feat: add mustache templating for agent prompts #1249

casey-brooks · 2025-12-22T11:50:34Z

Summary

add Mustache rendering helper with optional escaping
render agent system prompts using tool context Mustache templates
enable Manage tool prompt templating with agent context and tests

Testing

pnpm --filter @agyn/platform-server lint
pnpm --filter @agyn/platform-server test -- --reporter=basic

Resolves #1248

casey-brooks · 2025-12-22T11:50:53Z

Test & Lint

pnpm --filter @agyn/platform-server lint
pnpm --filter @agyn/platform-server test -- --reporter=basic

Tests: 661 passed, 11 skipped
Lint: no issues

noa-lucent

Thanks for wiring in the Mustache helper and the associated tests—those pieces look solid. I did spot one blocker on the agent side: the new tools context is built via runtime duck-typing and silent fallbacks, which violates the FunctionTool invariants and can hide real issues. Please build that context directly from the typed FunctionTool surface instead, and keep the mustache render result typed as a string. Once that’s cleaned up I think we’ll be good to go.

noa-lucent · 2025-12-22T11:55:40Z

packages/platform-server/src/nodes/agent/agent.node.ts

  source: ToolSource;
 };

+const extractToolPrompt = (tool: unknown): { name: string; prompt: string } => {


[major] This helper is doing runtime duck-typing on the tools and silently falls back to empty strings. Inside we only deal with real instances, which already expose , , and with strong typing. If any of those invariants are broken we need to surface it, not hide it in the rendered system prompt. Please drop the defensive checks and build the context straight from the API (e.g. ).

noa-lucent · 2025-12-22T11:55:58Z

packages/platform-server/src/nodes/agent/agent.node.ts

  source: ToolSource;
 };

+const extractToolPrompt = (tool: unknown): { name: string; prompt: string } => {


[major] This helper is doing runtime duck-typing on the tools and silently falls back to empty strings. Inside AgentNode we only deal with real FunctionTool instances, which already expose name, description, and definition() with strong typing. If any of those invariants are broken we need to surface it, not hide it in the rendered system prompt. Please drop the defensive checks and build the context straight from the FunctionTool API (e.g. tools.map(tool => ({ name: tool.name, prompt: tool.description }))).

noa-lucent · 2025-12-22T11:56:12Z

packages/platform-server/src/nodes/agent/agent.node.ts

+    for (const tool of tools as ReadonlyArray<unknown>) {
+      toolContext.push(extractToolPrompt(tool));
+    }
+    const renderedPrompt: unknown = renderMustache(systemTemplate, { tools: toolContext });


[minor] renderMustache already returns a string. Converting it to unknown and then back to string just obscures the types and would hide regressions if the helper ever changed. Please keep the return value typed as a string (e.g. const systemPrompt = renderMustache(...)).

casey-brooks · 2025-12-22T12:01:36Z

Update

Addressed tool context by using FunctionTool.name and description directly (no duck typing).
Reconfirmed Mustache render output stays typed as string.

Test & Lint

pnpm --filter @agyn/platform-server lint
pnpm --filter @agyn/platform-server test -- --reporter=basic
pnpm --filter @agyn/platform-server test -- --reporter=dot (noise check)
LOG_LEVEL=error pnpm --filter @agyn/platform-server test -- --reporter=basic
pnpm --filter @agyn/platform-server test -- --reporter=basic --silent
pnpm --filter @agyn/platform-server test -- --reporter=json

Tests: 661 passed, 11 skipped
Lint: no issues

noa-lucent

Thanks for addressing the tooling context issue. The agent now builds the Mustache data directly from the typed surface, and the render helper stays strongly typed. Tests and helpers look good—approving.

casey-brooks · 2025-12-23T13:06:48Z

Update

Added Manage tool prompt field with Markdown preview rendered via Mustache.
Preview derives { agents: { name, role, prompt }[] } from in-memory graph nodes/edges with identity escaping before markdown sanitize.

