@bastani/atomic 0.5.11-0 → 0.5.12-0

package/.agents/skills/workflow-creator/references/state-and-data-flow.md

@@ -0,0 +1,374 @@
+# State and Data Flow
+
+Data flows between sessions via `s.save()` (in the producing session) and `transcript()` / `getMessages()` (in consuming sessions or at the `.run()` level). Within a session, use plain TypeScript variables. This is the programmatic equivalent of `globalState` and reducers.
+
+## Between sessions: `s.save()` → `transcript()` / `getMessages()`
+
+**Completion rule:** `transcript()` and `getMessages()` can only access data from sessions whose callbacks have already returned (i.e., sessions in the `completedRegistry`). In a `Promise.all()` group, sibling sessions cannot read each other's output — only sessions that completed before the group started are available.
+
+### Saving output
+
+Each SDK has its own save pattern:
+
+```ts
+// Claude — pass session ID (auto-reads transcript)
+s.save(s.sessionId);
+
+// Copilot — pass SessionEvent[] from getMessages()
+s.save(await s.session.getMessages());
+
+// OpenCode — pass response { info, parts } from session.prompt()
+s.save(result.data!);
+```
+
+### Retrieving as rendered text
+
+`s.transcript(handle)` returns `{ path: string, content: string }`:
+- `path` — absolute file path to the transcript on disk
+- `content` — extracted assistant text, ready to embed in prompts
+
+Pass the session handle returned by a prior `ctx.stage()` call (handle-based, recommended). The string name `s.transcript("name")` also works when no handle is in scope.
+
+```ts
+.run(async (ctx) => {
+  const researchHandle = await ctx.stage({ name: "research" }, {}, {}, async (s) => {
+    await s.session.query("Research the topic.");
+    s.save(s.sessionId);
+  });
+
+  await ctx.stage({ name: "synthesize" }, {}, {}, async (s) => {
+    const research = await s.transcript(researchHandle);
+
+    // Use rendered text in a prompt
+    await s.session.query(`Synthesize this research:\n${research.content}`);
+
+    // Or reference the file path (useful for Claude file triggers)
+    await s.session.query(`Read ${research.path} and summarize the key findings.`);
+
+    s.save(s.sessionId);
+  });
+})
+```
+
+### Retrieving as raw messages
+
+`s.getMessages(handle)` returns `SavedMessage[]` — the native SDK messages exactly as stored:
+
+```ts
+.run(async (ctx) => {
+  const researchHandle = await ctx.stage({ name: "research" }, {}, {}, async (s) => {
+    // ... research work ...
+    s.save(s.sessionId);
+  });
+
+  await ctx.stage({ name: "analyze-results" }, {}, {}, async (s) => {
+    const messages = await s.getMessages(researchHandle);
+
+    // messages is SavedMessage[], where each entry is:
+    //   { provider: "copilot", data: SessionEvent }
+    //   { provider: "opencode", data: SessionPromptResponse }
+    //   { provider: "claude", data: SessionMessage }
+
+    // Process raw messages for detailed analysis
+    for (const msg of messages) {
+      if (msg.provider === "copilot") {
+        // Access Copilot-specific fields
+        const event = msg.data;
+      }
+    }
+  });
+})
+```
+
+### Returning values from session callbacks
+
+Session callbacks can return a value directly. The handle exposes it via `.result`:
+
+```ts
+.run(async (ctx) => {
+  const planHandle = await ctx.stage({ name: "plan" }, {}, {}, async (s) => {
+    // ... planning work ...
