+    type: "node:warning";
+    node: string;
+    reason: string;
+    fields: string[];
 } | {
     type: "route";
     from: string;
@@ -455,6 +481,8 @@ export interface Claude {
         onProgress?: (message: string) => void;
         /** Per-node turn limit. Overrides the client default when set. */
         maxTurns?: number;
+        /** Built-in SDK tool names to disallow for this node (e.g. ["Bash"]). */
+        disallowedTools?: string[];
     }): Promise<NodeResult>;
     /** Evaluate a routing condition — pick one of N choices */
     evaluate(opts: {

package/dist/workflows/browser.js

@@ -1,3 +1,3 @@
-export const triageWorkflow = {"id":"triage","name":"Alert Triage","workflow_type":"pr_review","description":"Investigate a production alert, determine root cause, create an issue, implement a fix, and notify the team","entry":"gather","nodes":{"gather":{"name":"Gather Context","instruction":"You are investigating a production alert. Gather all relevant context using the available tools:\n\n1. **Observability**: Pull error details, stack traces, recent logs, metrics, and active incidents around the time of the alert. Use whichever observability tools are available to you.\n2. **Source control**: Check recent commits, pull requests, and deploys that might be related.\n3. **Issue tracker**: Search for similar past issues or known problems.\n\nBe thorough — the investigation step depends on complete context. Use every tool available to you.","skills":["github","sentry","datadog","betterstack","linear"]},"investigate":{"name":"Root Cause Analysis","instruction":"Based on the gathered context, classify every distinct issue you found into one of two buckets: **novel** or **duplicate**.\n\nFor EACH issue found:\n1. Identify the root cause and affected code/service.\n2. Assess severity: critical (service down), high (major feature broken), medium (degraded), low (cosmetic/minor).\n3. Assess fix complexity: \"simple\" (a few lines, clear change), \"moderate\" (multiple files but well-understood), or \"complex\" (architectural, risky, or unclear).\n4. **Novelty check (REQUIRED — you MUST call tools for this step):** Filing a duplicate ticket is worse than under-reporting — the team inherits sprawl that humans have to clean up later. Complete EVERY step below before classifying anything as novel. Do NOT skip these searches even if the gather context already mentions existing issues or branches — you must verify independently.\n4a. **Search broadly — THIS REQUIRES TOOL CALLS.** You MUST call `linear_search_issues` (or `github_search_issues`) with multiple queries against the issue tracker (BOTH open AND closed): the affected file or module path, the error message or log string verbatim, the feature or service name, and the symptom family (e.g. \"startup crash\", \"data dropped silently\"). Do not rely on a single keyword — a narrow search is how duplicates slip through. If you skip this search, the eval check will fail.\n4b. **Read the full body of every promising candidate — do NOT match on title alone.** For each top match, open the issue and read its entire description, including any \"Recommended Fix\", \"Proposed Fix\", \"Action Items\", or \"Acceptance Criteria\" section. An existing open issue whose recommended fix already prescribes the change you were about to propose IS a duplicate, even if the symptom you just observed is a new instance of the same underlying bug.\n4c. **Check for meta-issues and bug families.** If an existing ticket names a pattern that your finding is another instance of (same root cause, same function/module, different input or caller), treat your finding as a sibling of that family and comment on the existing ticket. Do not file a parallel ticket per instance — that is the exact pattern that produces backlog sprawl.\n4d. **When uncertain, comment don't file.** If you are less than ~80% confident a finding is truly novel after steps 4a–4c, classify it as a duplicate and add a \"+1 — another instance\" comment to the closest existing issue with the new context. Over-reporting duplicates is cheap; ticket sprawl is expensive.\n\n   - If matched → **duplicate**. Record the existing issue ID in `duplicate_of`.\n   - If no match after 4a–4d → **novel**.\n\n**Output rules:**\n- `findings`: array of ALL issues found (both novel and duplicate).\n- `novel_count`: how many findings are novel (not duplicates).\n- `highest_severity`: the highest severity across ALL findings.\n- `recommendation`: what should happen next.\n\nDownstream nodes will act ONLY on novel findings. Duplicates will be +1'd automatically.\n\n**IMPORTANT — scope boundaries for this node:**\n- DO NOT create issues (Linear, GitHub, or any tracker). A downstream `create_issue` node handles that.\n- DO NOT create branches, commits, or pull requests. A downstream `implement` node handles that.\n- DO NOT call any write tools (`linear_create_issue`, `github_create_issue`, `github_create_pr`, `linear_add_comment`, `github_add_comment`, etc.). Downstream nodes handle all writes — issue creation, +1 comments, and PRs.\n- Your ONLY job is read, search, classify, and output. If you create an issue or add a comment here, the downstream eval check will fail or duplicate work.","skills":["github","linear"],"eval":[{"name":"novelty_search_was_performed","kind":"function","rule":{"any_tool_called":["linear_search_issues","github_search_issues"]}}],"output":{"type":"object","properties":{"findings":{"type":"array","items":{"type":"object","properties":{"title":{"type":"string","description":"Short description of the issue"},"root_cause":{"type":"string"},"severity":{"type":"string","enum":["critical","high","medium","low"]},"affected_services":{"type":"array","items":{"type":"string"}},"is_duplicate":{"type":"boolean"},"duplicate_of":{"type":"string","description":"Existing issue ID/URL if duplicate"},"fix_approach":{"type":"string"},"fix_complexity":{"type":"string","enum":["simple","moderate","complex"]}},"required":["title","root_cause","severity","is_duplicate"]}},"novel_count":{"type":"number","description":"Count of novel (non-duplicate) findings"},"highest_severity":{"type":"string","enum":["critical","high","medium","low"]},"recommendation":{"type":"string"}},"required":["findings","novel_count","highest_severity","recommendation"]}},"create_issue":{"name":"Create Issues & Triage Duplicates","instruction":"Process ALL findings from the investigation. The findings array contains both novel and duplicate issues.\n\n**For each NOVEL finding** (is_duplicate = false):\n1. **First, check if a prior triage run already created an issue for this exact bug.** Call `linear_search_issues` (or `github_search_issues`) with the error message or root cause. If a matching issue already exists with the same root cause:\n   - Treat it as a duplicate: DO NOT create a new issue.\n   - Populate `issueIdentifier`, `issueTitle`, and `issueUrl` from the existing issue.\n   - Add a \"+1\" comment if appropriate (using `linear_add_comment` or `github_add_comment`).\n   - Set the `action` to `updated` in the `issues` array.\n   - This prevents duplicate tickets when the same alert fires multiple times.\n\n2. If no matching issue exists, create a new one with a clear, actionable title by invoking `linear_create_issue` or `github_create_issue`. You MUST actually call the tool — do NOT synthesize an identifier from an alert ID, commit SHA, or your own imagination. If the tool errors, stop and report the failure; do not invent a fallback identifier. **Use the exact tool names above** — do NOT use generic alternatives like `create_issue` or `create_pull_request` from ToolSearch. The eval check requires the specific tool names listed here.\n3. Include: root cause, severity, affected services, reproduction steps, and recommended fix.\n4. Apply the `sweny`, `agent`, and `triage` labels. For Linear: call `linear_list_labels` to find the IDs for labels named `sweny`, `agent`, and `triage`, then pass those IDs in `labelIds` when calling `linear_create_issue` (skip any that don't exist in the team — don't invent IDs). For GitHub: pass `labels: [\"sweny\", \"agent\", \"triage\"]` to `github_create_issue`.\n5. Link to relevant commits, PRs, or existing issues.\n6. Copy the identifier returned by the tool verbatim into the `issueIdentifier` output field. The identifier must match `^[A-Z][A-Z0-9]*-\\d+$` (e.g. `OFF-1234`) — do not prefix it with anything else.\n\n**For each DUPLICATE finding** (is_duplicate = true):\n1. Find the existing issue (check duplicate_of field).