@bastani/atomic 0.8.29-alpha.2 → 0.8.29-alpha.4

package/dist/builtin/subagents/src/shared/types.ts

@@ -803,7 +803,6 @@ export interface RunSyncOptions {
 		schema: JsonSchemaObject;
 		schemaPath: string;
 		outputPath: string;
-		metadataPath: string;
 	};
 	acceptance?: AcceptanceInput;
 	acceptanceContext?: {
@@ -1008,8 +1007,9 @@ export function resolveWorkflowStageMaxSubagentDepth(
 ): number {
 	const maxDepth = resolveCurrentMaxSubagentDepth(configMaxDepth);
 	return isWorkflowStageOrchestrationContext(ctx)
-		// Workflow stages reserve one child-subagent hop; a 0-depth constraint would
-		// prevent the stage from delegating to its configured subagent at all.
+		// Workflow stages use the same two-hop default as main chat. A 0-depth
+		// constraint still preserves one child-subagent hop so configured workflow
+		// stages can delegate at least once.
 		? Math.min(maxDepth, Math.max(1, ctx.orchestrationContext?.constraints.maxSubagentDepth ?? 1))
 		: maxDepth;
 }
@@ -1030,7 +1030,7 @@ export function resolveSubagentDepthPolicy(
 }
 
 function workflowStageSubagentDepthMessage(depth: number, maxDepth: number, action: "call" | "resume" = "call"): string {
-	return `Nested subagent ${action} blocked (depth=${depth}, max=${maxDepth}). Sub-agents inside workflow stages cannot spawn nested sub-agents.`;
+	return `Nested subagent ${action} blocked (depth=${depth}, max=${maxDepth}). Sub-agents inside workflow stages are running at the maximum nesting depth.`;
 }
 
 export function subagentDepthBlockedMessage(

package/dist/builtin/subagents/src/slash/saved-chain-mapping.ts

@@ -1,36 +1,22 @@
 import * as fs from "node:fs";
 import * as path from "node:path";
 import type { ChainConfig } from "../agents/agents.ts";
-import { assertJsonSchemaDescriptor, assertStructuredOutputParameterSchema } from "../runs/shared/structured-output.ts";
 import { isDynamicParallelStep, isParallelStep, type ChainStep } from "../shared/settings.ts";
 import type { JsonSchemaObject } from "../shared/types.ts";
 
 function loadSavedOutputSchema(
 	chain: ChainConfig,
-	stepAgent: string,
+	_stepAgent: string,
 	outputSchema: unknown,
-	options: { schemaRole: "tool-parameters" | "collection" } = { schemaRole: "tool-parameters" },
 ): JsonSchemaObject | undefined {
 	if (outputSchema === undefined) return undefined;
-	const labelForSchema = (schemaPath?: string): string => schemaPath
-		? `outputSchema for chain '${chain.name}' step '${stepAgent}' (${schemaPath})`
-		: `outputSchema for chain '${chain.name}' step '${stepAgent}'`;
-	const validateSavedSchema = (schema: unknown, label: string): JsonSchemaObject => {
-		if (options.schemaRole === "collection") {
-			assertJsonSchemaDescriptor(schema, label);
-		} else {
-			assertStructuredOutputParameterSchema(schema, label);
-		}
-		return schema;
-	};
 	if (typeof outputSchema === "string") {
 		const schemaPath = path.isAbsolute(outputSchema)
 			? outputSchema
 			: path.join(path.dirname(chain.filePath), outputSchema);
-		const parsed = JSON.parse(fs.readFileSync(schemaPath, "utf-8")) as unknown;
-		return validateSavedSchema(parsed, labelForSchema(schemaPath));
+		return JSON.parse(fs.readFileSync(schemaPath, "utf-8")) as JsonSchemaObject;
 	}
-	return validateSavedSchema(outputSchema, labelForSchema());
+	return outputSchema as JsonSchemaObject;
 }
 
 export function mapSavedChainSteps(chain: ChainConfig, worktree = false): ChainStep[] {
@@ -50,7 +36,6 @@ export function mapSavedChainSteps(chain: ChainConfig, worktree = false): ChainS
 				chain,
 				`${step.collect.as} collection`,
 				step.collect.outputSchema,
-				{ schemaRole: "collection" },
 			);
 			return {
 				...step,

package/dist/builtin/web-access/CHANGELOG.md

@@ -6,7 +6,7 @@ All notable changes to this project will be documented in this file.
 