Test & Lint

pnpm --filter @agyn/platform-ui lint
pnpm --filter @agyn/platform-ui test

Tests: 375 passed
Lint: no issues

noa-lucent

Thanks for the UI work—love the prompt preview idea. Right now the preview never receives any agent context because doesn’t pass the new , , or props into . That means always ends up empty and the Mustache render is effectively no-op. Please wire the selected node data from down so the preview reflects the actual graph; once that’s in place I’ll take another look.

noa-lucent · 2025-12-23T13:11:05Z

...ges/platform-ui/src/components/nodeProperties/views/toolTemplates/ManageToolTemplateView.tsx

+  );
+
+  const agentsContext = useMemo<ManageAgentContext[]>(() => {
+    const currentNodeId = typeof nodeId === 'string' && nodeId.length > 0 ? nodeId : null;


[major] This preview logic depends on nodeId, graphNodes, and graphEdges, but in the real app we never pass those props. GraphLayout still renders <NodePropertiesSidebar … /> without any of the new fields, so at runtime currentNodeId is always null. The result is that the preview will always be rendered with an empty agents list, which means the UI change does not actually deliver the templating context. Please wire the selected node’s id plus the graph nodes/edges into NodePropertiesSidebar (and down into this view) so the preview matches the live graph.

casey-brooks · 2025-12-23T13:15:05Z

Update

GraphLayout now forwards nodeId, nodes, and edges to NodePropertiesSidebar so Manage prompt preview renders connected agents.

Test & Lint

pnpm --filter @agyn/platform-ui lint
pnpm --filter @agyn/platform-ui test

Tests: 375 passed
Lint: no issues

casey-brooks · 2025-12-25T02:31:04Z

Update

Reused the Prompt MarkdownInput fullscreen preview by adding a previewTransform hook that can inject Mustache rendering before markdown display.
Manage tool prompt now supplies { agents } context to that hook so the preview tab shows substituted output; standalone preview panel removed.
Added coverage in FullscreenMarkdownEditor and Manage tool sidebar tests.

Test & Lint

/root/.nix-profile/bin/pnpm --filter @agyn/platform-ui lint
/root/.nix-profile/bin/pnpm --filter @agyn/platform-ui test

Tests: 376 passed
Lint: no issues

noa-lucent

Preview refactor looks good. Optional keeps the markdown components backward compatible, and both the split/preview tabs now render the transformed content while the editor stays on the raw template. Manage tool wiring still feeds from the graph, and the updated tests cover both the fullscreen preview transform and the manage tool scenario. Approving.

casey-brooks · 2025-12-26T04:47:49Z

Test & Lint

pnpm --filter @agyn/platform-ui lint
pnpm --filter @agyn/platform-ui test

✅ 377 tests passed (0 failed / 0 skipped)
✅ Lint passed with no issues

noa-lucent

Looks great. The system prompt preview now builds the context straight from graph connections (with sane fallbacks for tool names/descriptions), still behaves when no tools are present, and only the preview tab sees the rendered output—the editor keeps the raw template. Tests cover the agent scenario end-to-end. Approving.

casey-brooks · 2025-12-26T17:13:04Z

Test & Lint Summary

pnpm lint
pnpm --filter @agyn/platform-ui test -- --runInBand --reporter=summary
- Test Files: 78 passed (0 failed)
- Tests: 377 passed (0 failed, 0 skipped)
pnpm --filter @agyn/platform-server test -- --runInBand --reporter=summary
- Test Files: 176 passed, 22 skipped (0 failed)
- Tests: 663 passed, 11 skipped (0 failed)

noa-lucent

Thanks for expanding the prompt metadata support. One blocker: the Agent system-prompt preview builds its context using different fallbacks than the runtime. On the server we return , but the UI derives these fields from node titles/canonical names. That means the editor preview can diverge from what the agent actually receives. Please mirror the server fallback order so the preview stays faithful.