\n2. Check the issue's comments — if the most recent comment is already from SWEny (contains \"+1\") within the last 24 hours, skip adding another comment.\n3. Otherwise add a SHORT comment: \"+1 — seen again {UTC timestamp}. {one sentence of new context}.\" (Keep it under 2 lines. No markdown headers, no emoji, no formatting.)\n4. If the existing issue is closed/done, reopen it.\n\nIf context.issueTemplate is provided, use it as the format for new issue bodies. Otherwise use a clear structure with: Summary, Root Cause, Impact, Steps to Reproduce, and Recommended Fix.\n\nUse the issue tracker specified in context (e.g. Linear, Jira, or GitHub Issues) — do NOT fall back to a different tracker. Output the created/updated issue identifiers.\n\n**IMPORTANT — scope boundaries for this node:**\n- DO NOT create branches, commits, or pull requests. The `implement` and `create_pr` nodes handle that.\n- DO NOT call `create_pull_request`, `github_create_pr`, or any PR/branch tools.\n- Your job is issue creation and duplicate management ONLY.","skills":["linear","github"],"eval":[{"name":"issue_tracker_tool_was_called","kind":"function","rule":{"any_tool_called":["linear_create_issue","github_create_issue","linear_search_issues","github_search_issues","linear_add_comment","github_add_comment"]}}],"output":{"type":"object","properties":{"issueIdentifier":{"type":"string","pattern":"^[A-Z][A-Z0-9]*-\\d+$","description":"Primary issue identifier (e.g. OFF-1234). MUST be the identifier returned by `linear_create_issue` / `github_create_issue` — do NOT invent a value from a Sentry short-ID or commit SHA. Used by downstream nodes for branch naming and PR linking."},"issueTitle":{"type":"string","description":"Title of the primary issue created"},"issueUrl":{"type":"string","pattern":"^https?://","description":"URL of the primary issue created, as returned by the create-issue tool"},"issues":{"type":"array","minItems":1,"description":"All issues created or updated during this step","items":{"type":"object","properties":{"identifier":{"type":"string"},"title":{"type":"string"},"url":{"type":"string"},"action":{"type":"string","enum":["created","updated","reopened"]}},"required":["identifier","action"]}}},"required":["issueIdentifier","issueTitle","issueUrl","issues"]},"retry":{"max":1,"instruction":{"auto":true}}},"skip":{"name":"Skip — All Duplicates or Low Priority","instruction":"Every finding from the investigation was either a duplicate or low-priority. No new issues need to be created.\n\nFor each **duplicate** finding (check the findings array for items where is_duplicate = true):\n1. Find the existing issue (check duplicate_of field).\n2. Check the issue's comments — if the most recent comment is already from SWEny (contains \"+1\") within the last 24 hours, skip adding another comment.\n3. Otherwise add a SHORT comment: \"+1 — seen again {UTC timestamp}. {one sentence of new context}.\" (Keep it under 2 lines. No markdown headers, no emoji, no formatting.)\n4. If the issue is closed/done, reopen it.\n\nFor **low priority** findings, log a brief note about why they were skipped.","skills":["linear","github"]},"implement":{"name":"Implement Fix","instruction":"Implement the fix identified during investigation, plus the tests that pin\nit as a regression guard. The issue created in the previous step is in your\ncontext — use its identifier for branch naming and commit messages.\n\n## Quality bar (non-negotiable)\n\nEvery change you make must clear all of these. If you cannot, STOP and\nreturn a failed status — do not ship a half-measure.\n\n- **Correctness**: the fix actually closes the failure mode in the issue.\n- **Completeness**: handles the obvious edge cases (nulls, empties, off-by-one,\n  concurrent paths). If the issue lists multiple symptoms, address them all\n  or explicitly state which are out of scope and why.\n- **Idiomatic for THIS repo, not in general.** Read at least one\n  neighboring file in the same module before you write any code, and\n  match the patterns you see: error-handling primitives the codebase\n  already uses, type discipline, naming, file layout. Examples to make\n  the shape concrete:\n  - In a NestJS service, throwing `NotFoundException` /\n    `BadRequestException` is idiomatic; `throw new Error(...)` is not.\n  - In a Go service, returning a wrapped `fmt.Errorf(\"...: %w\", err)`\n    beats panicking.\n  - In a TypeScript repo, narrow types properly — do not `as any` past\n    the type checker.\n  These are illustrations. Generalize: whatever the repo already does,\n  keep doing.\n- **No hacky shortcuts**: no `try/catch` to swallow errors, no commented-out\n  guards, no `// TODO: revisit`, no global flags to bypass validation, no\n  copy-pasted bodies. If the right fix needs a small refactor (e.g. extract\n  a helper), do it; if it needs a large refactor, file a follow-up issue\n  and ship the minimal correct fix without the refactor.\n- **Well-tested**: see Test Requirements below.\n\n## Steps\n\n0a. **Precondition — verify context.create_issue.issueIdentifier is well-formed.**\n   The identifier MUST match `^[A-Z][A-Z0-9]*-\\d+$` (e.g. `OFF-1234`). If it\n   is missing, empty, or matches a Sentry short-ID / slug / any other shape,\n   STOP immediately — do NOT invent a branch name from the Sentry event,\n   commit SHA, or alert payload. Return a failed status with a clear message\n   instead; the workflow will route to notify and a human will pick it up.\n\n0b. **Precondition — check for an existing open PR for this issue.** Call\n   `github_search_issues` (which hits GitHub's `/search/issues` endpoint\n   and returns both issues and PRs) with `type:pr state:open` and the\n   issue identifier wrapped in brackets — e.g. query\n   `[OFF-1234] in:title type:pr state:open` and pass `repo` for scoping.\n   If the response contains one or more PRs whose title starts with\n   `[<issueIdentifier>]`, STOP. Do NOT create a branch, commit, or any\n   new PR. Output:\n\n   ```\n   test_status: \"skipped\"\n   skipped_reason: \"open-pr-exists\"\n   existing_pr_url: \"<url of the existing PR>\"\n   test_files_changed: []\n   files_changed: []\n   branch: \"\"\n   commit_sha: \"\"\n   ```\n\n   The downstream `create_pr` node will be bypassed and the workflow will\n   route to `notify` so the team is told the prior PR is still pending.\n\n   Rationale: re-implementing a fix while an open PR already exists\n   produces a duplicate (or worse, the agent closes its own prior PR\n   mid-run and opens a new one with a different approach — exactly the\n   thrash observed in `letsoffload/permit-service` PR #81 → #82 and\n   `letsoffload/offload` PR #572 → #573, where two PRs for the same\n   identifier were opened-and-closed 12 to 16 seconds apart inside a\n   single workflow run).\n\n   Only proceed to Step 1 if no open PR exists for the identifier.\n\n1. Create a feature branch from the base branch (check context for\n   baseBranch, default `main`). **Branch name MUST include the issue\n   identifier** exactly as provided in `context.create_issue.issueIdentifier`.\n   Format: lowercase the identifier and use it as a prefix, e.g.\n   `off-1234-fix-null-check`. If the repository uses a different branch\n   convention (check context.branchPattern or the rules/CLAUDE.md for a\n   regex), follow that — but the identifier piece MUST still come from\n   `context.create_issue.issueIdentifier`, never from a Sentry short-ID.\n\n2. Read the relevant source files to understand the current code AND read\n   the nearest existing test file (e.g. `foo.service.spec.ts` next to\n   `foo.service.ts`, or `foo_test.py` next to `foo.py`) so your new tests\n   match the existing test conventions: framework, mock style, factory use,\n   describe/it shape, naming.\n\n3. Make the necessary code change. Keep it minimal and focused — do not\n   refactor surrounding code or add unrelated improvements.\n\n4. **Write tests for the fix.** See Test Requirements below.\n\n5. **Run the test suite for the affected package** and confirm green\n   (the package's `test` command from package.json, `pytest` for Python,\n   etc.). If the test runner is not obvious, read the project's\n   CLAUDE.md / README / package.json scripts. If you cannot run the\n   tests at all in this environment, say so explicitly in the output —\n   do not claim \"tests pass\" when you didn't run them.\n\n6. Stage and commit with a clear commit message. Start the message with the\n   issue identifier in brackets, e.g.