@bastani/atomic 0.9.11-alpha.4 → 0.9.11-alpha.6

package/dist/builtin/workflows/builtin/generate-and-filter-prompts.ts

@@ -1,15 +1,18 @@
+const GROUNDED_REPORTING = "Before reporting progress, audit each claim against a tool result from this session. Report only work you can point to evidence for; say so explicitly when something is unverified.";
+const READABLE_REPORT = "Lead with the outcome. Keep facts, decisions, caveats, and next steps; drop background and repetition. Use complete, readable sentences rather than compressed fragments.";
+
 export function renderGeneratorPrompt(task: string, ordinal: number): string {
-  return `<role>\nYou independently generate candidate ${ordinal}; do not imitate or assume other candidates.\n</role>\n\n<objective>\n${task}\n</objective>\n\n<requirements>\nProduce one distinct, concrete candidate. Explain its value, constraints, risks, and how it can be evaluated.\n</requirements>\n\n<output_format>\nA self-contained candidate artifact with title, proposal, rationale, risks, and evaluation evidence.\n</output_format>`;
+  return `<role>\nYou independently generate candidate ${ordinal}; do not imitate or assume other candidates.\n</role>\n\n<success_criteria>\nOne distinct, concrete candidate states its value, constraints, risks, and how it can be evaluated.\n</success_criteria>\n\n<stop_rules>\nStop after one self-contained, evaluable candidate; do not add alternatives.\n</stop_rules>\n\n<output_format>\nA candidate artifact with title, proposal, criteria-based rationale, risks, and evaluation evidence. ${READABLE_REPORT}\n${GROUNDED_REPORTING}\n</output_format>\n\n<objective>\n${task}\n</objective>`;
 }
 
 export function renderFilterPrompt(task: string, candidatePaths: readonly string[], shortlistSize: number): string {
-  return `<role>\nYou deduplicate and filter independently generated candidates.\n</role>\n\n<objective>\nSelect at most ${shortlistSize} strongest candidates for: ${task}\n</objective>\n\n<artifacts>\nRead every candidate: ${candidatePaths.join(", ")}\n</artifacts>\n\n<rubric>\nFirst collapse substantively equivalent candidates. Then score fit to the task, feasibility, evidence, distinctiveness, and risk. Near-duplicates must not gain weight by repetition. Record every discarded candidate and a concrete reason.\n</rubric>\n\n<output_format>\nCall structured_output with shortlist (candidate artifact paths in ranked order) and discarded entries containing path and reason.\n</output_format>`;
+  return `<artifacts>\nRead every candidate: ${candidatePaths.join(", ")}\n</artifacts>\n\n<role>\nYou deduplicate and filter independently generated candidates.\n</role>\n\n<rubric>\nFirst collapse substantively equivalent candidates. Then score fit to the task, feasibility, evidence, distinctiveness, and risk. Near-duplicates must not gain weight by repetition. Record every discarded candidate and a concrete reason.\n</rubric>\n\n<success_criteria>\nAt most ${shortlistSize} strongest distinct candidates remain, ranked by the rubric, and every discarded candidate has a concrete reason.\n</success_criteria>\n\n<stop_rules>\nStop after every candidate is shortlisted once or recorded as discarded.\n</stop_rules>\n\n<output_format>\nCall structured_output with shortlist (candidate artifact paths in ranked order) and discarded entries containing path and a concise, criteria-based reason.\n</output_format>\n\n<objective>\nSelect at most ${shortlistSize} strongest candidates for: ${task}\n</objective>`;
 }
 
 export function renderJudgePrompt(task: string, filterPath: string, shortlistSize: number): string {
-  return `<role>\nYou independently judge the filtered shortlist against the explicit rubric.\n</role>\n\n<objective>\nReturn at most ${shortlistSize} ranked candidate paths that best satisfy: ${task}\n</objective>\n\n<artifacts>\nRead the filter report at ${filterPath} and every candidate path it references.\n</artifacts>\n\n<rubric>\nCheck task fit, feasibility, evidence, distinctiveness, and material risk. Do not restore a duplicate merely because it is phrased differently.\n</rubric>\n\n<output_format>\nCall structured_output with shortlist and rationale.\n</output_format>`;
+  return `<artifacts>\nRead the filter report at ${filterPath} and every candidate path it references.\n</artifacts>\n\n<role>\nYou independently judge the filtered shortlist against the explicit rubric.\n</role>\n\n<rubric>\nCheck task fit, feasibility, evidence, distinctiveness, and material risk. Do not restore a duplicate merely because it is phrased differently.\n</rubric>\n\n<success_criteria>\nAt most ${shortlistSize} distinct candidate paths are ranked by rubric-grounded evidence.\n</success_criteria>\n\n<stop_rules>\nStop after evaluating every filtered candidate and ranking the qualifying paths.\n</stop_rules>\n\n<output_format>\nCall structured_output with shortlist and a concise, criteria-based rationale in complete sentences.\n</output_format>\n\n<objective>\nRank the candidates that best satisfy: ${task}\n</objective>`;
 }
 