 ### Changed
 
-- Published a synchronized Atomic 0.8.29-alpha.1 prerelease with the upstream pi TUI dependency aligned to `^0.79.3`; no functional changes were made in the web-access extension.
+- Published a synchronized Atomic 0.8.29-alpha.4 prerelease with the upstream pi TUI dependency aligned to `^0.79.3`; no functional changes were made in the web-access extension.
 
 ## [0.8.28] - 2026-06-11
 

package/dist/builtin/web-access/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/web-access",
-  "version": "0.8.29-alpha.2",
+  "version": "0.8.29-alpha.4",
   "private": true,
   "description": "Atomic extension for web search, URL fetching, GitHub repo cloning, PDF/video extraction. Fork of: https://github.com/nicobailon/pi-web-access",
   "contributors": [

package/dist/builtin/workflows/CHANGELOG.md

@@ -8,21 +8,28 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 
 ### Added
 
-- Added opt-in schema-backed workflow item results: `ctx.stage(..., { schema })`, `ctx.task(..., { schema })`, `ctx.chain` items, and `ctx.parallel` items now receive a schema-specific `structured_output` tool only for that item, require the final tool call, return the parsed value from `ctx.stage().prompt(...)`, and expose parsed task values as `result.structured` while preserving formatted JSON handoff text ([#1350](https://github.com/bastani-inc/atomic/issues/1350)).
+- Added opt-in schema-backed workflow item results: `ctx.stage(..., { schema })`, `ctx.task(..., { schema })`, `ctx.chain` items, and `ctx.parallel` items now receive a schema-specific `structured_output` tool only for that item, return the captured value from `ctx.stage().prompt(...)`, and expose parsed task values as `result.structured` while preserving formatted JSON handoff text ([#1350](https://github.com/bastani-inc/atomic/issues/1350)).
 
 ### Changed
 
+- Changed the builtin `ralph` workflow to start each iteration with `/skill:prompt-engineer` prompt-engineering and `/skill:research-codebase` research instead of an RFC/planner stage, pass the resulting research artifact to the orchestrator as primary implementation context, fork follow-up research from prior research session data, and feed unresolved reviewer findings into subsequent research passes ([#1371](https://github.com/bastani-inc/atomic/issues/1371)).
+- Changed builtin `goal`, `ralph`, and `open-claude-design` decision gates to use schema-backed workflow `structured_output` stages with TypeBox-native schema builders instead of registering bespoke terminating custom tools or wrapping plain JSON schemas with `Type.Unsafe`.
+- Changed the builtin `ralph` prompt-engineering stage to disable all tools while relying on the `/skill:prompt-engineer` skill prompt, keeping that first-pass rewrite focused and tool-free.
+- Changed builtin `goal` worker/reviewer prompts and `ralph` orchestrator/reviewer prompts to request end-to-end verification when practical, using browser-skilled subagents for web/frontend flows that may depend on backend/API behavior and tmux-skilled subagents for TUI or terminal-app scenarios.
 - Aligned the workflows extension with upstream pi TUI `^0.79.3` so workflow graph, custom UI, and prompt-broker integrations inherit the latest shared TUI compatibility fixes.
-- Documented the opt-in `structured_output` workflow path and clarified that ordinary workflow stages do not receive `structured_output` from the default tool registry; schema-enabled items auto-add the runtime tool to explicit `tools` allowlists ([#1350](https://github.com/bastani-inc/atomic/issues/1350)).
-- Clarified that workflow `structured_output` gate schemas must be top-level object tool-argument schemas, with arrays and primitives wrapped in object fields before being returned through the terminating tool, and documented the one-`prompt()` limit for schema-backed `StageContext` result contracts ([#1350](https://github.com/bastani-inc/atomic/issues/1350)).
-- Documented that terminating workflow-stage `structured_output` JSON stays inline even when large, while artifact-sized handoffs should still be saved to files when downstream stages do not need the full payload in context ([#1350](https://github.com/bastani-inc/atomic/issues/1350)).
+- Documented the opt-in `structured_output` workflow path and clarified that ordinary workflow stages do not receive `structured_output` from the default tool registry; schema-enabled items auto-add the runtime tool to explicit `tools` allowlists without adding extra workflow prompt text about the tool ([#1350](https://github.com/bastani-inc/atomic/issues/1350)).
+- Removed top-level-object restrictions from workflow `structured_output` gate schemas; Atomic now passes any plain JSON Schema object directly to the tool and documents the one-`prompt()` limit for schema-backed `StageContext` result contracts ([#1350](https://github.com/bastani-inc/atomic/issues/1350)).
 