\n   `[OFF-1234] fix: guard null projectId before database lookup`.\n\n## Scope boundary — do NOT open the PR here\n\nStop after the commit. Do NOT push the branch, do NOT run `gh pr create`,\ndo NOT call any PR-creation tool. The dedicated `create_pr` node opens the\nPR via `github_create_pr`. Opening it from this node bypasses the workflow's\neval gate and triggers a self-close-and-recreate workaround (PR opened via\nshell, then `github_create_pr` returns 422 \"already exists\" against the same\nbranch, then the agent closes the first PR to satisfy the eval). The\nidempotent fallback in `github_create_pr` will recover if you ignore this,\nbut the cleaner path is: commit only here, let `create_pr` push and open.\n\n## Test Requirements\n\nA code change without a test is unfinished work. Every fix this node ships\nMUST come with at least one new test case that:\n\n- **Would fail on the unfixed code.** A test that passes both before and\n   after the fix is decoration, not a regression guard. Mentally run it\n   against the previous version; if it would have passed, it is not the\n   right test.\n- **Asserts the user-facing contract**, not just an implementation\n   detail. For an API/handler bug: assert the response shape, status code,\n   or error type the client sees. For a data corruption bug: assert the\n   persisted row's final state, not just an internal call count.\n- **Covers the obvious adjacent edge cases.** If the bug was \"throws plain\n   Error on missing record\", add the missing-record case AND at least one\n   success-path case to confirm the fix didn't break the happy path.\n- **Lives in the right file** (next to the source under test, in the\n   project's existing test layout).\n\nThe eval gate on this node REQUIRES `test_status` to be either\n`pass` (tests ran and all passed) OR `no-framework` (the project\ngenuinely has no test runner — not \"I couldn't be bothered\").\n\nIf the project HAS a test framework, you must write tests and they\nmust pass. Empty `test_files_changed` with `test_status: pass` is a\nlie and eval will reject it via the schema's required-fields\ncontract; `not-run` and `fail` will retry once with reflection and\nthen halt for human attention.\n\nIf the fix turns out to be more complex than expected, stop and\nexplain why — do not force a bad fix.","skills":["github"],"output":{"type":"object","properties":{"branch":{"type":"string","description":"Feature branch name (must include the issue identifier)."},"commit_sha":{"type":"string","description":"SHA of the commit shipping the fix and tests."},"files_changed":{"type":"array","description":"Source files modified by the fix (NOT including test files).","items":{"type":"string"}},"test_files_changed":{"type":"array","description":"Test files added or modified to cover the fix. Should be\nnon-empty whenever `test_status: pass`. May be empty when\n`test_status: no-framework` (genuinely no test runner in\nthe repo).\n","items":{"type":"string"}},"test_command":{"type":"string","description":"Exact command run to validate (e.g. `npx nx test server`)."},"test_status":{"type":"string","enum":["pass","fail","no-framework","not-run","skipped"],"description":"Outcome of running the test suite.\n- `pass` — tests ran end-to-end and every assertion passed.\n  `test_files_changed` MUST be non-empty in this case (you\n  changed code; tests must cover it).\n- `no-framework` — the repo genuinely has no test runner\n  (e.g. a docs-only repo, a YAML-config repo, a research\n  notebook collection). Reserved for cases where there is\n  no test infrastructure to extend — not a permission slip\n  to skip tests in a normal codebase.\n- `fail` — tests ran and at least one failed. Do not push\n  a broken fix. Return this status, retry will fire once.\n- `not-run` — tests exist but the environment couldn't\n  execute them. Explain in `test_notes`. Retry will fire.\n- `skipped` — the node short-circuited before running tests\n  because a precondition was not met. Today the only\n  accepted reason is `open-pr-exists` (Step 0b found an\n  open PR with the issue identifier in its title).\n  `skipped_reason` and `existing_pr_url` MUST be populated.\n  Downstream `create_pr` is bypassed and the workflow\n  routes to `notify`.\n"},"skipped_reason":{"type":"string","description":"Why the node short-circuited. Only set when `test_status` is\n`skipped`. Today the only valid value is `open-pr-exists`.\n","enum":["open-pr-exists"]},"existing_pr_url":{"type":"string","description":"URL of the open PR that triggered the skip. Set when\n`skipped_reason` is `open-pr-exists` so the notify node\ncan link to the prior PR.\n","pattern":"^https?://"},"test_notes":{"type":"string","description":"One-paragraph explanation of which tests were added and what they assert."}},"required":["branch","commit_sha","files_changed","test_files_changed","test_status"]},"eval":[{"name":"test_status_is_pass_or_no_framework","kind":"value","rule":{"output_matches":[{"path":"test_status","in":["pass","no-framework","skipped"]}]}},{"name":"tests_present_when_pass_claimed","kind":"judge","rubric":"You are checking the contract between test_status and test_files_changed.\n\nRead result.data.test_status and result.data.test_files_changed.\n\nRules:\n- If test_status is \"pass\": test_files_changed MUST contain at least\n  one real test file path (paths with extensions like .spec.ts,\n  .test.ts, .test.tsx, .test.py, _test.go, _spec.rb, .test.rs,\n  etc., or paths in a __tests__/, tests/, or spec/ directory).\n  An empty array, or paths that are clearly not test files\n  (README.md, package.json, source files), is a contract violation.\n- If test_status is \"no-framework\": test_files_changed MAY be empty.\n  Pass vacuously.\n- If test_status is \"skipped\": test_files_changed MAY be empty.\n  The node short-circuited before writing any code (Step 0b\n  found an open PR for the issue). `skipped_reason` and\n  `existing_pr_url` MUST be populated; if they are missing,\n  return no.\n- If test_status is \"fail\" or \"not-run\": this evaluator should not\n  run (the prior value evaluator catches that). Pass vacuously.\n\nRespond yes if the contract is satisfied, no if it is violated.\n","pass_when":"yes"}],"retry":{"max":1,"instruction":{"auto":true}}},"create_pr":{"name":"Open Pull Request","instruction":"Open a pull request for the fix. The issue identifier and URL are in your context from the create_issue step.\n\n1. Push the branch to the remote.\n2. Create the PR using the `github_create_pr` tool. **Use this exact tool name** — do NOT use `create_pull_request` or any other variant from ToolSearch. The eval check requires `github_create_pr` specifically.\n**If `github_create_pr` returns a result with `reused: true`**, the implement node (or a prior run) already opened a PR for this branch and the tool recovered it instead of failing on 422. Treat this as success and return the PR's URL and number — do NOT close it and call `github_create_pr` again.\n3. PR title MUST start with the issue identifier in brackets. Format: `[{context.create_issue.issueIdentifier}] fix: {short description}`.\n4. In the PR body, include: summary of the bug, what the fix does, and a link to the issue (context.create_issue.issueUrl). If the investigation found a Sentry issue, include \"Fixes {sentryShortId}\" in the PR body so Sentry auto-resolves on merge.\n5. Pass `labels: [\"sweny\", \"agent\", \"triage\"]` to `github_create_pr`.\n\nIf context.prTemplate is provided, use it as the format for the PR body. Otherwise use a clear structure with: Summary, Changes, Testing, and Related Issues.\n\nIf context.cloudRunId is provided, append this HTML comment to the end of the PR body: <!-- sweny:run_id:{context.cloudRunId} -->\n\nReturn the PR URL and number.","skills":["github"],"eval":[{"name":"pr_was_created","kind":"function","rule":{"any_tool_called":["github_create_pr"]}}],"retry":{"max":1,"instruction":{"auto":true}}},"notify":{"name":"Notify Team","instruction":"Send a notification summarizing the triage result:\n\n1. Include: alert summary, severity, root cause (1-2 sentences), and links to any created issues or PRs.\n2. For critical/high severity, make the notification urgent.\n3. For medium/low, a standard notification is fine.\n\nUse whichever notification channel is available to you.","skills":["slack","notification"]}},"edges":[{"from":"gather","to":"investigate"},{"from":"investigate","to":"create_issue","when":"novel_count is greater than 0 AND highest_severity is medium or higher"},{"from":"investigate","to":"skip","when":"novel_count is 0, OR highest_severity is low"},{"from":"create_issue","to":"implement","when":"at least one novel finding has fix_complexity simple or moderate AND fix_approach is provided"},{"from":"create_issue","to":"notify","when":"all novel findings have fix_complexity complex, OR no clear fix_approach"},{"from":"skip","to":"notify"},{"from":"implement","to":"notify","when":"implement.test_status is \"skipped\" (Step 0b found an open PR already exists for the issue identifier)"},{"from":"implement","to":"create_pr","when":"implement.test_status is \"pass\" or \"no-framework\" (no existing PR; fix was implemented or there is no test framework)"},{"from":"create_pr","to":"notify"}],"skills":{}};