 export function renderFinalShortlistPrompt(task: string, decisionPath: string): string {
-  return `<role>\nYou present a concise, actionable final shortlist.\n</role>\n\n<objective>\nSummarize the selected candidates for: ${task}\n</objective>\n\n<artifact>\nRead the authoritative selection at ${decisionPath}; follow its order and do not add candidates.\n</artifact>\n\n<output_format>\nRanked markdown shortlist with candidate path, differentiator, evidence, tradeoffs, and recommended next evaluation.\n</output_format>`;
+  return `<artifact>\nRead the authoritative selection at ${decisionPath}; follow its order and do not add candidates.\n</artifact>\n\n<role>\nYou present a concise, actionable final shortlist so the reader can choose the next evaluation without reading the selection session.\n</role>\n\n<success_criteria>\nEvery selected candidate appears once in authoritative order with its differentiator, evidence, tradeoffs, and recommended next evaluation.\n</success_criteria>\n\n<stop_rules>\nStop after presenting every selected candidate once; do not add or reorder candidates.\n</stop_rules>\n\n<output_format>\nRanked markdown shortlist with candidate path, differentiator, evidence, tradeoffs, and recommended next evaluation. ${READABLE_REPORT}\n${GROUNDED_REPORTING}\n</output_format>\n\n<objective>\nSummarize the selected candidates for: ${task}\n</objective>`;
 }

package/dist/builtin/workflows/builtin/generate-and-filter-runner.ts

@@ -1,4 +1,4 @@
-import { readFile, writeFile } from "node:fs/promises";
+import { writeFile } from "node:fs/promises";
 import { join } from "node:path";
 import { Type, type Static } from "typebox";
 import type { WorkflowRunContext, WorkflowSerializableValue } from "../src/shared/types.js";
@@ -67,9 +67,12 @@ export async function runGenerateAndFilter(ctx: WorkflowRunContext<Inputs>): Pro
     decisionPath = judgePath;
   }
   const finalPath = join(root, "shortlist.md");
-  await ctx.task("final-shortlist", {
+  const finalShortlist = await ctx.task("final-shortlist", {
     prompt: renderFinalShortlistPrompt(ctx.inputs.prompt, decisionPath), context: "fresh",
     reads: [decisionPath, ...shortlist], output: finalPath, outputMode: "file-only",
   });
-  return { result: await readFile(finalPath, "utf8"), shortlist, candidate_artifact_paths: candidatePaths, filter_path: filterPath, judge_path: judgePath, final_path: finalPath, artifact_dir: root, manifest_path: manifestPath };
+  // `file-only` already returns a compact artifact reference. Reading the file
+  // back here would defeat that and push the whole report into the caller's
+  // context; `final_path` is returned for callers that want the contents.
+  return { result: finalShortlist.text, shortlist, candidate_artifact_paths: candidatePaths, filter_path: filterPath, judge_path: judgePath, final_path: finalPath, artifact_dir: root, manifest_path: manifestPath };
 }

package/dist/builtin/workflows/builtin/generate-and-filter.ts

@@ -1,5 +1,6 @@
 import { Type } from "typebox";
 import { workflow } from "../src/authoring/workflow.js";
+import { withSteeringPropagationContext } from "./steering-context.js";
 import { runGenerateAndFilter } from "./generate-and-filter-runner.js";
 
 export default workflow({
@@ -13,7 +14,7 @@ export default workflow({
     max_concurrency: Type.Integer({ minimum: 1, maximum: 12, default: 4, description: "Maximum simultaneous generator stages." }),
   },
   outputs: {
-    result: Type.String({ description: "Final human-readable shortlist report." }),
+    result: Type.String({ description: "Compact reference to the final shortlist report artifact; read `final_path` for the full report." }),
     shortlist: Type.Array(Type.String(), { description: "Ranked paths to selected candidate artifacts." }),
     candidate_artifact_paths: Type.Array(Type.String(), { description: "Paths to every generated candidate artifact." }),
     filter_path: Type.String({ description: "Path to the dedupe and filter decision." }),
@@ -22,5 +23,5 @@ export default workflow({
     artifact_dir: Type.String({ description: "Directory containing run artifacts." }),
     manifest_path: Type.String({ description: "Path to the candidate artifact manifest." }),
   },
-  run: runGenerateAndFilter,
+  run: async (ctx) => await runGenerateAndFilter(withSteeringPropagationContext(ctx)),
 });