 ### Fixed
 
-- Fixed direct workflow tool validation so schema-enabled `task`, `tasks`, `chain`, and `parallel` items reject array or primitive structured-output schemas at argument-validation time while accepting the same object-root contracts as runtime validation, including object-only `allOf` schemas ([#1350](https://github.com/bastani-inc/atomic/issues/1350)).
+- Fixed direct workflow tool validation so schema-enabled `task`, `tasks`, `chain`, and `parallel` items accept plain JSON Schema objects without additional object-root constraints ([#1350](https://github.com/bastani-inc/atomic/issues/1350)).
 - Fixed schema-backed workflow stages to fail with a clear stage-level error when `prompt()` is called more than once on the same `StageContext`, rather than surfacing the lower-level structured-output single-use guard ([#1350](https://github.com/bastani-inc/atomic/issues/1350)).
 - Fixed schema-backed workflow model fallback so an attempt that already captured a valid terminating `structured_output` result is treated as successful instead of retrying against fallback models and tripping the single-use result guard ([#1350](https://github.com/bastani-inc/atomic/issues/1350)).
 - Fixed the workflow graph overlay remaining interactive when the parent/main-chat agent opens `ask_user_question`: the graph keeps focus, the parent question stays pending behind it with a clear “Main chat needs input — exit graph to answer.” status hint, hiding/exiting the graph focuses the pending question, and host custom-UI state changes no longer hide, restore, remount, or repaint the overlay ([#1353](https://github.com/bastani-inc/atomic/issues/1353)).
+- Fixed builtin `ralph` skill-prompt stages to invoke bundled skills through `/skill:<name>` expansion so prompt engineering and research stages receive the intended skill instructions.
+- Fixed concurrent workflow stage resource reloads to serialize temporary subagent child environment isolation so parallel stage startup cannot leave parent process child flags accidentally cleared.
+- Fixed workflow stage sessions to keep bundled workflow package skills (`create-spec`, `impeccable`, `prompt-engineer`, `research-codebase`, and `skill-creator`) available while still disabling only the recursive workflows extension inside child sessions.
+- Fixed workflow stage resource discovery so bundled subagent definitions stay available, `subagent` is active by default with the same two-hop nesting budget as main chat, and explicitly allowlisted bundled extension tools such as `subagent`, `web_search`, `fetch_content`, and `intercom` remain visible even when a workflow is launched from a subagent child process.
 