-export const implementWorkflow = {"id":"implement","name":"Implement Fix","workflow_type":"pr_review","description":"Implement a code fix for an issue and open a pull request","entry":"analyze","nodes":{"analyze":{"name":"Analyze Issue","instruction":"Read the issue details and understand what needs to be fixed:\n\n1. Fetch the issue from the tracker (GitHub or Linear).\n2. **Check for an existing open PR for this issue.** Call\n   `github_search_issues` (GitHub's `/search/issues` endpoint returns\n   both issues and PRs) with `type:pr state:open` and the issue\n   identifier wrapped in brackets — e.g. query\n   `[OFF-1234] in:title type:pr state:open`. If the response contains\n   one or more PRs whose title starts with `[<issueIdentifier>]`,\n   STOP analysis. Output `existing_pr_url`\n   (the URL of the open PR) and a short `issue_summary` noting the\n   skip; leave `fix_plan` set to a one-line \"Skip: open PR\n   <url> already exists\". The workflow will route to `notify` and\n   the team will be told the prior PR is still pending.\n   Rationale: re-implementing while an open PR exists produces a\n   duplicate (or worse, the agent closes its own prior PR mid-run\n   and opens a new one — exactly the thrash seen in\n   `letsoffload/permit-service` PR #81 → #82 and\n   `letsoffload/offload` PR #572 → #573).\n3. Read the relevant source files to understand the current code.\n4. Identify the exact files and lines that need to change.\n5. Plan the fix approach.\n\nOutput a clear analysis of what needs to change and why.","skills":["github","linear"],"output":{"type":"object","properties":{"issue_summary":{"type":"string"},"files_to_change":{"type":"array","items":{"type":"string"}},"fix_plan":{"type":"string"},"risk_level":{"type":"string","enum":["low","medium","high"]},"existing_pr_url":{"type":"string","description":"URL of an open PR found in step 2 whose title starts with the\nissue identifier in brackets. Set ONLY when a prior PR exists\nand the analyze node is short-circuiting. Triggers the\n`analyze → notify` edge.\n","pattern":"^https?://"}},"required":["issue_summary","fix_plan"]}},"implement":{"name":"Implement Fix","instruction":"Implement the planned fix:\n\n1. Create a feature branch.\n2. Make the necessary code changes.\n3. Ensure changes are minimal and focused — fix the bug, nothing more.\n4. Stage and commit with a clear commit message referencing the issue.\n\nStop after the commit. Do NOT push or open the PR — the dedicated `create_pr`\nnode handles that via `github_create_pr` and its eval gate is what unblocks\nthe rest of the workflow.\n\nIf the fix is too risky or complex, explain why and skip.","skills":["github"]},"create_pr":{"name":"Open Pull Request","instruction":"Open a pull request for the fix:\n\n1. Push the branch to the remote.\n2. Create a PR with a clear title and description.\n3. Reference the original issue in the PR body.\n4. Pass `labels: [\"sweny\", \"agent\", \"task\"]` to `github_create_pr` (or `[\"sweny\", \"agent\", \"feature\"]` if the issue is a feature request).\n\nIf context.prTemplate is provided, use it as the format for the PR body. Otherwise use a clear structure with: Summary, Changes, Testing, and Related Issues.\n\nIf context.cloudRunId is provided, append this HTML comment to the end of the PR body: <!-- sweny:run_id:{context.cloudRunId} -->\n\nReturn the PR URL.","skills":["github"]},"notify":{"name":"Notify","instruction":"Send a notification about the implementation result:\n\n1. Include: issue reference, PR link, brief summary of changes.\n2. Keep it concise — one message, not a wall of text.","skills":["slack","notification"]},"skip":{"name":"Skip — Too Complex","instruction":"The fix was determined to be too complex or risky for automated implementation.\nAdd a comment to the issue explaining what was found and why it needs manual attention.","skills":["github","linear"]}},"edges":[{"from":"analyze","to":"notify","when":"analyze.existing_pr_url is set (an open PR for this issue already exists; skip the implement and create_pr nodes)"},{"from":"analyze","to":"implement","when":"Fix risk level is low or medium and a clear plan exists"},{"from":"analyze","to":"skip","when":"Fix is too complex, risky, or unclear"},{"from":"implement","to":"create_pr"},{"from":"skip","to":"notify"},{"from":"create_pr","to":"notify"}],"skills":{}};
+export const triageWorkflow = {"id":"triage","name":"Alert Triage","workflow_type":"pr_review","description":"Investigate a production alert, determine root cause, create an issue, implement a fix, and notify the team","entry":"gather","nodes":{"gather":{"name":"Gather Context","instruction":"You are investigating a production alert. Gather all relevant context using the available tools:\n\n1. **Observability**: Pull error details, stack traces, recent logs, metrics, and active incidents around the time of the alert. Use whichever observability tools are available to you.\n2. **Source control**: Check recent commits, pull requests, and deploys that might be related.\n3. **Issue tracker**: Search for similar past issues or known problems.\n\nBe thorough — the investigation step depends on complete context. Use every tool available to you.","skills":["github","sentry","datadog","betterstack","linear"]},"investigate":{"name":"Root Cause Analysis","instruction":"Based on the gathered context, classify every distinct issue you found into one of two buckets: **novel** or **duplicate**.\n\nFor EACH issue found:\n1. Identify the root cause and affected code/service.\n2. Assess severity: critical (service down), high (major feature broken), medium (degraded), low (cosmetic/minor).\n3. Assess fix complexity: \"simple\" (a few lines, clear change), \"moderate\" (multiple files but well-understood), or \"complex\" (architectural, risky, or unclear).\n4. **Novelty check (REQUIRED — you MUST call tools for this step):** Filing a duplicate ticket is worse than under-reporting — the team inherits sprawl that humans have to clean up later. Complete EVERY step below before classifying anything as novel. Do NOT skip these searches even if the gather context already mentions existing issues or branches — you must verify independently.\n4a. **Search broadly — THIS REQUIRES TOOL CALLS.** You MUST call `linear_search_issues` (or `github_search_issues`) with multiple queries against the issue tracker (BOTH open AND closed): the affected file or module path, the error message or log string verbatim, the feature or service name, and the symptom family (e.g. \"startup crash\", \"data dropped silently\"). Do not rely on a single keyword — a narrow search is how duplicates slip through. If you skip this search, the eval check will fail.\n4b. **Read the full body of every promising candidate — do NOT match on title alone.** For each top match, open the issue and read its entire description, including any \"Recommended Fix\", \"Proposed Fix\", \"Action Items\", or \"Acceptance Criteria\" section. An existing open issue whose recommended fix already prescribes the change you were about to propose IS a duplicate, even if the symptom you just observed is a new instance of the same underlying bug.\n4c. **Check for meta-issues and bug families.** If an existing ticket names a pattern that your finding is another instance of (same root cause, same function/module, different input or caller), treat your finding as a sibling of that family and comment on the existing ticket. Do not file a parallel ticket per instance — that is the exact pattern that produces backlog sprawl.\n4d. **When uncertain, comment don't file.** If you are less than ~80% confident a finding is truly novel after steps 4a–4c, classify it as a duplicate and add a \"+1 — another instance\" comment to the closest existing issue with the new context. Over-reporting duplicates is cheap; ticket sprawl is expensive.