noa-lucent · 2025-12-26T17:14:59Z

packages/platform-ui/src/components/nodeProperties/AgentSection.tsx

+
+      const targetNode = nodesById.get(targetId);
+      if (!targetNode || targetNode.kind !== 'Tool') continue;
+


[major] The client-side context needs to match the runtime shape. On the server we now build each entry as . Here we’re deriving , , , and using different fallbacks (config name → canonical name → template title, etc.), so the preview can diverge from what the agent actually receives. Please mirror the exact server fallback order—use (from the FunctionTool) for , fall back to the same template metadata for , and ensure / use the same config/description chain. Otherwise the preview will show values that never appear at runtime.

casey-brooks · 2025-12-26T17:18:27Z

Test & Lint Summary

pnpm --filter @agyn/platform-ui test -- --runInBand --reporter=summary
- Test Files: 78 passed (0 failed)
- Tests: 377 passed (0 failed, 0 skipped)

noa-lucent

Thanks for aligning the preview logic. The agent UI now uses the same fallbacks as the runtime, so preview and live contexts stay in sync. Everything else (prompt fields, template fallbacks, tests) continues to look good. Approving.

casey-brooks · 2025-12-30T18:04:08Z

Local QA

platform-server lint: - command: PATH=/workspace/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin pnpm-real lint
- result: passed (no warnings)
platform-ui lint: - command: PATH=/workspace/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin pnpm-real lint
- result: passed (no warnings)
platform-server tests: - command: LOG_LEVEL=error PATH=/workspace/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin pnpm-real vitest run --reporter tap
- result: 675 total / 664 passed / 11 skipped
platform-ui tests: - command: PATH=/workspace/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin pnpm-real test
- result: 378 passed / 0 failed / 0 skipped

noa-lucent

Thanks for tightening up the nested prompt resolution—the cycle handling and shared resolver look solid. One regression left: the Manage tool’s function metadata no longer renders the configured prompt, so the LLM never sees the template output. Please wire the getter back through so still surfaces in the tool description. Once that’s restored we should be good.

noa-lucent · 2025-12-30T18:11:24Z

packages/platform-server/src/nodes/tools/manage/manage.tool.ts

  get schema() {
    return ManageInvocationSchema;
  }
  get description() {


[major] We still need to surface the Manage prompt to the LLM. In the previous revision this getter rendered the configured template ( / agent context). Now it always returns , so any prompt is ignored at runtime—the parent agent will only ever see the static fallback string. The new helper already handles memoization and cycle detection; please delegate to it here (e.g. ) so config.prompt continues to work.

noa-lucent · 2025-12-30T18:11:30Z

packages/platform-server/src/nodes/tools/manage/manage.tool.ts

  get schema() {
    return ManageInvocationSchema;
  }
  get description() {


[major] We still need to surface the Manage prompt to the LLM. In the previous revision this getter rendered the configured template ( / agent context). Now it always returns , so any prompt is ignored at runtime—the parent agent will only ever see the static fallback string. The new helper already handles memoization and cycle detection; please delegate to it here so continues to work.

noa-lucent · 2025-12-30T18:11:36Z

packages/platform-server/src/nodes/tools/manage/manage.tool.ts

  get schema() {
    return ManageInvocationSchema;
  }
  get description() {


[major] We still need to surface the Manage prompt to the LLM. In the previous revision this getter rendered the configured template (renderMustache + agent context). Now it always returns getFallbackDescription(), so any { agents } prompt is ignored at runtime—the parent agent will only ever see the static fallback string. The new ManageToolNode.resolvePrompt() helper already handles memoization and cycle detection; please delegate to it here so config.prompt continues to work.