 ## [0.8.28] - 2026-06-11
 

package/dist/builtin/workflows/README.md

@@ -280,7 +280,9 @@ const decision = await ctx.stage("review-gate", { schema: Decision }).prompt(
 // decision.approved is typed from the schema.
 ```
 
-Atomic registers the canonical `structured_output` tool only for schema-enabled items, automatically adds it to explicit `tools` allowlists, and fails the item if the model completes without the final tool call. The schema is used directly as the tool argument contract, so wrap arrays or primitives in an object field such as `{ items: [...] }` or `{ value: ... }`. A schema-backed `StageContext` supports one `prompt()` call because the final-answer tool is an exact-once result contract; create another `ctx.stage(..., { schema })` for another structured prompt. `ctx.task`/`ctx.chain`/`ctx.parallel` results expose the parsed value as `result.structured` and keep `result.text` as formatted JSON for handoffs.
+Atomic registers the canonical `structured_output` tool only for schema-enabled items and automatically adds it to explicit `tools` allowlists. The schema is used directly as the tool argument contract. A schema-backed `StageContext` supports one `prompt()` call because the final-answer tool is a single result contract; create another `ctx.stage(..., { schema })` for another structured prompt. `ctx.task`/`ctx.chain`/`ctx.parallel` results expose the captured value as `result.structured` and keep `result.text` as formatted JSON for handoffs.
+
+`subagent` is available as a default workflow-stage tool with the same default two-hop nesting budget as main chat: a stage can launch a subagent, and that child can launch one nested subagent before the guard blocks further delegation. `tools` allowlists apply to bundled extension tools as well as built-ins; if a stage sets `tools`, list every tool it should see. Workflow stages can explicitly list `subagent`, `web_search`, `fetch_content`, `intercom`, and other loaded extension tools, while `excludedTools` and `noTools: "all"` still win. Bundled `@bastani/subagents` agent definitions are available to the `subagent` tool in workflow stages, including workflows launched from a subagent child process.
 
 ### Model fallbacks
 
@@ -589,7 +591,7 @@ Child workflow outputs: `result`, `findings`, `research_doc_path`, `artifact_dir
 
 ### `goal`
 
-Goal Runner workflow: initialize a persisted goal ledger with a per-run goal id and lifecycle events, render goal-continuation context, run bounded worker LM turns, append receipts, run three independent reviewers, and let a TypeScript reducer decide `complete`, `continue`, `blocked`, or `needs_human`. Token budget behavior is intentionally excluded.
+Goal Runner workflow: initialize a persisted goal ledger with a per-run goal id and lifecycle events, render goal-continuation context, run bounded worker LM turns, append receipts, run three independent reviewers, and let a TypeScript reducer decide `complete`, `continue`, `blocked`, or `needs_human`. Workers and reviewers are prompted to verify user-visible behavior end-to-end when practical with browser-skilled subagents for web/frontend flows that may depend on backend/API behavior and tmux-skilled subagents for TUI or terminal-app scenarios. Token budget behavior is intentionally excluded.
 
 ```text
 /workflow goal objective="Migrate the database layer to Drizzle ORM" base_branch=develop
@@ -607,22 +609,22 @@ Child workflow outputs: `result`, `status`, `approved`, `goal_id`, `objective`,
 