+    return { taskCount: 5, priority: "high" };
+  });
+
+  // Access the returned value on the handle
+  console.log(planHandle.result.taskCount); // 5
+})
+```
+
+## Within a session: TypeScript variables
+
+Use closures and variables for state within a single session:
+
+```ts
+.run(async (ctx) => {
+  await ctx.stage({ name: "review-fix" }, {}, {}, async (s) => {
+    // Local state — plain variables
+    let consecutiveClean = 0;
+    let priorOutput = "";
+    const findings: string[] = [];
+
+    for (let cycle = 0; cycle < 10; cycle++) {
+      const result = await s.session.query(
+        buildReviewPrompt((ctx.inputs.prompt ?? ""), priorOutput),
+      );
+
+      // Accumulate findings
+      const review = parseReviewResult(result.output);
+      if (review) {
+        findings.push(...review.findings.map(f => f.title));
+      }
+
+      // Track clean streak
+      if (!hasActionableFindings(review, result.output)) {
+        consecutiveClean++;
+        if (consecutiveClean >= 2) break;
+        continue;
+      }
+      consecutiveClean = 0;
+
+      // Apply fix
+      const fixResult = await s.session.query(buildFixSpec(review, (ctx.inputs.prompt ?? "")));
+      priorOutput = fixResult.output;
+    }
+
+    // All local state is available here
+    console.log(`Total findings across cycles: ${findings.length}`);
+    s.save(s.sessionId);
+  });
+})
+```
+
+## File-based persistence
+
+For data that needs to survive session restarts or be accessible outside the workflow, use file I/O:
+
+```ts
+import { readFile, writeFile, mkdir } from "fs/promises";
+import { join } from "path";
+
+.run(async (ctx) => {
+  await ctx.stage({ name: "plan" }, {}, {}, async (s) => {
+    // ... plan session ...
+  });
+
+  const planHandle = await ctx.stage({ name: "plan" }, {}, {}, async (s) => {
+    // Write artifacts to session directory
+    const artifactDir = join(s.sessionDir, "artifacts");
+    await mkdir(artifactDir, { recursive: true });
+
+    const report = { timestamp: Date.now(), status: "complete" };
+    await writeFile(
+      join(artifactDir, "report.json"),
+      JSON.stringify(report, null, 2),
+    );
+
+    s.save(s.sessionId);
+  });
+
+  await ctx.stage({ name: "generate-report" }, {}, {}, async (s) => {
+    // Read prior session's output via transcript (preferred over path traversal)
+    const planTranscript = await s.transcript(planHandle);
+
+    // Or read artifacts using the transcript's path to locate the session directory
+    const planSessionDir = join(planTranscript.path, "..");
+    const priorReport = JSON.parse(
+      await readFile(join(planSessionDir, "artifacts", "report.json"), "utf-8"),
+    );
+  });
+})
+```
+
+## Shared helper functions
+
+Extract SDK-agnostic logic into shared helpers. This is the key pattern for building workflows that work across all three SDKs:
+
+```
+.atomic/workflows/
+└── my-workflow/
+    ├── claude/index.ts             # Claude SDK code
+    ├── copilot/index.ts            # Copilot SDK code
+    ├── opencode/index.ts           # OpenCode SDK code
+    └── helpers/
+        ├── prompts.ts              # Prompt builders
+        ├── parsers.ts              # Response parsers
+        └── validation.ts           # Validation logic
+```
+
+### Prompt builders
+
+```ts
+// .atomic/workflows/my-workflow/helpers/prompts.ts
+export function buildPlanPrompt(spec: string): string {
+  return `Decompose into tasks:\n${spec}`;
+}
+
+export function buildReviewPrompt(spec: string, priorOutput?: string): string {
+  let prompt = `Review the implementation against:\n${spec}`;
+  if (priorOutput) {
+    prompt += `\n\nPrior fixes:\n${priorOutput}`;
+  }
+  return prompt;
+}
+```
+
+### Response parsers
+
+```ts
+// .atomic/workflows/my-workflow/helpers/parsers.ts
+export interface ReviewResult {
+  findings: Array<{ title: string; body: string; priority: number }>;
+  overall_correctness: string;
+}
+
+/** Layered fallback: direct parse → last fenced block → last balanced object.
