github-mirror/n8n
1
0
Fork 0
mirror of https://github.com/n8n-io/n8n.git synced 2026-10-17 05:29:44 +02:00
Code Issues Actions 14 Packages Projects Releases Wiki Activity Graph

fix(core): Point post-build guidance at the active skill (no-changelog) (#39203)

Browse Source 
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Riqwan Thamir 2026-09-28 15:23:22 +00:00 committed by GitHub
parent d3e846eb53
commit d0255c71ec
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 38 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
packages/@n8n/agents/docs/prompt-caching.md
View File

@ -74,6 +74,20 @@ splitting fragments by stability:
sees the instruction the moment the tool loads, just outside the cached
prefix.
Runtime skills follow the same principle. `ActiveSkills.modelMessages()`
appends the current skill body to the tool result that activated it: a
`load_skill` call, or any tool that calls `ctx.loadSkill`. Recorded `load_skill`
results are collapsed to `{ skillId, active }` first, so an obsolete persisted
body is never replayed. A programmatic activation stamps its tool call with the
skill id, so the skill re-anchors to the same result on later turns. The
top-level system prompt does not mention an anchored skill, so activating and
carrying a skill never rewrites the cached prefix. The `<active_skills>` block
is a recovery path. `instructions()` moves a skill into it only when no
visible, successfully resolved tool result can carry it: observational memory
masked the result, the activating call failed, or no record of the activation
exists. When observational memory masked the result, it already rewrote the
prefix, so the move costs no extra cache invalidation.
Other prefix-stability hygiene, already true or verified: tool ordering is
append-only (`getCurrentTools()` only ever appends), and none of the current
built-in `systemInstruction` sources (`delegate_subagent`, `write_todos`,

5
packages/@n8n/instance-ai/src/tools/workflows/__tests__/build-workflow.tool.test.ts
View File

@ -572,6 +572,11 @@ describe('createBuildWorkflowTool', () => {
expect(result.postBuildFlow?.instructions).toBe(
`Follow the active ${skillId} skill instructions.`,
);
// The guidance must point at the activated skill, not at an inline copy.
expect(result.postBuildFlow?.guidance).toContain(
`Follow the active ${skillId} skill instructions now`,
);
expect(result.postBuildFlow?.guidance).not.toContain('in `instructions` now');
});
it.each(['default', 'progressive'] as const)(

25
packages/@n8n/instance-ai/src/tools/workflows/build-workflow.tool.ts
View File

@ -344,11 +344,24 @@ export function autoImportMissingSdkSymbols(
const POST_BUILD_FLOW_SKILL_ID = 'post-build-flow';
const ONE_OFF_OPERATIONS_SKILL_ID = 'one-off-operations';
const ONE_OFF_OPERATIONS_GUIDANCE =
'This one-off build is not complete yet. Follow the one-off instructions in `instructions` now (do NOT load the one-off-operations skill — they are the same instructions). Simulated verification is NOT required and NOT the completion criterion: route setup if needed, then run the workflow live with the user’s approval, read back the actual node output, and report only what you read. Offer to keep or delete the workflow when the operation is done.';
/**
* Where the follow-up instructions live depends on the branch: the runtime
* delivers an activated skill with this tool result, and an inlined copy sits
* in `instructions`.
*/
function followInstructionsClause(skillId: string, label: string, activated: boolean): string {
return activated
? `Follow the active ${skillId} skill instructions now (do NOT call load_skill for it — it is already active)`
: `Follow the ${label} instructions in \`instructions\` now (do NOT load the ${skillId} skill — they are the same instructions)`;
}
const POST_BUILD_FLOW_GUIDANCE =
'This direct build is not complete yet. Follow the post-build instructions in `instructions` now (do NOT load the post-build-flow skill — they are the same instructions) before verification, setup, error-workflow follow-up, publishing, testing, or any final user-visible summary. Follow-up order is verification/setup first, then mocked/no-mock live-test when latest verification used mocks or simulations, then generic testing prompts. Until a non-simulated execution succeeds, never offer publishing as an alternative to the live test. A user-run execution counts only after `executions(action="list")` and `executions(action="get")` confirm that it succeeded and ran the required path; the user\'s statement alone is not execution evidence. Honor an explicit publish request before live execution only after warning that the live path remains untested. Offer the explicit error-workflow opt-in for direct new primary workflows only after the primary workflow is successfully published. Do not replace the error-workflow opt-in with a generic add-anything, publish, or test question.';
function oneOffOperationsGuidance(activated: boolean): string {
return `This one-off build is not complete yet. ${followInstructionsClause(ONE_OFF_OPERATIONS_SKILL_ID, 'one-off', activated)}. Simulated verification is NOT required and NOT the completion criterion: route setup if needed, then run the workflow live with the user’s approval, read back the actual node output, and report only what you read. Offer to keep or delete the workflow when the operation is done.`;
}
function postBuildFlowGuidance(activated: boolean): string {
return `This direct build is not complete yet. ${followInstructionsClause(POST_BUILD_FLOW_SKILL_ID, 'post-build', activated)} before verification, setup, error-workflow follow-up, publishing, testing, or any final user-visible summary. Follow-up order is verification/setup first, then mocked/no-mock live-test when latest verification used mocks or simulations, then generic testing prompts. Until a non-simulated execution succeeds, never offer publishing as an alternative to the live test. A user-run execution counts only after \`executions(action="list")\` and \`executions(action="get")\` confirm that it succeeded and ran the required path; the user's statement alone is not execution evidence. Honor an explicit publish request before live execution only after warning that the live path remains untested. Offer the explicit error-workflow opt-in for direct new primary workflows only after the primary workflow is successfully published. Do not replace the error-workflow opt-in with a generic add-anything, publish, or test question.`;
}
/** Tag-turn-only sections, stripped from the inline copy; follow-up turns load the full skill. */
const INLINE_SKIPPED_SECTIONS = [
@ -409,7 +422,7 @@ async function directPostBuildFlowHandoff(
required: true,
skillId: ONE_OFF_OPERATIONS_SKILL_ID,
reason: 'direct-one-off-build-succeeded',
guidance: ONE_OFF_OPERATIONS_GUIDANCE,
guidance: oneOffOperationsGuidance(activate !== undefined),
instructions: await getInlineSkillInstructions(ONE_OFF_OPERATIONS_SKILL_ID, skills, activate),
};
}
@ -418,7 +431,7 @@ async function directPostBuildFlowHandoff(
required: true,
skillId: POST_BUILD_FLOW_SKILL_ID,
reason: 'direct-build-succeeded',
guidance: POST_BUILD_FLOW_GUIDANCE,
guidance: postBuildFlowGuidance(activate !== undefined),
instructions: await getInlineSkillInstructions(POST_BUILD_FLOW_SKILL_ID, skills, activate),
};
}