package/dist/builtin/workflows/builtin/goal-artifacts.ts

@@ -45,7 +45,7 @@ export async function writeReviewRoundArtifact(
 ): Promise<string> {
   const artifactPath = join(artifactDir, "review-round-latest.json");
   const visibleReviews = reviews.map(withoutTurn);
-  // Deduplicated cross-reviewer findings batch so the next worker turn can
+  // Deduplicated cross-reviewer findings batch so the next orchestrator turn can
   // plan and repair the round's findings together instead of one at a time.
   const consolidatedFindings = consolidateFindingsBatch(
     reviews.map((review) => ({

package/dist/builtin/workflows/builtin/goal-models.ts

@@ -1,66 +1,66 @@
 import { reviewDecisionSchema } from "./goal-schemas.js";
 
-// Model chains are curated from Atomic's agentic-coding benchmark; see
-// ralph-models.ts for frontier data and drop rationale.
-export const workerModelConfig = {
-    model: "openai-codex/gpt-5.6-sol:medium",
+// Keep this model list identical to Ralph's orchestrator while preserving a
+// locally contained Goal configuration.
+export const orchestratorModelConfig = {
+    model: "anthropic/claude-opus-5:high",
     fallbackModels: [
-      "github-copilot/gpt-5.6-sol:medium",
-      "openai/gpt-5.6-sol:medium",
-      "openai-codex/gpt-5.5:medium",
-      "github-copilot/gpt-5.5:medium",
-      "openai/gpt-5.5:medium",
-      "anthropic/claude-fable-5:low",
-      "github-copilot/claude-opus-4.8 (1m):medium",
-      "anthropic/claude-opus-4-8:medium",
-      "cursor/gpt-5.6-sol:medium",
-      "cursor/gpt-5.5:medium",
-      "cursor/claude-fable-5:low",
-      "cursor/claude-opus-4-8-thinking:medium",
+      "github-copilot/claude-opus-5:high",
+      "openai-codex/gpt-5.6-sol:xhigh",
+      "github-copilot/gpt-5.6-sol:xhigh",
+      "openai/gpt-5.6-sol:xhigh",
+      "anthropic/claude-fable-5:high",
+      "github-copilot/claude-fable-5:high",
+      "kimi-coding/k3:max",
+      "moonshotai/kimi-k3:max",
+      "moonshotai-cn/kimi-k3:max",
+      "openai-codex/gpt-5.5:xhigh",
+      "github-copilot/gpt-5.5:xhigh",
+      "openai/gpt-5.5:xhigh",
+      "anthropic/claude-opus-4-8:high",
+      "github-copilot/claude-opus-4.8:high",
       "xai/grok-4.5:high",
-      "cursor/grok-4.5:high",
-      "zai/glm-5.2:high",
-      "zai-coding-cn/glm-5.2:high",
-      "cursor/glm-5.2",
-      "openrouter/openai/gpt-5.6-sol:medium",
-      "openrouter/openai/gpt-5.5:medium",
-      "openrouter/anthropic/claude-fable-5:low",
-      "openrouter/anthropic/claude-opus-4-8:medium",
+      "zai/glm-5.2:xhigh",
+      "zai-coding-cn/glm-5.2:xhigh",
+      "openrouter/anthropic/claude-opus-5:high",
+      "openrouter/openai/gpt-5.6-sol:xhigh",
+      "openrouter/anthropic/claude-fable-5:high",
+      "openrouter/moonshotai/kimi-k3:max",
+      "openrouter/sakana/fugu-ultra:high",
+      "openrouter/openai/gpt-5.5:xhigh",
+      "openrouter/anthropic/claude-opus-4-8:high",
       "openrouter/x-ai/grok-4.5",
       "openrouter/z-ai/glm-5.2:xhigh"
     ],
     excludedTools: ["ask_user_question"],
 };
 