+ *  Always extract the LAST block, not the first — see failure-modes.md §F4/§F8. */
+export function parseReviewResult(text: string): ReviewResult | null {
+  // 1. Direct parse
+  try {
+    const parsed = JSON.parse(text);
+    if (parsed?.findings) return parsed;
+  } catch {}
+
+  // 2. Last fenced block
+  const blockRe = /```(?:json)?\s*\n([\s\S]*?)\n```/g;
+  let lastBlock: string | null = null;
+  let m: RegExpExecArray | null;
+  while ((m = blockRe.exec(text)) !== null) {
+    if (m[1]) lastBlock = m[1];
+  }
+  if (lastBlock) {
+    try { return JSON.parse(lastBlock); } catch {}
+  }
+
+  return null;
+}
+```
+
+### Usage in workflows
+
+```ts
+// .atomic/workflows/my-workflow/claude/index.ts
+import { buildPlanPrompt, buildReviewPrompt } from "../helpers/prompts.ts";
+import { parseReviewResult } from "../helpers/parsers.ts";
+
+// ... use in run() callbacks
+```
+
+## Data flow patterns
+
+### Linear pipeline
+
+```
+Session A → s.save() → Session B reads via s.transcript(handleA)
+                       → s.save() → Session C reads via s.transcript(handleB)
+```
+
+### Fan-in (multiple prior sessions)
+
+```ts
+.run(async (ctx) => {
+  const researchHandle = await ctx.stage({ name: "research" }, {}, {}, async (s) => {
+    // ... research work ...
+    s.save(s.sessionId);
+  });
+  const analysisHandle = await ctx.stage({ name: "analysis" }, {}, {}, async (s) => {
+    // ... analysis work ...
+    s.save(s.sessionId);
+  });
+  const feedbackHandle = await ctx.stage({ name: "feedback" }, {}, {}, async (s) => {
+    // ... feedback work ...
+    s.save(s.sessionId);
+  });
+
+  await ctx.stage({ name: "merge" }, {}, {}, async (s) => {
+    const research = await s.transcript(researchHandle);
+    const analysis = await s.transcript(analysisHandle);
+    const userFeedback = await s.transcript(feedbackHandle);
+
+    await s.session.query(`Combine these inputs:
+Research: ${research.content}
+Analysis: ${analysis.content}
+Feedback: ${userFeedback.content}`);
+    s.save(s.sessionId);
+  });
+})
+```
+
+### Accumulating state across sessions
+
+Each session can read all prior completed steps (but not parallel siblings):
+
+```ts
+.run(async (ctx) => {
+  const h1 = await ctx.stage({ name: "session-1" }, {}, {}, async (s) => {
+    // ...
+    s.save(s.sessionId);
+  });
+  const h2 = await ctx.stage({ name: "session-2" }, {}, {}, async (s) => {
+    // ...
+    s.save(s.sessionId);
+  });
+
+  await ctx.stage({ name: "session-3" }, {}, {}, async (s) => {
+    // Read from any prior completed session via its handle
+    const s1 = await s.transcript(h1);
+    const s2 = await s.transcript(h2);
+
+    // Combine and process
+    const combined = `${s1.content}\n${s2.content}`;
+    // ...
+  });
+})
+```
+
+## Context-Aware Transcript Handoff
+
+When passing transcripts between sessions, compress at the boundary to prevent downstream context degradation. Use structured summaries that preserve actionable information while dropping verbose tool output (applies `context-compression` + `context-degradation`):
+
+```ts
+// helpers/compression.ts
+export function compressTranscript(content: string, maxTokenEstimate: number = 4000): string {
+  // Rough estimate: ~4 chars/token for English prose, ~2-3 for code.
+  // For precise budgeting, use the provider's tokenizer instead.