\n\n   - If matched → **duplicate**. Record the existing issue ID in `duplicate_of`.\n   - If no match after 4a–4d → **novel**.\n\n**Output rules:**\n- `findings`: array of ALL issues found (both novel and duplicate).\n- `novel_count`: how many findings are novel (not duplicates).\n- `highest_severity`: the highest severity across ALL findings.\n- `recommendation`: what should happen next.\n\nDownstream nodes will act ONLY on novel findings. Duplicates will be +1'd automatically.\n\n**IMPORTANT — scope boundaries for this node:**\n- DO NOT create issues (Linear, GitHub, or any tracker). A downstream `create_issue` node handles that.\n- DO NOT create branches, commits, or pull requests. A downstream `implement` node handles that.\n- DO NOT call any write tools (`linear_create_issue`, `github_create_issue`, `github_create_pr`, `linear_add_comment`, `github_add_comment`, etc.). Downstream nodes handle all writes — issue creation, +1 comments, and PRs.\n- Your ONLY job is read, search, classify, and output. If you create an issue or add a comment here, the downstream eval check will fail or duplicate work.","skills":["github","linear"],"eval":[{"name":"novelty_search_was_performed","kind":"function","rule":{"any_tool_called":["linear_search_issues","github_search_issues"]}}],"output":{"type":"object","properties":{"findings":{"type":"array","items":{"type":"object","properties":{"title":{"type":"string","description":"Short description of the issue"},"root_cause":{"type":"string"},"severity":{"type":"string","enum":["critical","high","medium","low"]},"affected_services":{"type":"array","items":{"type":"string"}},"is_duplicate":{"type":"boolean"},"duplicate_of":{"type":"string","description":"Existing issue ID/URL if duplicate"},"fix_approach":{"type":"string"},"fix_complexity":{"type":"string","enum":["simple","moderate","complex"]}},"required":["title","root_cause","severity","is_duplicate"]}},"novel_count":{"type":"number","description":"Count of novel (non-duplicate) findings"},"highest_severity":{"type":"string","enum":["critical","high","medium","low"]},"recommendation":{"type":"string"}},"required":["findings","novel_count","highest_severity","recommendation"]}},"create_issue":{"name":"Create Issues & Triage Duplicates","instruction":"Process ALL findings from the investigation. The findings array contains both novel and duplicate issues.\n\n**For each NOVEL finding** (is_duplicate = false):\n1. **First, check if a prior triage run already created an issue for this exact bug.** Call `linear_search_issues` (or `github_search_issues`) with the error message or root cause. If a matching issue already exists with the same root cause:\n   - Treat it as a duplicate: DO NOT create a new issue.\n   - Populate `issueIdentifier`, `issueTitle`, and `issueUrl` from the existing issue.\n   - Add a \"+1\" comment if appropriate (using `linear_add_comment` or `github_add_comment`).\n   - Set the `action` to `updated` in the `issues` array.\n   - This prevents duplicate tickets when the same alert fires multiple times.\n\n2. If no matching issue exists, create a new one with a clear, actionable title by invoking `linear_create_issue` or `github_create_issue`. You MUST actually call the tool — do NOT synthesize an identifier from an alert ID, commit SHA, or your own imagination. If the tool errors, stop and report the failure; do not invent a fallback identifier. **Use the exact tool names above** — do NOT use generic alternatives like `create_issue` or `create_pull_request` from ToolSearch. The eval check requires the specific tool names listed here.\n3. Include: root cause, severity, affected services, reproduction steps, and recommended fix.\n4. Apply the `sweny`, `agent`, and `triage` labels. For Linear: call `linear_list_labels` to find the IDs for labels named `sweny`, `agent`, and `triage`, then pass those IDs in `labelIds` when calling `linear_create_issue` (skip any that don't exist in the team — don't invent IDs). For GitHub: pass `labels: [\"sweny\", \"agent\", \"triage\"]` to `github_create_issue`.\n5. Link to relevant commits, PRs, or existing issues.\n6. Copy the identifier returned by the tool verbatim into the `issueIdentifier` output field. The identifier must match `^[A-Z][A-Z0-9]*-\\d+$` (e.g. `OFF-1234`) — do not prefix it with anything else.\n\n**For each DUPLICATE finding** (is_duplicate = true):\n1. Find the existing issue (check duplicate_of field).\n2. Check the issue's comments — if the most recent comment is already from SWEny (contains \"+1\") within the last 24 hours, skip adding another comment.\n3. Otherwise add a SHORT comment: \"+1 — seen again {UTC timestamp}. {one sentence of new context}.\" (Keep it under 2 lines. No markdown headers, no emoji, no formatting.)\n4. If the existing issue is closed/done, reopen it.\n\nIf context.issueTemplate is provided, use it as the format for new issue bodies. Otherwise use a clear structure with: Summary, Root Cause, Impact, Steps to Reproduce, and Recommended Fix.\n\nUse the issue tracker specified in context (e.g. Linear, Jira, or GitHub Issues) — do NOT fall back to a different tracker. Output the created/updated issue identifiers.\n\n**IMPORTANT — scope boundaries for this node:**\n- DO NOT create branches, commits, or pull requests. The `implement` and `create_pr` nodes handle that.\n- DO NOT call `create_pull_request`, `github_create_pr`, or any PR/branch tools.\n- Your job is issue creation and duplicate management ONLY.","skills":["linear","github"],"eval":[{"name":"issue_tracker_tool_was_called","kind":"function","rule":{"any_tool_called":["linear_create_issue","github_create_issue","linear_search_issues","github_search_issues","linear_add_comment","github_add_comment"]}}],"output":{"type":"object","properties":{"issueIdentifier":{"type":"string","pattern":"^[A-Z][A-Z0-9]*-\\d+$","description":"Primary issue identifier (e.g. OFF-1234). MUST be the identifier returned by `linear_create_issue` / `github_create_issue` — do NOT invent a value from a Sentry short-ID or commit SHA. Used by downstream nodes for branch naming and PR linking."},"issueTitle":{"type":"string","description":"Title of the primary issue created"},"issueUrl":{"type":"string","pattern":"^https?://","description":"URL of the primary issue created, as returned by the create-issue tool"},"issues":{"type":"array","minItems":1,"description":"All issues created or updated during this step","items":{"type":"object","properties":{"identifier":{"type":"string"},"title":{"type":"string"},"url":{"type":"string"},"action":{"type":"string","enum":["created","updated","reopened"]}},"required":["identifier","action"]}}},"required":["issueIdentifier","issueTitle","issueUrl","issues"]},"retry":{"max":1,"instruction":{"auto":true}}},"skip":{"name":"Skip — All Duplicates or Low Priority","instruction":"Every finding from the investigation was either a duplicate or low-priority. No new issues need to be created.\n\nFor each **duplicate** finding (check the findings array for items where is_duplicate = true):\n1. Find the existing issue (check duplicate_of field).\n2. Check the issue's comments — if the most recent comment is already from SWEny (contains \"+1\") within the last 24 hours, skip adding another comment.\n3. Otherwise add a SHORT comment: \"+1 — seen again {UTC timestamp}. {one sentence of new context}.\" (Keep it under 2 lines. No markdown headers, no emoji, no formatting.)\n4. If the issue is closed/done, reopen it.\n\nFor **low priority** findings, log a brief note about why they were skipped.","skills":["linear","github"]},"implement":{"name":"Implement Fix","instruction":"Implement the fix identified during investigation, plus the tests that pin\nit as a regression guard. The issue created in the previous step is in your\ncontext — use its identifier for branch naming and commit messages.\n\n## Quality bar (non-negotiable)\n\nEvery change you make must clear all of these. If you cannot, STOP and\nreturn a failed status — do not ship a half-measure.\n\n- **Correctness**: the fix actually closes the failure mode in the issue.\n- **Completeness**: handles the obvious edge cases (nulls, empties, off-by-one,\n  concurrent paths). If the issue lists multiple symptoms, address them all\n  or explicitly state which are out of scope and why.\n- **Idiomatic for THIS repo, not in general.