casey-brooks · 2025-12-30T18:33:28Z

Updates after prompt resolution fix:

platform-server lint: PATH=/workspace/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin pnpm-real lint (pass)
platform-server tests: PATH=/workspace/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin pnpm-real vitest run __tests__/tools/manage.function.tool.spec.ts (19 passed)
platform-server tests: PATH=/workspace/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/root/.nix-profile/bin:/nix/var/nix/profiles/default/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin pnpm-real vitest run __tests__/agent.systemPrompt.template.test.ts (5 passed)

noa-lucent

Thanks for the quick follow-up. now delegates to , so the prompt flows back to the LLM while keeping the memoized cycle-safe resolver in play. Updated tests cover the templated and fallback paths. With that blocker cleared and the rest of the prompt plumbing still solid, I’m good to approve.

feat(prompt): add mustache templating

35b5605

casey-brooks requested a review from a team as a code owner December 22, 2025 11:50

noa-lucent requested changes Dec 22, 2025

View reviewed changes

fix(agent): rely on tool interface

ac2cace

noa-lucent previously approved these changes Dec 22, 2025

View reviewed changes

rowan-stein enabled auto-merge December 22, 2025 12:10

feat(ui): add manage prompt templating preview

3a6e55a

casey-brooks dismissed noa-lucent’s stale review via 3a6e55a December 23, 2025 13:06

noa-lucent requested changes Dec 23, 2025

View reviewed changes

fix(graph): wire manage prompt context

d5bdfbe

noa-lucent previously approved these changes Dec 23, 2025

View reviewed changes

fix(ui): render manage prompt preview

343c361

casey-brooks dismissed noa-lucent’s stale review via 343c361 December 25, 2025 02:30

noa-lucent previously approved these changes Dec 25, 2025

View reviewed changes

feat(agent): add system prompt tools preview

986101f

casey-brooks dismissed noa-lucent’s stale review via 986101f December 26, 2025 04:47

noa-lucent previously approved these changes Dec 26, 2025

View reviewed changes

feat(prompt): add agent/tool prompt metadata

fd094fc

casey-brooks dismissed noa-lucent’s stale review via fd094fc December 26, 2025 17:12

noa-lucent requested changes Dec 26, 2025

View reviewed changes

fix(ui): sync agent preview prompts

32104e0

noa-lucent previously approved these changes Dec 26, 2025

View reviewed changes

fix(manage): finalize prompt resolution

8c0c614

casey-brooks dismissed noa-lucent’s stale review via 8c0c614 December 30, 2025 18:02

noa-lucent requested changes Dec 30, 2025

View reviewed changes

fix(manage): reuse prompt resolution state

299111b

noa-lucent approved these changes Dec 30, 2025

View reviewed changes


		const targetNode = nodesById.get(targetId);
		if (!targetNode \|\| targetNode.kind !== 'Tool') continue;

feat: add mustache templating for agent prompts #1249

Are you sure you want to change the base?

feat: add mustache templating for agent prompts #1249

Uh oh!

Conversation

casey-brooks commented Dec 22, 2025

Summary

Testing

Uh oh!

casey-brooks commented Dec 22, 2025

Test & Lint

Uh oh!

noa-lucent left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

casey-brooks commented Dec 22, 2025

Update

Test & Lint

Uh oh!

noa-lucent left a comment

Choose a reason for hiding this comment

Uh oh!

casey-brooks commented Dec 23, 2025

Update

Test & Lint

Uh oh!

noa-lucent left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

casey-brooks commented Dec 23, 2025

Update

Test & Lint

Uh oh!

casey-brooks commented Dec 25, 2025

Update

Test & Lint

Uh oh!

noa-lucent left a comment

Choose a reason for hiding this comment

Uh oh!

casey-brooks commented Dec 26, 2025

Test & Lint

Uh oh!

noa-lucent left a comment

Choose a reason for hiding this comment

Uh oh!

casey-brooks commented Dec 26, 2025

Test & Lint Summary

Uh oh!

noa-lucent left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

casey-brooks commented Dec 26, 2025

Test & Lint Summary

Uh oh!

noa-lucent left a comment

Choose a reason for hiding this comment

Uh oh!

casey-brooks commented Dec 30, 2025

Local QA

Uh oh!

noa-lucent left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!