-// Keep this model list identical to Ralph reviewer-a while preserving an
-// independent Goal configuration and Goal's richer decision schema.
-// Reviewer-a leads its fallbacks with Kimi K3 so reviewer A and B decorrelate.
+// Keep Goal's reviewer configuration independent so GPT-5.6 precedes Kimi K3
+// within both the leading direct-provider group and the OpenRouter group.
 export const reviewerModelConfig = {
-    model: "anthropic/claude-fable-5:high",
+    model: "anthropic/claude-opus-5:high",
     fallbackModels: [
-      "kimi-coding/k3:max",
-      "moonshotai/kimi-k3:max",
-      "moonshotai-cn/kimi-k3:max",
+      "github-copilot/claude-opus-5:high",
+      "anthropic/claude-fable-5:high",
+      "github-copilot/claude-fable-5:high",
       "openai-codex/gpt-5.6-sol:xhigh",
       "github-copilot/gpt-5.6-sol:xhigh",
       "openai/gpt-5.6-sol:xhigh",
+      "kimi-coding/k3:max",
+      "moonshotai/kimi-k3:max",
+      "moonshotai-cn/kimi-k3:max",
       "openai-codex/gpt-5.5:xhigh",
       "github-copilot/gpt-5.5:xhigh",
       "openai/gpt-5.5:xhigh",
-      "github-copilot/claude-opus-4.8 (1m):high",
       "anthropic/claude-opus-4-8:high",
-      "cursor/claude-fable-5:high",
-      "cursor/gpt-5.6-sol:xhigh",
-      "cursor/gpt-5.5:high",
-      "cursor/claude-opus-4-8-thinking:high",
+      "github-copilot/claude-opus-4.8:high",
       "xai/grok-4.5:high",
-      "cursor/grok-4.5:high",
       "zai/glm-5.2:xhigh",
       "zai-coding-cn/glm-5.2:xhigh",
-      "cursor/glm-5.2",
+      "openrouter/anthropic/claude-opus-5:high",
       "openrouter/anthropic/claude-fable-5:high",
-      "openrouter/moonshotai/kimi-k3:max",
       "openrouter/openai/gpt-5.6-sol:xhigh",
+      "openrouter/moonshotai/kimi-k3:max",
       "openrouter/sakana/fugu-ultra:high",
       "openrouter/openai/gpt-5.5:xhigh",
       "openrouter/anthropic/claude-opus-4-8:high",