** Read at least one\n  neighboring file in the same module before you write any code, and\n  match the patterns you see: error-handling primitives the codebase\n  already uses, type discipline, naming, file layout. Examples to make\n  the shape concrete:\n  - In a NestJS service, throwing `NotFoundException` /\n    `BadRequestException` is idiomatic; `throw new Error(...)` is not.\n  - In a Go service, returning a wrapped `fmt.Errorf(\"...: %w\", err)`\n    beats panicking.\n  - In a TypeScript repo, narrow types properly — do not `as any` past\n    the type checker.\n  These are illustrations. Generalize: whatever the repo already does,\n  keep doing.\n- **No hacky shortcuts**: no `try/catch` to swallow errors, no commented-out\n  guards, no `// TODO: revisit`, no global flags to bypass validation, no\n  copy-pasted bodies. If the right fix needs a small refactor (e.g. extract\n  a helper), do it; if it needs a large refactor, file a follow-up issue\n  and ship the minimal correct fix without the refactor.\n- **Well-tested**: see Test Requirements below.\n\n## Steps\n\n0a. **Precondition — verify context.create_issue.issueIdentifier is well-formed.**\n   The identifier MUST match `^[A-Z][A-Z0-9]*-\\d+$` (e.g. `OFF-1234`). If it\n   is missing, empty, or matches a Sentry short-ID / slug / any other shape,\n   STOP immediately — do NOT invent a branch name from the Sentry event,\n   commit SHA, or alert payload. Return a failed status with a clear message\n   instead; the workflow will route to notify and a human will pick it up.\n\n0b. **Precondition — check for an existing open PR for this issue.** Call\n   `github_search_issues` (which hits GitHub's `/search/issues` endpoint\n   and returns both issues and PRs) with `type:pr state:open` and the\n   issue identifier wrapped in brackets — e.g. query\n   `[OFF-1234] in:title type:pr state:open` and pass `repo` for scoping.\n   If the response contains one or more PRs whose title starts with\n   `[<issueIdentifier>]`, STOP. Do NOT create a branch, commit, or any\n   new PR. Output:\n\n   ```\n   test_status: \"skipped\"\n   skipped_reason: \"open-pr-exists\"\n   existing_pr_url: \"<url of the existing PR>\"\n   test_files_changed: []\n   files_changed: []\n   branch: \"\"\n   commit_sha: \"\"\n   ```\n\n   The downstream `create_pr` node will be bypassed and the workflow will\n   route to `notify` so the team is told the prior PR is still pending.\n\n   Rationale: re-implementing a fix while an open PR already exists\n   produces a duplicate (or worse, the agent closes its own prior PR\n   mid-run and opens a new one with a different approach — exactly the\n   thrash observed in `letsoffload/permit-service` PR #81 → #82 and\n   `letsoffload/offload` PR #572 → #573, where two PRs for the same\n   identifier were opened-and-closed 12 to 16 seconds apart inside a\n   single workflow run).\n\n   Only proceed to Step 1 if no open PR exists for the identifier.\n\n1. Create a feature branch from the base branch (check context for\n   baseBranch, default `main`). **Branch name MUST include the issue\n   identifier** exactly as provided in `context.create_issue.issueIdentifier`.\n   Format: lowercase the identifier and use it as a prefix, e.g.\n   `off-1234-fix-null-check`. If the repository uses a different branch\n   convention (check context.branchPattern or the rules/CLAUDE.md for a\n   regex), follow that — but the identifier piece MUST still come from\n   `context.create_issue.issueIdentifier`, never from a Sentry short-ID.\n\n2. Read the relevant source files to understand the current code AND read\n   the nearest existing test file (e.g. `foo.service.spec.ts` next to\n   `foo.service.ts`, or `foo_test.py` next to `foo.py`) so your new tests\n   match the existing test conventions: framework, mock style, factory use,\n   describe/it shape, naming.\n\n3. Make the necessary code change. Keep it minimal and focused — do not\n   refactor surrounding code or add unrelated improvements.\n\n4. **Write tests for the fix.** See Test Requirements below.\n\n5. **Run the test suite for the affected package** and confirm green\n   (the package's `test` command from package.json, `pytest` for Python,\n   etc.). If the test runner is not obvious, read the project's\n   CLAUDE.md / README / package.json scripts. If you cannot run the\n   tests at all in this environment, say so explicitly in the output —\n   do not claim \"tests pass\" when you didn't run them.\n\n6. Stage and commit with a clear commit message. Start the message with the\n   issue identifier in brackets, e.g.\n   `[OFF-1234] fix: guard null projectId before database lookup`.\n\n## Scope boundary — do NOT open the PR here\n\nStop after the commit. Do NOT push the branch, do NOT run `gh pr create`,\ndo NOT call any PR-creation tool. The dedicated `create_pr` node opens the\nPR via `github_create_pr`. Opening it from this node bypasses the workflow's\neval gate and triggers a self-close-and-recreate workaround (PR opened via\nshell, then `github_create_pr` returns 422 \"already exists\" against the same\nbranch, then the agent closes the first PR to satisfy the eval). The\nidempotent fallback in `github_create_pr` will recover if you ignore this,\nbut the cleaner path is: commit only here, let `create_pr` push and open.\n\n## Test Requirements\n\nA code change without a test is unfinished work. Every fix this node ships\nMUST come with at least one new test case that:\n\n- **Would fail on the unfixed code.** A test that passes both before and\n   after the fix is decoration, not a regression guard. Mentally run it\n   against the previous version; if it would have passed, it is not the\n   right test.\n- **Asserts the user-facing contract**, not just an implementation\n   detail. For an API/handler bug: assert the response shape, status code,\n   or error type the client sees. For a data corruption bug: assert the\n   persisted row's final state, not just an internal call count.\n- **Covers the obvious adjacent edge cases.** If the bug was \"throws plain\n   Error on missing record\", add the missing-record case AND at least one\n   success-path case to confirm the fix didn't break the happy path.\n- **Lives in the right file** (next to the source under test, in the\n   project's existing test layout).\n\nThe eval gate on this node REQUIRES `test_status` to be either\n`pass` (tests ran and all passed) OR `no-framework` (the project\ngenuinely has no test runner — not \"I couldn't be bothered\").\n\nIf the project HAS a test framework, you must write tests and they\nmust pass. Empty `test_files_changed` with `test_status: pass` is a\nlie and eval will reject it via the schema's required-fields\ncontract; `not-run` and `fail` will retry once with reflection and\nthen halt for human attention.\n\nIf the fix turns out to be more complex than expected, stop and\nexplain why — do not force a bad fix.","skills":["github"],"disallowed_tools":["WebFetch","WebSearch"],"output":{"type":"object","properties":{"branch":{"type":"string","description":"Feature branch name (must include the issue identifier)."},"commit_sha":{"type":"string","description":"SHA of the commit shipping the fix and tests."},"files_changed":{"type":"array","description":"Source files modified by the fix (NOT including test files).","items":{"type":"string"}},"test_files_changed":{"type":"array","description":"Test files added or modified to cover the fix. Should be\nnon-empty whenever `test_status: pass`. May be empty when\n`test_status: no-framework` (genuinely no test runner in\nthe repo).\n","items":{"type":"string"}},"test_command":{"type":"string","description":"Exact command run to validate (e.g. `npx nx test server`)."},"test_status":{"type":"string","enum":["pass","fail","no-framework","not-run","skipped"],"description":"Outcome of running the test suite.\n- `pass` — tests ran end-to-end and every assertion passed.\n  `test_files_changed` MUST be non-empty in this case (you\n  changed code; tests must cover it).\n- `no-framework` — the repo genuinely has no test runner\n  (e.g. a docs-only repo, a YAML-config repo, a research\n  notebook collection). Reserved for cases where there is\n  no test infrastructure to extend — not a permission slip\n  to skip tests in a normal codebase.\n- `fail` — tests ran and at least one failed. Do not push\n  a broken fix. Return this status, retry will fire once.\n- `not-run` — tests exist but the environment couldn't\n  execute them. Explain in `test_notes`. Retry will fire.