 ### `ralph`
 
-Plan → orchestrate → simplify → review workflow with optional final-stage PR handoff: write an RFC-style technical design document under `specs/`, delegate implementation through sub-agents, simplify recent changes, run parallel reviewers, and iterate until approval or the loop limit. Ralph skips PR creation by default; prompt text alone does not opt in. Pass `create_pr=true` to authorize only the final `pull-request` stage to inspect provider credentials and attempt provider-appropriate PR/MR/review creation (for example GitHub `gh`, Azure Repos `az repos pr create`, or Sapling/Phabricator tooling). Ralph's own PR-creation instructions live in that final stage. Reviewers inspect repository infrastructure directly as needed; Ralph no longer runs separate `infra-*` discovery stages.
+Prompt-engineering → research → orchestrate → review workflow with optional final-stage PR handoff: transform the user prompt into a codebase and online research question with `/skill:prompt-engineer`, run `/skill:research-codebase` against it, write findings under `research/`, delegate implementation through sub-agents from that research, run parallel reviewers, and iterate until approval or the loop limit. Ralph's orchestrator and reviewers are prompted to verify user-visible behavior end-to-end when practical with browser-skilled subagents for web/frontend flows that may depend on backend/API behavior and tmux-skilled subagents for TUI or terminal-app scenarios. Follow-up iterations pass unresolved review artifacts into prompt-engineering/research and fork research from prior research session data when available. Ralph skips PR creation by default; prompt text alone does not opt in. Pass `create_pr=true` to authorize only the final `pull-request` stage to inspect provider credentials and attempt provider-appropriate PR/MR/review creation (for example GitHub `gh`, Azure Repos `az repos pr create`, or Sapling/Phabricator tooling). Ralph's own PR-creation instructions live in that final stage. Reviewers inspect repository infrastructure directly as needed; Ralph no longer runs separate `infra-*` discovery stages.
 
 ```text
-/workflow ralph prompt="Plan and migrate the database layer to Drizzle ORM" max_loops=3 base_branch=develop
-/workflow ralph prompt="Plan and migrate the database layer to Drizzle ORM" max_loops=3 base_branch=develop create_pr=true
+/workflow ralph prompt="Migrate the database layer to Drizzle ORM" max_loops=3 base_branch=develop
+/workflow ralph prompt="Migrate the database layer to Drizzle ORM" max_loops=3 base_branch=develop create_pr=true
 ```
 
 | Input              | Type      | Required | Default       | Description                                                   |
 | ------------------ | --------- | -------- | ------------- | ------------------------------------------------------------- |
-| `prompt`           | `text`    | ✓        | —             | Task, feature request, issue summary, or spec path to plan, execute, refine, and review. |
-| `max_loops`        | `number`  | —        | `10`          | Maximum plan/orchestrate/review iterations before completion or optional final handoff. |
+| `prompt`           | `text`    | ✓        | —             | Task, feature request, issue summary, or spec path to research, execute, refine, and review. |
+| `max_loops`        | `number`  | —        | `10`          | Maximum research/orchestrate/review iterations before completion or optional final handoff. |
 | `base_branch`      | `string`  | —        | `origin/main` | Branch reviewers and the optional final stage compare the current delta with; also used to create a missing worktree. |
 | `git_worktree_dir` | `string`  | —        | `""`          | Optional reusable Git worktree root. Empty runs in the invoking checkout; non-empty values run Ralph stages in the created/reused worktree. |
 | `create_pr`        | `boolean` | —        | `false`       | Safe-by-default PR creation flag. Omitted or `false` skips the final `pull-request` stage and omits `pr_report`; prompt text alone does not opt in, and only strict `true` authorizes the final `pull-request` stage to attempt provider-appropriate PR/MR/review creation. |
 
-Child workflow outputs: `result`, `plan`, `plan_path`, `implementation_notes_path`, `approved`, `iterations_completed`, `review_report`, and `review_report_path`. `pr_report` is included only when `create_pr=true` and the final `pull-request` stage runs.
+Child workflow outputs: `result`, `plan` (latest transformed research question), `plan_path` (compatibility alias for `research_path`), `research`, `research_path`, `implementation_notes_path`, `approved`, `iterations_completed`, `review_report`, and `review_report_path`. `pr_report` is included only when `create_pr=true` and the final `pull-request` stage runs.
 