+  const maxChars = maxTokenEstimate * 4;
+  if (content.length <= maxChars) return content;
+
+  // Preserve first and last sections (recency + primacy bias)
+  const headSize = Math.floor(maxChars * 0.4);
+  const tailSize = Math.floor(maxChars * 0.4);
+  const head = content.slice(0, headSize);
+  const tail = content.slice(-tailSize);
+
+  return `${head}\n\n[... ${content.length - headSize - tailSize} chars compressed ...]\n\n${tail}`;
+}
+```
+
+```ts
+await ctx.stage({ name: "synthesize" }, {}, {}, async (s) => {
+  const research = await s.transcript("research");
+  // Compress before injecting into prompt to stay within token budget
+  const compressed = compressTranscript(research.content, 4000);
+  await s.session.query(`Synthesize this research:\n${compressed}`);
+  s.save(s.sessionId);
+});
+```
+
+## File-Based Coordination
+
+Use the filesystem as a coordination layer instead of inlining large data into prompts. This applies `filesystem-context`:
+
+```ts
+.run(async (ctx) => {
+  await ctx.stage({ name: "plan" }, {}, {}, async (s) => {
+    await s.session.query(`Create a plan for: ${(s.inputs.prompt ?? "")}\n\nWrite it to plan.md.`);
+    s.save(s.sessionId);
+  });
+
+  await ctx.stage({ name: "execute" }, {}, {}, async (s) => {
+    // Reference the file by path — lets the agent read selectively
+    await s.session.query(`Read plan.md and implement each task. Mark tasks done as you go.`);
+    s.save(s.sessionId);
+  });
+})
+```

package/.agents/skills/workflow-creator/references/user-input.md

@@ -0,0 +1,206 @@
+# User Input (mid-workflow)
+
+This reference covers **mid-workflow** user interaction — pausing a running stage to ask the user a question, approve a permission, or confirm a decision. It's the programmatic equivalent of `.askUserQuestion()`.
+
+For **invocation-time** inputs (the values the user supplies when they launch the workflow from the CLI or the picker), see `workflow-inputs.md` instead. Invocation-time inputs are declared on `defineWorkflow({ inputs: [...] })` and arrive in `ctx.inputs` before any stage starts — the workflow author reads them via `ctx.inputs.<name>`.
+
+## Claude
+
+### Via `canUseTool` callback
+
+The Claude Agent SDK provides a `canUseTool` callback for runtime approval and user interaction:
+
+```ts
+import { query } from "@anthropic-ai/claude-agent-sdk";
+
+// Inside a ctx.stage() callback:
+async (s) => {
+  const result = query({
+    prompt: "Implement the feature, but ask me before making any database changes.",
+    options: {
+      canUseTool: async (toolName, toolInput, options) => {
+        if (toolName === "Write" && toolInput.file_path?.includes("migration")) {
+          // Prompt the user for approval
+          const approved = await promptUser("Allow database migration?");
+          return approved
+            ? { behavior: "allow" }
+            : { behavior: "deny", message: "User declined migration" };
+        }
+        return { behavior: "allow" };
+      },
+    },
+  });
+  for await (const msg of result) { /* process */ }
+},
+```
+
+### Via `AskUserQuestion` tool
+
+Allow the agent to ask the user questions by including `AskUserQuestion` in `allowedTools`:
+
+```ts
+const result = query({
+  prompt: (ctx.inputs.prompt ?? ""),
+  options: {
+    allowedTools: ["Read", "Write", "Edit", "Bash", "AskUserQuestion"],
+  },
+});
+```
+
+### Via streaming input
+
+For interactive sessions, use streaming mode to feed user input:
+
+```ts
+const q = query({ prompt: (ctx.inputs.prompt ?? ""), options: { ... } });
+
+// Feed additional input while the agent is running
+q.streamInput("Here's the additional context you asked for...");
+```
+
+## Copilot
+
+Session callbacks (`onUserInputRequest`, `onElicitationRequest`,
+`onPermissionRequest`) are passed as `sessionOpts` — the third argument to
+`ctx.stage()`. The runtime forwards them to `client.createSession()`.
+`onPermissionRequest` defaults to `approveAll` when not specified.