\n- `skipped` — the node short-circuited before running tests\n  because a precondition was not met. Today the only\n  accepted reason is `open-pr-exists` (Step 0b found an\n  open PR with the issue identifier in its title).\n  `skipped_reason` and `existing_pr_url` MUST be populated.\n  Downstream `create_pr` is bypassed and the workflow\n  routes to `notify`.\n"},"skipped_reason":{"type":"string","description":"Why the node short-circuited. Only set when `test_status` is\n`skipped`. Today the only valid value is `open-pr-exists`.\n","enum":["open-pr-exists"]},"existing_pr_url":{"type":"string","description":"URL of the open PR that triggered the skip. Set when\n`skipped_reason` is `open-pr-exists` so the notify node\ncan link to the prior PR.\n","pattern":"^https?://"},"test_notes":{"type":"string","description":"One-paragraph explanation of which tests were added and what they assert."}},"required":["branch","commit_sha","files_changed","test_files_changed","test_status"]},"eval":[{"name":"test_status_is_pass_or_no_framework","kind":"value","rule":{"output_matches":[{"path":"test_status","in":["pass","no-framework","skipped"]}]}},{"name":"tests_present_when_pass_claimed","kind":"judge","rubric":"You are checking the contract between test_status and test_files_changed.\n\nRead result.data.test_status and result.data.test_files_changed.\n\nRules:\n- If test_status is \"pass\": test_files_changed MUST contain at least\n  one real test file path (paths with extensions like .spec.ts,\n  .test.ts, .test.tsx, .test.py, _test.go, _spec.rb, .test.rs,\n  etc., or paths in a __tests__/, tests/, or spec/ directory).\n  An empty array, or paths that are clearly not test files\n  (README.md, package.json, source files), is a contract violation.\n- If test_status is \"no-framework\": test_files_changed MAY be empty.\n  Pass vacuously.\n- If test_status is \"skipped\": test_files_changed MAY be empty.\n  The node short-circuited before writing any code (Step 0b\n  found an open PR for the issue). `skipped_reason` and\n  `existing_pr_url` MUST be populated; if they are missing,\n  return no.\n- If test_status is \"fail\" or \"not-run\": this evaluator should not\n  run (the prior value evaluator catches that). Pass vacuously.\n\nRespond yes if the contract is satisfied, no if it is violated.\n","pass_when":"yes"}],"retry":{"max":1,"instruction":{"auto":true}}},"create_pr":{"name":"Open Pull Request","instruction":"Open a pull request for the fix. The issue identifier and URL are in your context from the create_issue step.\n\n1. Push the branch to the remote.\n2. Create the PR using the `github_create_pr` tool. **Use this exact tool name** — do NOT use `create_pull_request` or any other variant from ToolSearch. The eval check requires `github_create_pr` specifically.\n**If `github_create_pr` returns a result with `reused: true`**, the implement node (or a prior run) already opened a PR for this branch and the tool recovered it instead of failing on 422. Treat this as success and return the PR's URL and number — do NOT close it and call `github_create_pr` again.\n3. PR title MUST start with the issue identifier in brackets. Format: `[{context.create_issue.issueIdentifier}] fix: {short description}`.\n4. In the PR body, include: summary of the bug, what the fix does, and a link to the issue (context.create_issue.issueUrl). If the investigation found a Sentry issue, include \"Fixes {sentryShortId}\" in the PR body so Sentry auto-resolves on merge.\n5. Pass `labels: [\"sweny\", \"agent\", \"triage\"]` to `github_create_pr`.\n\nIf context.prTemplate is provided, use it as the format for the PR body. Otherwise use a clear structure with: Summary, Changes, Testing, and Related Issues.\n\nIf context.cloudRunId is provided, append this HTML comment to the end of the PR body: <!-- sweny:run_id:{context.cloudRunId} -->\n\nReturn the PR URL and number.","skills":["github"],"eval":[{"name":"pr_was_created","kind":"function","rule":{"any_tool_called":["github_create_pr"]}}],"retry":{"max":1,"instruction":{"auto":true}}},"notify":{"name":"Notify Team","instruction":"Send a notification summarizing the triage result:\n\n1. Include: alert summary, severity, root cause (1-2 sentences), and links to any created issues or PRs.\n2. For critical/high severity, make the notification urgent.\n3. For medium/low, a standard notification is fine.\n\nUse whichever notification channel is available to you.","skills":["slack","notification"]}},"edges":[{"from":"gather","to":"investigate"},{"from":"investigate","to":"create_issue","when":"novel_count is greater than 0 AND highest_severity is medium or higher"},{"from":"investigate","to":"skip","when":"novel_count is 0, OR highest_severity is low"},{"from":"create_issue","to":"implement","when":"at least one novel finding has fix_complexity simple or moderate AND fix_approach is provided"},{"from":"create_issue","to":"notify","when":"all novel findings have fix_complexity complex, OR no clear fix_approach"},{"from":"skip","to":"notify"},{"from":"implement","to":"notify","when":"implement.test_status is \"skipped\" (Step 0b found an open PR already exists for the issue identifier)"},{"from":"implement","to":"create_pr","when":"implement.test_status is \"pass\" or \"no-framework\" (no existing PR; fix was implemented or there is no test framework)"},{"from":"create_pr","to":"notify"}],"skills":{}};
+export const implementWorkflow = {"id":"implement","name":"Implement Fix","workflow_type":"pr_review","description":"Implement a code fix for an issue and open a pull request","entry":"analyze","nodes":{"analyze":{"name":"Analyze Issue","instruction":"Read the issue details and understand what needs to be fixed:\n\n1. Fetch the issue from the tracker (GitHub or Linear).\n2. **Check for an existing open PR for this issue.** Call\n   `github_search_issues` (GitHub's `/search/issues` endpoint returns\n   both issues and PRs) with `type:pr state:open` and the issue\n   identifier wrapped in brackets — e.g. query\n   `[OFF-1234] in:title type:pr state:open`. If the response contains\n   one or more PRs whose title starts with `[<issueIdentifier>]`,\n   STOP analysis. Output `existing_pr_url`\n   (the URL of the open PR) and a short `issue_summary` noting the\n   skip; leave `fix_plan` set to a one-line \"Skip: open PR\n   <url> already exists\". The workflow will route to `notify` and\n   the team will be told the prior PR is still pending.\n   Rationale: re-implementing while an open PR exists produces a\n   duplicate (or worse, the agent closes its own prior PR mid-run\n   and opens a new one — exactly the thrash seen in\n   `letsoffload/permit-service` PR #81 → #82 and\n   `letsoffload/offload` PR #572 → #573).\n3. Read the relevant source files to understand the current code.\n4. Identify the exact files and lines that need to change.\n5. Plan the fix approach.\n\nOutput a clear analysis of what needs to change and why.","skills":["github","linear"],"output":{"type":"object","properties":{"issue_summary":{"type":"string"},"files_to_change":{"type":"array","items":{"type":"string"}},"fix_plan":{"type":"string"},"risk_level":{"type":"string","enum":["low","medium","high"]},"existing_pr_url":{"type":"string","description":"URL of an open PR found in step 2 whose title starts with the\nissue identifier in brackets. Set ONLY when a prior PR exists\nand the analyze node is short-circuiting. Triggers the\n`analyze → notify` edge.\n","pattern":"^https?://"}},"required":["issue_summary","fix_plan"]}},"implement":{"name":"Implement Fix","instruction":"Implement the planned fix:\n\n1. Create a feature branch.\n2. Make the necessary code changes.\n3. Ensure changes are minimal and focused — fix the bug, nothing more.\n4. Stage and commit with a clear commit message referencing the issue.\n\nStop after the commit. Do NOT push or open the PR — the dedicated `create_pr`\nnode handles that via `github_create_pr` and its eval gate is what unblocks\nthe rest of the workflow.\n\nIf the fix is too risky or complex, explain why and skip.","skills":["github"],"disallowed_tools":["WebFetch","WebSearch"]},"create_pr":{"name":"Open Pull Request","instruction":"Open a pull request for the fix:\n\n1. Push the branch to the remote.\n2. Create a PR with a clear title and description.\n3. Reference the original issue in the PR body.\n4. Pass `labels: [\"sweny\", \"agent\", \"task\"]` to `github_create_pr` (or `[\"sweny\", \"agent\", \"feature\"]` if the issue is a feature request).\n\nIf context.prTemplate is provided, use it as the format for the PR body. Otherwise use a clear structure with: Summary, Changes, Testing, and Related Issues.\n\nIf context.cloudRunId is provided, append this HTML comment to the end of the PR body: <!-- sweny:run_id:{context.cloudRunId} -->\n\nReturn the PR URL.","skills":["github"]},"notify":{"name":"Notify","instruction":"Send a notification about the implementation result:\n\n1. Include: issue reference, PR link, brief summary of changes.\n2. Keep it concise — one message, not a wall of text.","skills":["slack","notification"]},"skip":{"name":"Skip — Too Complex","instruction":"The fix was determined to be too complex or risky for automated implementation.\nAdd a comment to the issue explaining what was found and why it needs manual attention.","skills":["github","linear"]}},"edges":[{"from":"analyze","to":"notify","when":"analyze.existing_pr_url is set (an open PR for this issue already exists; skip the implement and create_pr nodes)"},{"from":"analyze","to":"implement","when":"Fix risk level is low or medium and a clear plan exists"},{"from":"analyze","to":"skip","when":"Fix is too complex, risky, or unclear"},{"from":"implement","to":"create_pr"},{"from":"skip","to":"notify"},{"from":"create_pr","to":"notify"}],"skills":{}};