package/dist/builtin/workflows/builtin/goal-orchestrator-prompts.ts

@@ -0,0 +1,94 @@
+import type { GoalLedger } from "./goal-types.js";
+import {
+  RECEIPT_EXPECTATIONS,
+  renderGoalContinuationPrompt,
+  renderLatestReviewArtifacts,
+  renderReceiptHistory,
+  taggedPrompt,
+} from "./goal-prompts.js";
+import { WORKER_PREFLIGHT_CONTRACT } from "./shared-prompts.js";
+
+export const GOAL_ORCHESTRATOR_RECEIPT_CONTRACT = [
+  "Complete the objective before claiming readiness. Leave work only at a true blocker or impossibility; do not redefine success around a partial state.",
+  "Map every explicit requirement, artifact, command, test, gate, invariant, and deliverable to authoritative current evidence whose scope matches the claim.",
+  "Unless the objective/criteria forbid committing, have an implementation agent commit in this checkout with a descriptive message, confirm a clean tree with `git status --porcelain`, and include the commit identifier. Do not defer committing.",
+  "Return delegations, files changed, commands and outcomes, evidence, blockers, residual risks, readiness, and verification remaining.",
+].join("\n");
+
+export const GOAL_ORCHESTRATION_GUIDANCE = [
+  "You supervise implementation, investigation, edits, and validation through the `subagent` tool rather than implementing directly.",
+  "Delegate only work that is genuinely independent and too large to finish in a handful of tool calls. Do not assign subagents as a check on work you already own. Prefer one subagent over several.",
+  "Delegate implementation with its relevant objective, cwd, files, constraints, findings, and validation. Use focused locator/analyzer/pattern or shell-heavy delegation when that work meets the delegation threshold.",
+  "Keep overlapping work with one owner; parallelize only independent tasks. While an agent runs, prepare dependent follow-up work rather than duplicating its assignment.",
+  "Coordinate follow-ups for all required implementation, tests, docs, validation, and cleanup before reporting readiness.",
+].join("\n");
+
+export const GOAL_ORCHESTRATOR_BEST_PRACTICES = [
+  "The output is an orchestrator receipt produced after reading current goal/review artifacts and incorporating delegated results.",
+  "Distinguish completed, evidenced changes from recommendations and blockers. If goal context or required subagent capability is unavailable, report the blocker rather than success.",
+  "If the final paragraph would be a plan, a question, or a promise to act next, make the appropriate tool calls instead of ending the turn.",
+].join("\n");
+
+export const GOAL_SUBAGENT_TRACKING_GUIDANCE = [
+  "Use `todo` as the active delegation ledger when work is meaningfully multi-step: record owner, purpose, and expected output; mark starts, append results, and close only after incorporation or explicit rejection.",
+  "Before the receipt, resolve each pending/in_progress item as completed, blocked, or deferred with a reason so parallel work and follow-ups remain visible.",
+].join("\n");
+
+type GoalOrchestratorPromptArgs = {
+  readonly ledger: GoalLedger;
+  readonly ledgerPath: string;
+  readonly blockerThreshold: number;
+  readonly latestReviewArtifactPaths: readonly string[];
+  readonly workflowStartCwd: string;
+};
+
+export function renderGoalOrchestratorPrompt(
+  args: GoalOrchestratorPromptArgs,
+): string {
+  return [
+    renderGoalContinuationPrompt(
+      args.ledger,
+      args.ledgerPath,
+      args.blockerThreshold,
+      args.latestReviewArtifactPaths,
+    ),
+    taggedPrompt([
+      ["context", [`Current working directory: ${args.workflowStartCwd}`, "Use it for repository work and relative paths unless an explicit cwd is intentional; pass it to delegated agents."].join("\n")],
+      ["project_setup", WORKER_PREFLIGHT_CONTRACT],
+      ["orchestration_guidance", GOAL_ORCHESTRATION_GUIDANCE],
+      ["subagent_tracking", GOAL_SUBAGENT_TRACKING_GUIDANCE],
+      ["receipt_contract", [GOAL_ORCHESTRATOR_RECEIPT_CONTRACT, RECEIPT_EXPECTATIONS].join("\n")],
+      ["constraints", [
+        "Do not submit a PR; a later authorized PR/MR/review action handles that external write after approval.",
+        "For the requested change/build/fix, make in-scope local edits and non-destructive validation through subagents without asking. Confirm destructive actions, other external writes, and scope expansion first.",
+        "Preserve repository architecture and conventions unless the literal contract and repository evidence justify changing them; add no features or abstractions beyond the task.",
+      ].join("\n")],
+      ["output", "Return readable Markdown headed: Delegations performed, Progress made, Files changed, Commands run, Evidence, Blockers, Ready for review, Remaining work."],
+      ["role", "You are the sub-agent orchestrator; supervise the complete objective through the `subagent` tool rather than implementing directly."],
+      ["objective", [
+        `Read the goal ledger at ${args.ledgerPath} and latest review artifacts from the workflow read hint.`,
+        "Perform the initialization preflight, then delegate the smallest coherent work that satisfies the literal objective, acceptance criteria, current state, and consolidated findings.",
+        "Run or delegate repository-relevant validation, including end-to-end playwright-cli or tmux validation for executable user scenarios. Incorporate results and follow-ups through completion; report a true blocker and safest partial state without inventing success.",
+        GOAL_ORCHESTRATOR_BEST_PRACTICES,
+      ].join("\n")],
+    ]),
+  ].join("\n\n");
+}
+
+export function renderForkedGoalOrchestratorPrompt(
+  ledger: GoalLedger,
+  ledgerPath: string,
+  latestReviewArtifactPaths: readonly string[],
+): string {
+  return taggedPrompt([
+    ["receipts", [`Goal ledger artifact: ${ledgerPath}`, renderReceiptHistory(ledger), renderLatestReviewArtifacts(latestReviewArtifactPaths)].join("\n\n")],
+    ["orchestration_guidance", GOAL_ORCHESTRATION_GUIDANCE],
+    ["subagent_tracking", GOAL_SUBAGENT_TRACKING_GUIDANCE],
+    ["receipt_contract", [GOAL_ORCHESTRATOR_RECEIPT_CONTRACT, RECEIPT_EXPECTATIONS].join("\n")],
+    ["constraints", "The established literal contract, acceptance matrix, contract-fidelity audit, findings batch, regression evidence, closure, worktree, PR handoff, setup, E2E, and blocked-threshold rules remain in force. Do not shrink the ledger objective."],
+    ["objective", [
+      "Continue the same goal-runner orchestrator thread as supervisor, using the `subagent` tool for implementation and validation through completion.",
+      "Read the current ledger and latest artifacts, coordinate the smallest coherent remaining delegation, incorporate results, and return the established readable receipt. If the ending would only promise or plan more work, make the tool calls instead.",
+    ].join("\n")],
+  ]);
+}