 ### `open-claude-design`
 

package/dist/builtin/workflows/builtin/deep-research-codebase.ts

@@ -44,27 +44,6 @@ interface DeepResearchCodebaseResult {
 
 const FILE_ONLY_OUTPUT = "file-only" satisfies WorkflowOutputMode;
 
-const CODEBASE_SKILLS = {
-  locator:
-    "codebase-locator — use this skill's search-first discipline when mapping where files, symbols, docs, tests, and configuration live.",
-  analyzer:
-    "codebase-analyzer — use this skill's evidence-driven deep-read style when explaining behavior, architecture, control flow, data flow, and edge cases.",
-  patternFinder:
-    "codebase-pattern-finder — use this skill's example-mining approach when separating reusable conventions from one-off details.",
-  researchLocator:
-    "codebase-research-locator — use this skill's historical-discovery approach when finding prior research, specs, ADRs, issues, and TODOs.",
-  researchAnalyzer:
-    "codebase-research-analyzer — use this skill's synthesis approach when extracting decisions, constraints, stale assumptions, and open questions from prior research.",
-  onlineResearcher:
-    "codebase-online-researcher — use this skill's source-citing approach when external documentation or ecosystem behavior materially affects the answer.",
-} as const;
-
-function codebaseSkillGuidance(
-  ...skills: readonly (keyof typeof CODEBASE_SKILLS)[]
-): string {
-  return skills.map((skill) => CODEBASE_SKILLS[skill]).join("\n");
-}
-
 function taggedPrompt(sections: readonly PromptSection[]): string {
   return sections
     .map(([tag, content]) => {
@@ -446,11 +425,7 @@ export default defineWorkflow("deep-research-codebase")
               "role",
               "You are a senior codebase research scout preparing work for specialist agents.",
             ],
-            ["objective", `Map the repository. Research question: ${prompt}`],
-            [
-              "codebase_skills",
-              codebaseSkillGuidance("locator", "analyzer", "patternFinder"),
-            ],
+            ["objective", `Map the repository using parallel codebase-locator, codebase-analyzer, and codebase-pattern-finder subagents. Research question: ${prompt}`],
             [
               "instructions",
               [
@@ -480,10 +455,9 @@ export default defineWorkflow("deep-research-codebase")
             ["role", "You locate prior project research and decision history."],
             [
               "objective",
-              "Find existing docs, specs, ADRs, issues/PR notes, TODOs, and research artifacts relevant to the task.",
+              "Find existing docs, specs, ADRs, issues/PR notes, TODOs, and research artifacts relevant to the task using parallel codebase-research-locator subagents.",
             ],
             ["task", "{task}"],
-            ["codebase_skills", codebaseSkillGuidance("researchLocator")],
             [
               "instructions",
               [
@@ -520,10 +494,9 @@ export default defineWorkflow("deep-research-codebase")
             ],
             [
               "objective",
-              `Extract reusable historical context. Research question: ${prompt}`,
+              `Extract reusable historical context using parallel codebase-research-analyzer subagents. Research question: ${prompt}`,
             ],
             ["prior_research_locator_output", "{previous}"],
-            ["codebase_skills", codebaseSkillGuidance("researchAnalyzer")],
             [
               "instructions",
               [
@@ -558,13 +531,9 @@ export default defineWorkflow("deep-research-codebase")
         ["role", "You turn scout research into clean work partitions."],
         [
           "objective",
-          `Return at most ${partitionCap} independent partitions for this research question: ${prompt}`,
+          `Return at most ${partitionCap} independent partitions for this research question: ${prompt}. Use parallel codebase-locator, codebase-analyzer, and codebase-pattern-finder subagents.`,
         ],
         ["scout_output", "{previous}"],
-        [
-          "codebase_skills",
-          codebaseSkillGuidance("locator", "analyzer", "patternFinder"),
-        ],
         [
           "instructions",
           [
@@ -607,11 +576,11 @@ export default defineWorkflow("deep-research-codebase")
                 "scout_context",
                 `Read the scout artifact before making evidence claims: ${displayWorkflowPath(scoutPath)}\nCompact saved-output reference: {previous}`,
               ],
-              ["codebase_skills", codebaseSkillGuidance("locator")],
               [
                 "instructions",
                 [
                   "Find the highest-signal files, tests, docs, commands, configs, and symbols for this partition.",