 export const seedContentWorkflow = {"id":"seed-content","name":"Content Generation Pipeline","workflow_type":"content_generation","description":"Audit content gaps, generate educational content, validate quality, and publish to Supabase","entry":"audit","nodes":{"audit":{"name":"Audit Content Gaps","instruction":"You are auditing an educational platform's content library.\n\nUse supabase_count to check how many items exist for each combination of:\n- Content types: worksheets, games, lessons\n- Subjects: math, science, english, history, geography, economics, music, art, life-skills, foreign-language\n- Difficulties: beginner, easy, medium, hard\n\nAlso check the input parameters — if the user specified a subject, topic, difficulty, and count,\nskip the full audit and just report what they want generated.\n\nReturn a structured content plan: what to generate, how many, at what difficulty and grade level.\nPrioritize gaps where a subject has ZERO content of a given type.","skills":["supabase"],"output":{"type":"object","properties":{"plan":{"type":"array","items":{"type":"object","properties":{"contentType":{"type":"string"},"subject":{"type":"string"},"topic":{"type":"string"},"difficulty":{"type":"string"},"gradeLevel":{"type":"string"},"count":{"type":"number"},"reason":{"type":"string"}},"required":["contentType","subject","topic","difficulty","count"]}},"summary":{"type":"string"}},"required":["plan","summary"]}},"generate":{"name":"Generate Content","instruction":"You are an expert educator creating content for children ages 5-14.\n\nFor each item in the content plan from the previous step, generate the content directly.\nThe content must be high-quality, engaging, age-appropriate, and educationally sound.\n\n**For worksheets:**\nGenerate a JSON array of problems. Each problem needs: \"problem\", \"answer\", \"format\" (answer-line | multiple-choice | write-on).\nRequirements:\n- Every answer MUST be correct. Double-check math, facts, spelling.\n- Problems must progress in difficulty within the set.\n- Use at least 2 different problem structures (e.g., equations + word problems for math).\n- Vary phrasing — no two problems should feel copy-pasted.\n- For the target grade level, use appropriate vocabulary and complexity.\n\n**For lessons:**\nGenerate complete HTML + CSS + JS for an interactive lesson.\nRequirements:\n- Must have substantial educational content (not just a shell with navigation).\n- Include interactive elements: quizzes, click-to-reveal, drag-and-drop.\n- Must be visually engaging with colors, icons, and clear typography.\n- Include practice exercises with immediate feedback.\n- Minimum 5 teaching sections plus 2 practice activities.\n\n**For games:**\nGenerate a description and educational requirements only (actual game code is generated\nby the platform's Phaser-based generator).\n\nUse supabase_insert to write each piece of content to the appropriate table.\n\nTable schemas:\n- worksheets: id (auto), title, subject, topic, difficulty, grade_level, num_questions, problems (jsonb), status='published', source='ai-seed'\n- lessons: id (auto), title, subject, difficulty, grade_level, lesson_type, html_code, css_code, js_code, status='published', source='ai-seed'\n- games: id (auto), title, subject, difficulty, grade_level, game_type, html_code, css_code, js_code, status='published', source='ai-seed'\n\nReturn a summary of what was generated and inserted.","skills":["supabase"],"output":{"type":"object","properties":{"generated":{"type":"array","items":{"type":"object","properties":{"contentType":{"type":"string"},"title":{"type":"string"},"subject":{"type":"string"},"id":{"type":"string"},"status":{"type":"string"}}}},"totalGenerated":{"type":"number"},"errors":{"type":"array","items":{"type":"string"}}},"required":["generated","totalGenerated"]}},"validate":{"name":"Validate Content Quality","instruction":"Review the content that was just generated and inserted.\n\nFor each item, use supabase_query to fetch it back and verify:\n\n**Worksheets:**\n- Parse the problems JSON — are all answers correct?\n- Are problems diverse (not repetitive)?\n- Is difficulty appropriate for the stated grade level?\n- Are there the right number of questions?\n\n**Lessons:**\n- Is the HTML substantial (> 500 characters)?\n- Does it contain actual educational content (not just navigation shells)?\n- Are there interactive elements?\n\n**Games:**\n- Does the HTML contain a working game structure?\n\nFlag any issues. If critical issues are found (wrong answers, empty content),\nuse supabase_update to set status='draft' and add a note.\n\nReturn a quality report.","skills":["supabase"],"output":{"type":"object","properties":{"totalReviewed":{"type":"number"},"passed":{"type":"number"},"flagged":{"type":"number"},"issues":{"type":"array","items":{"type":"object","properties":{"id":{"type":"string"},"title":{"type":"string"},"issue":{"type":"string"},"severity":{"type":"string"},"action":{"type":"string"}}}},"summary":{"type":"string"}},"required":["totalReviewed","passed","flagged","summary"]}},"report":{"name":"Generate Report","instruction":"Compile a final report of the content generation run.\n\nInclude:\n- Total content generated by type and subject\n- Quality validation results\n- Any issues found and actions taken\n- Recommendations for next content generation run (what gaps remain)\n\nFormat as a clean markdown summary that a solo operator can scan in 30 seconds.","skills":[],"output":{"type":"object","properties":{"markdown":{"type":"string"},"nextPriorities":{"type":"array","items":{"type":"string"}}},"required":["markdown"]}}},"edges":[{"from":"audit","to":"generate"},{"from":"generate","to":"validate"},{"from":"validate","to":"report"}],"skills":{}};

package/dist/workflows/implement.yml

@@ -79,6 +79,12 @@ nodes:
       If the fix is too risky or complex, explain why and skip.
     skills:
       - github
+    # Same rationale as the triage workflow: keep the implement node focused
+    # on the repo. Bash stays available because `git add`/`git commit` and
+    # test runners are essential.
+    disallowed_tools:
+      - WebFetch
+      - WebSearch
   create_pr:
     name: Open Pull Request
     instruction: >-

package/dist/workflows/triage.yml

@@ -478,6 +478,18 @@ nodes:
       explain why — do not force a bad fix.
     skills:
       - github
+    # Defense-in-depth: the implement node should read repo files, edit
+    # them, run tests, and commit — not browse the web. Disallowing the
+    # built-in web tools removes them from the model context so the agent
+    # cannot drift off-task into external research mid-implementation.
+    # Note: Bash is intentionally NOT in this list because the agent needs
+    # it for `git add`/`git commit` and to run test commands. Restricting
+    # *which* Bash invocations are allowed (e.g. blocking `gh pr create`
+    # / `git push`) belongs to a follow-up that wires the SDK's
+    # canUseTool callback to a per-node Bash deny pattern list.
+    disallowed_tools:
+      - WebFetch
+      - WebSearch
     output:
       type: object
       properties:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sweny-ai/core",
-  "version": "0.1.100",
+  "version": "0.1.102",
   "type": "module",
   "bin": {
     "sweny": "./dist/cli/main.js"