+
+### Via `onUserInputRequest`
+
+Handle `ask_user` tool requests from the agent:
+
+```ts
+await ctx.stage({ name: "plan" }, {}, {
+  onUserInputRequest: async (request) => {
+    // request.question contains the agent's question
+    const answer = await promptUser(request.question);
+    return answer;
+  },
+}, async (s) => {
+  await s.session.send({ prompt: (ctx.inputs.prompt ?? "") });
+  s.save(await s.session.getMessages());
+});
+```
+
+### Via `onElicitationRequest`
+
+For form-style UI with structured options:
+
+```ts
+await ctx.stage({ name: "plan" }, {}, {
+  onPermissionRequest: approveAll,
+  onElicitationRequest: async (request) => {
+    // request contains form fields and options
+    return {
+      action: "submit",
+      values: { strategy: "conservative", confirm: true },
+    };
+  },
+}, async (s) => {
+  await s.session.send({ prompt: (ctx.inputs.prompt ?? "") });
+  s.save(await s.session.getMessages());
+});
+```
+
+### Programmatic approval
+
+For fully autonomous workflows, use `approveAll` to skip all permission prompts.
+This is the **default** when `onPermissionRequest` is not specified in `sessionOpts`:
+
+```ts
+import { approveAll } from "@github/copilot-sdk";
+
+// Explicit (same as the default):
+await ctx.stage({ name: "plan" }, {}, { onPermissionRequest: approveAll }, async (s) => {
+  await s.session.send({ prompt: (ctx.inputs.prompt ?? "") });
+  s.save(await s.session.getMessages());
+});
+```
+
+### Custom permission handling
+
+For fine-grained control over permissions:
+
+```ts
+await ctx.stage({ name: "plan" }, {}, {
+  onPermissionRequest: async (request) => {
+    // request.kind: "shell" | "write" | "read" | "mcp" | "custom-tool" | "url" | "memory" | "hook"
+    if (request.kind === "shell" && request.command?.includes("rm -rf")) {
+      return { kind: "denied-permanently", reason: "Dangerous command" };
+    }
+    return { kind: "approved" };
+  },
+}, async (s) => {
+  await s.session.send({ prompt: (ctx.inputs.prompt ?? "") });
+  s.save(await s.session.getMessages());
+});
+```
+
+## OpenCode
+
+The `s.client` and `s.session` are auto-created by the runtime. Use them
+directly — no manual client creation needed.
+
+### Via TUI control endpoints
+
+OpenCode uses TUI control endpoints for user interaction:
+
+```ts
+// Inside a ctx.stage() callback:
+async (s) => {
+  // Wait for the next TUI control request
+  const controlRequest = await s.client.tui.next();
+
+  // Respond to the control request
+  await s.client.tui.response({
+    requestID: controlRequest.data!.id,
+    response: "User's answer here",
+  });
+},
+```
+
+### Via permission handling
+
+Handle permission requests programmatically:
+
+```ts
+// Inside a ctx.stage() callback:
+async (s) => {
+  // Subscribe to events and handle permission requests
+  const unsubscribe = await s.client.event.subscribe((event) => {
+    if (event.type === "permission.requested") {
+      s.client.session.permission({
+        sessionID: event.sessionID,
+        permissionID: event.permissionID,
+        approved: true,
+      });
+    }
+  });
+},
+```
+
+## Combining user input with control flow
+
+Use user input results in conditional logic. This Claude example uses
+`AskUserQuestion` by including it in `allowedTools` — the agent asks the
+user directly, and you parse the response to branch:
+
+```ts
+import { query } from "@anthropic-ai/claude-agent-sdk";
+
+// Inside a ctx.stage() callback (Claude example):
+async (s) => {
+  const plan = await s.transcript("plan");
+
+  // Let the agent ask the user for approval via AskUserQuestion
+  const result = await s.session.query(
+    `Here is the plan:\n${plan.content}\n\nAsk the user if they approve this plan. ` +
+    `If they approve, execute it. If not, revise it based on their feedback.`,
+  );
+
+  s.save(s.sessionId);
+},
+```
+
+For Copilot, use `onUserInputRequest` in `sessionOpts` (see above). For
+OpenCode, use the TUI control endpoints.