+                  "Use parallel codebase-locator subagents to explore different areas of the partition.",
                   "Explain why each path matters for the research question.",
                   "Prioritize exact paths and symbol names over broad descriptions.",
                   "Flag areas that look relevant but could not be verified.",
@@ -643,11 +612,10 @@ export default defineWorkflow("deep-research-codebase")
                 "scout_context",
                 `Read the scout artifact before making evidence claims: ${displayWorkflowPath(scoutPath)}\nCompact saved-output reference: {previous}`,
               ],
-              ["codebase_skills", codebaseSkillGuidance("patternFinder")],
               [
                 "instructions",
                 [
-                  "Identify recurring implementation patterns, abstractions, naming conventions, and anti-patterns in this partition.",
+                  "Identify recurring implementation patterns, abstractions, naming conventions, and anti-patterns in this partition using parallel codebase-pattern-finder subagents.",
                   "Use concrete examples with paths, symbols, or test names.",
                   "Distinguish established conventions from one-off implementation details.",
                   "Avoid generic advice that is not grounded in the repository.",
@@ -711,11 +679,10 @@ export default defineWorkflow("deep-research-codebase")
                 "context",
                 `Read these artifacts before analyzing: ${displayWorkflowPaths(analyzerReads)}\nCompact saved-output reference: {previous}`,
               ],
-              ["codebase_skills", codebaseSkillGuidance("analyzer")],
               [
                 "instructions",
                 [
-                  "Analyze behavior, control flow, data flow, lifecycle, error handling, and test coverage for this partition.",
+                  "Analyze behavior, control flow, data flow, lifecycle, error handling, and test coverage for this partition using parallel codebase-analyzer subagents.",
                   "Build on the locator output; do not repeat file discovery except where needed as evidence.",
                   "Call out edge cases, invariants, and coupling to other partitions.",
                   "If evidence is incomplete, explain what remains unknown and how to verify it.",
@@ -747,11 +714,11 @@ export default defineWorkflow("deep-research-codebase")
               ["assignment", `Partition ${i}/${partitions.length}: ${partition}`],
               ["research_question", prompt],
               ["local_context", onlineResearcherLocalContext],
-              ["codebase_skills", codebaseSkillGuidance("onlineResearcher")],
               [
                 "instructions",
                 [
                   "Identify external library/framework behavior, standards, or docs that materially affect the local interpretation.",
+                  "Use parallel codebase-online-researcher subagents to explore different angles of external research.",
                   "Cite sources, package names, API names, versions, or documentation titles when available.",
                   "Explain how each external fact applies to this repository.",
                   "If external research is unnecessary or unavailable, say so and focus on local implications.",
@@ -829,14 +796,6 @@ export default defineWorkflow("deep-research-codebase")
           "specialist_reports",
           `Read the complete explorer handoff artifact(s) at ${displayWorkflowPaths(explorerPaths)}. They preserve every partition's Locator, Pattern Finder, Analyzer, and Online Researcher output from the original inline specialist handoff while keeping this prompt bounded.`,
         ],
-        [
-          "codebase_skills",
-          codebaseSkillGuidance(
-            "analyzer",
-            "researchAnalyzer",
-            "onlineResearcher",
-          ),
-        ],
         [
           "instructions",
           [
@@ -845,6 +804,7 @@ export default defineWorkflow("deep-research-codebase")
             "Prioritize claims supported by concrete paths, symbols, tests, docs, or cited external references.",
             "Resolve contradictions explicitly and preserve important uncertainty.",
             "Avoid inventing facts not supported by the supplied reports; state unknowns instead.",
+            "Use parallel codebase-analyzer, codebase-research-analyzer, and codebase-online-researcher subagents as needed to verify claims or fill critical gaps in the supplied reports.",
             "End with actionable next steps for a developer who will use this research.",
           ].join("\n"),
         ],